Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Energy-efficient device architecture and technologies for the internet of everything Mahapatra, Chinmaya 2018

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


24-ubc_2019_february_mahapatra_chinmaya.pdf [ 6.1MB ]
JSON: 24-1.0376032.json
JSON-LD: 24-1.0376032-ld.json
RDF/XML (Pretty): 24-1.0376032-rdf.xml
RDF/JSON: 24-1.0376032-rdf.json
Turtle: 24-1.0376032-turtle.txt
N-Triples: 24-1.0376032-rdf-ntriples.txt
Original Record: 24-1.0376032-source.json
Full Text

Full Text

Energy-Efficient Device Architecture and Technologiesfor the Internet of EverythingbyChinmaya MahapatraB.Tech., National Institute of Technology, Rourkela, India, 2009M.A.Sc., The University of British Columbia, Vancouver, Canada, 2013A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDoctor of PhilosophyinTHE FACULTY OF GRADUATE AND POSTDOCTORALSTUDIES(Electrical and Computer Engineering)The University of British Columbia(Vancouver)December 2018© Chinmaya Mahapatra, 2018The following individuals certify that they have read, and recommend to the Faculty ofGraduate and Postdoctoral Studies for acceptance, the thesis entitled:Energy-Efficient Device Architecture and Technologies for the Inter-net of Everythingsubmitted by Chinmaya Mahapatra in partial fulfillment of the requirements for thedegree of Doctor of Philosophy in Electrical and Computer Engineering.Examining Committee:Victor CM Leung, Electrical and Computer EngineeringSupervisorShahriar Mirabbasi, Electrical and Computer EngineeringSupervisory Committee MemberJuri Jatskevich, Electrical and Computer EngineeringUniversity ExaminerBhushan Gopaluni, Chemical and Biological EngineeringUniversity ExaminerAdditional Supervisory Committee Members:Peyman Servati, Electrical and Computer EngineeringSupervisory Committee MemberiiAbstractAround the globe, integrating information and communication technologies with phys-ical infrastructure is a top priority in pursuing smart, green living to improve energyefficiency, protect the environment, improve the quality of life, and bolster economycompetitiveness. Internet-of-Everything (IoE) is a network of uniquely identifiable, ac-cessible, and manageable smart things that are connected through a network of hetero-geneous devices and people, usually consisting of battery-operated nodes, and mostlyworking at remote places, without human intervention. This leads us to issues con-cerning IoE Systems such as network lifetime, battery efficiency, carbon emissions,low-power security and efficient data transmission, which have been analysed in thisthesis and solutions have been proposed for them.First, we investigate wireless energy harvesting (WEH), wake-up radio (WUR)scheme, and error control coding (ECC) as enabling solutions to enhance the perfor-mance of sensor networks-based IoE systems while reducing their carbon footprints.Specifically, a utility-lifetime maximization problem incorporating WEH, WUR, andECC, is formulated and solved using a distributed dual sub gradient algorithm based onthe Lagrange multiplier method. Discussion and verification through simulation resultsshow how the proposed solutions improve network utility, prolong the lifetime, andpave the way for a greener IoE by reducing their carbon footprints.iiiNext, we introduce active radio frequency identification tags based cluster head se-lection, data-awareness and energy harvesting in IoE systems. The results show thatsuch IoE systems are better equipped to deal with energy efficiency and data deliveryproblems. Simulation results support our data aware energy saving approach and showsignificant improvement over state-of-the art techniques. To design an energy-efficientand low-resource consuming security solution for IoE systems, we propose a Physi-cally Unclonable Function based security scheme that exploits variations of physicalsensor characteristics through a prototype printed circuit board design and challenge-response pair generation using the quadratic residue property. Through simulations andmeasurements, we show that our design scheme is better in terms of energy and com-putation requirements and provides a two-fold secure data transfer. Finally, we applyour solutions to a home energy management system and find an optimal model to saveenergy in a broad IoE system application.ivLay SummaryIt is projected that there will be more than 50 billion smart objects connected to theInternet of Everything (IoE) within the coming decade. These smart objects connectthe physical world with the world of computing and people are expected to revolu-tionalize all aspects of our daily lives and transform a number of application domainssuch as healthcare and transportation, etc. In this thesis, we present an overview of thechallenges involved in designing and implementing energy-efficient IoE devices andpropose promising solutions to address these challenges. Our solution takes a holisticsystem design approach considering all the critical elements of the system architecture,by implementing lightweight networking layer on sensor devices, which has energy-efficient cross-layer data driven architecture, power-efficient security and error resilientschemes. Most of the data will be stored and fetched through the cloud, thus concen-trating on enhancing the system’s performance and saving energy.vPrefaceI am the primary researcher and author for all the research contributions made in thisthesis. I conducted the literature review to identify the research problems. I formu-lated the research problems, gathered data, performed mathematical analysis and ex-periments, and carried out the numerical simulations, lab and field measurements. Ialso wrote the manuscripts for each publication.The contributions of the co-authors of my papers are as follows. Prof. Victor C.M.Leung is my supervisor. He has provided valuable guidance, technical suggestions,and constructive feedback for identifying the research problem, making the researchprogress, and preparing the associated manuscripts. Prof. Shahriar Mirabbasi, Prof.Thanos Stouraitis, Prof. Y.L. Guan and Prof. Zhenguo Sheng helped me in my researchduring my PhD. Prof. Shahriar Mirabbasi is a committee member of my PhD supervi-sory committee whereas Prof. Thanos Stouraitis and Prof. Zhenguo Sheng are researchcollaborators in my supervisor’s Natural Sciences and Engineering Research Council(NSERC) projects related to my research. I have consulted them during all my researchworks and they have provided valuable guidance, constructive technical feedback, andalso helped in editorial corrections while preparing the corresponding manuscripts forpublication. Prof. Y.L. Guan suggested improvements to improve technical contents inmy chapter 2.viDr. Pouya Kamalinejad was a former post-doctoral fellow in my lab. His PhDwork on wireless energy harvesting integrated circuit design helped me formulate andtest my model for energy harvesting and network lifetime. Peter Woo is a SystemValidation Design Engineer at Microsemi Corporation, Vancouver, Canada. He wasthe one who helped me fabricate my circuit board for my energy-efficient PUF securitycircuit in Chapter 4. Dr. Roberto Rosales is a test lab manager at the University ofBritish Columbia(UBC), he helped me with my lab setup and PUF prototype circuitmeasurements. Dr. Akshaya Moharana is an engineer at PowerTech Labs, BC, Canada.He helped me with the data for power consumption of British Columbia’s residents forthe last 15 years, which aided my research in Chapter 5.Publications that resulted from the research presented in this thesis are as follows:[1] C. Mahapatra, Z. Sheng, V. C.M. Leung, and T. Stouraitis, “A reliable and energyefficient iot data transmission scheme for smart cities based on redundant residuebased error correction coding,” in Sensing, Communication, and Networking -Workshops (SECON Workshops), 2015 12th Annual IEEE International Confer-ence on, June 2015, pp. 1–6. (Linked to Chapter 2)[2] P. Kamalinejad, C. Mahapatra, Z. Sheng, S. Mirabbasi, V. C.M. Leung, and Y. L.Guan, “Wireless energy harvesting for the internet of things,” CommunicationsMagazine, IEEE, vol. 53, no. 6, pp. 102–108, June 2015. (Linked to Chapter 2and Chapter 3)[3] C. Mahapatra, Z. Sheng, P. Kamalinejad, V. C.M. Leung, and S. Mirabbasi, “Op-timal power control in green wireless sensor networks with wireless energy har-vesting, wake-up radio and transmission control,” IEEE Access, vol. 5, pp. 501–518, 2017. (Linked to Chapter 2)vii[4] C. Mahapatra, Z. Sheng, and V. C.M. Leung, “Energy-efficient and distributeddata-aware clustering protocol for the internet-of-things,” in Electrical and Com-puter Engineering (CCECE), 2016 IEEE Canadian Conference on. IEEE, 2016,pp. 1–5. (Linked to Chapter 3)[5] C. Mahapatra, P. Kamalinejad, T. Stouraitis, S. Mirabbasi, and V. C.M. Le-ung, “Low-complexity energy-efficient security approach for e-health applica-tions based on physically unclonable functions of sensors,” in 2015 IEEE Inter-national Conference on Electronics, Circuits, and Systems (ICECS), Dec 2015,pp. 531–534. (Linked to Chapter 4)[6] C. Mahapatra, S. P. Woo, R. Rosales, T. Stouraitis, V. C.M. Leung, and S. Mirab-basi, “Energy-efficient, puf-based security design for Internet-of-Things (iot) in-frastructure,”, in IEEE Internet-of-Things Journal, 2nd revision 2018. (Linked toChapter 4)[7] C. Mahapatra, A. K. Moharana, and V. C.M. Leung, “Energy management insmart cities based on internet of things: Peak demand reduction and energy sav-ings,” Sensors, vol. 17, no. 12, p. 2812, 2017. (Linked to Chapter 5)viiiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xivList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvList of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xixAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xxivDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxv1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Elements of an Internet of Everything System . . . . . . . . . . . . . . 31.2 Open Issues in IoE Systems . . . . . . . . . . . . . . . . . . . . . . . . 61.2.1 Hardware-Based Issues in IoE . . . . . . . . . . . . . . . . . . 8ix1.2.2 Issues Related to Wireless Energy Harvesting . . . . . . . . . . 81.2.3 Poilcy Based Issues . . . . . . . . . . . . . . . . . . . . . . . . 101.2.4 Data Related Issues . . . . . . . . . . . . . . . . . . . . . . . . 101.2.5 CO2 Emissions in IoE Systems . . . . . . . . . . . . . . . . . . 111.3 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121.3.1 Prior Work on Optimal Energy Control in Wireless Sensor Net-works Based IoE Systems . . . . . . . . . . . . . . . . . . . . 121.3.2 Prior Work on Energy-efficient and Distributed Data-Aware Rout-ing and Clustering Protocol . . . . . . . . . . . . . . . . . . . . 141.3.3 Prior Work on Energy-Efficient, Security Design for IoE Infras-tructure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161.3.4 Prior Work on Policy-based Energy Management in Smart HomeIoE Ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . 191.4 Research Focus and Goals . . . . . . . . . . . . . . . . . . . . . . . . 211.4.1 Broad Goals of the Thesis . . . . . . . . . . . . . . . . . . . . 231.4.2 Key Contributions and Results . . . . . . . . . . . . . . . . . . 241.5 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282 Optimal Energy Control in Wireless Sensor Networks Based IoE Systems 322.1 System Model and Problem Formulation . . . . . . . . . . . . . . . . . 332.1.1 Routing and Flow Conservation . . . . . . . . . . . . . . . . . 342.1.2 Energy Cost Model . . . . . . . . . . . . . . . . . . . . . . . . 352.1.3 Packet Loss and Data Re-transmission . . . . . . . . . . . . . . 362.1.4 Problem Definition : Network Lifetime Maximization throughEnergy Cost Model . . . . . . . . . . . . . . . . . . . . . . . . 41x2.2 Wireless Energy Harvesting and Wake-Up Radio Scheme . . . . . . . . 432.3 Joint Utility & Network Lifetime Trade-off and Distributed Solution . . 492.3.1 Dual Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 512.3.2 Solution to GWSN Distributed Algorithm and Its ConvergenceAnalysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 532.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552.4.1 Convergence Plots . . . . . . . . . . . . . . . . . . . . . . . . 562.4.2 Utility and Lifetime Trade-off with WEH and WUR Constraints 582.4.3 Impact of Error Control Coding on Performance and Lifetime . 602.4.4 Effect of Energy Harvesting and Error Correcting Codes onTelosB Sensor Node . . . . . . . . . . . . . . . . . . . . . . . 612.4.5 Green Networking : Reduction in Carbon Footprint . . . . . . . 653 Energy-efficient and Distributed Data-Aware Routing and ClusteringProtocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683.2 IoT Network Model: Cluster Head Selection and Energy Cost Formu-lation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693.2.1 Active RFID Tag Based Cluster Head Allocation . . . . . . . . 703.2.2 Data Aware Processing . . . . . . . . . . . . . . . . . . . . . . 723.2.3 RF Energy Harvesting . . . . . . . . . . . . . . . . . . . . . . 733.2.4 Power Management Unit . . . . . . . . . . . . . . . . . . . . . 743.3 Data Aware Energy Efficient Distributed Clustering Protocol for IoT . . 763.4 Results and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 76xi4 Energy-Efficient, PUF-Based Security Design for Internet-of-Things (IoT)Infrastructure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 814.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 814.1.1 Security Requirements . . . . . . . . . . . . . . . . . . . . . . 834.2 Proposed Security Approach and Protocol . . . . . . . . . . . . . . . . 854.2.1 Quadratic Residue based Device Authentication . . . . . . . . . 874.2.2 IoT PUF Security Encryption through Digital Signatures . . . . 894.3 Energy-Efficient Circuit Design . . . . . . . . . . . . . . . . . . . . . . 904.3.1 PUF Prototype Circuitry Design with Consideration of LeakageCurrent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 934.3.2 Circuit Design Optimization through Noise Analysis . . . . . . 944.3.3 Circuit Design Optimization . . . . . . . . . . . . . . . . . . . 1004.4 Results and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 1074.4.1 Output Measurements of the PUF Prototype Circuit . . . . . . . 1074.4.2 Security Characteristics . . . . . . . . . . . . . . . . . . . . . . 1124.4.3 Threat Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 1154.4.4 Energy Consumption . . . . . . . . . . . . . . . . . . . . . . . 1165 Energy Management in Smart Cities: Peak Demand Reduction and En-ergy Savings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1195.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1195.2 Home Energy Management as a Service . . . . . . . . . . . . . . . . . 1235.2.1 The Hardware Architecture . . . . . . . . . . . . . . . . . . . . 1235.2.2 The Software Architecture and Communication Interface . . . . 1255.3 HEM as a Markov Decision Process and Its Solution . . . . . . . . . . 126xii5.3.1 State-Action Modelling of Appliances . . . . . . . . . . . . . . 1305.3.2 User Convenience and Reward Matrix . . . . . . . . . . . . . . 1325.4 Neural Fitted Q-based Home Energy Management . . . . . . . . . . . . 1355.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 1385.5.1 Case I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1415.5.2 Case II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1476 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . 1496.1 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 1496.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1536.2.1 Highly Efficient, Low-cost, and Small-Form-Factor WirelessEnergy Harvesting System . . . . . . . . . . . . . . . . . . . . 1546.2.2 Channel Statistics for IoE Systems . . . . . . . . . . . . . . . . 1546.2.3 Cross-Layer Design . . . . . . . . . . . . . . . . . . . . . . . . 1556.2.4 Security and Privacy Concerns . . . . . . . . . . . . . . . . . . 1556.2.5 Home Energy Management . . . . . . . . . . . . . . . . . . . . 156Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157A Proof of the Lemma 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174B Proof of the Proposition 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 176xiiiList of TablesTable 2.1 Notations used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34Table 2.2 WSN simulation parameters . . . . . . . . . . . . . . . . . . . . . . 56Table 2.3 Energy cost for TelosB mote w.r.t Ecomm = 1mW . . . . . . . . . . . 62Table 2.4 LTE micro base station based sink node power model parameters . . 66Table 3.1 IoT simulation parameters . . . . . . . . . . . . . . . . . . . . . . . 78Table 4.1 Amplified voltage response for different challenge voltages between0 to 5V with respect to various silicon photodiodes. . . . . . . . . . 111Table 4.2 Energy cost of various hash functions compared to our design . . . . 117Table 4.3 Energy cost of various asymmetric encryption algorithms as imple-mented in different sensor motes . . . . . . . . . . . . . . . . . . . 117Table 5.1 Maximum load rating of home appliances . . . . . . . . . . . . . . . 131Table 5.2 User preference of appliances (Pa) . . . . . . . . . . . . . . . . . . . 132Table 5.3 Actions taken by MCCU . . . . . . . . . . . . . . . . . . . . . . . . 143xivList of FiguresFigure 1.1 Interconnections in components of Internet-of-Everything.[1] ©CISCO-IBSG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2Figure 1.2 Layered IoT architecture. . . . . . . . . . . . . . . . . . . . . . . . 4Figure 1.3 Energy gap generated with decreasing size of IoT nodes with re-duced energy availability and increased security vulnerabilities aswell as increased data generation. . . . . . . . . . . . . . . . . . . 7Figure 1.4 Energy efficiency models on the basis of their technologies used. . . 7Figure 2.1 Connectivity graph . . . . . . . . . . . . . . . . . . . . . . . . . . 34Figure 2.2 Analytical results of different coding schemes for IEEE 802.15.4based sensor. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40Figure 2.3 WUH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45Figure 2.4 Feasible energy bound for harvested energy . . . . . . . . . . . . . 48Figure 2.5 WSN topology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56Figure 2.6 Simulation plots of convergence of GWSN algorithm. . . . . . . . . 57Figure 2.7 Simulation plots of network aggregate utility - lifetime trade-off fordifferent α . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58xvFigure 2.7 Simulation plots of network aggregate utility - lifetime trade-off fordifferent α . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59Figure 2.8 Energy harvesting profile and allocated energy plots. . . . . . . . . 60Figure 2.9 Impact of ECC on lifetime of sensor nodes. . . . . . . . . . . . . . 63Figure 2.10 Impact of WEH & ECC on TelosB mote. . . . . . . . . . . . . . . 64Figure 2.11 Plot of packet loss rate versus carbon footprint . . . . . . . . . . . . 66Figure 3.1 Data-aware RFID tag based IoT architecture. . . . . . . . . . . . . 69Figure 3.2 Proposed power management unit. . . . . . . . . . . . . . . . . . . 75Figure 3.3 Number of alive nodes in an IoT system versus number of rounds. . 78Figure 3.4 Total residual energy of nodes in the IoT architecture. . . . . . . . . 79Figure 3.5 Total packets delivered to the base station server from nodes . . . . . 79Figure 4.1 The dark current vs. reverse voltage of different silicon PIN photo-diodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87Figure 4.2 Quadratic residue cycles with seeds S0 = 16 & S0 = 146 . . . . . . 88Figure 4.3 Proposed sensor PUF authentication protocol architecture. . . . . . 90Figure 4.4 Circuit to measure the dark current . . . . . . . . . . . . . . . . . . 92Figure 4.5 The PUF prototype circuit for measuring dark current from a photo-diode. (Top: with a guard ring around the input of TIA to minimizethe leakage current, Bottom: without a guard ring). . . . . . . . . . 95Figure 4.6 Noise model of TIA circuit with a photodiode. . . . . . . . . . . . . 96Figure 4.7 Plots for different CF values with RF = 10MOhm for TIA LMV793. 103Figure 4.8 Plots for different CF values with RF = 6MOhm for TIA LMP2231. 104Figure 4.9 Plots for different RF values with CF = 2pF for TIA with LMP2231. 105xviFigure 4.10 Plots for various CF values with RF = 50MOhm for TIA withLMP2231. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106Figure 4.11 TIA circuit with input offset of the OpAmp. . . . . . . . . . . . . . 108Figure 4.12 TIA amplified output voltage (response) vs. reverse voltage (chal-lenge) of different silicon PIN photodiodes. . . . . . . . . . . . . . 110Figure 4.13 Measured and stored values of dark currents (ID) in Everlight sili-con photodiode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111Figure 4.14 Optical IoT sensor PUF’s resilience towards various adversary at-tacks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114Figure 5.1 HEMaaS hardware architecture of a typical Canadian condo . . . . 124Figure 5.2 Software architecture and communication framework of HEMaaSplatform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125Figure 5.3 User interface design. . . . . . . . . . . . . . . . . . . . . . . . . . 139Figure 5.4 HEM interface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140Figure 5.5 Plot of the total demand versus time during a typical Canadian win-ter month in Ontario . . . . . . . . . . . . . . . . . . . . . . . . . 141Figure 5.6 Plot of sample episodic run NFQbHEM learning process . . . . . . 141Figure 5.7 Plot of the total demand versus time for different peak reductionpercentages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145Figure 5.8 Plot of the user convenience (uc) versus time . . . . . . . . . . . . 146Figure 5.9 Plot of the user convenience (uc) versus time for (20% Good, 60%Medium and 20% Bad) and (10% Good, 40% Medium and 50%Bad) robustness measure. . . . . . . . . . . . . . . . . . . . . . . . 146xviiFigure 5.10 Comparison of peak reduction energy savings and carbon-footprintreductions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148xviiiList of Abbreviationsµp MicroprocessorACK AcknowledgementADC Analog-To-Digital ConverterARQ Automatic Repeat RequestBCH Bose-Chaudhuri-HocquenghemBRL Batch Reinforcement LearningBS Base StationCCM Community Cloud Management PanelCH Cluster HeadCIPK Carbon Intensity Per KWhCOAP Constrained Application ProtocolCRC Cyclic Redundancy CheckDAC Digital-To-Analog ConverterxixDAEECI Data Aware Energy Efficient Clustering Protocol For IoTDC Direct CurrentDEEC Distributed Energy Efficient ClusteringDoS Denial Of ServiceDR Demand ResponseDyR Dynamic RangeDSM Demand Side ManagementECC Error Correction CodingEDEEC Enhanced Distributed Energy Efficient ClusteringEH Energy HarvestingFIPS Federal Information Processing StandardsGWSN Green Wireless Sensor NetworkHEED Hybrid Energy-Efficient Distributed ClusteringHEMaaS Home Energy Management As A ServiceHMAC Hash Message Authentication CodeIBSG Internet Business Solutions GroupIC Integrated CircuitICT Information And Communication TechnologyxxIETF Internet Engineering Task ForceIoE Internet Of EverythingIoT Internet Of ThingsKNN K-Nearest NeighborKRLE Krun-Length EncodingLEACH Low-Energy Adaptive Clustering HierarchyLTE Long Term EvolutionLUT Look-Up-TableM2M Machine-To-MachineMCCU Main Command And Control UnitMDP Markov Decision ProcessMQTT Message Queue Telemetry TransportMRC Mixed Radix ConversionMWh Mega-Watt-hourNAT Network Address TranslationNFQI Neural Fitted Q-IterationOOK On-Off KeyingP2M Person-To-MachinexxiP2P Person-To-PersonPCB Printed Circuit BoardPCE Power Conversion EfficiencyPEGASIS Power Efficient Gathering In Sensor Information SystemsPMU Power Management UnitPUF Physically Unclonable FunctionQoS Quality Of ServiceQR Quadratic ResidueRAS Resume All ServicesRBF Radial Basis FunctionsRBFNN Radial Basis Function Neural NetworkRF Radio-FrequencyRFID Radio Frequency IdentificationRL Reinforcement LearningRNS Residue Number SystemRPL Routing Protocol For Low Power And Lossy NetworksRRNS Redundant Residue Number SystemRSA Rivest-Shamir-AdlemanxxiiSAS Stop All ServiceSHA-1 Secure Hash Algorithm 1SoC Sensors On ChipTIA Trans-Impedance AmplifierTOU Time-Of-UseUC ConvenienceUI User InterfaceUIP User Input PreferencesWEH Wireless Energy HarvestingWSN Wireless Sensor NetworkWU Wake-Up CommandWUR Wake Up RadioxxiiiAcknowledgmentsFirstly, I would like to express my sincere gratitude to my advisor Prof. Victor Leungfor the continuous support during the course my Ph.D. His knowledge, work ethic andperseverance are truly inspiring and his guidance has helped me in every stage of myresearch and in the writing of this thesis. I could not have imagined having a betteradvisor and mentor for my Ph.D work.I would like to express my regards to Prof. Shahriar Mirabbasi, Prof. ThanosStouraitis and Dr. Zhenguo Sheng for helping me with their immense technical knowl-edge and experience on specific subject matter related to my PhD. reasearch and thesis.I would also like to thank Dr. Roberto Rosales, Dr. Pouya Kamalinejad and Peter Woofor help with my lab measurements, PUF prototype circuit design and journal writing.I an very grateful to Dr. Akshaya Moharana for helping me with power-demand datafor British Columbia homes. I thank my labmates for their insightful discussions andall my friends for their encouragement. I would like to thank my family: my parentsand my wife, who have been the biggest support pillars of my life.I am very grateful to Natural Sciences and Engineering Research Council of Canada(NSERC) and University of British Columbia for their support for my doctoral workthrough Canada Graduate Scholarships-Doctoral (CGS D) scholarship and Four-yearfellowship (FYF) scholarship respectively.xxivDedicationTo my parents and my wife.xxvChapter 1IntroductionThe Internet-of-Things (IoT) started in 2009 with a vision of connecting devices to de-vices and persons to devices. Technologies like Radio Frequency Identification (RFID)and wireless sensor networks (WSNs) form the backbone of such interactions. The in-dustrial sector estimates that by 2020 more than half billion devices will be connectedwith each other [1–3]. When virtually every device is connected with each other andall manual commands are replaced by intelligent machines and automation, the sys-tem will be enormous and complex spanning across a varied range of protocols andstandards. IoT aims to make the Internet ubiquitous and pervasive, and has the poten-tial to affect many aspects of users’ quality of life. To monitor their environment andsend/receive data, the networked heterogeneous devices connected in an IoT structureare typically equipped with sensors, controlling processors, wireless transceivers, andan energy source (e.g., a battery) . Applications envisioned for IoT span a wide rangeof fields including home automation, healthcare, surveillance, transportation, smart en-vironments, and many more [4, 5].The Internet of Everything (IoE) as a concept first came out from the CISCO Inter-1Figure 1.1: Interconnections in components of Internet-of-Everything.[1]©CISCO-IBSGnet Business Solutions Group (IBSG) [1]. It is an extension of IoT that encompassespeople, data, things and processes to give a meaningful, energy-efficient, intelligent,relevant and secure insight to connections between the layers interconnected togetherin an agile and iterative flow. Its technologies, including heterogeneous WSNs, areused to monitor many aspects of an ecosystem ranging from a small office space to acity, in real time. In this thesis, we will be using the terms IoE and IoT interchangeably.Fig. 1.1 shows the flow process of IoE and a brief summary of the individual compo-nents of its ecosystem is as given below.People: People are an integral part of any IoE ecosystem as they are the ones who gen-erate the enormous amount data through the usage of devices. Nowadays, connectivityto the world of Internet can be established through numerous devices including personalcomputers, smartphones, tablets, wearables and many more. Apart from the traditionalways, there are other paths of connectivity with the world through social networkingsites such as Facebook and Twitter, entertainment on demand hubs like Youtube, Net-2flix and Amazon Prime. As the Internet evolves toward IoE, we will be connected inmore relevant and valuable ways. This way people themselves are the most importantnodes of this ecosystem for whom the IoE exists and the research is geared towardsproviding smooth technological flow for them.Data: The widespread proliferation of internet-connected devices as described aboveused by people in the era of IoE coupled with increasing fidelity and data acquisitionmodality generates 2.5 quintillion bytes of data each day [6]. As the devices used togenerate data become more intelligent, these vast amounts of data will produce deeperinsights into managing the relevant data for the people.Things: These are physical devices including smart sensors, connected objects, con-sumer devices and many more. In IoE, these things will sense more data, becomecontext-aware, and provide more experiential information to help people and machinesmake more relevant and valuable decisions.Process: Processes are the ways in which the people, data and things work with eachother to provide meaningful insights for the overall structure of the IoE. Following theright process will make sure that the right information is delivered to the right personat the right time in an appropriate way.In the next section we describe the layered architecture of an IoE system, its openissues and possible solutions.1.1 Elements of an Internet of Everything SystemWith the rapid development of big data and IoE, the number of networking devices anddata volume are increasing dramatically. Since portable and battery operated systemslike smartphones, tablets, and cameras will always be connected, enormous amountsof user data will be generated and their energy consumption will dramatically increase.3One of the important challenges is supplying adequate energy to operate the networkin a self-sufficient manner without compromising quality of service (QoS). In order toFigure 1.2: Layered IoT architecture.tackle these challenges, the Internet Engineering Task Force (IETF) has taken the leadin standardizing protocols for resource constrained devices such as Routing Protocol forLow Power and Lossy Networks (RPL) and Constrained Application Protocol (CoAP)[3]. But, to develop them in a large scale, a considerable insight and development isrequired. IEEE P2413 [7], the standard for an architectural framework for the IoT,aims to provide an architecture framework which captures the commonalities acrossdifferent domains and provides a basis for instantiation of concrete IoT architectures.Fig. 1.2 shows a layered architecture of a resource-constrained IoT system. The4layers of the IoT system represent different processes through which data pass beforebeing sent to cloud servers via the wireless/wired media [8, 9]. The layers of IoTrepresent the different stages of processing the data coming from the interconnectedsystem as described in Fig. 1.1.Physical layerThe physical layer consists of end-devices of the IoT system such as sensors, smart-phones, smart devices. These are energy-constrainted, small in size and have limitedhardware capability. Enormous amount of data is generated from these devices. Asmore and more devices are connected to the Internet, data generation has reached theorder of thousands of exabytes. These devices consist of a limited energy source (e.g.,a battery) to monitor their environment and send/receive data.Monitoring and PreprocessingMonitoring and pre-processing are essential parts of energy-efficient data management.Monitoring of user data is important from the point of view of data management aswell as network security. Routing and clustering of data from the lower layer to theupper layer also needs constant monitoring of data, which consumes energy and needsnetwork resources. Hence, preprocessing of data is important to extract the relevantdata for transmission, thereby reducing the transmission delay of the network.SecurityAs we gradually move toward using some of the smart devices for critical operations,security will become a primary driver. Due to the IoT devices carrying critical datain many applications and being in an unsecured environment security is paramount forIoT devices. This is achieved in this layer and it applies to both the wired as well as5wireless networks.Gateway LayerThe gateway architecture is designed in a way to support many operating systems andseveral versions of other similar operating system types. Gateways connect the networkof the end-devices and core networks to the cloud servers. When the end nodes generateresource requirements for IoT applications, they will send the data processing or storagetasks to the cloud servers.1.2 Open Issues in IoE SystemsSeveral open issues in the layered architecture are related to limited battery capacityof the devices, their network lifetime, secure data transfer in limited energy devicesand energy-efficient data management. Since energy efficiency is of utmost importanceto the battery constrained IoT devices, IoT-related standards and research works havefocused on the device energy conserving issues [10–12]. Although the size of endphysical nodes is falling fast, the energy-storage devices are improving in a slowerpace as shown in Fig. 1.3, leading to a reducing amount of available energy in smallernodes. Including a battery means increased deployment cost and more importantlymaintenance cost (to change the battery periodically). Since the node’s lifetime is gen-erally significantly higher than the battery-lifetime, it is desirable to develop energysensor nodes with increased node lifetime, that perpetually run on harvested energy, isdata-aware and uses energy resources intelligently. The goal of IoT systems is to packmore and more functionality for energy-constrained nodes in a wireless environment.This leads to an energy-gap and calls for significant improvements in energy-efficiencyfor computing and communication in energy-constrained nodes [10, 11].6Figure 1.3: Energy gap generated with decreasing size of IoT nodes with re-duced energy availability and increased security vulnerabilities as well asincreased data generation.Figure 1.4: Energy efficiency models on the basis of their technologies used.7In Fig. 1.4, we categorize the energy efficiency models on the basis of the tech-nologies used in them. Building an energy-efficient architecture for IoE systems isa gigantic task and is not limited to only areas defined in Fig. 1.4. However, in thisthesis we have analyzed and proposed solutions for the blocks and scopes of the IoEsystem that can lead to a significant energy improvement. The specific issues consid-ered are divided into hardware-related, harvested energy related, policy and user based,data awareness and carbon emission reduction based. Below, these open issues in thecontext of energy efficiency in IoE systems are reviewed.1.2.1 Hardware-Based Issues in IoEDesign of integrated circuit (IC) in an IoE network is vital in conserving energy. Aconcept of energy-efficient sensors on chip (SoC) [13, 14] improves the design of IoEnetworks by combining sensors, processing power on a single chip to reduce the datatraffic, increase security, reduction in carbon footprint as well as the energy consump-tion of the overall infrastructure. Energy-sparse, size-constrained end nodes have lim-ited resources to guarantee strong security and hence are often considered as the weak-est link in an end-to-end system. While the resource available for security is reducing(Fig. 1.3) with reducing size, the security requirements of these leaf nodes are increas-ing, creating a strong need for research in lightweight, resource-constrained securitytechnologies [15, 16]. Embedded hardware security techniques could be a potentialsolution to preserve the highest level of security within this infrastructure.1.2.2 Issues Related to Wireless Energy HarvestingThe large scale growth in the number of wirelessly connected devices however comeat the cost of a critical challenge in large scale implementation of WSNs technology8and in a greater scope, IoE, in providing energy to the nodes. In most applications,wireless nodes which solely rely on an energy storage device (e.g., battery) need tobe deployed in very large numbers and in hard-to-reach locations. Maintaining such anetwork through replacing the batteries is a cumbersome process and is uneconomicalespecially when a long life-time is desired. Energy harvesting is a promising remedyto cope with the energy challenge. A wireless node can harvest energy from differentforms of environmental sources such as thermal, wind, solar, vibration [17]. Amongthese resources, wireless energy harvesting is an attractive candidate and provides keyadvantages in virtue of being controllable and having lower cost and smaller form factorimplementations [18, 19]. Incorporation of energy harvesting is a promising remedy tocope with the energy challenge. Energy harvesting enables easier deployment of nodesin remote areas aiding in virtually maintenance-free operation and significant reductionin the carbon footprint associated with manufacturing and replacing batteries. Scaveng-ing energy form the aforementioned environmental sources is an opportunistic process,i.e., it highly relies on the presence of the source and environmental conditions. In thecontext of our system, the wireless energy sources fall into two categories of dedicatedsources and Ambient sources [18]. A dedicated RF source is deliberately deployedto supply energy to the nodes at a designated rate and optimum frequency (e.g., sinknode). An ambient source, on the other hand, is a less predictable energy source hap-pens to exist within the operation area of the network, but are not designed as a part ofthe network. Examples of ambient sources include TV and radio towers (static ambientsource) and WiFi access points (dynamic ambient source). Due to their unpredictablenature, harvesting energy from ambient sources is an opportunistic process which re-quires some level of adaptivity and entails a more sophisticated design both at circuitand system levels.91.2.3 Poilcy Based IssuesPolicies and techniques based on real time usage data in IoE systems can help reducethe energy consumption significantly [20]. Monitoring, preprocessing, making intel-ligent decisions based on user feedback and behaviour can play an important role inmaking these policies for energy-efficiency a success. The biggest challenge in man-aging such an ecosystem also known as a smart IoE based environment is to makeefficient and informed decisions from user data, behaviour and feedback. This task isnot easy to implement in a big ecosystem interconnect such as cities, homes, industries.Automation alone would not be enough and require the models for user feedback andanalysing behavioural patterns [21, 22]. This can save the energy in the range of 3-6%.Management of smart systems with optimized policy for saving energy often requiresanalyzing IoT data to optimize efficiency, comfort, safety, and to make decisions fasterand in a more precise manner.1.2.4 Data Related IssuesData collected from different sources in IoE systems have a huge amount of informa-tion. Processing these vast amounts of data for analysis can be resource intensive andtime consuming and hence, a large amount of energy is required. A challenging taskfor IoE systems is the low power data acquisition of sensed data. The main challengeis due to the fact that different query-driven user command generate varied sized data.Some of them are sparse in nature, some have higher rates and some are periodic. Therehave been several lossy compression algorithms devised specifically for resource con-strained wireless motes (sensor nodes). These algorithms include: Krun-length encod-ing (KRLE) [23], lightweight temporal compression (LTC) [24], wavelet quantizationthresholding and RLE (WQTR) [25], and compressive sampling (CS) [26], [27]. Since10the radio on a wireless device consumes orders of magnitude more power than othercomponents (e.g., ADC, CPU) [28], streaming all the data may consume too muchpower to be viable. As such, using data-awareness i.e dividing the data demand be-tween critical and non-critical data, to reduce radio transmissions will help increasesystem longevity, decrease overall system power requirements, and decrease systemcosts. Not only their size but their behavior also varies. Some are random in naturewith no correlation to the previous datasets while others are heavily correlated versionsof their previous time samples. Several solutions like K-nearest neighbor (KNN) andRadial basis functions (RBF) have been investigated to predict the behavior of data[29]. But the data variability and different service quality requirements of IoE systemsare not taken into consideration yet. Thus their is a need to analyze, investigate anddevelop models to utilize the data efficiently in low power motes.1.2.5 CO2 Emissions in IoE SystemsToday, 15 billion interactive devices are exchanging information about many aspectsof our lives, and the IoT is bound to become even more ingrained in our world as 200billion devices are expected to be actively used by 2030 [30]. If the growth of sensorsand IoT-enabling technology continues at today’s pace, 30% of the information andcommunication technology (ICT) market will be made up of IoT, data, and devices in2030. The internet releases around 300m tonnes of CO2 a year – as much as all the coal,oil and gas burned in Turkey or Poland, or more than half of the fossil fuels burned inthe UK. Enourmous amount of data generation in IoT systems, use of multiple batteries,electricity accounts for around 40% of the total ICT energy demand and 0.8% of globalCO2 emissions [30]. With ever increasing IoE devices, CO2 emissions are bound toincrease. Hence, a challenging task is to efficiently handle the factors affecting the11carbon-footprint increase to save on carbon emissions, thereby making the enviromentgreen.1.3 Related WorksHere, the prior works regarding the open issues related to energy efficiency as describedin the Section 1.2 are reviewed. The shortcomings in the existing literature are high-lighted to provide the motivations and objectives of our research.1.3.1 Prior Work on Optimal Energy Control in Wireless SensorNetworks Based IoE SystemsIn this section, we discuss the existing works in the literature concerning the problemsand solutions related to the increase in a WSN system lifetime. WSN system forms thebackbone of an IoE system. Energy efficiency with traffic dynamics have been an activearea of research in the WSN community since last two decades. Hence we focus ondeveloping a complete energy efficient framework for WSN based IoE systems throughthe existing work.Optimization methods have been extensively used in previous research works tosolve for network lifetime of wireless sensor networks. Network lifetime maximiza-tion with flow rate constraint have been studied in many prior works. Kelly et al. wasthe first to propose two classes of distributed rate control algorithms for communica-tion networks [31]. Madan et al. [32] solved the lifetime maximization problem witha distributed algorithm using the subgradient method. In [33], Ehsan et al. proposean energy and cross-layer aware routing schemes for multichannel access WSNs thataccount for radio, MAC contention, and network constraints, to maximize the networklifetime. But, the problems formulated and solved in all these approaches neither does12take into account a proper energy model incorporating all the transceiver resources norit involves the application performance trade-off due to increase in lifetime by decreas-ing rate flows.System utility and network lifetime are problems that are related to each other ina reciprocal relationship meaning maximizing one will degrade the other. Chen et al.[34] analyzed the utility-lifetime trade-off in wireless sensor network for flow con-straints. He et al. [35] followed a cross-layer design approach. Both of these paperstake transmission rate as the sole indicator of the system throughput, which is not trueas the reliability plays a vital role in determining the system performance. Reliabilityin the system can be improved by introducing error control schemes into the sensornodes with multipath routing introduced by lun et al. [36]. In [37], Yu et al. analy-ses the automatic repeat request (ARQ) as well as a hybrid ARQ scheme for WSNs.The ARQ scheme requires re-transmission if there is a failure of packet delivery whichincreases energy consumption of node. Xu et al. [38] describes a rate-reliability andlifetime trade-off for WSNs by taking theoritical end to end error probability of pack-ets. Similarly, Zou et al. [39] has taken a joint lifetime-utility-rate-reliability approachfor WSNs taking a generic error coding processing power model. Both [38] and [39]lack the inclusion and analysis of an error control scheme with their encoding/decod-ing powers as well as the delay performance of the overall system with error correctionemployed.Energy harvesting is proposed as a possible method to improve the network lifetimeand rechargeable batteries in WSNs by He et al. [40] ,Magno et al. [41] ,Deng etal. [42] and Kamalinejad et al. [43]. Practically, energy can be harvested from theenvironmental sources, namely, thermal, solar, vibration, and wireless radio-frequency(RF) energy sources [17]. While harvesting from the aforementioned environmental13sources is dependent on the presence of the corresponding energy source, RF energyharvesting provides key benefits in terms of being wireless, readily available in theform of transmitted energy (TV/radio broadcasters, mobile base stations and hand-held radios), low cost, and small form factor implementation. Recently, dynamics oftraffic and energy replenishment incorporated in the network power model has beenan active research topic. Some of the challenges are addressed by [44], [45] and [46].They assume battery energy to be zero at start, which may not be practical for manyapplication scenarios that has sensors with rechargeable batteries. challenges causedby packet loss due to interference has also not been addressed.Green networking of late in the past four to five years has attracted a lot of attention.Koutitas et al. [47] has analyzed a maximization problem based on carbon footprintsgenerated in terrestrial broadcasting networks. In [48] Naeem et al. have maximizedthe data rate while minimizing the CO2 emissions in cognitive sensor networks. But itis yet to be seen how much carbon emissions can be minimized while maximizing theutility and lifetime with reliability and energy harvesting constraints.1.3.2 Prior Work on Energy-efficient and Distributed Data-AwareRouting and Clustering ProtocolAs explained in the beginning of this Chapter, IoE plays an important role by bringingtogether people, process, data, and things to make networked connections more rele-vant and valuable. Its technologies, including heterogeneous WSNs, are used to mon-itor many aspects of an ecosystem ranging from a small office space to a city, in realtime. Routing is one of the critical technologies in IoE as opposed to traditional ad-hocWSNs. It is more challenging due to constrained resources in terms of energy supply,processing capability, frequent topology changes and reliable data delivery within a14limited time period. Based on network structure, routing protocols can be sub-dividedinto two categories, flat routing and hierarchical routing. In a flat topology, all nodesperform the same tasks and have the same functionalities in the network. Whereas, ina hierarchical topology, nodes perform different tasks and are typically organized intolots of clusters according to specific metrics. In clustering, members of the clusterselect a cluster head (CH) [49]. All nodes belonging to the same cluster send their datato CH, where, CH aggregates data and sends aggregated data to base station (BS).Clustering algorithms in the literature are divided based on their energy efficiencyin two types of networks i.e., homogeneous and heterogeneous WSNs. HomogeneousWSNs considers that the all sensor nodes in the system have the same energy leveland all the nodes takes turn according to a given probability to become CH. Low-Energy Adaptive Clustering Hierarchy (LEACH) [50], Power Efficient Gathering inSensor Information Systems (PEGASIS) [51] and Hybrid Energy-Efficient DistributedClustering (HEED) [52] are examples of cluster based protocols which are designedfor homogenous WSNs. However, these techniques perform poorly in heterogeneousWSNs scenario as nodes having less energy expire faster than higher energy nodes.Heterogeneous WSN topology takes into account that the nodes have different ini-tial energy. Thus they perform better than homogeneous WSNs in a real applicationscenario with variety of sensors such as warehouses, home monitoring and surveillance.Distributed Energy Efficient Clustering (DEEC) [53], Developed DEEC (DDEEC) andEnhanced DEEC (EDEEC) [54] are some of the heterogenous WSN protocols. Thesedistributed clustering algorithms for heterogeneous WSNs have similar topologicalstructure to an IoT system. Although multi-hop routing and residual energy for se-lecting CHs are considered, they neither incorporate the intricacies nor the benefit of adiversified and event driven IoT system.151.3.3 Prior Work on Energy-Efficient, Security Design for IoEInfrastructureAs one of the most crucial blocks of the system (Section 1.1), which provides authen-tication, authorization, and data integrity, energy-efficient security implementation isone of the major concerns for the wide adaptation of IoE [15, 16]. As IoE systemsare typically portable and energy and/or hardware-resource limited, and thus requirelow-complexity and energy-efficient implementation of security protocols in the hard-ware which would work on its own. This exposes IoE systems to a number of attacks,like frequency prediction, replay, denial-of-service, and eavesdropping. These attackscan compromise the system security, in terms of its confidentiality, privacy, and dataintegrity. without much human-intervention [55, 56]. Several cryptographic mecha-nisms and protocols have been proposed and successfully implemented in conventionalsystems [15, 55–58] without any stringent energy, cost, speed, memory, or comput-ing resource restrictions. There are a number of energy-efficient implementations ofcryptography in sensor systems [59, 60], but they are relatively easy to compromise.Physically unclonable functions (PUFs) are among the potential solution to datasecurity and counterfeiting problems [61, 62]. On-chip security can be implementedduring chip production utilizing chip integration techniques. A physically unclonablefunction (PUF) refers to a structure’s physical characteristic that is usually easy to mea-sure but hard to model or predict [62, 63]. Instead of storing the secret key into a digitalsystem, a PUF-based security approach derives its keys from inherent natural featuresof the system. A PUF-based output behaves like a random function and is unpredictableeven for an attacker with physical access to the device. Furthermore, in contrast withconventional digital architectures, the PUF-based approaches intertwine cryptographyand sensor properties, making the attack to such systems more challenging. Various16types of PUFs, each with its challenge/response pair generation capability have beencategorized broadly into three categories in the existing literature [63, 64]. These areWeak PUFs, Strong PUFs, and Controlled PUFs.Weak PUFsThey have a small number of challenge/response pairs. The response RC to a givenchallenge C is used to derive a secret key, which is never shared with anyone in public.Once an attacker gains full access to the physical device, all the challenge/responsepairs can be modeled in a short time and the security of the device can be compromised.Some common Weak PUF designs are include SRAM PUF [65], Butterfly PUF [66],and Coating PUF [67].Strong PUFsThey have a complex hardware mapping to generate challenge/response pairs in a waythat makes it hard for the adversary to easily predict their behavior in a short time.Some applications of Strong PUFs are in device authentication [68] and key formation[69]. Typical security features of Strong PUFs are:(a) Impossible to be cloned or physically duplicated. This means it is impossible todesign a PUF with same physical imperfections that are originally present in the PUFto be cloned.(b) The challenge/response pairs generated by the PUF should come in large numbers,making it difficult for an adversary to launch a brute-force attack on the PUF to deter-mine the challenge/response pairs in limited time.(c) Even with known challenge/response pairs, if the distribution of the responsescomes from a polynomial distribution, the adversary will not be able to predict theresponses to a given challenge.17One strong-PUF design [70] described a physical one-way random function-basedoptical PUF. Although PUF-based, this design does not integrate a PUF into its chal-lenge/response system. Also, it requires a large external setup to validate the systemand it is difficult to integrate into a resource-constrained sensor circuit. In [71], the au-thors proposed an Arbiter PUF (APUF) implementation that uses the XOR of responsesfrom the Arbiter PUFs implemented on the same chip to decrease the predictability ofthe responses. However, APUFs are susceptible to modeling attacks [64]. To addressthe problems in APUFs, several other PUF-based designs were introduced to counterthe modeling attacks; these are XOR PUF [72], feed-forward PUF [73], and ROPUF[74], which addresses stability issues with APUF outputs.Controlled PUFsControlled PUFs satisfy all the unique features of Strong PUFs, and, in addition,implement a controlled logic based on those features to formulate a more advancedfunctionality on the system, making it more secure. Recently, public PUFs, SIMula-tion Possible but Laborious (SIMPL) PUF, device-aging- and process-variation-basedsecurity primitives and public key protocols have been proposed that provide secu-rity by exploiting the difference between actual execution and simulation times [61].However, they generally require large computational efforts that result in high energyrequirements.State-of-the art PUF designs have been proposed in recent years for the resourceconstrained IoT systems. Authors in [75] present a way to use the fuzzy commitmenton unmanned IoT devices that utilizes two noisy factors from the inside and outside ofthe IoT device. This work is based on input and output noise data and is very differentfrom our proposed method which utilizes physical variations. Another recent work in18[76] utilizes the TERO-PUF metastable structure and is implemented in FPGA. In the[77] paper, it is argued with experimental results that model-building machine learningattacks can be successful in compromising security of FPGA-based PUFs. In [77] and[63], it has been described that controlled PUFs, where physical variations are usedto hide challenge-response pairs successfully from the attackers, are able to provide astable and long term security solution for the IoE systems.1.3.4 Prior Work on Policy-based Energy Management in SmartHome IoE EcosystemIn this section, we explain the existing state-of-art about the energy management in asmart home IoE ecosystem. The literature shows the various solutions as to how and towhat extent user policies and feedbacks affect the IoE system.Recent developments in the area of information and communication technologieshave provided an advanced technical foundation and reliable infrastructures for thesmart house with a home energy management system [78, 79]. Development of lowpower, cost-efficient and high performance smart sensor technologies have providedus with the tools to build smart homes [80, 81]. As a result, a service platform canbe implemented in a smart home to control the demand Response (DR) intelligently.This type of system should also give the users enough flexibility to input their choiceswhile deciding on control of home devices [82] This makes the system more coherent,user friendly and scalable. While different hardware, software, communication archi-tectures have been proposed and compared by their power consumption, performance,etc. [83–85], the cost of implementing the infrastructure like: hardware devices, soft-ware framework, communication interfaces, etc. are still high enough that hinder theprocess of implementing the smart home technology for ordinary users. Moreover, the19hardware and software architectures may not be able to handle the growing number ofsensors and actuators with their heterogeneity.Many authors have attempted to address the way to reduce peak energy basedon an agent-learning framework using multiple tools such as model predictive con-trol [86], particle swarm optimization [87], iterative dynamic programming based [88]and gradient-based methods [89]. However, these models are probabilistic and do notconstitute learning from interaction with the environment. Further, these models aremostly price based, where cost saving instead of user preferences is a predominant fac-tor. Some other solutions proposed in [90] and [91] consider Q-learning based agentinteraction system, however they target only particular appliances like air conditionersand LED lights.In [92], authors have proposed a fully-automated energy management system basedon the classical Q-learning based Reinforcement Learning (RL). The modelling is delaybased, where users have a way of inputting their energy requests via time-schedulingand the agent learns gradually with time to find the optimal solution. However, this ap-proach has several limitations. The author assumes mathematical disutility fuction andconsumer initiated energy usage. Finding disutility function for each home or residenceis costly and difficult and too much user interaction is not desired for a interoperableenergy management system. [93] focuses on applying a batch RL algorithm to controla cluster of electric water heaters. A more relevant work is reported in [94], whichproposes device-based Markov Decision Process (MDP) models. It assumes that theuser behaviour and grid control signals are known. However, these assumptions arenot realistic in practice. In [95], authors use a discrete-time MDP based framework tofacilitate the use of adaptive strategies to control a population of heterogenous thermo-statically controlled loads to provide DR services to the power grid using Q-learning.20Again the application here is specific to load controlled by ambient temperature.1.4 Research Focus and GoalsIn this section, we summarize the inferences and shortcomings from the existing liter-ature to clarify the focus of our thesis. The summary is drawn with respect to variouslayers of energy-efficient architecture of IoE infrastructure.On Optimal Energy Control in Wireless Sensor Networks Based IoE SystemsAs evident from the existing literature, achieving energy savings through battery replen-ishment and traffic dynamics optimization in a network power model of sensor systemsis an active research problem. The shortcomings of the existing literature which moti-vated us to provide solutions to address them in our thesis are oulined as below:• Network lifetime and utility formulation in the existing work neither takes intoaccount the energy consumption model nor the system performance trade-offwith lifetime increase.• Joint lifetime- utility-rate-reliability approach for WSNs in state-of-art incorpo-rates a generic error coding processing power model without re-transmission en-ergy requirements or energy savings due to enhanced error correction capability.• The existing work assumes the battery energy to be zero at start. This assumptionwill not work for the scenario of an IoE system which contains rechargeablebatteries.• The utilization of the network varies with listening power of the receiver block.This is a major energy consuming block and whose analysis have been missingfrom the existing literature.21• Existing solution models address the needs of a narrow class of applications inspecific areas, and are not suitable for a broad range of applications.On Energy-efficient and Distributed Data-Aware Routing and ClusteringProtocolIoT and heterogeneous WSNs systems are similar in being equipped with sensors, basestation (data gathering and decision making node) and wireless transceivers. But IoTsystem is more diversified in involving some notable variations like interaction betweenmultiple protocols, sensing systems having varied energy values, asynchronous eventdriven processing and gateway node in between sensors and BS to route data moreefficiently. Moreover, due to the evolution of active RFID tags [96] with reading ca-pability in the range of meters and various energy harvesting mechanisms [43, 97],prudent techniques in IoE systems using them are better equipped to handle the energyefficiency and network lifetime problem.On Energy-Efficient, Security design for IoE InfrastructureAs evident from the existing literature, design of energy-efficient and resource-optimizedsecurity system is a challenge for IoE systems. We address the shortcomings of the ex-isting literature to design such a energy-efficient security system in Chapter 4. Theinferencs and challenges from the existing literature are oulined below :• The existing security solutions for the IoE systems focuses on incorporating soft-ware oriented solutions. This demands extra hardware resources for the alreadyresource constrained systems.• The PUF solutions for the IoE systems focuses on FPGA-based system im-plemetation and is not an optimised solution for a broad range of IoE systems.22• Other existing solutions have minimal circuitry implementation but compromiseon the security and energy-efficiency problems.On Policy-based Energy Management in Smart Home IoE EcosystemAs evident from the existing literature, peak energy demand reduction by maximizinguser convenience in a smart home based IoE ecosystem is the major goal of Chapter 5.The smart home system is a broad application scenario for an IoE implementation. Theexisting literature lacks a comprehensive IoE system analysis. Specifically,• Models proposed in state-of-art literature are probabilistic and do not constitutelearning from interaction with the environment. Further, these models are mostlyprice based, where cost saving instead of user preferences is a predominant factor.• The cost of implementing the infrastructure like: hardware devices, softwareframe- work, communication interfaces, etc. are still high enough that hinder theprocess of implementing the smart home technology for ordinary users.• Security is a major issue in the successful implementation of an IoE system,analysis and model of which is missing from the smart home systems.• Time-of-Use (TOU) models of smart meters in the existing technology mostlyhelp the local distribution company and in order to take advantages of the TOU,each household has to adopt a change in the use of the appliances which maycause signicant discomfort to the consumers.1.4.1 Broad Goals of the ThesisTo fulfill the shortcomings of the existing literature, here we broadly define the objec-tive of our thesis.23• Data is generated at a rapid pace and nodes of an IoE system are diminishing insize. The energy resources are insufficient with decreasing size of nodes, increas-ing number of nodes, volume of data and demand for embedded security. Hence,the major goal of this thesis is to fill the energy gap required for an IoE systemas depicted in Fig. 1.3.• The prime objective to fulfil our goal is to find an energy efficient model imple-mentation which would consume the least amount of hardware resouces whilemaintaining a high quality of service for the end users.• To analyze the effect different techniques such as error control coding, wirelessenergy harvesting and event driven data listening, has on the IoE system. Andvalidating their effects through simulations and experiments.• Analyze and validate through design and meaurements, the effect of designinga energy-efficient security block for the IoE system. As this is one of the mostimportant blocks for the successful implementation of the IoE system, it is im-perative to analyze this blocks through the trade-offs of energy and security.• Testing and validating the designed energy-efficient model for the IoE systemthrough a broad application scenario is also incorporated into the objective. Thisis to give the users a viable and practical criteria along with its pros and cons fortheir own implementation of the system.1.4.2 Key Contributions and ResultsThe contributions of the thesis are described in this section for each chapter whichfollows our broad goals.24Optimal Energy Control in Wireless Sensor Networks Based IoE SystemsOur work in Chapter 2, focuses on solving the research problems mentioned above.We achieve our goal of increasing network lifetime through incorporating a wirelessenergy harvesting, error correction coding and wake-up-radio model into our systemwhile maintaining the quality of service requirements. We substantiate our systemthrough thorough simulations of various network lifetime-utility trade-offs. The detailsof our objectives are as follows:• We solve the data-utility lifetime trade-off problem by taking an approximatedlifetime function as well as the energy harversting, wake up radio duty cyclingand retransmissions into the utility function. This solves the problem of incor-porating a proper energy model for the system. This also focuses on reducingthe reciever power (the major power hungry block of sensing system) throughwake-up radio based duty cycling approach. Through a system parameter vari-ation in the simulation of data-utility lifetime trade-off problem, we provide theuser more flexibility in chosing the appropriate trade-off for a broad range ofapplications.• We incorporate a redundant residue number system based error correcting tech-nique and compare it with ARQ and Bose-Chaudhuri-Hocquenghem (BCH) tosolve the problem of achieving better retransmission rate, thus enhancing energysavings of the netowrk. Innovatively, the packet error rate and delay are beingincluded while computing lifetime and performance of the sensor network. Thissolves the re-transmission problem in the existing literature and through simu-lation the error-correction coding schemes’ network lifetime enhancing benefitshave been established in common sensor nodes.25Energy-efficient and Distributed Data-Aware Routing and Clustering ProtocolOur main solutions with respect to the shortcomings in the existing literature are asfollows:• The system is distributed in two-levels based on their initial energy as normalnodes with standard battery energy and advanced nodes with a times more energythan normal nodes [54]. We use the RFID tagging and reading mechanism toreduce the energy consumption during the cluster head (CH) selection phase tillall the advanced nodes (also called gateway nodes) have their energy exhausted.Thereby prolonging lifetime of the network.• We validate data awareness by dividing the sensor based on urgent and regulardata demand and switch nodes between high/low power state based on data re-quirement at the user side. The solution expects to save energy in the nodes andimproving battery life.• We additionally incorporate RF energy harvesting for normal nodes with a powermanagement unit (PMU) to further improve network lifetime.Energy-Efficient, Security design for IoE InfrastructureOur solution described in Chapter 4 falls in the controlled-PUF category which tries toprovide long term energy efficient secure solution for the IoE systems. The focus ofour thesis is in the energy efficiency and minimal resource design. Thorough measure-ments and testing leads to a Strong PUF design with integrated control logic to furtherconsolidate the security of the system. Specifically, our objectives are summarized asbelow:26• Our approach focuses on finding a challenge/response pair to authenticate thesystem with minimal circuitry addition to the already resource constrained sys-tem.• Energy efficient implementation is the main focus of our approach. Hence a so-lution is proposed which is hardware based instead of traditional software basedsolutions.• Rather than implementing complex computations and hardware circuitry, we fo-cus on building a simple circuit which provides the desired security solutions.Policy-based Energy Management in Smart Home IoE EcosystemIn summary the contributions addressing the issues described in Section 1.4 are asfollows:• User interface : Using a node-red development framework1 and message queuetelemetry protocol secure broker, a user interface has been designed. It incor-porates intelligent energy management capability and provides user input op-tions. Temperature control of appliances, operation rescheduling and On/Offcommands are initiated through the interface.• Peak demand reduction : Using the proposed HEMaaS methodology, a rewardmatrix is generated for each peak reduction threshold. There are four peak re-duction thresholds considered in Chapter 5: 5%,10%,15% and 20%. Based onthe user convenience suitable load reduction decisions are obtained.1Node-RED is a web-based programming tool for wiring together hardware devices, APIs and onlineservices. [Online] Available :• Fault tolerance and user privacy : Taking different random combinations ofrobustness measure, it has been shown how the user convenience is affected whenuser privacy is compromised and system has hardware fault.• Energy saving and Carbon-footprint reduction : The energy savings and car-bon emmission reduction has been shown for a community of 85 houses over ayear.1.5 Thesis OutlineBelow, we summarize the achieved solutions for the thesis objectives in different chap-ters of the thesis:• In Chapter 2, we formulate and solve a joint maximization problem of system per-formance (measured by data utilization) and lifetime for wireless sensor network.Apart from throughput, packet loss and retransmissions and data utilization of anetwork also has a major impact on the performance of a WSN system. Retrans-missions affects the throughput of the system depending on the amount of packetloss a network suffers in a given time slot. Data utilization for a node is depen-dent on the time frame in which the node is active. Therefore packet loss anddata utilizations are incorporated in the system model to provide a more realisticdata loss and utilization model for the WSN system. As energy is scarce resourcefor a WSN system, energy harvesting is adapted in the system model to increaseits lifetime. Energy harvesting is dynamic and varies as to how can be harvestedin each time slot. We model the harvesting as a stochastically varying Gaussiani.i.d process. The problem throws challenges in finding an optimal solution as thetime-variation combined with retransmissions, packet loss and harvesting makes28it complex. We, then provide a distributed solution to the problem by solving thedata-utility and network lifetime separately. We consider retransmissions as dis-crete, packet loss is varied as the system utility and the optimal energy is foundout as a function of utility and lifetime of the network.• In Chapter 3, we have proposed a Data Aware Energy Efficient distributed Clus-tering protocol for IoT (DAEECI) by saving cluster head (CH) selection energyusing active RFID tags, cutting processing energy by incorporating data aware-ness factor in the system and improving lifetime by inculcating RF energy har-vesting. The system is distributed in two-levels based on their initial energy asnormal nodes with standard battery energy and advanced nodes with a times moreenergy than normal nodes. We use the RFID tagging and reading mechanism toreduce the energy consumption during the CH selection phase till all the ad-vanced nodes (also called gateway nodes) have their energy exhausted. Therebyprolonging lifetime of the network. We propose data awareness by dividing thesensor based on urgent and regular data demand and switch nodes between high-/low power state based on data requirement at the user side. The solution expectsto save energy in the nodes and improving battery life. We additionally incorpo-rate RF energy harvesting through a power management unit for normal nodes tofurther improve network lifetime. Our simulation depict substantial improvementin lifetime of network and data delivery to the base station.• In Chapter 4, we propose an IoT sensor security scheme that utilizes a physicallyunclonable function (PUF) of the sensor. As a proof of concept, we present theapproach in the context of silicon photo diodes and use their dark current vari-ations as a PUF. The challenge used for system authentication is generated by29quadratic residues. In an effort to build a system prototype, we measure the darkcurrent of photo-diodes in terms of noise and energy consumption, in order toidentify an optimal configuration of the circuit. A prototype PUF circuit of thesensor node incorporating the current amplification circuitry was designed andtested to prove the feasibility of dark current measurements in a portable envi-ronment. We have proposed, implemented, and tested an authentication protocolusing PUF and the quadratic residues. We have also proposed an asymmetricdigital signature-based encryption scheme, using the PUF response-generatedprivate key, and simulated it using parameters of the PUF circuit and the au-thentication protocol. Our approach is validated by using measured, simulated,and analyzed the currents, adversary attacks, and energy requirements, to val-idate the approach. This concept can be extended to IoT applications that usealternative types of sensors (beyond photo-diodes), as long as the sensors exhibitrandom-like physical property variations.• In Chapter 5, a new method named as Home Energy Management as a Service(HEMaaS) is proposed which is based on neural network based Q-learning algo-rithm. Although several attempts have been made in the past to address similarproblems, the models developed do not cater to maximize the user convenienceand robustness of the system. Here, we have proposed an advanced Neural FittedQ-learning method which is self-learning and adaptive. The proposed methodprovides an agile, flexible and energy efficient decision making system for homeenergy management. A typical Canadian residential dwelling model has beenused to test the proposed method. Based on analysis, it was found out that theproposed method offers a fast and viable solution to reduce the demand and con-30serve energy during peak period. It also helps in reducing the carbon footprintof residential dwellings. Once adopted, city blocks with significant residentialdwellings can significantly reduce the total energy consumption by reducing orshifting their energy demand during peak period. This would definitely help IoEnetwork administrators to optimize their resources and keep the tariff low due tocurtailment of peak demand.• In Chapter 6, summary and concluding remarks are provided and possible futureresearch directions are discussed.31Chapter 2Optimal Energy Control in WirelessSensor Networks Based IoE SystemsIn this chapter, we formulate and solve a joint maximization problem of system per-formance (measured by data utilization) and lifetime for wireless sensor network. Thepacket loss and data utilizations are incorporated to provide a more realistic data lossand utilization model for the WSN based IoE system. As energy is scarce resource fora WSN system, energy harvesting is adapted in the system model to increase its life-time. We model the harvesting as a stochastically varying. Contrary to articles [44–46],our model assumes that the battery starts with a initial energy and the network opera-tions has to be sustained using harvesting and wake up radio (WUR), using harvestingfrom ambient RF energy rather than using a solar energy harvester which needs ex-tra circuitry. The overall problem throws challenges in finding an optimal solution asthe time-variation combined with retransmissions, packet loss and harvesting makes itcomplex. We, then provide a distributed solution to the problem by solving the data-utility and network lifetime separately. Motivated by the emerging concept of Green32Wireless Sensor Network (GWSN) in which the lifetime and throughput performanceof the system is maximized while minimizing the carbon footprints, our goal is to buildan sustainable WSN system by supplying adequate energy to improve the system life-time and providing reliable/robust transmission without compromising overall qualityof service.The rest of this chapter is organized as follows. System model formulation is de-scribed in Section 2.1. In Section 2.2, we propose the WEH and WUR schemes forWSN system. In Section 2.3, we formulate the joint utility-lifetime trade-off prob-lem and formulate a distributed solution based on subgradient method and Section 2.4shows our simulation plots.2.1 System Model and Problem FormulationWe consider a network of non-mobile and identical sensor nodes denoted by N. Sen-sor nodes collect data from the surrounding information field and deliver it to the sinknode/collector node denoted by S. As in [98], sensors communicate either in an uni-formly distributed ring topology or randomly in a multi-hop ad-hoc topology. We as-sume that the sensor devices in an WSN system are transmitting over a set of links L.We model the wireless network as a {edge, link} connectivity graph G(Z,L), where theset, Z = N∪S, represents the source and sink nodes. The set of links, L, represent thecommunication link between the nodes. Two nodes i and j are connected if they cantransmit packets to each other with i∈N and j∈Ni. Fig. 2.1 shows a sample connectiv-ity graph with three sensor nodes (i1, i2, i3), one sink node (s1) and six communicationlinks (l1, l2, l3, l4, l5, l6). The communication between node i1 and s1 is a single-hoptransmission whereas between i3 and s1 denotes a multi-hop transmission with nodei2 acting as relay for data of node i3. The set of outgoing links and the set of incom-33Figure 2.1: Connectivity graphing links corresponding to a node i are denoted by O(i) and I(i) respectively. Thus, inFig. 2.1, O(i2) = (l3, l6) and I(i2) = (l4, l5). Table 2.1 delineates the parameters usedfor the analysis of our scenarios in Chapter 2.Table 2.1: Notations usedSymbol Description Symbol Description Symbol Description‖.‖∞ ∞-norm ET X Transmit energy [J/bit] Pe Packet Loss Rate‖.‖p p-norm ERX Receive energy [J/bit] Ps Packet Success RateN Set of Sensor Nodes EPR Processing energy [J/bit] Pb Bit error rateS Set of Sink Nodes ESN Sensing energy [J/bit] LP Length of packeti Outgoing Sensor Node PLS Ideal Listening power [W] E(T ) Expected no. retransmissionsj Incoming Sensor Node EB Battery energy of Sensor h Number of hopsri j Rate of Information Flow PH Harvested power GF(2b) Galois Field of b-bitsRi j Source rate W′U Wake-up-radio on-off signal U(.) Utility functionCl Capacity of Link γ Path loss exponent α System design parameterTnetwork Lifetime of Network d Communication distance ε Lifetime approx. constant2.1.1 Routing and Flow ConservationWe model the data transmission rates and routing of data in the network using flowconservation equation. Let ri j denote the rate of information flow from nodes i tonode j. Let Ri j denote the total information rate generated at source node i to becommunicated to sink node j∈Ni. It is assumed that no compression is performedat the source node and data transmission is lossless. Thus satisfying flow conservation34constraint, we have the flow equations at the nodes for time slot t as∑j∈Ni(r ji(t)− ri j(t))= Ri j(t),∀i ∈ N, j ∈ Ni (2.1)The maximum transmission rate of a link is also known as its capacity Cl . For a giventransmit power of node and bandwidth of the channel, this value is fixed and is a uppedbound of ri j as 0≤ ri j ≤Cl .2.1.2 Energy Cost ModelThe network lifetime is dependent on the power consumption of the sensor node Pi peractive duty cycle slot Ti of a node. This involves the combined operations of sensing,processing and communication (receive/transmit). If a sensor node goes out of the ser-vice due to energy deficiency, then all the sensing services from that node are affectedtill the battery is replaced. Radio transceiver is the one of the most power hungry blockof a sensor device. The communication energy per bit per time slot Ecomm(t) consists ofERX(t) (receiver energy per bit per time slot) and ET X(t) (transmitter energy per bit pertime slot). The computation energy includes EPR(t) (processing energy per bit per slot)and ESN(t) (sensing energy per bit per time slot). Let, EB(t) ≥ 0 is the total residualenergy left in a sensor node operated by battery at time slot t. The power consumptionin a time slot t is modeled asPi(t) = ∑i∈N, j∈Niri j(t)ET X(t)+ ∑i∈N, j∈Nir ji(t)ERX(t)+ ∑i∈N, j∈NiRi j(t)EPR(t)+ ∑i∈N, j∈NiRi j(t)ESN(t)+ ∑l∈O(i)PLS(t)(2.2)From the communication energy model in [32], we modify our transmitter energy fortransmitting one bit of data from i∈N to j∈Ni across distance d as35ET X = a1+a2 ·dγi j (2.3)Where γ is the path loss exponent varying from γ∈[2,6], a1 and a2 are constants de-pending on the characteristics of the transceiver circuit.2.1.3 Packet Loss and Data Re-transmissionAs often as the packets are failed to be delivered to the sink node, the re-transmissionconsumes extra energy from the battery source of the sensor node, thereby decreasingits lifetime substantially. Therefore, a fundamental approach to reduce the packet loss isnecessary to be integrated together with upper layer protocols to deliver reliable WSNmanagement. Thus, we propose to use the approach of Error Correction Coding (ECC)to improve transmission reliability. ECC adds redundancy to improve the transmissionreliability thereby reducing the efficiency, it is still a more preferable solution, becauseit helps to improve both reliability and latency. We derived a error coding scheme onthe theoretical basis of Redundant residue number systems (RRNS) which have beenintroduced in [99, 100]. The performance is evaluated in terms of the packet error rateand compared with the state of the art Automatic Repeat reQuest (ARQ) scheme that iswidely used in IEEE 802.15.4 radio. A preliminary analysis has been done in [19] thathas been extended into our system model in this Chapter.Analysis of packet error in ARQ schemeIn ARQ scheme, data is decoded by cyclic redundancy check (CRC) codes and theerroneous data is re-transmitted from the sender. Here we consider stop and wait ARQmethod. Assuming the ACK bits are received without error, the packet error rate of theARQ scheme is given byPARQe = 1− (1−Pb)LP (2.4)36where LP is the packet length of the payload transmitted in a single transmission, Pb isthe bit error rate. Pb for sensor nodes in IEEE 802.15.4 is given in [101].Analysis of packet error in ECC schemesFor BCH and RRNS codes, let us assume that we use a (n,k,e) e-error control methodwith n− k redundant bits appended to the k-data bits. We further assume that thetransmission of the packets between the sensor node and sink node is in bursts of n-bit data. Therefore, the packet loss rate at the sink node is given asPECCe = 1−1− n∑i=e+1 niPib(1−Pb)n−i⌈LPk⌉(2.5)Where d.e is the ceiling function. We assume that due to poor channel conditions andinterference, when a packet is unsuccessful in reaching its destination, it is counted asloss of packet and a re-transmission is required. The packet is assumed to be success-fully delivered when the acknowledgement (ACK) for the delivery is received. Thus ittakes one complete trip for the packet to be assured as successfully delivered. Let Pebe the probability of an event where the packet is lost in being delivered from sensorto sink or the ACK failed to reach the sensor from sink. Thus, for a single hop theexpected number of re-transmissions is given by [32]E(Tr) =1(1−Pe) (2.6)37Where, Pe is the packet loss rate of ARQ or ECC schemes. Accordingly, packet lossrate for end-to-end in a h-hop scenario is given asE(Tr,h) =h(1−Pe) (2.7)Lemma 1. Let Pe be the probability of an event where the packet is lost in being deliv-ered from sensor to sink or the ACK failed to reach the sensor from sink. Thus, for asingle hop the expected number of re-transmissions is given byE(Tr) =1(1−Pe) (2.8)Where, Pe is the packet loss rate of ARQ or ECC schemes. Accordingly, packet loss ratefor end-to-end in a h-hop scenario assuming each node transmission is independent ofthe other as per the TDMA based MAC protocol.E(Tr,h) =h(1−Pe) (2.9)Proof. See [32].Redundant residue arithmetic based error correction schemeA residue number system (RNS) is a non-weighted number system that uses relativelyprime bases as moduli set over GF (2b) [102]. Owing to the inherent parallelism ofits structure and its fault tolerance capabilities, shows fast computation capability andreliability. RNS is defined by a set of β moduli m1,m2, . . . ........mβ , which are relativelyprime to each other. Consider an integer data A, which can be represented in its residues38Γ1,Γ2, . . . ...ΓβΓi = A mod mi, i=1,2,....l (2.10)Θ=β∏i=1mi (2.11)The maximum operating range of the RNS is Θ given by (2.11). The correspondinginteger A can be recovered at the decoder side from its β residues by using the ChineseRemainder Theorem [102] asA =l∑i=1Γi×M−1i ×Mi (2.12)where Mi = Θ/mi and the integers M−1i are the multiplicative inverses of Mi and com-puted apriori. One common modulus set (2b−1−1,2b−1,2b−1+1) with a power of twoin the set makes it relatively easy to implement efficient arithmetic units. A redundantresidue number system (RRNS) is defined as a RNS system with redundant moduli. InRRNS, the integer data X is converted in β non-redundant residues and δ -β redundantresidues. The operating range Θ remains the same and the moduli satisfy the condi-tion m1 < m2 < . . . .... < mβ < mβ+1 < mβ+2 < . . . .... < mδ . RRNS can correct upto b(δ −β )/2c errors. If we consider the popular modulus set, mentioned above, andadd the redundant modulus (2b+1) to it, becomes the (2b−1−1,2b−1;2b−1+1,2b+1)RRNS with capability to detect one error, it is explained extensively in [102]. Since theChinese Remainder Theorem approach require processing large-valued integers, a suit-able method for avoiding this is invoking the so-called base-extension (BEX) methodusing mixed radix conversion (MRC)[103] that reduces the computation overhead byminimum distance decoding.Based on RRNS, we propose an online error detection and correction scheme for the390 2 4 6 8 10 12 14 16 18 2000. [dB]Packet Loss Rate (Pe)  ARQBCH (128,58,11)RRNS (128,60,32)(a) Packet loss vs. SNR for IEEE802.15.4 based sensor at different codingschemes.0 0.05 0.1 0.15 0.205101520Packet Loss RateE(T)  ARQBCH(128,57,11), RRNS(128,60,32)(b) Expected No. of Packet Re-transmissions vs. packet loss rate at differ-ent coding schemes for IEEE 802.15.4 basedsensor.Figure 2.2: Analytical results of different coding schemes for IEEE 802.15.4based sensor.GWSN systems. A parallel to serial converter changes A into its decimal representation.In a look-up-table (LUT), we store the modulus values of numbers 0− 9 and 10χ ( χ∈ 1,2, .......κ) with respect to the δ moduli (β non redundant moduli and δ -β redundantmoduli). All operations are performed in parallel modulo channels without the need oftransmission of information from one modulo channel to another. So, for l moduli, wehave δ modulo channels operating in parallel, all operations in each performs modulo ofthe particular modulus till δ . Finally, we append the respective MAC IDs of the sensordevices at the front end of each set of packet data and transmit it to the gateway/sinknode. RRNS Algorithm in [19] shows the decoding process at the sink node/gateway.It first receives the packet and tries to recover the data. After the recovery of the dataand the error moduli, it appends a 1-bit TRUE flag with the ACK signal and sends itto the sensor node to notify the reception of data, else it sends a 1-bit FALSE flag withACK to the sensor node signifying to resend the packet data again. The sensor nodein turn transmits the δ -β redundant residues again instead of sending the full n bits ofdata again.40Packet loss statistics for different error correction schemesWe perform an analysis to find out the packet loss rate of the IEEE 802.15.4 basedsensor. The systems signal to noise ratio is varied from 0dB to 20dB. The packet errorrate is generated for BCH (128,57,11) and RRNS (128,60,32). These values of n aretaken to correlate with the packet load of 133 bits (payload of 127 bits and 6 bits ofheader). From Fig. 2.2a, it can be inferred that ECC schemes provide approximatelya gain of 4 dB in SNR as compared to ARQ scheme for the same packet loss rate.This is equivalent to a power gain of around 2 watts, which is essential savings incase of energy constrained GWSN systems. RRNS code provides slightly better gainof around 2 dB, owing to its better error correction capability compared to BCH code.Accordingly, in Fig. 2.2b we plot the values of re-transmissions required for ARQ, BCHcodes, and RRNS codes. The plot depicts a similar nature as predicted in (2.5). As wecan see, simple ARQ scheme in a packet loss rate varying from 0 to 20% requiresexpected number of re-transmissions of ∼ 1 to 17, whereas expected number of re-transmissions in BCH and RRNS coding schemes is ∼ 1 to 4 . The figure of merit forboth BCH and RRNS shows average number of expected packet re-transmissions, evenfor a packet loss of 20% as ≈ 4, significantly outperforms the simple ARQ scheme.This can save a tremendous amount of energy leading to network lifetime enhancement.2.1.4 Problem Definition : Network Lifetime Maximizationthrough Energy Cost ModelThe processing energy EPR in (2.13) increases with redundancy P′ = (n− k)/k. There-transmissions consumes extra energy resources apart from the original transmis-sion which is mandatory, hence incorporating the expected number of retransmissions41E(Tr,hi) for hi-hops into (2.13), we get power consumption as in time slot tPi(Pe,hi, t) = ∑i∈N, j∈Niri j(t)ET X(t)(1+E(Tr,hi))+ ∑i∈N, j∈Nir ji(t)ERX(t)(1+E(Tr,hi))+ ∑i∈N, j∈NiRi j(t)EPR(t)(1+E(Tr,hi)P′)+ ∑i∈N, j∈NiRi j(t)ESN(t)+ ∑l∈O(i)PLS(t)(2.13)packet success rate Ps(t) affects the sample rate in the rate flow constraint as∑j∈NiTi∑t=1(ri j(t)− r ji(t)+Ps(t)Ri j(t))≤ 0, ∀i ∈ N, j ∈ Ni (2.14)Let Tnetwork be the total number of active duty cycle slots representing the life-timeof the network. We focus on maximizing the operation time of the whole network(Tnetwork) until the first node fails,Tnetwork = minn∑x=1Tx,(n ∈ 1,2, ....i) (2.15)The problem of maximizing the network lifetime can be stated asmaxt≥0,EB(t)>0Tnetworksubject toTi∑t=1(Pi(Pe,hi, t)− 1Ti ·EB(t))≤ 0,∑j∈NiTi∑t=1(r ji(t)− ri j(t)−Ps(t)Ri j(t))≤ 0, ∀i ∈ N, j ∈ NiET X = a1+a2 ·dγ , γ ∈ [2,6]0≤ ri j ≤Cl(2.16)In our model, we have considered a battery with a finite maximum capacity EBmax,where EB(t)≤EBmax. Further, due to hardware limitations the total power consumption42is upper bounded by maximum consumption Pmax (i.e Pi(t)<Pmax, ∀ j ∈ Ni,∀t ∈ Ti).Problem in (2.16) is not convex. By substituting s = 1/Tnetwork, we obtain a convexproblem.mins≥0ssubject toTi∑t=1(Pi(Pe,hi, t)− si ·EB(t))≤ 0,1≤ t ≤ Ti0 < EB(t)≤ EBmax,1≤ t ≤ TiConstraints in (2.16)(2.17)2.2 Wireless Energy Harvesting and Wake-Up RadioSchemeBlock diagram of a generic wireless energy harvesting (WEH) enabled sensor node isshown in Fig. 2.3a. As shown in the figure, the nodes consists a rectifier, transceiver(RX, TX), sensors and sensor interface, storage unit (rechargeable battery), power man-agement unit (PMU) 1 and the processor. An RF-to-DC converter (also known as recti-fier) constitute the core of the wireless energy harvesting unit. The rectifier is in chargeof converting the received RF power (by the receiver antenna) to a usable DC supply.This stable DC energy can be used to charge a battery and/or drive the electronic cir-cuitry of the node. The conversion from RF to DC comes with some energy loss inthe internal circuitry of the rectifier which quantified in terms of power conversion effi-ciency (PCE) of the rectifier. PCE being the ratio of the converted DC power to the RFinput power, has significant implications on the overall performance of the power har-vesting unit. This is further highlighted by reference to Friis free space equation whichgives the available harvested power by [104]: The PCE is optimized for a designated in-14.2.443put power which corresponds to an specific communication distance. For longer (thanoptimal) distances (d2i j), the rectified power abruptly drops. When a receiver node iis in the energy harvesting mode, the power harvested (PHi) from base station serversource in a time slot t can be calculated as followsPHi(t) =η ·PT X · |Hi(t)|2d2i j,1≤ t ≤ Ti (2.18)Where, η is PCE and Hi denotes the channel gain between between source and receiverat time slot t. As shown, the PCE is optimized for a designated input power (receivedform the antenna) which corresponds to an specific communication distance. Beyondthis optimal point, the rectifier provides sufficient energy for storage or to drive the nodecircuitry. However, for longer distances from the sink node, the rectified power abruptlydrops. In WEH-enabled nodes, PMU is in charge of managing the flow of energy to thestorage unit, node circuitry and to the main receiver (RX). Aside from high efficiency,other key performance metrics of a WEH unit include high sensitivity (i.e., ability toharvest energy from small levels input power), wide dynamic range (i.e., maintaininghigh efficiency for a wide range of input powers), multi-band operation (i.e., ability toharvest wireless energy from wireless transmissions at different frequencies). Extensivestudies exist in the literature investigating on techniques to improve the performance ofWEH unit [18, 104]. The design presented in [105] studies techniques to enhance theefficiency of WEH unit and a muliti-band approach to enable harvesting and differentfrequencies.In a wireless sensor node, although the receiver is practically called in to action onlywhen its service is required, it has to constantly keep listening to the communicationchannel for the commands from the sink node. This so called idle listening mode power44(a)(b)Figure 2.3: WEH-enabled wireless sensor node . (a) Block diagram of WEH-enabled sensor node, (b) Timing diagram.(PLS) consumption when integrated over the lifetime of the node makes the receiver asignificant energy consumer and is dependent on the amount of network utilized forgiven duty cycle. Let α ∈ (0,1) be the system parameter that defines the amount ofnetwork utilization. The amount of energy consumption modeled in terms of α in45(2.13) isPi(Pe,hi, t) = ∑i∈N, j∈Niri j(t)ET X(t)(1+E(Tr,hi))+ ∑i∈N, j∈Nir ji(t)ERX(t)(1+E(Tr,hi))+ ∑i∈N, j∈NiRi j(t)EPR(t)(1+E(Tr,hi)P′)+ ∑i∈N, j∈NiRi j(t)ESN(t)+ ∑l∈O(i)α(t)PLS(t)(2.19)In a wireless sensor node, the receiver unit despite not being the most power hungryblock, constitutes a significant portion of the overall energy consumption of the system.While similar to other building blocks, the receiver is practically called in to action onlywhen its service is required, it has to constantly keep listening to the communicationchannel for the commands from the sink node. This so called idle listening modepower consumption when integrated over the life time of the node makes the receiver asignificant energy consumer.An efficient solution to tackle the energy consumption during the idle listeningmode is duty cycling (also known as rendez-vous scheme) in which the receiver main-tains in deep sleep mode and only wakes up when there is a message to be received fromthe main transmitter (TX). There are three main classes of duty cycling, namely, syn-chronous, pseudo-asynchronous and asynchronous [106]. In the synchronous scheme,the transmitter and all the receivers pre-schedule designated time slots in which the re-ceivers wake up for to receive the commands and fulfill the transmission. Such schemeimposes considerable overhead in terms of complexity and power consumption in orderto establish time synchronization and leads to idle energy consumption if there is nodata to be received during the pre-scheduled time slots. In the pseudo-asynchronousscheme, the receivers wake up at designated time but a synchronization between thetransmitter and receiver is not required. In the asynchronous scheme which the mostenergy efficient approach among the duty-cycling classes, the receivers spends most of46their time in deep sleep mode and only wake up when interrupted by the transmitter.This interrupt message is generated by a wake-up radio (WUR). WUR is a simple andlow-power receiver which keeps listening to the channel and only wakes up the mainreceiver when the is a request for transmission to the associated node [107]. Fig. 2.3bschematically compares the energy profile of a conventional transceiver versus that of aWUR-enabled transceiver. As shown in the figure, the main receiver (RX) in the WUR-enabled transceiver is activated less frequently and only upon receipt of the wake-upcommand (WU) which is followed by the interrupt message generated by the WUR.Fig. 2.3b schematically compares the energy profile of a conventional transceiverversus that of a WUR-enabled transceiver. As shown in the figure, as compared tothe conventional method, the main receiver (RX) in the WUR-enabled transceiver isactivated only upon receipt of the wake-up command (WU) which is followed by theinterrupt message generated by the WUR. The infrequent activation of RX facilitatesa substantial energy conservation over the life-time of the wireless node. Obviously,WUR scheme is favourable only if the power consumption of the WUR is much smallerthan that of RX (i.e., PWUR << PRX in Fig. 2.3b).WEH-enabled nodes provide a good opportunity for a very efficient implementationof WUR [108]. Fig. 2.3, shows the block diagram of one such implementation for on-off keying (OOK) WU message. As shown in the figure, the rectifier block of the WEHunit can be re-utilized to perform as a simple envelope detector while also providingenergy supply for the rest of WUR circuitry [108].Let PCHi(t), denotes the cumulated harvested energy in all the slots of node i. Forsimplicity, we assume the harvested energy is available at the start of each interval t. Wealso assume that the battery has finite capacity and harvested energy can only recharge47Time SlotsTotal Energy0PHi(t)Optimal path (PHi(t))*PHi(t) -EBmaxFigure 2.4: Feasible energy bound for harvested energytill the maximum capacity of battery EBmax.PCHi(t) =t∑x=1PHi(x),(t ∈ 1,2, ....Ti) (2.20)PCHi(t) is a continuous increasing function that lies between points (0,0) and (Ti,PCHi(Ti))as shown in Fig. 2.4. The cumulative node energy PCi (t) for all (t ∈ 1,2, ....Ti) cannotbe more than PCHi(t). Using this constraint, the dynamic charging and discharging ofbattery can be modeled asEB(t+1) = EB(t)−Pi(t)+PHi(t)PCi (t)≤ PCHi(t),∀t ∈ 1,2, ....Ti(2.21)To find an optimal energy consumption (PCi (t))∗, we need to find the upper and lowerbound of consumed energy. (2.21) gives the upper bound on the consumed energy.Further, (PCi (t))∗ must satisfy that, the residual energy of nodes at all time slots i.e.(PCi (t))∗−PCHi(t) cannot exceed the battery maximum capacity EBmax, forms the lower48bound of (PCi (t))∗ . Thus the problem in (2.17), can be reformulated asmins≥0sisubject toTi∑t=1(Pi(Pe,hi, t)− si ·EB(t)−PHi(t))≤ 0,1≤ t ≤ Ti0 < EB(t)≤ EBmax,1≤ t ≤ TiPCHi(t)−EBmax ≤ PCi (t)≤ PCHi(t),∀t ∈ 1,2, ....TiConstraints in (2.15), (2.16), (2.18), (2.19) and (2.20)(2.22)2.3 Joint Utility & Network Lifetime Trade-off andDistributed SolutionSolving standalone maximization of network lifetime problem by varying the sourcerates will result in allocation of zero source rates to the node. Thus, it results in applica-tion performance of the system to be worst. Therefore, it is optimal to jointly maximizethe network lifetime with the system’s application performance. We associate the net-work performance with the utility function Ui(.). In [31], it has shown that each nodei∈N is related to a utility function and achieve different kind of fairness by maximizingthe network utility. Thus the utility is a function of the node source rate Ri j. Apart fromsource rates, packet success rate Ps also affects the overall system performance. Thus,the utility function has to be modified to accommodate the packet success rate and thepayload data efficiency as Ui(Ri j,Ps). Max-Min fairness maximizes the smallest ratein the network whereas the Proportional fairness favors the nodes nearer to the sinknode. As given in [31], by aggregating the utility, the network lifetime can be solvedin a distributed way with an approximated approach as Fεs (.) =( 1ε+1) · sε+1i . Thus, the49network lifetime problem in (2.22) becomesmins≥0(1ε+1)· sε+1isubject to constraints in (2.22), (2.19) & (2.14)(2.23)Using (2.23), we can now formulate a joint trade-off between maximizing utilityand network lifetime simultaneously. Our method differs from other approaches inChapter 1-Section 1.3 as we consider a more realistic scenario, incorporating path loss,fairness, packet loss statistics for error control schemes as well as energy harvesting anda event driven radio wake-up scheme. Thus the cross-layer joint maximization problemis given asmax(s,Ri j,ri j)≥0Ti∑t=1(α(t)∑i∈N∑j∈NiUi(Ri j(t),Ps(t))− (1−α(t))(1ε+1)· sε+1i)subject to constraints in (2.22), (2.19) & (2.14)(2.24)We have introduced a system parameter α∈[0,1] in (2.19). It gives the trade-off be-tween the utility and network lifetime. For α=0, the utility is zero and for α=1, networklifetime is maximum with worst application performance. The maximization objectivefunction is concave as U(.) is concave and network lifetime problem Fεs (.) is convex.We try to solve the primal problem via solving the dual problem [98]. We keep theexpected number of transmissions E(Tr,hi) in hops hi as constant and vary the rate ri j.The constraint set in (2.24) represents a convex set. According to slater’s condition forstrong duality, if the non-linear constraints are strictly positive, duality gap betweenprimal and dual problem is small. Thus the primal can be solved by solving the dualproblem and the desired primal variables can be obtained. The dual-based approach50leads to an efficient distributed algorithm.2.3.1 Dual ProblemTo solve the problem in a distributed manner, we formulate the Lagrangian in terms ofthe Lagrange Multipliers λ and µ by relaxing the inequality constraints in (2.24).L(λ ,µ,s,rij,Rij,U(Rij,Ps), t) =Ti∑t=1(α(t) ∑i∈N∑j∈NiUi(Ri j(t),Ps(t))− (1−α(t))( 1ε+1) · sε+1i)+ ∑j∈NiTi∑t=1λl(t)(ri j(t)− r ji(t)+Ps(t)Ri j(t))+ ∑i∈N∑j∈NiTi∑t=1µi(t)(Pi(Pe,hi, t)− si ·EB(t)−PHi(t))(2.25)The corresponding Lagrange dual function D(λ ,µ) is given byD(λ ,µ) = sups,rij,Rij,UL(λ ,µ,s,rij,Rij,U(Rij,Ps), t)subject to constraints in (2.22), (2.19) & (2.14)(2.26)The solution is given by F∗F∗ = minλ>0,µ>0D(λ ,µ) (2.27)51The dual problem can be decomposed further into two different subproblems D1(λ ,µ)and D2(λ ,µ).D1(λ ,µ) = max(Ri j,ri j)≥0 ∑i∈N ∑j∈NiTi∑t=1α(t) ·Ui(Ri j(t),Ps(t))+∑l∈LTi∑t=1λl(t)(ri j(t)− r ji(t)+Ps(t)Ri j(t))+∑i∈N∑j∈NiTi∑t=1µi(t) · (ri j(t)ET X(t)(1+E(Tr,hi))+ r ji(t)ERX(t)(1+E(Tr,hi))+Ri j(t)EPR(t)(1+E(Tr,hi)P′)+Ri j(t)ESN(t)+α(t)PLS(t))subject to ET X = a1+a2 ·dγ , γ ∈ [2,6]0≤ ri j ≤ClPCHi(t)−EBmax ≤ PCi (t)≤ PCHi(t),∀t ∈ 1,2, ....Ti(2.28)D2(λ ,µ) =−{ max(s,EB)≥0∑i∈N ∑j∈NiTi∑t=1µt (si ·EB(t)+PHi(t))+Ti∑t=1(1−α(t))(1ε+1)· sε+1i }subject to 0 < EB(t)≤ EBmax,1≤ t ≤ TiPCHi(t)−EBmax ≤ PCi (t)≤ PCHi(t),∀t ∈ 1,2, ....Ti(2.29)Subproblem D1(λ ,µ) is a rate control problem in the network and transport layer of thesensor networks. For all active links l ∈ L, we substituted ∑i∈Lwith ∑i∈N∑j∈Ni. SubproblemD2(λ ,µ) gives the bound on the inverse lifetime. The objective function of the pri-mal problem is not strictly convex in all its primal variables {s,Ri j,ri j}. The sub-dualproblems D1(λ ,µ) is only piecewise differentiable. Therefore, the gradient projectionmethod cannot be used to solve the problem. We use the subgradient method [98] tosolve the problem iteratively till a desirable convergence is reached.522.3.2 Solution to GWSN Distributed Algorithm and ItsConvergence AnalysisLet, {s∗(λ ,µ),R∗i j(λ ,µ),r∗i j(λ ,µ),(PCHi(t))∗,P∗s (t),P∗LS(t)} be the optimal solutions forproblems (2.28) and (2.29). The Lagrange multipliers (λl,µi) have cost interpretationto them. λl represents the link capacity cost and µi denotes the battery utilization cost ofsensor node i. The gradients∇λD(λ ,µ) and∇µD(λ ,µ) denote the excess link capacityand battery energy respectively. Problems D1(λ ,µ) in (2.28) represent the maximiza-tion of the aggregate utility of the network in presence of flow constraints and energyspent in the network. The network lifetime problem D2(λ ,µ) in (2.29) maximizes therevenue from battery capacities subtracting the lifetime-penalty function, resulting inreduction of lifetime.The procedure for solving the algorithm is outlined as follows:• Initialize all the inputs (ET X ,ERX ,ESN ,EPR,PLS,EB) and step sizes ϕτ ← 0.01, ψτ ←0.01/√τ , ε ← 20• Although the problem in D1(λ ,µ) and D2(λ ,µ) is convex, the solution is complexand difficult to implement due to the intricacies introduced by incorporation of optimalenergy consumption ((PCi (t))∗), packet loss (P∗s (t)) and WUR (P∗LS(t)). From (2.28)and (2.29), it is evident that (PCi (t))∗ is dependent on optimal lifetime (s∗i j) and samplerate (R∗i j). Therefore we take (PCHi(t))∗ as some function g of lifetime and sample rate.g(s∗i j,R∗i j) = f ((PCHi(t))∗) (2.30)• We model PCHi(t) w.r.t the channel gain Hi(t) distributed as i.i.d with mean 0. Oncethe optimal s∗i j,R∗i j is found, PCHi(t) is found using f−1(g(s∗i j,R∗i j)).• The packet success rate Ps(t) is varied ∈ [80,100] and system utility parameter α(t)and overall node utilization Ui(Ri j(t),Ps(t)) determines the optimal listening power53P∗LS(t).• Thus from all the previous assumptions mentioned above, the time coupling propertyof the node can be excluded and finding solution for limt→1λ (t),µ(t) would be good∀t ∈ (1,2,3, ....Ti).• The Lagrange multipliers can be updated byλl (t,τ+1) =[λl (t,τ)+ϕτ ∑j∈Ni(ri j(t,τ)− r ji(t,τ)+Ps(t,τ)Ri j(t,τ))]+,µi (t,τ+1) =[µi (t,τ)+ψτ ∑i∈N∑j∈Ni(Pi(Pe,hi, t,τ)− si ·EB(t,τ)−PHi(t,τ))]+(2.31)• From (2.31) , it can be seen that as the flow ri j exceeds the capacity of link Cl ,the link cost and node energy cost increases. Thus higher link and node-battery pricesresult in greater penalty in the objective function in (2.28) forcing source rates Ri j &flows ri j to reduce. Although higher node-battery cost (2.29) allow greater revenuefor the same increase in battery capacities (by increasing ’s’), there is a correspondingpenalty incurred due to the consequent lower lifetimes.Lemma 2. When ε→∞, the network lifetime Tnetwork determined by the optimal solu-tion s∗ of problem (2.24) approximates the maximum network lifetime of the wirelesssensor network.Proof. See Appendix AFurther, let us make the following two assumptions as below:• Assumption 1: Let Ui(Ri j,Ps) be defined as log2(Ri jPs) which is an increasing andconcave function, and its inverse and hessian exists.• Assumption 2: Hessian of Ui(Ri j,Ps) is negative semidefinite and rmini j ≤ri j≤rmaxi j .54Define L = max L as the maximum number of links that a sensor node uses. Let U = maxU′i (Ri j,Ps) and R = max ri j, be the maximum rate flow of the node when transmittinginformation from i→ j.Proposition 1. If the assumptions 1 and 2 above hold and the step size satisfies0<ϕτ ,ψτ<2L1/2U R. Then starting from any initial rates rmini j ≤ri j≤rmaxi j , & price λl,µi≥0,every limit point of the sequence {s(λ ,µ),Ri j(λ ,µ),ri j(λ ,µ)} generated byGWSN Algorithm,is primal-dual optimal.Proof. See Appendix BProposition 2. By the above distributed algorithm, dual variables (λl ,µi) converge tothe optimal dual solutions (λ ∗l ,µ∗i ), if the stepsizes are chosen such thatϕτ(i)→ 0,∞∑i=1ϕτ(i) = ∞,ψτ(i)→ 0,∞∑i=1ψτ(i) = ∞ (2.32)2.4 Simulation ResultsTo show the joint trade-off between maximizing utility and network lifetime in terms ofsystem parameter α , path loss γ , packet loss statistics {Pie},energy harvesting PHi , weconsider a WSN as shown in Fig. 2.5 with seven nodes distributed over a square regionof 100m × 100m. The node at the middle of the network is taken as the sink node andthe other six nodes are either source or source/relay nodes. Nodes {i1, i2, i4, i5} act assource nodes whereas nodes {i3, i6} act as source node to deliver its own data and relaynodes for delivering nearest neighbor’s data to the sink node. The parameters taken forthe simulation are depicted in Table 2.2. The value of ET X , {a1,a2} are chosen from[32] with γ=4. ERX and ESN are taken from [109]. Processing energy EPR is assumed55−60 −40 −20 0 20 40 60−60−40−200204060  i1i5i3i6i4i2i  : Sensor NodeS : Sink NodeS1Figure 2.5: WSN be same as the sensing energy ESN . Also, at start t0 the initial battery energy EB inall the nodes is taken as 1 J. We run our simulations till 500 iterations to get a desiredsolution for the system.Table 2.2: WSN simulation parametersParameter Description Value Parameter Description Valuea1, a2 Transceiver Constant 10−7, 0.1 ·a1 J/bit ESN Sensing energy 5 ·10−8 J/bitγ Path Loss Exponent 4 EPR Processing energy 5 ·10−8 J/bitε Lifetime Approx. Constant 20 PLS Idle listening power 1 mWERX Receiver energy 1.35 ·10−7 J/bit EB Battery energy 1 J2.4.1 Convergence PlotsTo show the convergence of our GWSN algorithm, we plotted in Fig. 2.6a, the con-vergence of source node rates for different sensor nodes with respect to the numberof iterations. We have chosen sensor node {i1, i3, i5, i6}, where {i1, i5} act as onlysensor nodes and {i3, i6} act as both sensor and relay node. The step size is taken asϕτ = 0.01, where τ is the index of iteration. It can be observed that the step size playsa vital role as it controls the magnitude of oscillations near the optimal solution. Thelarger the step size, the faster the convergence but with more variations near the pointof optimality whereas smaller step size reach a stable optimal solution with lesser fluc-tuations near the optimal. As predicted by our algorithm, sensor nodes that have lower56No. of Iterations (=)0 50 100 150 200 250 300 350 400 450 500Source Rate (Kbps)0100200300400500600700Sensor Node 1Sensor Node 5Sensor Node 6Sensor Node 3(a) Convergence of source node rates for dif-ferent sensor nodes with respect to the numberof iterations with α = 0.1.0 10 20 30 40 500.εAccuracy of Approximating Lifetime [0 − Lowest ; 1 − Highest]  (b) Error in measuring the lifetimewith respect to the lifetime approxima-tion coefficient.Figure 2.6: Simulation plots of convergence of GWSN algorithm.lifetime {i1, i5} are assigned higher rates, whereas nodes with higher lifetime {i3, i6}have lower rates being assigned to them. Fig. 2.6b shows the error in measuring thelifetime with respect to the coefficient ε .Errror in Approximating Li f etime =∣∣∣∣s− 1ε+1sε+1∣∣∣∣ (2.33)According to Appendix A, if the coefficient ε is large enough then the lifetime approx-imated by (2.24) is the maximum lifetime. Fig. 2.6b validates the point, as it can beseen that at ε = 10, we get less than 10% error in measurement of lifetime. For our570 24 68 1012 140.80.850.90.95100. Lifetime (seconds)Percent Utilization of NetworkSystem Parameter (α)(a) Network aggregate utility - lifetime trade-off withoutWER, WUR and ECC.0 24 68 1012 140.80.850.90.95100. Lifetime (seconds)Percent Utilization of NetworkSystem Parameter (α)(b) Network aggregate utility - lifetime trade-off withWER and without WUR & ECC.Figure 2.7: Simulation plots of network aggregate utility - lifetime trade-off fordifferent α .Algorithm, we have initialized the value of ε as 20 with less than 5% error in lifetimeprediction.2.4.2 Utility and Lifetime Trade-off with WEH and WURConstraintsThe impact of the system design parameter α(t) is shown in Fig. 2.7a, 2.7b & 2.7c.α(t) is varied between 0.1 to 0.9. The network utility is computed as (6∑i=1log2(Ri jPs))which is the aggregate utility of all the nodes not including the sink node s1. The aggre-580 24 68 1012 140.80.850.90.95100. Lifetime (seconds)Percent Utilization of NetworkSystem Parameter (α)(c) Network aggregate utility - lifetime trade-off with WER& WUR without ECC.0 24 68 1012 140.80.850.90.95100. Lifetime (seconds)Percent Utilization of NetworkSystem Parameter (α)(d) Network aggregate utility - lifetime trade-off withWER, WUR & ECC.Figure 2.7: Simulation plots of network aggregate utility - lifetime trade-off fordifferent α .gate utility have been normalized with respect to the maximum utility of the network.Fig. 2.7a shows that the network lifetime decreases and the utility increases as the in-crement of α . On the contrary, we can observe that as the weighted system parameterα decreases, the corresponding optimal network lifetime increases. It can be seen inFig. 2.7b that the lifetime increases to 8.5s from 4.5s. Fig. 2.8a shows the harvestedenergy profile from (2.18) for the farthest node in the network. Replacing the optimals∗i j,R∗i j in (2.30), PCHi(t) is found using f−1(g(s∗i j,R∗i j))as shown in Fig. 2.8b. Further,if wake-up radio scheme is applied with energy harvesting, the lifetime increases to59Time Slots0 50 100 150 200Energy Replenisment (W)00.511.52(a) Replenishment profile for har-vested energy.Time Slots0 20 40 60 80 100 120 140 160 180 200Energy (J)00.511.52(b) Energy resource allocation.Figure 2.8: Energy harvesting profile and allocated energy plots.∼10s as in Fig. 2.7c. The network utility of the system also increases to 0.87 withenergy harvesting and 0.97 with both harvesting and WUR. Hence, based on the de-sired performance, designer can chose the value of α and solve the set of equations foroptimal lifetime and source node rates.2.4.3 Impact of Error Control Coding on Performance andLifetimeFig. 2.7d shows the utility-lifetime trade-off with error coding applied. The systemlifetime is further increased as compared to Fig. 2.7a-(c), to 14s and the network ismore utilized at 91%. To visualize the impact of error coding on the performance of thesystem, we plot the network lifetime versus the packet loss rate Pie at α = 0.1. Fig. 2.9ashows the plot of network lifetime for different cases with packet loss rate varying from0 to 20%. For a packet loss rate between 10% to 20% ,the network lifetime increasesmore than 3 times with only energy harvesting and wake-up radio scheme. Whereaswith the coding scheme applied, it doubles further giving a 6 times improvement. Weevaluate the network lifetime of nodes {i1, i3, i5, i6}, where {i1, i5} act as only sensornodes and {i3, i6} act as both sensor and relay node. The network lifetime is shown inFig. 2.9b versus the system parameter α incorporating harvesting and coding at packet60loss rate of 20%. As expected from (2.25), the lifetime of node i1 is the least. Relayingof data from i5→i6 improves the lifetime of node i5. Nodes i3 and i6 have a hugeimprovement in their lifetime owing to their proximity to the sink node from wherethey harvest energy according to (2.18). Even though the total energy consumption isincreased, the harvested energy increase is sufficient enough to boost its lifetime.2.4.4 Effect of Energy Harvesting and Error Correcting Codes onTelosB Sensor NodeFor analyzing the effect of our error correcting codes performance on node lifetime,we have taken real time sensor energy cost from [110] for different sensors as shownin Table 2.3. The Table shows different commonly used sensing devices, their EPRand ESN energy cost normalized w.r.t communication energy Ecomm for common sensormote TelosB (TelosB is a IEEE 802.15.4 compliant sensor mote that runs a TinyOSoperating system with a CC2420 radio.). The battery power is taken as 9000 milli-Amphere-Hour (capacity of 2 standard 1.5− volt batteries used in sensors). Fig. 2.10ais drawn for RRNS, BCH, and ARQ for a packet loss rate of 20% showing the estimatedlifetime in days for the TelosB mote versus the total average power consumption Pifrom (2.13). For low power sensors i.e acceleration, pressure, light, proximity given inTable 2.3, TelosB motes lifetime increases by ∼1.7 times with BCH error scheme andmore than doubles with RRNS error scheme. Whereas for power hungry sensor suchas Temperature, the processing energy is higher, thus overpowering the effect of smallnumber of retransmissions in error coding schemes. One of the major overheads oferror correcting codes in addition to transmission and reception of redundant bits is thedelay associated with encoding and decoding of packets. Let us assume that tARQ is thetotal time required for sending the packets to the sink node and receiving an ACK back.61Table 2.3: Energy cost for TelosB mote w.r.t Ecomm = 1mWSensors Type & Model No. EPREcommESNEcomm Sensors Type & Model No.EPREcommESNEcommAcceleration (MMA72600Q) 0.044 0.000027 Proximity (CP 18) 0.047 0.267Pressure (2200/2600 Series) 0.044 0.00013 Humidity (SHT 1X) 0.043 0.4Light (ISL 29002 18) 0.047 0.00068 Temperature (SHT 1X) 0.94 1.5Further, if the decoding latency of a block code like (n,k,e) BCH is tBCHdec . From [101],the decoding latency is given bytBCHdec = (2ne+2e2)(tadd + tmult)⌈bbm⌉(2.34)Here, tadd and tmult are time required for additions and multiplications in GF (2b),and bm is the number of bits of micro controller used in sensor nodes. In an 8-bit microcontroller, tadd take one cycle and tmult takes two cycles as computation time. The num-ber of cycles depends on the frequency of the micro controller. RRNS codes of form((2b−1−1,2b−1;2b−1+1,2b+1)) needs tARQ(k/n) as the total time required for sending thepackets to the sink node and receiving ACK back. The decoding latency depends on thetotal additions and multiplications in the number of iterations(δβ). Depending on thevalue of β for each step there are 2β multiplications and β additions involved. Further,there are(δβ)number of moduli operations involved. Thus, the decoding latency forRRNS codes istRRNSdec =((δβ )tadd +(δβ )tmult)⌈ bbm⌉+((δβ )e)⌈ ebm⌉(2.35)To analyze the effectiveness of the coding schemes, we have plotted the delay in send-ing one packet of data versus the packet loss rate of 10% and 20%. If we take tARQ =50ms, from (2.34) and (2.35), delays of BCH(127,57,11) and RRNS(128,60,32) can620 2% 4% 6% 8% 10% 12% 14% 16% 18% 20%051015202530Packet Loss RateNetwork Lifetime (seconds)  α=0.1, Uncoded without EH & WURα=0.1, Uncoded with EH & WURα=0.1, Coded with EH & WUR(a) Plot of network lifetime versus packet loss rate.0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.90102030405060708090100System Parameter (α)Network Lifetime (seconds)  Sensor Node 1Sensor Node 5Sensor Node 6Sensor Node 3(b) Network lifetime of different sensor nodes versus system parameter α .Figure 2.9: Impact of ECC on lifetime of sensor found as tBCHdelay = tARQ ∗ (n/k)+ tBCHdec and tRRNSdelay = tARQ ∗ (n/k)+ tRRNSdec . TelosB has a16-bit microcontroller and its clock frequency is 8MHZ. Fig. 2.10b shows the delay inmilliseconds. It can be inferred that the coding schemes outperforms the ARQ shemein terms of total transmission delay. RRNS scheme has less delay compared to BCHcoding due to its better coding rate and faster decoding. It can also be seen that as thepacket loss rate increases, the delay gap between the three schemes increases. ThusRRNS has better performance in terms of lifetime improvement as well as lower delay631 1.5 2 2.5 3 3.5050100150Average Power Consumption [Pi]Network Lifetime [Days]  α=0.1, RRNS coded with EH & WURα=0.1, BCH coded with EH & WURα=0.1, uncoded ARQ with EH & WURα=0.1, uncoded ARQ without EH & WUR(a) Network lifetime prediction of ARQ, BCH and RRNS schemes.10% 20%0100200300400500600700800900Packet Loss RateDelay [ms]  ARQBCHRRNS(b) Transmission delay performance of ARQ, BCH and RRNSschemes.Figure 2.10: Impact of WEH & ECC on TelosB mote.64as the packet loss rate increases in bad channel conditions.2.4.5 Green Networking : Reduction in Carbon FootprintFor network to be green, the carbon emissions has to be reduced. The index of measureof carbon emissions is Xgr of CO2 per year. For each packet loss in the network causesthe data server station or the sink node to transmit back NACK to sensor node. Thetransmitting power (PST X) of the data station depends on the fuel type from which thestation derives its electrical power, can be either coal or gas. Thus value of X can beeither 870 or 370 [47]. (PST X) depends on the type of technology used. If we assumethat the sink node data station runs on the Long Term Evolution (LTE) network anduses the static micro cell topology with radius 100m. Then from [111] and [47], thecarbon footprint generated by sink node isFSCO2 = PST X · (ET,hi +1) ·8.64 ·10−3 ·X [KgCO2/Year]PST X =(PDT XµPACT X ,static+PSP,static)(1+CPS)(2.36)Where, the notations are described in Table 2.4. Apart from the sink node, the batteryis also responsible for generation of carbon footprint. Typical AA batteries used insensors have a end of life carbon emission of 4.3 KgCO2 per 30 batteries[112]. Thus,the carbon footprint [KgCO2/Year] generated by number of batteries used is directlyproportional to the total batteries used in a year (Byearu ) and is given asFBCO2 = Byearu(4.330)[KgCO2/Year], Byearu =365Tnetwork(2.37)650% 10% 20%0100020003000400050006000700080009000Packet Loss RateF CO2[kgCO 2/year]  ARQ BCHRRNSFigure 2.11: Plot of packet loss rate versus carbon footprintTable 2.4: LTE micro base station based sink node power model parametersParameter Description ValuePDT X Power consumed by sink node base station server 2 WµPA Power Amplifier efficiency 20%CT X ,static Static transmitted power 0.8PSP,static Static signal processing power 15 WCPS Power supply loss 0.11The total carbon footprint (FCO2) is therefore the sum of carbon footprints in (2.36)and (2.37). To show the effectiveness of using ECC, WEH & WUR, we plot FCO2for different packet loss rate of (0,10,20). We take X=370, the fuel for production ofelectricity as gas. The Tnetwork for different schemes ARQ, RRNS and BCH are takenfrom Fig. 2.10a at Pi=1mW. Fig. 2.11 shows the carbon footprint at different schemes.It can be seen that as the packet loss rate increases, the carbon footprint is tremendouslyreduced for RRNS and BCH. It is∼2.5 times lesser kgCO2 per year at 10% packet loss66and ∼4 times lesser kgCO2 per year at 20% packet loss. So, as the channel goes bad,the carbon emissions for normal scheme like ARQ increases tremendously, whereasincorporation ECC and harvesting the network becomes more greener.67Chapter 3Energy-efficient and DistributedData-Aware Routing and ClusteringProtocol3.1 IntroductionIn this chapter, we investigate a cross-layer approach that will provide interaction be-tween different layers in terms of energy efficient transmissions w.r.t data-awareness,energy harvesting and varied data-demand topology. The main pitfalls of the algorithmsdelineated in Chapter 1-Section 1.3 w.r.t energy consumption and network lifetime are,energy consumed in cluster head (CH) selection at each round, assuming nodes in al-ways ON state [113], and limited battery capacity of energy constrained sensors [114].Thus it is required to come up with a protocol specifically for IoT systems. Hence,we have proposed a distributed data-aware energy-efficient clustering protocol for IoT(DAEECI) which includes data awareness, RFID based CH selection and RF energy68harvesting using a Power Management Unit(PMU). The rest of the Chapter is orga-Figure 3.1: Data-aware RFID tag based IoT architecture.nized as follows. In Section 3.2, we describe IoT network model. Section 3.3 describesour DAEECI protocol. Section 3.4 analyzes our simulation results.3.2 IoT Network Model: Cluster Head Selection andEnergy Cost FormulationThe IoT system taken here is depicted in Fig. 3.1. The network is a random distributionof Ntot sensor nodes in a square area of side X meters and gateway nodes K used for dataaggregation and routing to BS. The nodes are differentiated based on their initial energyas advanced and normal sensor nodes. Advanced nodes have a times more energy thannormal nodes and are also known as gateway nodes (K), as they route data to the basestation. Thus the total nodes in the system are (Ntot +K). Each cluster has one gatewaynode based on minimum distance, as their CH. Let EB be the initial battery energy ofthe normal node. Let K be the number of distributed clusters that service all nodesand have one gateway node per cluster for data routing to base station. All the clusterheads send the aggregated data from sensor nodes they service to BS server. The BSserver is user driven based on data request from different user generated applications.69Let Etot−cls is the total energy of the clusters given asEtot−cls = Ntot ∗EB+K ∗EB ∗ (1+a) (3.1)3.2.1 Active RFID Tag Based Cluster Head AllocationThe cluster head selection is one of the major drawbacks of current clustering algo-rithms. In LEACH [50] and DEEC [53] algorithm, cluster head selection is divided intorounds, where each node randomly decides whether to become a cluster head based ona threshold Ti(s) computed by apriori decided probability pi.T (si) =pi1−pi·(r mod(1pi)) , i f Si ∈ G0 Else(3.2)where, r is the current round number, and G is the set of nodes that have not beencluster-heads in the last ni rounds (pi = 1ni ). Let the energy dissipated in a round(Eround) is adopted from the radio model in [53] asEround = L (Ntot +K)(Erx+Etx)+NtotEDA+Kεampd4toBS+Ntotε f sd2toCH (3.3)where, EDA is the data aggregation cost expended in CH, dtoBS is the average dis-tance between CH to BS, dtoCH is the average distance between cluster members toCH, L is the number of bits to be transmitted, εamp is the energy consumption of trans-mitter amplifier circuit, Etx is the transmitted energy consumed per bit and Erx is thereceived energy per bit, ε f s is the free space parameter. From (3), it can be inferredthat L∗(Ntot +K)∗(Etx+Erx) and L∗Ntot ∗ε f s ∗d2toBS are the energy consumed for CH70selection and routing data from nodes to CH, respectively.Therefore, to save the energy consumed in CH selection, we propose to incorporateactive RFID tags coupled to member nodes and a tag reader at the gateway node. Theconceptual topology is depicted in the expanded view of the WSN in Fig. 3.1. RFIDis an emerging automatic identification technology in which information is carried byradio waves. RFIDs are classified as passive, semi-passive, or active [96]. Passive RFIDtags function without a battery, has almost infinite lifetime but can operate in the rangeof only couple of centimeters. Whereas, an active RFID [115] can be read at distancesof 100 m or more, greatly improving the utility of the device, but it is battery poweredand has shorter life. The use of active tags with sensor nodes and a tag reader at thegateway will eliminate the need of choosing the CH till the gateway nodes are exhaustedof their energy. Nodes collect data from the environment and send them to the RFIDreader which in turn sends it to the BS. From the BS data are sent to the cloud in order toprovide it to the user through the services initiated by the cloud. With the evolution oftags like CC2650 SensorTag1 which operate with 2.4GHz transmission and supportingtechnologies such as Bluetooth, ZigBee and IPv6, it is feasible to incorporate the modelfor IoT WSN systems. Using our proposed method, as the tag reader reads the senseddata from the tags, computation for routing data to the CH is not required. The energyconsumed for CH selection becomes L∗ (Ntot ∗Erx+K ∗ (Etx+Erx)). This happens tillall the gateway nodes die in which case the routing follows energy consumption in (3)again.1Available[online]: Data Aware ProcessingSensors in IoT systems are not always active. There are two types of data request fromusers, one is periodic monitoring type of application such as warehouses and industrialcontrol and another is on demand processing such as home survielience, temperaturecontrol, smoke and water detectors. Thus data awareness of sensors is critical to itslongetivity. Sensors that service users periodically have to be in active state all thetime whereas the sensors sending data sporadically can be kept in sleeping state formost of the time. They can be woken up from sleep by asynchronous triggering ontheir pins when a certain threshold is crossed. An efficient approach to address this isduty− cycling, in which the receiver on-demand switches between active and sleep-ing states. Among the different categories of duty− cycling, namely synchronous,pseudo− asynchronous and pure asynchronous, latter provides the most efficient so-lution in terms of energy consumption [43]. In the asynchronous approach, the sen-sor device is in deep sleep mode and only wakes up when signalled by the BS or itsneighbouring devices through an interrupt command generated by a low-power wake-up radio (WUR). Let the transmitted energy consumption of sleeping nodes is only αpercent of Etx, where ζ≤α≤1, ζ is a small number close to 0. Let there be ns numberof sleeping nodes in the system. Therefore, the Eround is as followsEround = L (Ntot +K)(Erx+αEtx)+NtotEDA+Kεampd4toBS+Ntotε f sd2toCH (3.4)When α=1, all the nodes are awake and transmitting data read by the tag reader. Butwhen the data demand is low, the α value is small providing tremendous energy savingin the system.723.2.3 RF Energy HarvestingEnergy harvesting is a promising remedy to cope with the energy challenge. The re-cent technology trend in energy harvesting provides a fundamental method to prolongbattery longevity of sensor devices [116]. In RF energy harvesting (EH) circuit, theantenna receives the transmitted radio waves and converts the received RF energy intoa stable direct current (DC) energy source to supply the sensor device. Energy harvest-ing depends on the distance from the harvesting source. If the EH circuit is deployedon the sensor devices with a power management unit, it can harvest RF energy fromthe transmitted electromagnetic waves of the transmitter circuit of its own as well asnearby nodes, gateway nodes and BS [116][43]. In practice, the conversion from thereceived RF power to the usable DC supply comes with a certain amount of power lossin the matching circuit and in the internal circuitry of the power converter. The powerconversion efficiency (η) of the converter is the ratio of the generated usable DC out-put power to the input RF power. State-of-the-art RF-to-DC converters (also known asrectifiers) can achieve high η values, up to 70% or more [116]. η is an indication ofthe amount of harvested energy that is available for the sensor device. Here, we assumethat the energy harvested by the nodes vary randomly between 0<β≤1 of total har-vested energy EH(t). EH(t) is the maximum harvested energy and is taken as η timesthe battery energy per unit time t. At short range, it is possible to harvest a tiny amountof energy from a typical WiFi gateway router transmitting at a power level of 50 to 100mW. The RF energy which reaches the sensor node is efficiency η multiplied by theenergy harvest factor β and is approximately 0 to 5% of the total transmitted power ofan antenna for a distance in the range of tens of meters [117]. This amount of power isbest used for devices with low-power consumption and long or frequent charge cycles.Typically, devices that operate for weeks, months, or years on a single set of batteries73are good candidates for being wirelessly recharged by RF energy. In some applicationssimply augmenting the battery life or offsetting the sleep current of a microcontroller isenough to justify adding RF-based wireless power and energy harvesting technology.[117]3.2.4 Power Management UnitPower management unit (PMU) is an integral part of any energy harvesting system.PMU is in charge of controlling the storage of the harvested energy. It also manages thedistribution of the available energy among different consumers in an effort to maximizethe lifetime of the device while maintaining a high quality of service (QoS). We extendthe architecture of the PMU proposed in [118] to enable effective cooperation with theEH unit. The architecture proposed, is an event triggered/asynchronous scheme basedon the signal generated by a wake-up radio 2. The PMU architecture also detects/pre-empts the failure of a node in the event of energy deficiency.The detailed block diagram of the PMU for the EH sensor device is shown inFig. 3.2. The PMU starts its operation by a trigger signal generated by the WUR unitof WEH unit (INT ERRUPT ). The PMU first activates the main transceiver through(ON/OFF) and then sends a wake up signal (WAKE UP) to the sensing unit to start itsoperation. The sensing unit toggles the STOP/RUN to high, signifying the PMU thatit is in running mode. The REQ signal indicates the amount of energy required by thesensing unit. The signals BAT and SE indicate the amount of energy left in the batterydevice and the EH unit storage element respectively. Accordingly, the PMU activatesswitches SW1 through signal SENSE to fulfill the power requirements of the sensingunit. The sensor unit is in charge of sensing, data processing via a microprocessor (µp)2Section 2.274and finally transmitting them to a low-power transceiver based on Bluetooth, WiFi,IEEE 802.15.4, Zigbee, etc. The sensor devices require a minimum power of PDmin tooperate in sensing mode. When the energy in the battery device goes below a certainthreshold PT H < 1.5PDmin, the PMU sends a RECHARGE command to the storage el-ement by activating switch SW2 of WEH unit to charge the battery. When the energylevel of the device remains 1.1PDmin, the device sends out of service (OUS) commandFigure 3.2: Proposed power management the sink node, signaling that it goes out of the service till it recharges itself againto more than 1.5 PDmin. The sink node in turn sends a stop all service (SAS) signal tothe device. The sink node/gateway puts the device out of the sensing service loop butkeeps transmitting RF energy for harvesting. As the device is ready for service again,it sends a READY signal to the sink node which in turn gives resume all services (RAS)signal to the device.753.3 Data Aware Energy Efficient DistributedClustering Protocol for IoTIn this section, we present the detail of our Data aware energy efficient distributedclustering protocol for IoT (DAEECI) protocol. DAEECI uses similar function of initial(Etot−cls) and residual energy (Ei(r)) level as in [53] of the nodes to select the cluster-heads at each round. To avoid that each node needs to know the global knowledge of thenetworks, it estimates the ideal value of the network life-time to compute the referenceenergy (E¯(r)) consumed by a node in a round. Our DAEECI divides the problem intodifferent user cases based on data awareness (i.e. either α is 1 for periodic data sensingor 0≤α<1 for sparse data sensing) and percentage of gateway nodes present in the IoTsystem (K is high or low). The normal nodes are assumed to have their dedicated RFenergy harvesting circuit. The algorithm is summarized as in Algorithm 1.3.4 Results and AnalysisIn this section we provide performance evaluation of our DAEECI algorithm. We definea network area of 100 ∗ 100 m2. The simulation parameters are provided in Table 3.1.The performance metrics taken in the simulations are number of Alive nodes, Residualenergy of nodes and Packets sent to BS. We used Matlab for evaluating our algorithmwith other known protocols. In our scenario, we have evaluated the system with fourdifferent cases based on α and K for 10000 rounds. For all the cases, we assume thatthe advanced nodes (gateway nodes) have a = 3 times the more energy than the sensornodes. The cases are as follows:Case 1 : a = 3, Ntot = 100, K = 30, 0.8 ≤ α ≤ 1, noEh. Here, the data demand onsensor nodes is high with no energy harvesting present.Case 2 : a = 3, Ntot = 100, K = 30, 0.8≤ α ≤ 1, Eh. Here, the data demand on sensor76Algorithm 1: Data aware energy efficient distributed clustering protocol1 Initialize :2 Uniformly distributed region X*X.3 Ntot , K, EDA, Etx, Erx, ε f s, εamp, Eh(t), L.4 dtoCH=X√2∗K ∗pi , dtoBS=0.765∗X2.5 Start :6 The average energy of rth round is given asE (r) =1(Ntot +K)Etot−cls(1− rR)(3.5)where, R denotes the total rounds and is defined asR =Etot−clsEround(3.6)If nodes have different amounts of energy, pi of the nodes with more energyshould be larger than popt (optimum probability of choosing a cluster head).7pi =(1+a)poptEi(r)E(r) ,∑KEK (r)> 0poptEi(r)E(r) ,∑KEK (r)≤ 0(3.7)The energy dissipated in a round Eround , incorporating total cluster energy∑KEK (r), data awareness factor α and RF energy harvesting factor η is given as8 −→for ∑KEK (r) > 0Eround = L(NtotErx+K (Erx+αEtx)+NtotEDA+Kεampd4toBS)−NtotβEH(t) (3.8)−→for ∑KEK (r) ≤ 0Eround = LNtot (Erx+αEtx)+K (Erx+αEtx)+NtotEDA+Kεampd4toBS+Ntotε f sd2toCH−NtotβEH(t) (3.9)Thus we can find the lifetime of network R by putting (1), (8) and (9) in (6).9 End77Table 3.1: IoT simulation parametersParameters ValueNetwork Size 100x100 m2Sensor nodes Ntot in each Cluster 100Initial battery energy of nodes EB 0.5 JPacket Size L 4000 bitsEtx and Erx 50 nJ/bitε f s 10 nJ/bit/m2εamp 0.0013 pJ/bit/m4EDA 5 nJ/bit/signalpopt 0.1α and β rand(0,1)η 0.4nodes is high with energy harvesting present.Case 3 : a = 3, Ntot = 100, K = 30, 0.2 ≤ α ≤ 0.4, Eh. Here, the data demand onsensor nodes is low with energy harvesting present.Case 4 : a = 3, Ntot = 100, K = 50, 0.8 ≤ α ≤ 1, Eh. Here, the data demand on sen-sor nodes is high with energy harvesting present. Moreover, there are higher numberof gateway nodes present compared to previous three cases. Fig. 3.3 represents the0 2000 4000 6000 8000 10000x(rounds)0102030405060708090100y(nodes alive)LEACHDDEECEDEECDAEECI,,-high, no Eh, Medium KDAEECI,,-high, Eh, Medium KDAEECI,,-low, Eh, Medium KDAEECI,,-low, Eh, High KFigure 3.3: Number of alive nodes in an IoT system versus number of rounds.780 2000 4000 6000 8000 10000X(rounds)020406080100120Residual Energy of NodesLEACHDDEECEDEECDAEECI,,-high, no Eh, Medium KDAEECI,,-high, Eh, Medium KDAEECI,,-low, Eh, Medium KDAEECI,,-low, Eh, High KFigure 3.4: Total residual energy of nodes in the IoT architecture.0 2000 4000 6000 8000 10000x(rounds)0246810y(packets sent)#105LEACHDDEECEDEECDAEECI,,-high,no Eh,Medium KDAEECI,,-high,Eh,Medium KDAEECI,,-low,Eh,Medium KDAEECI,,-low,Eh,High KFigure 3.5: Total packets delivered to the base station server from nodes .number of nodes alive during the lifetime of the network. It clearly shows that by intro-ducing RFID based cluster selection and data aware processing, the lifetime improvessignificantly of the IoT network. LEACH and DDEEC perform poorly as all its nodesare dead by the end of 4000 rounds. Our DAEECI algorithms performance without en-ergy harvesting is comparable to the EDEEC algorithm. With the introduction of EH ,our method outperforms the EDEEC method as around 20% nodes are still alive at theend of 10000 rounds. It can also be inferred that the low data demand of the sensors incase of sparse sensor data requirement almost boosts up the lifetime of the system by79100%.Fig. 3.4 represents the sum of residual energy of all the nodes in the network. TheDAEECI algorithm incorporating RFID tags and data awareness again allows nodesto have higher residual energy compared to the LEACH, DDEED and EDEEC meth-ods. The EDEEC and higher data demand systems DAEECI almost perform similarly.Fig. 3.5 represents the packets sent to the BS from the cluster heads. The notable thingto infer is that low data demand reduces the amount of packet sent to the BS, whereasirrespective of the high data demand (high α), our algorithm still delivers more packetsto the BS than other state-of-the-art methods.80Chapter 4Energy-Efficient, PUF-Based SecurityDesign for Internet-of-Things (IoT)Infrastructure4.1 IntroductionAs IoT devices are deployed in unmonitored, unsecure environments, secure IoT sys-tems are needed, based on security algorithms, whose hardware implementation pro-vides a balance between security and energy efficiency, in order to also support com-munication among large number of IoT nodes. Embedded security implementation ofa physically unclonable function (PUF) is described in this chapter. A PUF models a(partly) physical system S that can be challenged with randomly generated challengesc ∈ C, upon which it reacts with corresponding responses r ∈ RC. Furthermore, incontrast with conventional digital architectures, the PUF-based approaches intertwinecryptography and sensor properties, making the attack to such systems more challeng-81ing.Metrics like uniqueness, randomness, and bit-aliasing [119][120] have been usedto evaluate PUF performance. They are described below.Uniqueness: Given two PUFs (i and j) of the same optical sensor type, each havingan l-bit response, let ri and r j define their responses to a given challenge c. The meanuniqueness among a group of p PUFs isUniqueness =2p(p−1)p−1∑i=1p∑j=i+1HD(ri,r j)l×100% (4.1)where HD is the hamming distance from different instances of PUF chips of the samesensors and p is the number of PUFs. Ideally, this value is around 50%.Randomness: This is measured by the probability of obtaining a ′0′ or ′1′ in thePUF response to a challenge. Ideally, it should be 50%, for a PUF response to beconsidered as unbiased. It is obtained by computing the bth bit (′0′|′1′) of the l-bitresponse of the pth PUF.Randomness =1ll∑b=1rp×100% (4.2)Bit-aliasing: Due to changes in ambient operating conditions, like temperature andpower supply voltage, the response to a challenge may be slightly varying. To computethis, the mean of flipped bits among responses is found. Generally, it is best to calculatethe worst-case scenarios at the boundary conditions of operating parameters. Ideally,the error should be 0%, but in actual measurements this should be as small as possible.824.1.1 Security RequirementsIoT has various applications in military, e-health, banking etc. Therefore, security andprivacy issues are becoming major concerns in the operation of IoT. Attacks on IoTmay include eavesdropping, Denial of Service (DoS), and Man-in-Middle [16]. DoShappens when unauthorized access to the system occurs. Eavesdropping is an attack onconfidentiality, where intruders may listen to the communication between sender andreceiver. Adversaries who get hold of the key may appear as genuine senders. Thistype of attack is known as Man-in-Middle attack and may occur when the key can bepredicted easily, thus helping the adversary to break into the system. Various param-eters are used to measure the robustness, reliability, and integrity of a IoT system. Asour design uses a sensor PUF, we will first define metrics related to the robustness ofPUF implementation against adversaries [15]. Then, we will define metrics related tothe security of an IoT system secure [16]. Let F : C→ R : F(C) = R, c ∈C,r ∈ RC bea function that maps the challenge/response pairs of a PUF system. Next, we discuss aset of possible attacks that can be attempted against a PUF.Frequency Prediction Attack ( Also known as modeling attack): In this attack, theadversary collects previous output values and tries to make a probability distributionmodel to predict the future outputs Oi as P(Oi = 0/1) [121]. The randomness propertyof the PUF has to be satisfied to resist this type of attacks.Replay Attack: In this attack, the adversary tries to predict the output by studying theoutputs that have similar inputs. If the hamming distance between subsequent inputoutput pairs form a polynomial distribution, as defined in the uniqueness PUF property,generally the associated cipher would resist this attack [121].Cloning Attack: In this type of attack, the adversary tries to clone the original PUF83to replicate the challenge/response behavior, as in the original PUF. So, if F is trans-formed to F ′ with a very high probability, but F(RC) 6=F(RC′), then the PUF is resistivefrom such type of attacks.Side-Channel Attack: A PUF may be attacked passively by using side-channel infor-mation, such as power consumption or electromagnetic radiation emanated from a chipcontaining a PUF. Generally, this type of attacks changes the physical properties of thePUF.Apart from successfully resisting the attacks on the PUF circuits, the IoT systemalso needs to provide successful protection from cyber attacks on the internet. Require-ments for implementing security in IoT-based applications [122] include:Device Authentication: When the device is plugged into the network, it should authen-ticate itself prior to receiving or transmitting data. Just as user authentication allows auser to access a network based on user name and password, machine authentication al-lows a device to access a network based on a similar set of credentials stored in a securegateway server. Once the device is recognized as authentic, the data transfer happens.Confidentiality and Privacy: Latest technologies can be used to gain access to themessage sent from the source to the destination. Therefore, it needs to be hidden fromthe adversaries and an end-to-end secrecy of the data must be maintained.Data Integrity: In addition to confidentiality and privacy, integrity is also an importantsecurity factor during the transmission of data in IoT. An adversary can insert somemalicious code or data into the system to corrupt it. The altered data may reach thedestination node and prove fatal to safety-critical IoT systems, like e-health and bank-ing. Therefore, an integrity mechanism is crucial in protecting the original data fromexternal attacks.Access Control: Access control ensures that the intruder has as minimal access to other84parts of the system as possible. It is important to authenticate the device at regular in-tervals to stop an intruder from gaining access to the system. Sensors which are notauthenticated are excluded from the system network.In a previous paper[123], we proposed a PUF-based cryptographic system that isbased on optical sensors used in e-health systems, such as heart rate monitors and pulseoximeters [124], using photodiodes. In this Chapter, we analyze the intricate detailsof the internal circuitry and have built the prototype of the system. A generic crypto-graphic system is proposed for IoT applications, as infrared optical sensors find appli-cations in various fields, like defense, e-health, banking, and home automation[125].The design we propose is different from all designs mentioned in the introduction as itdoes not exploit digital variations [72–74]. Instead, our approach leads to Strong PUFwith integrated control logic to further consolidate the security of the system.The rest of the Chapter is organized as follows. In Section 4.2, we describe theproposed method and algorithms. Section 4.3 presents a prototype PUF circuit designand the effects of noise constraints and energy consumption in reaching an optimaldesign point. In Section 4.4, we analyze the simulation and measurement results.4.2 Proposed Security Approach and ProtocolAmong the biggest challenges in the design and implementation of a PUF-based sys-tems are:1. Finding a physical property of the device, whose variations are difficult to predict.2. Designing large cycles of random-like binary challenges to authenticate the sys-tem, so that the authentication process cannot be easily predicted in a short timeframe.853. Encrypting the system through PUF-based digital signature generation, and ver-ifying only through public key information.4. Requiring minimal circuitry changes in the existing system hardware.Note that we are using here the optical sensor as a proof-of-concept. The techniquecan be also applied to other sensors with random-like physical properties. The physicalproperty that we have chosen is the dark current of the photodiode and the large cy-cles of binary pseudo-random challenge/response pairs (i.e., input-output pairs to andfrom the target sensing system) generated by quadratic residue property to validate theapproach.The dark current in silicon-based photodiodes depends on the doping concentrationof carriers as well as changes in temperature. For a given temperature and operatingconditions, the dark current varies due to inherent variation in doping concentration. Notwo photodidodes can have exactly the same dark current. We have measured and plot-ted dark current variations for different silicon-based (Si-based) photodiodes at roomtemperature and operating at wavelength λ = 940 nm. Fig. 4.1 shows the measuredvalue of dark currents in nA (nano-amperes) versus the reverse voltage of photodiodes,varying from 0 to 25 Volts. Our region of measurement spans from 0 V upto the diodes’breakdown voltage. The measurements are shown for photodiodes of the same manu-facturer as well as those of different vendors (two from Vishay, two from Everlight, andone from Fairchild). The measurement procedure is explained in detail in the resultssection.As the IoT devices are commonly resource-constrained, the computational resourcesof gateway nodes (cloud server) as mentioned in previous section about IoT architec-ture are utilized. After the authentication of the device the encryption of the data using86Figure 4.1: The dark current vs. reverse voltage of different silicon PIN signatures follows.4.2.1 Quadratic Residue based Device AuthenticationAs the dark current of each photodiode is uniquely varying, we use a quadratic residuebased scheme to propose a device authentication protocol. Let a and N be positiveintegers, with N 6= 0. We note the following definition [126].Definition 1. Given an integer number a ∈ Z and a natural number n ∈N that are rel-atively prime, then a is called the quadratic residue (QR) modulo n, if the congruencyx2 ≡ a mod n has a solution, that is, a is a perfect square modulo n.In general, the quadratic residues follow a residue cycle starting with an initial seedS0, which is also a QR modulo n itself. To illustrate this concept, we provide an examplein Fig. 4.2 by computing the QRs mod 319 (note that in this example, 319 = 29 × 11,that is, we have p = 29 and q = 11). If we take initial seed S0 as 16 or 146 which areQRs, i.e., S20 mod 319, we achieve the cycles shown in Fig. 4.2.87S0=16 256 141 103 82 25257 53 190 170 169 306S0=146 26259291Figure 4.2: Quadratic residue cycles with seeds S0 = 16 & S0 = 146Algorithm 2: Sensor PUF authentication protocolchallenge/response Pair Generation(1) The cloud server generates two large prime numbers p and q, and initial seedsSi∈QR mod [p ∗ q(= n)], i ∈ 0,1,2, ...,k. The value of k is determined by theiterations needed to validate the system as decided by the authenticating party.(2) The server then fills up a mapping table starting from S0 as shown in Fig. 4.3.After the first seed cycle is complete, it completes the table S1 as in step 1 and startsfilling the rest of the table. It continues generating seeds till the table is completedfor all seeds from S0 to Sk.(3) The server randomly chooses values from the mapping table to send to thesensor node and forms a response voltage output pair corresponding to the sentvalues, using its pre-measured (voltage) values (nv(table)) of dark currents of thephotodiode (or in general, any other PUF variable) of the circuit.Key Generation(1) Let g < n be a randomly chosen generator of the multiplicative group of integersmodulo n (Z∗n).(2) Let H be a collision-resistant hash function.(3) The sensor node’s voltage (nv(measured)) w.r.t. dark current is measuredcorresponding to the QR value sent from the cloud server and hash function of anXOR operation with its ID and the data is performed to generate the public key asPkey=H(nv(measured)⊕ID⊕Data).(4) The public key Pkey is then sent back to the cloud through a wireless channel.Authenticating and Pairing(1) Upon receiving the Pkey, the cloud decodes the message using the challenge/re-sponse pair table D̂ata=H−1{Pkey}⊕ID⊕nv(table) by comparing the values ofnv(measured) with its own values nv(table) . If a satisfying level of confidence isfound in the correlation between challenge/response pairs, it sends acknowledgmentto the sensor node.(2) Upon receiving the acknowledgment, the device is authenticated and the pairingis formed. The process is repeated after every defined time interval to re-enforce theintegrity of the device.88It can be seen that the two seeds form a residue cycle of lengths 12 and 4, respec-tively. Note that these two cycles are disjoint, that is, the intersection of the two setsof residue cycles is a null set. This is true in general for any two residue cycles. Sim-ilarly, other QR seeds can be randomly chosen and the residue cycles can be formedthat would be difficult to predict because of the random nature of the choice [127]. Inour approach, since we are using n = p×q, where p and q are odd prime numbers, todetermine whether a is a quadratic residue modulo n, one has to know how we havefactorized n in terms of its odd primes p and q and then find the solution of the con-gruence x2 mod n. We will randomly change p and q and thus due to the difficulty offinding p and q, it is computationally time consuming to find x satisfying x2 ≡ a mod n[128] [129][130].The authentication process consists of three different stages. The first stage is thechallenge number generation based on QR at the cloud server and sending it to thesensor node using a secure transmission channel. Second stage is the key generationat the sensor end and transmitting it back to the cloud server. The final stage involvesdecoding of the key at the server and sending back the acknowledgment for pairing.Algorithm 2 explains the three stages in detail.4.2.2 IoT PUF Security Encryption through Digital SignaturesWe will describe the QR based PUF security variant of the ElGamal digital signatureprotocol [131] to do encryption of the data. An adversary that can introduce maliciousdata may cause the system to respond inappropriately. The receiver of message needsassurance that the message belongs to the sender and he should not be able to repudiatethe origination of that message. This requirement is very crucial in IoT applications.As many IoT devices are sensors broadcasting observations, cryptographic digital sig-89S0S1SKCloud ServerQR Seed TableResponseE-health Sensor 1E-health Sensor 2E-health Sensor 3ChallengeA-PDF Page Crop DEMO: Purchase from to remove the watermarkFigure 4.3: Proposed sensor PUF authentication protocol architecture.natures ensure the integrity of the device’s data stream. Algorithm 3 describes thesignature generation process in detail.4.3 Energy-Efficient Circuit DesignFig. 4.4a shows the simplified circuit diagram of the sensor node, involved in mea-surement of dark current for a given challenge value generated by the cloud. Fig. 4.4bshows the details of current-to-voltage conversion circuitry. The sensor node consistsof mainly three parts : an IoT sensor processor, a photodiode, and a trans-impedanceamplifier (TIA). When the challenge value is received by a processor, the digital valueis converted to its corresponding analog voltage (reverse bias of the photodiode) by adigital-to-analog converter (DAC). Then, the dark current generated by the photodiodeis converted to voltage by the TIA circuit, which consists of low-offset/low-leakageOpAmp with a gain resistor RF connecting the input and output of the OpAmp. Thedark current (typically in the range of 100pA to 1nA) flows through this resistor, cre-ating the voltage (in the range of 1mV to 100mV) at the output of the TIA, which isthen amplified by a simple non-inverting low-offset OpAmp, and the amplified voltage90Algorithm 3: PUF-based Digital SignaturesSystem parameters(1) The cloud server generates two large prime numbers p and q, and initial seedsa∈QR mod [p∗q(= n)].(2) Let g < n be a randomly chosen generator of the multiplicative group of integersmodulo n (Z∗n).(3) Let H be a collision-resistant hash function.Key Generation(1) The server fills up a mapping table starting from a0 as shown in Fig. 4.3. Afterthe first seed cycle is complete, it completes the table a1 and starts filling the rest ofthe table. It continues generating seeds till the table is completed for all seeds froma0 to ak.(2) Generate the private key XA. The server randomly chooses values from themapping table to send to the sensor node and forms a response voltage outputpair corresponding to the sent values, using its pre-measured (voltage) values(XA = nv(table)) of dark currents of the photodiode (or in general, any other PUFvariable) of the circuit.(3) Generate the public key YA as gXA mod n.(4) Thus the sender has the set of keys {XA,YA}.Signature generation(1) Message m is ID⊕Data.(2) Choose a random a such that 1<a<p-1 and gcd(a, p−1) = 1.(3) Compute r = gk(mod n). Compute s = (H(m)−XA.r)a−1(mod n−1).(4) If s=0, start over again. Then the pair (r,s) is the digital signature of m. Thesigner repeats these steps for every signature.Verification(1) Look Up public key YA for device d.(2) Verify gm = Y rA ∗ rs(mod n)91is converted by an analog-to-digital converter (ADC) to digital values, which the pro-cessor uses to convert to a key that is then sent to the cloud server. The DAC of thecircuit can be either external components or integrated parts of the processor.(a) Simplified internal circuitry of an IoT sensor to measurethe dark current.(b) Current-to-voltage conversion circuit with TIA.Figure 4.4: Circuit to measure the dark current92One of main challenges of using dark current in a portable sensor node is that themagnitude of the dark current is in the order of pA to nA, and any measurement attemptof such tiny current requires careful attention to the effect of noise on the measurementresult. It requires an accurate measurement method with limited energy budget. In ourprototype, the current is first converted to voltage by low-noise/high-gain TIA which isamplified by a simple non-inverting voltage amplifier before being measured by ADC.Depending on the level of output voltage from TIA, this voltage amplifier can be alsoomitted from the design, but in this prototype version we have included it for measure-ment. In IoT applications like the ones proposed here, the energy budget of the sensornode is limited, and the amount of energy consumption is mainly determined by theminimum amount of time required for the circuit to process the challenge voltage re-ceived from the processor of the sensor node. The speed bottleneck of the sensor nodeis on the bandwidth of the TIA circuit, and the trade off between the bandwidth andnoise relationship of TIA are studied through measurements described in the followingsections.4.3.1 PUF Prototype Circuitry Design with Consideration ofLeakage CurrentAs the range of dark current of photodiode is in the order of pA to nA, special designconsideration is required in order to prevent leakage current from interfering with accu-rate measurements of such small amounts of current. Several factors could contributeleakage current in the design, and one of main factors that can have direct impact onour measurement is the leakage current through the inverting input of the OpAmp usedin TIA. When the impedance of the input of OpAmp is not large enough, a significantfraction of the dark current can leak through the OpAmp instead of flowing through the93gain resistor of TIA, degrading the measurement accuracy. For our initial measurement,we chose the LMV793 OpAmp from Texas Instruments[132], which is low-noise/low-leakage and CMOS-based. The LMV793 has input bias current of 100fA at 5V Vdd ,Vcm=2V at 25oC, which is much smaller than the typical range of dark current of pho-todidoes. Another contributor of leakage current is the material used in the PCB. Thesolder mask used in the PCB generally helps to reduce moisture infiltration on to thePCB material, but too much area of the solder mask could build up surface charge,affecting measurement accuracy. Therefore, the solder mark was removed near thesensitive region of our circuit (near the photodiode and the input of TIA). The dust andmoisture accumulated between the inverting input of TIA and low-impedance supplytrace on the PCB could also induce leakage current flow. This leakage between tracescould get worse when the voltage difference between adjacent traces is large. In orderto address this issue, we implemented a guard ring, whose potential is driven by thesame voltage applied to the inverting input of TIA. This guard ring surrounds the traceconnected to the inverting input of TIA to remove potential difference. Fig. 4.5 showsthe PUF prototype layout of the circuit used to measure the dark current of the photo-diode, amplify it and derive the desired voltage in the measurement range of millivolts.The figure shows two different versions of circuit without and with a guard ring tominimize leakage current.4.3.2 Circuit Design Optimization through Noise AnalysisDue to the small magnitude of the dark current from photodiodes, it is important toinvestigate the effects of noise on measurements. Fig. 4.6 shows the noise model of theTIA part of the circuit. RD is the output impedance of photodiodes and CIN is the sumof capacitances of the photodiode and the input of the OpAmp. RF is the feedback gain94Figure 4.5: The PUF prototype circuit for measuring dark current from a photo-diode. (Top: with a guard ring around the input of TIA to minimize theleakage current, Bottom: without a guard ring).resistor and CF is the compensation capacitor of TIA. The compensation capacitor(CF ) is required in the TIA to create a zero to stabilize the circuit since without thecompensation capacitor the circuit is essentially a differentiator, which is inherentlyunstable. Since the feedback factor (β ) of the TIA is defined as VIN /VOUT , β can beexpressed as below:β =VINVOUT=(1+ sRFCF1+ RFRD + sRF(CIN +CF))(4.3)The zero and pole frequency from 1β with the feedback capacitor are95Figure 4.6: Noise model of TIA circuit with a photodiode.fZ = 12piRF (CIN+CF )fP = 12piRFCF(4.4)In order to investigate the magnitude of the voltage noise spectral density appearingat the TIA output (eo), we need to consider both the feedback factor β of the TIA andthe noise spectral density of the OpAmp itself (en). The input referred voltage spectraldensity of OpAmp (en) is scaled by a factor of 1/β to become the voltage spectraldensity at the output of TIA (eo). The voltage spectral density profile of the OpAmp(en) can be obtained from the datasheet of the component [132], and the total outputnoise due to en can be calculated by integrating the square of the output noise densityover the entire frequency and taking the square-root of the value.Eo,rms =√∫ ∞−∞e2o( f )d f =√∫ ∞−∞1β 2( f )e2n( f )d f (4.5)96Since the magnitude of the dark current is quite small (in the order of hundreds ofpA to a few nA), the TIA needs to have very large V/I gain, which is defined by thefeedback resistor. However, as RF becomes comparable to the output impedance of thephotodiode (RD), some of the dark current would leak through RD. Therefore, we chose10MOhm for RF , considering the typical value of 100MOhm of shunt resistance RD ofthe photodiode for initial calculation in this section [133]. The compensation capacitorCF was set to be 10pF to filter out the high frequency noise and to make the TIA stable.The value of input capacitance CIN was set to 15 pF considering typical capacitance ofphotodiodes, and the output impedance of the photodiode RD was chosen to 100MOhm.The zero and pole frequency of TIA are calculated as fZ = 700Hz and fP = 1592Hz.Using the datasheet of LMV793 [132], we calculate the corner frequency ( fc), wherethe transition from 1/f flicker noise to white noise occurs, as fc = 340.2Hz. Now, wecan proceed to calculate the voltage noise of the TIA circuit.For f < fc, the rms-noise in this frequency region becomesEo,rms,1 =√∫ fcfo(1β)2e2n( f )d f = 0.347µVrms (4.6)For fc < f < fz, both 1/β and en are constant, and the rms-noise in this region becomesEo,rms,2 =√∫ fzfc(1β)2e2nd f = 0.125µVrms (4.7)For fz < f < fp, the rms noise isEo,rms,3 =√∫ fpfz(1β ( f ))2e2nd f = 0.330µVrms (4.8)97Finally, for fp < f , the noise spectral density isEo,rms,4 =√∫ f∞fp(1β ( f ))2e2nd f = 111.536µVrms (4.9)The voltage noise density at the output of TIA due to the current noise of theOpAmp is equal to the input referred current noise of OpAmp multiplied by the impedanceof feedback resistor and capacitor. Therefore, when a very large gain resistor is used,it is important to include the contribution from the current noise of the OpAmp forcomplete noise analysis. The rms output noise due to the current noise isEo,rms =√∫ ∞−∞e2nid f =√∫ ∞−∞i2n×(RF || 1sCF)2d f (4.10)From the datasheet of LMV793, we assumed that the corner frequency ( fi) where thecurrent noise starts increasing is 30kHz in this calculation. For f < fp region, both inand RF are constant, and the rms-noise from current noise source becomesEo,rms,5 =√∫ fpfoi2nR2Fd f = 3.99µVrms (4.11)For fp < f < fi and fi < f , the rms-noise becomesEo,rms,6 = 3.88µVrmsEo,rms,7 = 39.44µVrms(4.12)The thermal noise of the feedback resistor also needs to be considered for completenoise calculation of TIA. The noise of feedback resistor RF at the room temperature98can be calculated asEo,rms,RF =√4kT RF(pi212piRFCF)=√kT(1CF)Eo,rms,RF = 20.28µVrms(4.13)Note that the rms-thermal noise of the RF appearing at the output of TIA does notdepend on the value of resistor any more. This is because as RF increases the bandwidthof TIA decreases at the same rate as the increase in thermal noise. Therefore, thevalue of RF doesn’t have any impact on the thermal noise contribution. However, thesignal is amplified by the gain of TIA, which is equal to RF . Thus, it is important tomaximize RF to obtain the best SNR ratio. The contribution due to OpAmp’s noiseis a little bit less intuitive. One might think that as the bandwidth of TIA decreasesby increasing RF , the output noise due to the OpAmp should be reduced since thehigh frequency noise would not appear at the output of TIA. However, the calculationshows that this is not the case. Also, the current noise also increases slightly since ZFincreases by (4.10). Despite this increase in OpAmp noise, however, the amount ofadded noise is not significant compared to the amount of amplification on the signalachieved by increasing RF . Therefore, it is still valid to say that using a largest possibleRF maximizes SNR. Finally, the total rms output noise of TIA can be calculated byadding squared values of rms noise from different sources of noise, and taking a square-99root of the final value.Eo,rms,Total =√(7∑i=1(Eo,rms,i)2)+(Eo,rms,RF )2Eo,rms,Total = 120.16µVrms(4.14)If the maximum range of dark current from a photodiode is 200pA, the maximumoutput voltage swing appearing at the TIA output with the gain of 10MOhm isVmax−pp,T IAout ={200pA×(RDRF+RD)}×10MΩ= 1.82mVpp(4.15)Therefore, we find the dynamic range of the TIA from (4.14) and (4.15)DyR =Vmax−pp,T IAoutVnoise,rms,T IAout=1820µVpp120.16µVrms= 15.15 (4.16)4.3.3 Circuit Design OptimizationSince the key generation is done by measuring the dark current of the photodiode usinga TIA circuit, the role of the sensor node processor is minimized to a few tasks – (a)receive challenge/ response from cloud server, (b) convert digital challenge values toanalog bias voltage to the photodiode, (c) convert received analog voltage from TIA todigital and encrypt it into a key using Algorithm 2, and (d) send the response back tocloud server. The speed of processing these tasks is much faster than the speed bottle-neck of the TIA circuit. Since other computation effort required by the processor is notsignificant compared to the amount of time TIA needs to process challenge voltages,100the energy consumed by the sensor node is estimated to be inverse-proportional to thebandwidth of TIA. As a result, the total energy required for the sensor node dependson three major factors -(a) the time required for TIA to convert dark current to corre-sponding voltage response for each challenge, (b) the number of challenges requiredfor one authentication by the proposed protocol and (c) the energy consumed by thecomponents, such as the OpAmp required for dark current measurement. Since onlythe processor part of the sensor node needs to be powered when no challenge/responseis received from the cloud server, the energy consumed by the other part of the sensornode, including the TIA circuitry, during idle time, is assumed to be zero. In short,the energy saving of the sensor node can be achieved by maximizing the speed of TIAcircuitry, minimizing the number of challenges required by the security algorithm, andchoosing power-efficient components for dark current measurement.As shown earlier, the bandwidth of the TIA depends on the feedback capacitance(CF ) and the gain (RF ) of TIA. The speed of the TIA, however, is closely related to thenoise at the TIA output. In general, as the bandwidth increases, the maximum allowablespeed of varying challenge voltages increases, reducing the amount of computationtime, but the TIA with large bandwidth also allows the high frequency noise to appearat the output, degrading the SNR and the dynamic range (DyR) of the TIA. From thepole frequency we obtained in the previous section, the minimum time required for theTIA to process each challenge voltage can be calculated astp =1fp=11592Hz= 628.3µs (4.17)Therefore, the energy required to process each challenge voltage by the TIA is thepower consumed by the TIA multiplied by tp. From the datasheet of LMV793, the101power consumption for the chosen OpAmp is 7mW with 5V supply voltage [132].Therefore, the estimated energy is calculated asEbit,T IA = tp×PT IA = 628.3µs×7mW = 4.40µJ (4.18)One of the options for further reducing the energy consumption of the TIA is toincrease the bandwidth of TIA by either decreasing the gain (RF ) or decreasing thecompensation capacitance (CF ). However, larger bandwidth inevitably increases thenoise floor due to high frequency noise of the OpAmp that starts appearing at the outputof TIA, degrading the dynamic range of the TIA. Fig. 4.7a and Fig. 4.7b show thetrade-off between energy consumption and dynamic range of TIA with LMV793, fordifferent values of RF and CF . The maximum range of dark current was assumed to be200pA, and dynamic range was calculated by taking the ratio of the estimated outputvoltage to the output voltage noise in rms as defined in (4.16). As CF increases, theoutput signal voltage remains the same but the overall noise decrease, improving DyR.As RF increases, the noise floor stays almost the same while the output signal voltageincreases; improving DyR. However, for both cases, the reduced bandwidth decreasesthe speed of TIA and worsens the energy consumption.Another option for reducing the energy consumption is to choose more power-efficient OpAmp for TIA. The LMV793 we used for earlier energy estimation con-sumes 7mW which is not very power efficient. Choosing the OpAmp for battery-powered applications could significantly reduce the energy consumption. To improvethe energy savings, we used OpAmp LMP2231 that is designed for battery-poweredapplications which draws only 50uW of power from 5V supply [134]. The noise cal-culation for LMP2231 is redone and the energy/noise trade-off is shown in Fig. 4.8.102(a) Dynamic range vs. energy for various CFvalues for TIA LMV793.(b) Dynamic range vs. energy trade-off forvarious RF values for TIA LMV793.Figure 4.7: Plots for different CF values with RF = 10MOhm for TIA LMV793.Compared to LMV793, the energy consumption is significantly reduced down to anorder of tens of nanoJoules. For example, with RF=6MOhm and CF=8pF, a dynamicrange of 18.0 can be achieved with energy consumption of 0.0151 uJ. It is noticeablethat as CF increases, the energy consumption increases linearly, while the rate of in-crease in dynamic range eventually slows down. Fig. 4.8 shows the magnitude of totalTIA noise for different CF values. Since the gain is fixed to 6MOhm, DyR in Fig. 4.8aand total noise in Fig. 4.8b are inverse-proportional to each other.Fig. 4.9 shows the trade-offs between dynamic range and energy consumption asthe value of RF changes from 4 to 100 MOhm. As RF increases, energy consumptionof TIA increases linearly, but the slope of dynamic range slightly decreases as morecurrent starts flowing through RD instead of RF . Fig. 4.9b shows that the amount of to-tal noise is relatively constant for different RF values, and Fig. 4.9c shows the source ofnoise. As mentioned earlier, changing the value of RF did not affect the thermal noiseappearing at the output of TIA since the rate of thermal noise increase is the same asthe rate of bandwidth decrease of TIA, canceling each other. The noise contributed by103(a) Dynamic range vs. energy trade-off for var-ious CF values for TIA LMP2231.(b) Total noise vs. energy trade-off for variousCF values for TIA LMP2231.Figure 4.8: Plots for different CF values with RF = 6MOhm for TIA LMP2231.the OpAmp slightly increases as RF increases. The above results show that changingthe values of RF and CF affects the amount of energy consumption linearly. However,changing RF has more drastic impact on the dynamic range than changing CF , as thevalue of RF and CF become larger. Therefore, it is better to increase the dynamic rangeof TIA by increasing RF , and decrease the energy consumption by lowering CF . Notethat as RF approaches the output impedance of photodiode, a fraction of dark currentcould start leaking through the internal resistance of the photodiode. However, thismay not be an issue in implementing the proposed idea in terms of true dark currentmeasurement since the purpose of measuring dark current is to identify a unique pho-todiode for security purpose, and the output impedance of individual photodiodes canbe pre-measured and stored in the cloud along with the pre-measured dark-current pro-file required for key verification. Therefore, the variance of output impedance betweenphotodiodes can be compensated in the cloud during the authentication process, addingan extra layer of complexity and making the system more secure.Fig. 4.10a and Fig. 4.10b show the trade-off between dynamic range and energy104(a) Dynamic range vs. energy trade-off fordifferent RF values with CF = 2pF for TIA withLMP2231.(b) Total noise vs. energy trade-off for dif-ferent RF values with CF = 2pF for TIA withLMP2231.(c) Noise contribution for different RF val-ues with CF = 2pF for TIA with LMP2231.Figure 4.9: Plots for different RF values with CF = 2pF for TIA with LMP2231.105(a) Dynamic range vs. energy trade-offfor different CF values with RF = 50MOhmand CD = 15pF for TIA with LMP2231.(b) Total noise vs. energy trade-off fordifferent CF values with RF = 50MOhm andCD = 15pF for TIA with LMP2231.(c) Noise contribution for different CFvalues with RF = 50MOhm and CD = 15pFfor TIA with LMP2231.Figure 4.10: Plots for various CF values with RF = 50MOhm for TIA withLMP2231.106consumption for larger gain values of RF =50MOhm. The results show that a dynamicrange of 119.6 can be achieved with smaller energy consumption of 0.126uJ with RF=50MOhm and CF = 8pF. It clearly indicates that increasing RF gives better energysaving for the same dynamic range, as long as the leakage current through the internalshunt resistance of the photodiode can be compensated accurately. Fig. 4.10c showsthe contribution to the total noise for different CF values. As CF increases, both thethermal noise of RF resistor and OpAmp noise decrease. The decrease in thermal noisecomes from the decreased pole frequency by (4.13), and the decrease in OpAmp noisecomes from decreased noise gain 1/β at the high frequency region by (4.3).The compensation capacitance values used in all Figures in this section are at least10 times greater than the minimum capacitance required for guaranteeing stability ofthe TIA. The above results suggest when configuring the TIA for IoT sensor nodeschoosing RF = 50MOhm and CF= 10pF. This leads to a dynamic range of 130.3 (7bits)with energy consumption of 0.157uJ.4.4 Results and Analysis4.4.1 Output Measurements of the PUF Prototype CircuitFig. 4.11 shows the TIA circuit with OpAmps having offset at their inputs. The currentfrom the photodiode and the output voltage of TIA can be expressed asID =VT IA OUT−V ∗T IA+RFVT IA OUT = IDRF +V ∗T IA+(4.19)where V∗T IA is the voltage appearing at the “–“ input of the first OpAmp when the107Figure 4.11: TIA circuit with input offset of the OpAmp.output voltage is zero, and it can be expressed as a sum of VT IA+ and the input offsetvoltage of the OpAmp asV ∗T IA+ =VT IA++Vo f f set opamp1 (4.20)The voltage output of a non-inverting amplifier can be written asVOUT = G(V ∗T IA OUT −VT IA+)+VT IA+ (4.21)where the gain of non-inverting amplifier (G) is 1+ R1R2 . The voltage at the “-“ inputof the second OpAmp when the output voltage is zero is described asV ∗T IA OUT =VT IA OUT +Vo f f set opamp2 (4.22)where Vo f f set opamp2 is the input offset voltage of the second OpAmp. Then, the108voltage at the output of the second OpAmp can be written asVOUT = G(VT IA OUT +Vo f f set opamp2−VT IA+)+VT IA+= G(IDRF +V ∗T IA++Vo f f set opamp2−VT IA+)+VT IA+= G(IDRF +Vo f f set opamp1+Vo f f set opamp2)+VT IA+(4.23)Then, the current can be re-written asID =VOUT −VT IA+RFG− Vo f f set opamp1+Vo f f set opamp2RF(4.24)The input offset voltage of the OpAmp comes from the mismatch between two in-puts of OpAmp due to manufacturing variation, and it is constant for a given devicewhen the common-mode voltage is constant at a fixed temperature. Therefore, we canexpect that the real value of dark current might differ from the measured one by theconstant value due to the input offset voltage of OpAmp. Fig. 4.12 shows the measureddark current of various surface-mount photodiode devices with LMV2231 OpAmp witha TIA gain of 50MOhm at 24 Celsius. As the reverse bias voltage applied to the pho-todiode increases, the amount of dark current also increases. Some photodiodes, suchas Vishay and Everlight, appear to have less than 20pA of dark current variation whenthe reverse bias voltage changes from 5 to 25V. However, this could be also due to thedark current leaked through the internal shunt resistor of photodiodes, which could hap-pen when the shunt impedance of photodiode is comparable to the gain resistor value(50MOhm). The dark current measurement can be affected by two main factors: inputoffset voltage of the two opamps in our circuit, and tolerance of gain resistance. The109Figure 4.12: TIA amplified output voltage (response) vs. reverse voltage (chal-lenge) of different silicon PIN photodiodes.input offset voltage of an opamp is caused by mismatch of differential input transistorscreated during manufacturing process. Although the range of offset values can be foundfrom the datasheet of the opamp vendors, the exact value is not known. Also, the gainresistor has its own tolerance. As shown in (4.24), the exact amount of dark currentcan be shifted from the true value based on these factors. Through the experiments, theresistance of a resistor is found to be a constant here and the input offset voltage is alsoconstant at a fixed input voltage level; therefore, the difference between the measuredand true dark current value is constant. Since knowing the exact dark current valueis not important to identify between different sensor nodes as long as pre-measureddark current profile matches with the ones from a sensor node, these variation from1100 2 4 6 8 106810IndexQR Challenge(a) QR challenge voltages with seedS0=16 and 146.0 2 4 6 8 1000.51IndexI D [nA](b) ID response stored in server.0 2 4 6 8 1000.51IndexI D Measured [nA](c) ID measured from Fig. 4.0 2 4 6 8 10−1050510IndexError [pA](d) Error between stored ID & measuredID.Figure 4.13: Measured and stored values of dark currents (ID) in Everlight siliconphotodiode.manufacturing process does not play a negative roles for our security system.Table 4.1: Amplified voltage response for different challenge voltages between 0to 5V with respect to various silicon photodiodes.ReverseBiasVoltage (V)QSB34GR(mV)(Fairchild)PD93-21C(mV)(Everlight)VEMD2020X01(mV)(Vishay)PD70-01B(mV)(Everlight)VBP104S(mV)(Vishay)0.01 547 32 53 462 2501 625 41 57 511 2812 699 49 60 557 3053 770 56 62 602 3244 839 62 64 643 3405 908 68 66 684 354In Fig. 4.12, we have plotted the output voltage measurements using (4.23) withrespect to the reverse bias voltage for the photodiode dark currents shown in Fig. 4.1.The measurements shown in Fig. 4.12 are for an ambient room temperature of 25oC andan operating voltage range of 0 to 25V. The TIA amplified voltages are the measuredresponses to each challenge in the form of reverse voltage generated as in Algorithm 2.111Measurements in Fig. 4.12 are used in testing the IoT-PUF circuit in the lab environ-ment before deployment. Where not possible to maintain laboratory voltage range, theproposed circuit also works well in the low voltage range on 0 to 5V. Table 4.1 depictsthe TIA amplified voltage response for different challenge voltages between 0 to 5Vwith various silicon photodiodes taken in this Chapter as samples for validation. Ascan be interpreted from the table that with a 12-bit ADC, the resolution can be as smallas 1.22 milli-Volts. Hence all the photodiodes with amplified voltage w.r.t their darkcurrents are in the range much higher than the ADC resolution. Further, the photo-diodes are distinguishable even with an error in measurement accuracy of 1%-2% forvery low values as in the VEMD2020X01 and PD93-21C photodiodes and ≈20 timesmore accurate for other photodiodes.Fig. 4.13 depicts the measured and stored values of dark current (ID) of an Everlightsilicon photodiode. The measured value is the response to the corresponding challengeto the IoT sensor. The plot in Fig. 4.13a shows the QR voltage challenge generatedby the server with two different seeds and sent to the sensor. Fig. 4.13b shows thevalue of (ID) stored in the serve as a response to the corresponding challenge, whichis then compared to the measured ID, i.e., the response from sensor in Fig. 4.13c. Theerror,shown in Fig. 4.13d, is small and the correlation between the two current valuesis ∼ 0.9996. This means that using Algorithm 2 the device can be authenticated.4.4.2 Security CharacteristicsEarlier, we presented various PUF-IoT security metrics. Here, we will evaluate howour proposed algorithm and design satisfy the key metrics of a secure system from ourmeasurements and algorithmic simulations. First, we will analyze our design for PUF-targeted attacks on the system.112Frequency Prediction Attack : We challenged the circuit with 10000 challenges andmeasured the response Oi of the system. From Fig. 4.14a, it can be inferred that theprobability of Oi being a particular value is always around 0.5, which makes the ran-domness parameter to be ≈ 50%. Therefore, our method is resilient to this kind ofattack.Replay Attack : In this attack, the adversary tries to predict the output by studyingthe outputs that have similar inputs. In the proposed design, the randomness of input-output in the sensor PUF is enhanced by using a QR challenge/response pair generator.From Fig. 4.14b, it is evident that the proposed method generates a polynomial distri-bution with respect to the input-output hamming distance, i.e., the distance between theoutput vectors by changing one bit of input vectors in every iteration, for two differentPUFs, i and j. This shows resiliently to this type of attack.Cloning Attack : As the cipher generator knows the ID issued to the sensor, as wellas the measured values of challenge/response pairs, it is difficult for the attacker to au-thenticate its own PUF. Further, as the mapping is a PUF function, it is impossible toexactly replicate the physical variations of the system.Side-Channel Attack : This type of attack changes the physical properties of the orig-inal PUF. Once the inherent PUF properties are changed, challenge c will not generatethe same response rc as before. Rather, it will generate r′c. So, the adversary will notbe able to validate its system. Once that will fail, the system will remove access for theaffected IoT node. The node will be flagged as malicious and will not be able to send1130 5 10 1500. distanceRelative Frequency(a) Input-output hamming distance distri-bution.0 2000 4000 6000 8000 1000000. (Oi)P(O i = 0/1)(b) probability of output Oi = 1.Figure 4.14: Optical IoT sensor PUF’s resilience towards various adversary at-tacks.any further data.Device authentication is performed using Algorithm 2. Only when successful pairingis done and access is granted to the a particular IoT node, the data transfer happens.Depending on the vulnerability of the network, authentication and pairing are done atregular intervals to maintain data integrity and authorized access to the system. Thismakes the system robust to DoS type of attacks. Also a bit-aliasing error is observedfor various optical sensors. The error is the worst-case error, at boundary temperatureconditions. We have taken a temperature range of 5oC−40oC for our measurements,although a larger range is possible as per the datasheet. The worst-case error is foundto be 5.39% and the best case is 0.8%. To ensure integrity in the system, a public key isgenerated using the private key generated by the challenge/response pair of the IoT-PUFcircuit. As shown in Algorithm 3, the encryption algorithm resist any eavesdroppingattacks on the system.1144.4.3 Threat AnalysisSecurity threat to the PUF-based IoT system is through the user, manufacturer andexternal adversary. All the threats posed by external adversaries are explained in theprevious section with the corresponding various types of attacks on the system. In thissection, we will explain our system’s defense to the other two threats.Malicious user : It is the owner of the IoT device with potential to perform attacksto learn the secrets to gain access to restricted functionality. By uncovering the flaws inthe system the malicious user tries to sell secrets to third parties, or even attack similarsystems. Our proposed system will not be able to stop such an user to model somesystems, but as our dark current property is unique it can stop the malicious user frommodeling the system altogether. Although such a threat can be initially successful, butit will not give long term results for the user.Bad Manufacturer : Is the producer of the device with the ability to exploit the tech-nology to gain information about the users, or other IoT devices. Such a manufacturercan deliberately introduce security holes in its design to be exploited in the future foraccessing the user’s data and exposing it to third parties. Again the manufacturer’sattack cannot be successful on our proposed system due to its random like properties.It is impossible to change the physical properties of a photo-diode using external re-sources. By doing so the diode will be corrupted and is unusable for the PUF circuit.Thus making this threat ineffective.External threat : External adversary does not have access to the physical device. Thesecure cloud used in our authentication protocol if intercepted by the external adver-115sary, can only give them access to the public key. The public key in turn is generatedfrom the private key derived by PUF variations of our proposed system. Hence, this cangive adversary initial success. But, without knowing the actual voltage used to generatethe dark current, it is impossible for the adversary to decode the message. And suchan interception will alert the receiving server system, which will in turn block the ma-licious node. Thus, the adversary would have to start again. In the worst case scenario,the nodes may become unusable by the system, but still the message will remain safefrom being decoded.4.4.4 Energy ConsumptionEnergy is a central concern in the deployment of IoT nodes having limited battery sizeand computational resources. Here, we investigate and compare the energy cost of var-ious cryptographic protocols with our IoT PUF, from a computation at energy point ofview. The energy consumption is linearly proportional to the processing time as de-scribed in (4.18) . The design proposed in this Chapter, uses RF = 50MOhm and CF=10pF with a dynamic range of 130.3 (7 bits), having energy consumption of 0.157uJ.Depending on the length of the encryption bits, the energy consumption can be com-puted from the TIA measured data. Our Op-Amp draws only 50uW of power from a5V supply voltage. PUFs provide lightweight hardware fingerprints just like hash func-tions and can be used alternatively for authentication of the device [135]. Some of thehash functions found in the literature are MD5 (Message Digest 5)[136], SHA-1 (Se-cure Hash Algorithm 1)[137], and HMAC (Hash Message Authentication Code)[138].MD5 is a cryptographic hash function to derive the authentication token, also calledwhite list. SHA-1 is a 160-bit hash function, which resembles the MD5 algorithm.This was designed by the US National Security Agency (NSA) to be part of the Digital116Signature Algorithm. The standard for implementing hash-based authentication is theHMAC as in FIPS (Federal Information Processing Standards) [138]. HMAC is used incombination with an approved cryptographic hash function and needs a secret key forthe calculation and the verification of the MACs. In Table 4.2, we of show the energyconsumption the hash functions. MD5 and SHA-1 are lightweight hash functions andconsume less energy as compared to HMAC. HMAC is a keyed function, for bit rangesof 0 to 128 bits, and the energy consumption fluctuates by a very minute amount. SHA1has more steps of computation than MD5, hence it consumes more energy than MD5.All the values are shown per Byte of data. Our design consumes 0.1794 uJ/Byte, theleast of all other hashing functions.Table 4.2: Energy cost of various hash functions compared to our designS.No. Hashing Method Energy Consumption (uJ/Byte)1 Our IoT PUF 0.182 MD5 0.593 SHA-1 0.764 HMAC 1.16Table 4.3: Energy cost of various asymmetric encryption algorithms as imple-mented in different sensor motesAlgorithm MICAz mote TelosB moteCycles Energy Cycles EnergyECC-160 [139] 15.6 M 55 mJ 14.0 M 17 mJOur IoT PUF 128-bit 4640 16 µJ 3480 4 µJRSA-1024 [139] 108.1 M 378 mJ 60.4 M 73 mJOur IoT PUF 1024-bit 5.9 M 21 mJ 4.4 M 5 mJAlgorithm 3, uses an asymmetric approach to encrypt our IoT-PUF system. Thissecures the system from intruders and provides resistance to malicious attacks. It alsoprovides confidentiality, privacy, and integrity to the IoT node. Rivest-Shamir-Adleman117(RSA) and Elliptic Curve Cryptography (ECC) are two lightweight secure asymmetricalgorithms for IoT[139–142]. Both work by generating public and private keys. Publickeys are published openly, whereas private keys are made secure. ECC has faster com-putation times and bit-shifting operations instead of multiplications, to save energy forlow power devices. In Table 4.3, we have compared our method with ECC and RSA interms of energy cost, and number of computation cycles, using tiny sensors, such as theMICAz and TelosB. Note that the computation energy values are taken from [139]. Itcan be inferred from the table that the proposed method consumes less energy as com-pared to other public cryptographic algorithms. This is due to the fact that our privateand public keys are generated by PUF functions rather than complex multiplications asin RSA and ECC. Further, the decryption is done at the receiving end in a cloud server,thus relieving the low power IoT devices of computation burdens.118Chapter 5Energy Management in Smart Cities:Peak Demand Reduction and EnergySavings5.1 IntroductionSmart cities in brief can be defined as a city which uses information and communicationtechnologies (ICT) such as smart sensors, cognitive learning, and context awareness tomake lives more comfortable, efficient, and sustainable [2]. Cities today face mul-tifarious challenges, including environmental sustainability, low carbon solutions andproviding better services to their citizens. Given these trends, it is critical to understandhow ICT can help make future cities more sustainable. As microcosms of the smartcities, smart and green buildings and homes stand to benefit the most from connect-ing people, process, data, and things. The Internet of Things (IoT) is a key enablerfor smart cities, in which sensing devices and actuators are major components along119with communication and network devices. Management of smart homes often requiresanalyzing IoT data from the interconnected networked devices to optimize efficiency,comfort, safety, and to make decisions faster and more precise [143].The significant efficiency gains from home automation can make cities sustainablein terms of resources. Importantly, the IoT ambitions and scope are designed to respondto the need for real-time, context-specific information intelligence and analytics to ad-dress specific local imperatives [144]. Further, realization of smart, energy-efficientand green home infrastructure would allow the development of ’livable’ interconnectedcommunities, which will form the backbone of a futuristic green city architecture [145].Hence, energy management in smart homes is a key aspect of building efficient smartcities [146]. Energy management consists of demand side management (dsm), peakload reduction and reducing carbon emissions[147]. In an industrialized country, res-idential and commercial loads in urban centers consume a significant amount of elec-trical energy. As per the survey report [148] nearly 39%− 40% of the total energyconsumption in Canada is consumed by the residential and commercial complexes. Itis evident from various load surveys that the demand of electricity in these residences ishighly variable and changes throughout the day. Therefore, finding suitable strategiesfor efficient management of home energy demand and to help reduce the energy con-sumption during peak period will make the communities’ more energy efficient. TheCanada Green Building Council is working towards finding ways of making buildingsgreener and community sustainable1. Therefore, the need for energy efficient buildingsis growing rapidly.The power systems require equilibrium between electricity generation and demand[149]. Power system operators dispatch generating units primarily based on operating1[Online] Available : http://www.cagbc.org120cost or market bid price. In order to meet the increased demand during peak period,more resources are often required to increase the generation capacity. Since additionof resources to meet the peak demand is an expensive investment, distribution systemplanners and utility engineers very often consider the reduction in peak load as a fea-sible solution to the problem. However, peak load reduction is mostly valuable forutilities and most popular only in a purely market-driven energy management environ-ment. Under these circumstances, Demand Response (DR) [150] and [151] offers anopportunity for consumers to play a significant role in the operation of the electric gridby reducing or shifting their electricity consumption during peak periods in responseto time-based rates or other forms of financial incentives. In most of the cases, DR isa voluntary program that compensates the consumers. There are many modern meth-ods that reduce the peak load and load at peak time which is referred as Demand SideManagement (DSM) [152]. Current market framework and lack of experience and un-derstanding of the nature of demand response are the most common challenges in DSMnowadays [152].Newer technologies like energy management using smart meters are now becomingpopular in places like Ontario, Canada where few utilities have introduced energy tariffbased on the Time-Of-Use (TOU) model in which a consumer pays differently for theenergy consumption at the different time of the day. This has been possible due tothe implementation of smart meters which track the energy usage in a home on anhourly basis [153] and then consumption information is bundled into multiple TOUprice brackets. However, all these processes mostly help the local distribution companyand in order to take advantages of the TOU, each household has to adopt a change inthe use of the appliances which may cause significant discomfort to the consumers. Inthis scenario home appliance scheduling with electrical energy services for residential121consumers is useful.In this Chapter, a home energy management system named as Home Energy Man-agement as a Service (HEMaaS) is proposed which provides intelligent decisions, isinteractive with the environment, scalable and user friendly. Wi-Fi connected smartsensors with centralised decision-making mechanism can identify peak load conditionsand employ the automatic switching to divert or reduce power demand during peakperiod, thereby reducing the energy consumption. Therefore, by implementing moni-toring and controlling sections of the HEMaaS platform using web services, one mayachieve the agility, flexibility, scalability, and other features required for a feasible andaffordable HEMaaS platform.We have based our experimental findings on a typical Canadian residential apart-ment IoE system to investigate the effectiveness of the proposed home energy manage-ment service. The main objective of HEMaaS is to shift and curtail household applianceusages so the peak demand and total energy consumption can be reduced. A new neuralnetwork based reinforcement learning algorithm has been proposed in this Chapter toachieve the objectives. The classical Q-learning problem of the reinforcement learninghas been formulated as a neural fitted supervised learning problem here and is namedNeural Fitted Q-based Home Energy Management (NFQbHEM) algorithm. We designa node-red framework based user interface for controlling home appliance action basedon NFQbHEM algorithm. The reward matrix incorporates user convenience parametersfor state- action transition and includes user preference, power cost savings, robustnessmeasure and user input preferences to initialize the algorithm.Rest of the Chapter is organized as follows: Section 5.2 describes the HEMaaSplatform and its architecture. Section 5.3 formulates the home energy managementproblem as a markov problem and its possible solution strategy is described using vari-122ous modelling parameters. The NFQbHEM algorithm is explained in Section 5.4. Theexperimental results are shown in Section 5.5 for different cases.5.2 Home Energy Management as a ServiceHome energy management is a service platform for the users to efficiently perform de-mand side management and control. It consists of home appliances connected througha grid of interconnected network of devices with preference given to the user conve-nience. The platform may be used for different types of community houses (condo andtown homes) to manage their energy consumption. The systems may be categorizedinto hardware and software architectures.5.2.1 The Hardware ArchitectureA typical home consists of various appliances. These appliances establish a connectionwith the user and provide them with the monitoring and controlling capabilities. Theyare to be monitored and controlled locally or remotely by a HEMaaS platform using aSonoff wireless switch 2. Most of the common home devices fall within the (current,voltage) range of Sonoff currently commercially available in the market.The architectural diagram is shown in Figure 5.1. It consists of a Main Commandand Control Unit (MCCU), Sonoff wireless switch, Smart meter, Gateway router and aCommunity Cloud Management panel (CCM). The MCCU is the main intelligence ofthe network which is responsible for triggering grid signals based on the output of themachine learning algorithms. It also has an input port which monitors for user inputsignals and accordingly provides user input to the controller. Sonoff Switch receives2The Sonoff is a device that is to be put in series with the power lines allowing it to turn any device onand off remotely. Its voltage range is 90-250V and it can handle a max current of 10A.[Online] Available: 5.1: HEMaaS hardware architecture of a typical Canadian condothe trigger at its input port from the MCCU and turns the appliance Off/On accordingly.Smart meter provides power consumption data to the power station for overall efficientcommunity energy management. Gateway router translates the MCCU messages usingnetwork address translation (NAT) in order to translate from a private network address(like 192.168.x.x, 10.0.x.x) to a public facing one. The smart meter and CCM areoutside of gateway router and are separated by a secured firewall. CCM is monitored bythe city power substation. The substation according to its generation and distributionhas a set amount of available power for the community to use. CCM receives inputfrom the substation and sends those commands to the each home’s MCCU which interm updates its power management strategy.124Figure 5.2: Software architecture and communication framework of HEMaaSplatform5.2.2 The Software Architecture and Communication InterfaceThe HEM MCCU needs to process the NFQbHEM algorithm integrating historicaldata as well as the user input preferences. Thus a decision needs to be formed quickly.Moreover, the state-action pair and user preferences change rapidly throughout the dayand HEMaaS platform needs to provide service in a timely manner. Therefore, herea Linux-based fast microcontroller has been used, namely Raspberry Pi33. RaspberryPi3 runs the NFQbHEM algorithm using python programming language and plots thecharts with its matplot library. Figure 5.2 shows the software architecture and com-munication framework of HEMaaS platform. The web-based node-red programmingmodel have been choosen to implement the controlling structure of the HEMaaS plat-form. It is easy to implement with a flow and is easily explandable if more appliances3[Online] Available : the network. The user input is modelled inside the flow with a switch. Userinput manually can cause either a delay in the operation of the appliance or it willreset its temperature. These settings can also be changed via the smart MCCU algo-rithmic decision. A lightweight, low-power and secure protocol has been used in theChapter to communicate between home appliances and the MCCU over Wi-Fi. Theprotocol is called Message Queue Telemetry Transport (MQTT) [154] and it is opti-mized for high-latency or unreliable networks. MQTT provides three level security forthe data over the network. It uses a broker to publish messages to clients who sub-scribe to a particular topic. Topic are in the form of a hierarchy of devices in the home[Home/(Room)/(Device)/RaspberryPi GPIO Pin]. Mosquitto4 broker has been usedin this architecture. Broker performs authentication via username and password, clientID and X.2 certification to validate the clients in the HEM network. Thus intrusion canbe prevented. A dashboard user interface (UI) for desktop and mobile have been de-signed to give users ample interaction opportunities. The design of the UI is describedin detail in the result section.5.3 HEM as a Markov Decision Process and ItsSolutionWe formulate our HEM problem as a set of discrete states, where each state representsa binary formulation of the power levels of home appliances. The MCCU issues com-mand to switch these power states. We model the power states as a Markov DecisionProcess (MDP) and derive its solution using reinforcement learning (RL) based NeuralFitted Q-Iteration (NFQI) algorithm. The reason for choosing RL with neural networkfunction classifier is based on the type of system being modeled and its behavior. As4[Online] Available : [155] and [156], the machine learning algorithms are divided into unsupervisedand supervised learning. For unlabeled data algorithms like k-means, gaussian mix-ture models are applied to the data. However, as we have historical data [148] to beused for our modeling, these algorithms will not be the best suited for our scenario.For labeled data training and fitting, algorithms like regression, decision trees, supportvector machines, naive Bayes classifier and neural networks are used. As the HEMaaSsystem has user interaction and feedback from wireless access point of the appliances,only using supervised learning algorithms to fit the data for maximum accuracy/mini-mizing cost will be time and resource consuming. The algorithm has to interact withthe environment and objects, learn from their feedback and should update its goalsaccordingly. Thus reinforcement learning (RL) [157], which starts from a particularstate, learns from the environment and update its goals is the best suited for our appli-cation. As the algorithm will pass through multiple states in order to reach its optimumgoal, a supervised classifier can be used in conjunction with the RL algorithm. Neuralnetwork is slow, but classifies accurately in comparison to other supervised learningmethods [158]. Hence it is chosen as the modeler for our system.MDP [159] is a set of discrete time stochastic control process where outcomes areobtained with a combination of partly random events and partly by a decision makingprocess. At each time step, the MDP is modeled as a sequence of finite states si ∈ S,the agent action ai ∈ A that are evaluated based on a random process to lead the agentto another state. For each action performed, the agent receives an award R. As in [159],MDP is formulated as a set of four-tuple <S,A,P,R>, where P is the state transitionprobability when agent moves from state (s(t)→s(t + 1)) ∈ S. From the current statesi(t) ∈ S to state s j(t) ∈ S in response to action a ∈ A, the transition probability isP(si,a,s j) and an award R(si,a) is received. Let sk denote the state of the system127just before the kth transition. In an infinite horizon problem (s→ 1...∞), maximumaverage discounted reward received is found using the action executed at each stateusing a reward policy pi(s). RL [157] is a machine learning approach that solves theMDP problem. It learns the policy online with real-time interaction with the dynamicenvironment and adjusts the policy accordingly. After a certain set-up time, the optimalpolicy can positively be found.Q-learning is an online algorithm that performs reinforcement learning [160]. Thealgorithm calculates the quality of a state-action pair which is denoted by Q and is ini-tialized to zero at the beginning of the learning phase. At each step of environmentinteraction, the agent observes the environment and decides on an action to changestate based on the current state of the system. The new state gives the agent a rewardwhich indicates the value of the state transition. The agent keeps a value functionQpi(s(t),a(t)) according to an action performed which maximizes the long-term re-wards. The Q-factor update equation with discounted reward is as followsQt+1(s(t+1),a(t)) = Qt(s(t),a(t))+α(s(t),a(t))[R(t)+γ ·MAX (Qt(s(t+1),a(t)))−Qt(s(t),a(t))] (5.1)Where, α(s(t),a(t)) is the learning rate (0<α<1) and γ is the discount factor withinthe range 0 and 1. If γ is close to 0, the agent chooses immediate rewards, else it willchoose to explore and aim for long-term rewards. In [160] it is proved that the learningrate α is a function of k, where k is the number of state transitions. It satisfies thecondition asαk =AB+ k(5.2)128Where, A and B need to be found out using simulations.Online learning methods like Q-learning are good from a conceptual point of viewand are very successful when applied to problems with small, discrete state spaces.But for more realistic systems, the ‘exploration overhead’, stochastic approximationinefficiencies and stability issues cause the system to get stuck in sub-optimal policies.Updating the Q-value of state-action pair (s(t),a(t)) in time step t this may influencethe values (s(t−1),a(t)) for all a ∈ A of a preceding state st−1. However, this changewill not back-propagate immediately to all the involved preceding states. Batch Rein-forcement Learning (BRL) typically address all three issues and come up with specificsolutions. It performs efficient use of collected historical data and yield better policies[161]. It consists of three phases, which are exploration, learning and application. Ex-ploration has an important impact on the quality of the policies that can be learned.The distribution of transitions in the provided batch must resemble the ‘true’ transitionprobabilities of the system in order to allow the derivation of good policies. For achiev-ing this, training of samples is done from the system itself, by simply interacting withit. When samples cover the state spaces closed to the goal state, policy achieved willbe closed to the optimal policy and convergence would be faster. NFQI algorithm isone of the popular algorithms described in [162]. Given a set of transition samples over(s(t),a(t),R(t),s(t+1)) and an initial Q-value q0s,a = 0, derive an initial approximationQ0 with Q0 = q0s,a. Update the value of qks,a at each iteration. Define a training set Tkand convert the update problem into a supervised neural network based learning prob-lem. Finally, find the resulting function approximator Qˆi using the pattern trained usingset T k. At the end, a greedy policy is used to define the policy pi(s).pi(s) = argmaxa∈AQ(s,a) (5.3)1295.3.1 State-Action Modelling of AppliancesThe software architecture of the homes in communities shown in Section 5.2.2 de-scribes a typical condo home architecture with living room, bedroom, kitchen andwashroom. Each of the sections have various common home appliances having var-ied peak load power rating as in Table 5.1 as taken from [163]. The states s(t) definedin the Algorithm 4 are different combinations of power levels derived from the peakpower rating of the appliances. Apart from refrigerator all other appliances can beturned On/Off in a smart home as the refrigerator needs to continuously run throughoutthe day and should not be stopped. Usage pattern of all other appliances vary through-out the day and can be controlled through the MCCU. Therefore, in total there are 10appliances and a 2n−1 transition states depicting various combination of power levels(n = 9) which results in 511 states. Lets depict each appliance in ascending order oftheir peak power level from Table I with level ’pl’. Thus Lighting will be symbolizedby p1 and WasherDryer by p9. The power values are coded as binary states i.e. 0represents the Off state and 1 represents On state. For example, 001001010 meansMicrowave, Heater−2(Bedroom) and Stove are in On state and rest all are in Off con-dition. The total power consumed at that instant t is 7600 Watts if every On applianceis operating at peak load.130Table 5.1: Maximum load rating of home appliancesAppliances Peak Power Rating [Watts]Heater - 1 (Living Room) 2500Heater - 2 (Bedroom) 2000Heater - 3 (Kitchen) 1500Iron Center 1000Microwave 1100Dishwasher 1300Lighting 600Stove 5000Washer Dryer 5500Refrigerator 150There are four different actions that can be performed based on the states. Turningthe appliance Off, turning it On, pausing the operation and postponing the operation.For the case of simplicity, turning the appliance Off is considered as an required action.Also pausing and postponing the operation of the appliance can be selected for the sym-bolic Off state through the MCCU control based on the situation. The representationremains the same but power level changes. Moreover, we also define User Input Prefer-ences (UIP) as a user input control which changes the decision of the MCCU controlleralgorithm as desired by the user at a certain time interval. After the scheduling task isover, the control is shifted back to MCCU algorithm. Agent can move from one stateto another state after performing an action. The user inconvenience is modeled in thereward matrix and the goal of the strategy is to minimize the user inconvenience.1315.3.2 User Convenience and Reward MatrixIn this section, user convenience UC(t) is modeled at a time instant t and the goal is tomaximize the UC(t). The reward values for turning off an appliance is based on the userinconvenience. The parameters taken to model UC(t) are user preference (Pa(t)) ofthe appliances, power consumption energy cost saving (Ca(t)), and robustness (Sa(t)).Maximum inconvenience is caused by turning off an user preffered appliance at a giventime. The time slot is discretized for every 15 minutes regarding the preferences and isdivided into four times of the day i.e. Morning(MR), Afternoon(AF), Evening(EV) andNight(NT). Table 5.2 depicts the (Pa(t)) values of the appliances for different times ofthe day and the preferences are set according to a typical winter usage in Canada.Table 5.2: User preference of appliances (Pa)Appliances Morning(MR) Afternoon(AF) Evening(EV) Night(NT)Heater - 1 (Living Room) 1 0.3 1 0.3Heater - 2 (Bedroom) 1 0.3 0.4 1Heater - 3 (Kitchen) 0.6 0.3 0.7 0.1Iron Center 0.6 0.1 0.1 0.1Microwave 1 0.1 0.8 0.1Dishwasher 0.5 1 0.3 0.7Lighting 0.4 0.1 0.7 0.1Stove 0.7 0.1 1 0.1Washer Dryer 0.6 0.6 0.3 0.5User inconvenience UIC(t) due to turning off an appliance with preference repre-sented by Table 5.2 will becomeUIC(t) =C1 ·Pa(t) (5.4)C1 is a constant and is set to 1 to give user preference maximum importance whilechoosing the agent action. As appliances are turned off, energy savings in terms ofthe cost is achieved. So turning Off the maximum power consuming appliance at agiven time t will give the maximum convenience to the users in terms of cost savings.132Rest all appliances’ energy cost is normalized w.r.t the maximum power load of themaximum power consuming appliance at t. User inconvenience UIC(t) due to turningoff an appliance is also dependent on the cost saving (Ca(t)).UIC(t) =C2 · (1−Ca(t)) (5.5)The more the cost saving, lesser the user inconvenience. But cost cannot be savedsacrifising preference comfort for users. Hence, constant C2 will have lower contribu-tion to the UIC(t). We take C2 as 0.5 here for our case. Emergency (Ea(t)) gives usersoptions for choosing to start an appliance regardless of the time of the day, power con-sumed and preference control. When the user choose to run an appliance, it becomesa don’t care condition in the state for that instant t. Hence the number of state-actionpair for the reward matrix decreases. The appliance power is substracted fom the goalusage power.Section 5.2.2 describes how MQTT handles broker security with Password Authen-tication, Client ID Authentication, SSL/T LS Certification and firewalls. Robustness ofa system shows how it is immune to security threats and fault tolerant. Less robustsystem also creates inconvenience to the users. Robustness of the system is modelledbehaviouraly has been categorized as {Good, Medium and Bad}. For each behaviourof the system a constant C3 value have been assigned to the UIC(t) function asUIC(t) =C3 ·Sa(t)C3 =0.2,Good0.3,Medium0.5,Bad(5.6)133The user experiences more inconvenience for a Bad system as compared to a Goodsystem in terms of their robustness measure. User convenience UC(t) is calculatedfrom (5.4), (5.5) and (5.6) asUC(t) = 1−{C1Pa(t)+C2 (1−Ca(t))+C3Sa(t)3}(5.7)Algorithm 4: Reward Matrix (R) Computation AlgorithmInitialize: n,R(s,a) = Zeros(2n−1)Initialize: Threshold Power (Th)Load : Actual power consumption curve1 while (s,a) ! = (2n−1) do2 R(s,a)←−1; (State Transition Not possible)3 if cumulative power ≤ Th (Goal State) then4 R(s,a)←0; (Turning Off an Appliance)5 R(s,a)←1; (transition to the same state)6 else7 R(s,a)←0; (transition to the same state)8 R(s,a)←UC(t); (Otherwise)9 end10 endReward matrix (R) is based on the user convenience values for each appliance usingAlgorithm 4. Size of the reward matrix depends on the number of appliances and thenumber of power levels the house agent can occupy. The size of the reward matrix forthis problem is 255×255. Power level zero is not taken into consideration as it is im-possible for the power to reduce to zero level in a home throughout the day. Algorithm4 depicts the steps to formulate the reward matrix and is true for any number of state134Algorithm 5: Reward Matrix (R) Computation Algorithm1 {1: function UC(C1,C2,C3,Pa,Ca,Sa)if Goal reached without turning Off appliance thenUC(t)← 1endif Goal reached after turning Off appliance thenUC(t)← 1−UIC(t)endif Goal not reached after turning Off appliance thenUC(t)← 1−UIC(t)−0.2endif Goal reached but resulting power level ≤ 0.6·Th thenUC(t)← 1−UIC(t)−0.1endUC(t)←02: return UC(t)}transitions. According to required threshold power (T h) to be achieved, reward matrixR(s,a) is computed as per Algorithm 4. T h is the goal state where the optimization ofpower stops. Algorithm 5 depicts the process to compute the user convenience. Thegoal state may be reached with or without turning off an appliance. According to thepower level where the goal state is reached, user convenience value is penalized. Themost penalty is for goal state not being reached even after turning off an appliance. Thepenalties are 0.1 at goal state power less than or at 60% of threshold power and 0.2 forgoal not being reached even after turning off an appliance.5.4 Neural Fitted Q-based Home Energy ManagementThe proposed Neural Fitted Q-based Home Energy Management (NFQbHEM) algo-rithm is described in this section. The algorithm is based on RL based NFQI method135as in Section 5.3. The algorithm works in three phases: exploration, training and appli-cation. In the exploration phase, NFQbHEM captures the historical demand data basedon different seasons [163]. Winter month data has been chosen in our application. Thealgorithm is defined in Algorithm 6 and the steps are listed as follows,Algorithm 6: NFQbHEM AlgorithmInput : Define Q0 = q0s,a = 0, sk = T hOutput : pi(s)1 IT ER←2002 T k←empty set3 θ←randomweight4 while (||Q(s+1,a)−Q(s,a)||< 10−4) do5 qis,a = r(t) + γ·MAX Qi−1(s(t+1),a(t)) ;6 for 1:ITER do7 T k←T k−1∪(s,a;qi+1s,a );8 δ←Qˆ(s(t+1),a(t))−Qi(s(t+1),a(t));9 φk(s,a)←exp−||s− sk||22∗σ2k ;10 θ←θ+αδφ(s,a);11 end12 P(s,a)← eExplorationCount ·age ·Q(s,a)Σ(eExplorationCount ·age ·Q(s,a))13 end14 return pi(s)Exploration Phase:Step 0 (Inputs): Set the Q-factors to some arbitrary values (e.g. 0).Step 1 : For each state s, the set of admissible actions, a is defined, and an action a ∈ Ais chosen randomly and applied. After applying a(t) in s(t), the next state s(t + 1) isreached and the immediate reward r(t) from Algorithm 4 is calculated.Step 2 : The set of (s(t),a(t),R(t),s(t +1)) is inserted from the environment as a new136sample F . Repeating the process, sufficient samples are found to train the algorithm.Training Phase:Step 1 : The training initializes Q0 = q0s,a = 0, and tries to find a function approximatorQˆi.Step 2 : Similar to the Q-update process, append a corresponding pattern set T k to theset (s,a;qi+1s,a ).Step 3 : As our historical data is a curve fitting problem, Radial Basis Function NeuralNetwork (RBFNN) [164] is chosen to approximate the function Q(s,a).Step 4 : The feature function φ : S x A maps each state-action pair to a vector of featurevalues.Step 5 : θ is the weight vector specifying the contribution of each feature across allstate-action pairs. The weight is updated at each iteration. The training is done for 200iterations in our case.Execution Phase:Step 1 : Current data determine the state of the system.Step 2 : A greedy policy is used to find the policy pi(s) as in (5.3).Step 3 : Later in learning with more episodes, exploitation makes more sense because,with experience, the agent can be more confident about what it knows.Step 4 : Stopping criterian with absolute error||Q(s+1,a)−Q(s,a)||< 10−4.1375.5 Experimental ResultsThis section describes the results of HEMaaS platform with the NFQbHEM algorithmto control 10 appliances in a sample condo home in a smart community. The samplecondo home is a one bedroom condo with 10 appliances connected wirelessly to theMCCU. Power measurements have been taken consistently for a month and NFQb-HEM algorithm have been applied to the measured load. Due to the experimentalnature of the setup and results, the results have been presented in the context of thesample condo home. Comparison with other architectures in literature have not beendrawn as the method described here is unique to the setup and it would be unfair tocompare algorithms with different setup. Due to hardware complexity, it is very hardto implement other algorithms for the setup explained in this Chapter. As explainedin the software architecture and communication interface in Section 5.2.2, the MCCUconsists of a Raspberry-Pi3 deploying a node-red platform. MQTT (Mosquitto) is usedas the broker between the MCCU publisher and subscribing home appliances. Custompython code with the NFQbHEM algorithm deployed on it runs on the Raspberry-Pi3 tocontrol the home appliances’ Delay/Pause/On/Off operation via sonoff wi-fi switchesthrough a MQTT gateway. The node-red dashboard interface designed in this Chap-ter offers an easy and convenient user interface (UI) for a homeowner to interact withthe HEMaaS system. Figure 5.3 and Figure 5.4 illustrates our user interface (UI) flowdesign and dashboard control respectively. The UI shows the node-RED flow of thedifferent appliances as connected to the MCCU. Each appliance is controlled througha GPIO pin and follows the Home/(Room)/(Device)/Pin hierarchy. The proposed UIalso offers several visualization features to a user. They can have access to real-timeand historical appliance usage information with graphs via the sonoff accumulated data138information of real time power usage. User Input preferences (UIP) can be set via thedashboard. The different options which are available include setting a temperature forHeater-1, Heater-2, Heater-3 and washer-dryer, rescheduling washer-dryer operationand starting necessary appliances immediately bypassing the automated control for aparticular duration. The UI is also accessible from anywhere in the world via the smart-phone app. If for any reason there is a communication failure, the local settings of theappliances will take precedence.Figure 5.3: User interface design.139Figure 5.4: HEM interface.Matplot library of python gives us the tool to analyse the power demand data fordifferent cases. Two different cases have been discussed here for analyzing and plottingour results.Case I : A sample day’s total power consumption data is compared with different peakpower reduction of 5%,10%,15% and 20% of the total peak demand. The user conve-nience is also shown as a comparison.Case II : The user convenience in terms of random (Good, medium and bad) behaviorof the system is analyzed in this case.For the Case I above, the energy in KWh savings and reduction in carbon-footprint fora community consisting of 85 condos of our typical architecture as in Section 5.2.1 isalso plotted.140Figure 5.5: Plot of the total demand versus time during a typical Canadian wintermonth in Ontario5.5.1 Case IFigure 5.6: Plot of sample episodic run NFQbHEM learning process141In this section, the actual power consumption plot is generated using 10 smart appli-ances. The plot in Figure 5.5 shows the peak demand in watts versus time of the day.The interval of time duration is 15 minutes. Starting with initial Q(s,a), the HEMaaSplatform has to learn to find the optimal path when peak demand power during a cer-tain interval exceeds the available power. The available power is taken as a percentagereduction of the peak power. 5%,10%,15% and 20% are taken as the peak reductionpercentages to test and validate our methodology. Algorithm 6 has been initializedwith starting parameters of learning α = 0.5, discount γ = 0.8, A and B as 90 and 100respectively. The center state sk is taken as the median power consumed at a particularinterval. The peak power is 6300 watts and Algorithm 4 depicts the reward matrix ini-tial computation. The total energy consumption historical data of a typical condo hasbeen taken from national resources canada [165] for a typical winter month in Canadianontario province. The feature function φ is derived from approximating the curve ofthe historical data and is used to train the weight vector θ . When the total power con-sumption is greater than the peak power power reduction, it selects actions (randomly)and moves from current state to a new state, receives reward and then it starts issuingcontrol signals (Delay/Pause/On/O f f ) to other appliances until one of the goal statesis reached.Figure 5.6 depicts the learning process of the NFQbHEM algorithm. The graphis plotted between number of episodes algorithm running for look-up table based Q-learning and the neural network Q-learning based NFQbHEM algorithm. The pro-posed algorithm learns faster and reaches a stable value in only about 570 episodes ascompared to look-up only based Q-learning. The algorithm is stopped at 570 episodesas the error achieved is 10−5. Thus neural network modeler helps the Q-learningachieve its goal state faster. Figure 5.7 shows the total demand versus time for different142peak reduction percentages. Once the optimal policy is found, the MCCU will executethe sequence of rules (turning off appliances, rescheduling their timing of operation andtemperature control one by one) until the goal state with maximum user convenience isreached. At the optimal policy, MCCU determines when the power goes above the de-sired reduction, it modifies its power as in Figure 5.7. Table 5.3 shows the appropriateactions taken by the MCCU unit at varied time intervals for different appliances.Table 5.3: Actions taken by MCCUTimeRequiredLoad ReductionRequired Action5% Reduction Threshold10:15-10:30 am 400 W Turn off the Heater-1 and Heater-310% Reduction Threshold10:15-10:30 am 650 W Turn off the Heater-1, Heater-2 and Heater-36:00-6:15 pm 600 W Reduce the temp. setting of Heater-115% Reduction Threshold6:00-6:15 am 500 W Reduce the temp. setting of Heater-1 and Heater-210:00-10:15 am 1500 WThe temperature setting of the washer-dryer may bechanged to reduce the power demand or washer-dryeroperation may be rescheduled to another time.10:15-10:30 am 1500 WThe temperature setting of the washer-dryer maybe changed to reduce the power demand or washer-dryeroperation may be rescheduled to another time.5:00-5:15 pm 250 W Turn off the Heater-15:15-5:30 pm 300 W Turn off the Heater-26:00-6:15 pm 600 W Reduce the temp. setting of Heater-1143TimeRequiredLoad ReductionRequired Action20% Reduction Threshold6:00-6:15 am 500 W Reduce the temp. setting of Heater-1 and Heter-26:30-6:45 am 500 W Turn off the Heater-210:00-10:15 am 1500 WThe temperature setting of the washer-dryer maybe changed to reduce the power demand or washer-dryeroperation may be rescheduled to another time.10:15-10:30 am 1500 WThe temperature setting of the washer-dryer may bechanged to reduce the power demand or washer-dryeroperation may be rescheduled to another time.11:15-11:30 am 150 W Refrigerator Turned Off4:45-5:00 pm 150 W Turn off the Refrigerator5:15-5:30 pm 500 W Turn off the Heater-35:30-5:45 pm 800 W Turn off the Heater-2 and Heater-36:00-6:30 pm 600 W Reduce the temp. setting of Heater-1The user convenience (UC), is shown in Figure 5.8 for the four peak reductionthreshold values. It can be inferred from the figures that the UC decreases with theincrease in the threshold for power saving. Some of the peak load consumption whichlies during the afternoon and evening time slots are affected severely. One suggestionof improvement in the user convenience could be having a variable thresholds for theNFQbHEM algorithm. Therefore the times of day having maximum user utility powerconsumption, the available power threshold can be increased and can be compensatedwith a lower available power threshold during other Off peak times while maintainingthe overall average power threshold at the same level. If the user convenience level canbe maintained more than 70% for most times of the day, then the HEMaaS system can144Figure 5.7: Plot of the total demand versus time for different peak reduction per-centages145Figure 5.8: Plot of the user convenience (uc) versus timeFigure 5.9: Plot of the user convenience (uc) versus time for (20% Good, 60%Medium and 20% Bad) and (10% Good, 40% Medium and 50% Bad) ro-bustness measure.146be successful in delivering a coherent and inter-operable platform.In this section, the behavioral modeling of the system is considered in terms itsrobustness. Robustness measure evaluates a systems quality in terms of security andfault tolerance. To simulate this behavior in the system, the UC(t) from (5.6) has beenchosen with randomly assigning measure of robustness C3 at different time intervals.The power level of peak reduction at 15% is taken as the threshold. Figure 5.9 depictsthe UC w.r.t the time of the day and is compared for two different situations. Thefirst situation has (20% Good, 60% Medium and 20% Bad) robustness measure andthe second situation has (10% Good, 40% Medium and 50% Bad) robustness measurerespectively. The user convenience is severely affected for both cases specifically inthe second situation, due to presence of more faulty/malicious channel. Thus securityand fault tolerance is shown to have significant effect to the users. Once the UC goesbelow 50%, the system is considered as a very poorly managed system where users areforced to save energy sacrificing their comfort, which is highly undesirable.5.5.2 Case IIThe carbon intensity per KWh (CIPK) is a fundamental measure of a sustainable soci-ety. The lesser the CIPK, the better the society in terms of its environment and livabilityindex. The energy savings that are obtained from results in Section 5.5.1 can be seenas potential price saving for the community as well as a means of reducing the CO2gas emissions. As Canada is progressing towards a sustainable green building infras-tructure, it is a healthy sacrifice to have some inconvenience to achieve the greaterbenefit of having a greener environment in terms of achieving lesser carbon emissions.From [165], the Ontario province’s CIPK is obtained as 125 gr−CO2 per KWh. Usingthe CIPK, the energy savings and carbon emission savings have been computed for a147community consisting of 85 condos. Figure 5.10 shows the energy savings in Mega-Watt-hour (MWh) per year. It also shows the carbon-footprint savings in Kg−CO2per year. The improvement is nearly 14 times from 5% to 15% peak power reduction,which is quite substantial.(a) Energy savings with peak demand reduction(b) Plot of carbon-footprint reductionFigure 5.10: Comparison of peak reduction energy savings and carbon-footprintreductions148Chapter 6Conclusions and Future Work6.1 Summary and ConclusionsIn this thesis, the major challenges of energy efficient implementation of architecturesand technologies with respect to an IoE network have been discussed and differentsolutions to solve these problems have been critically evaluated. In Chapter 1, wehave discussed the layered architecture of IoE systems and showed how physical layer,monitoring and security layers are linked to each other. Our thesis tackles the issue ofenergy-efficient implementation in these layers. In Section 1.2, we categorize the prob-lem of energy-efficiency based on hardware design, wireless energy harvesting, energysaving policies, data transmission, management and control, and carbon-footprint gen-eration for IoE networks. We have proposed solutions to the issues mentioned in Sec-tion 1.2 through Chapter 2-5. In Chapter 2 and Chapter 3, issues related node batteryrelated issues are discussed and solutions were found out to increase the network life-time through wireless energy harvesting, data transmission , error correction codingand data awareness. In summary the major take aways of the work in this thesis are as149follows:• The major contribution of the work in Chapter 2 is to provide a solution fordata-utility lifetime trade-off problem by incorporating a detailed energy modelcombining various strategies of maximum network utilization and network life-time increase by error correction, proper duty cycling and wireless battery energyreplenishment. We provide user the flexibilty of choosing their system trade-offparameters by showing sumulation results for varied cases. This caters to a broadapplication scenario for the IoE systems having wireless sensing objects.• Continuing our objective of saving energy, in Chapter 3, we have applied ourenergy model solution from Chapter 2 to save energy through data awareness inan event driven IoE system as compared to a traditional WSN system. Our firstgoal is to apply the energy saving problem with respect to a IoE system and thenuse the diversified nature of the IoE systems to solve the problem and save batteryenergy and increase network lifetime.• Our major contribution in the Chapter 4, is the design of a low energy, resourcelimited PUF prototype current amplification circuit that mitigates the key attacksaimed at the system such as man in middle, cloning, and modelling attacks.Replicating and authenticating the system for the intruder is specifically blockedby our proposed solution. The results are verified by measurements and simula-tions. This provides the solution for an energy-efficient security design.• After proposing, testing, validating and implementing energy-efficient design forindividual layers and blocks in Fig. 1.2 through Chapter 2 to Chapter 3, in Chap-ter 5 we have implemented the algorithmic models in to smart homes as micro-150cosms of smart cities based on IoE systems (which has a system level implemen-tation and application). This implementation deals with the policy-based issuesthat have an impact throughout the system architecture. Through measurements,we validated our energy and data awareness model incorporating security for atypical IoE application scenario. This gives the users of this work flexibility inchoosing their system. It also shows the merits and demerits of applying eachcriteria of energy-efficient models of Chapter 2-4 to their overall convenienceand system’s QoS.Specifically in Chapter 2, Wireless energy harvesting is investigated as a remedyto prolong the lifetime of sensor nodes and enable maintenance-free operation. Wake-up radio scheme is incorporated as an efficient solution to address the idle listeningenergy dissipation of sensor nodes. RRNS Error control coding is proposed to improvethe reliability of the transmission and reduce re-transmission, hence, reducing energyconsumption. A utility-lifetime maximization problem incorporating WEH, WUR andECC schemes is formulated and solved using distributed dual subgradient algorithmbased on Lagrange multiplier method. Simulation results verify the effectiveness ofthe proposed schemes in reducing the energy consumption and accordingly, carbonfootprint of wireless sensor nodes, providing the means for a greener wireless sensornetwork.In Chapter 3, we propose a Data aware energy efficient distributed clustering pro-tocol for IoT (DAEECI) by saving CH selection energy using active RFID tags, cuttingprocessing energy by incorporating data awareness factor in the system and improv-ing lifetime by inculcating RF energy harvesting. We propose a PMU architecture thataccommodates a battery charging scheme using the harvested energy through a WEHunit. We formulate energy consumption models in each round data is sent from sensor151nodes to BS through gateway nodes. Our simulation depict substantial improvement inlifetime of network and data delivery to the BS.The hardware-based related energy efficiency issues are dealt in Chapter 4. Theseissues are discussed in terms of security layer implementation of an IoE system. Theenergy-efficient security hardware design is an important part of the IoE system whichcan’t be neglected. A hardware based energy-efficient PUF current amplication proto-type have been developed and tested. Specifically, in Chapter4, we have proposed anIoT sensor PUF-based security design that exploits variations of physical sensor char-acteristics (e.g., dark current, is presented in this work) and challenge/response pairgeneration using the quadratic residue property. We have proposed algorithms for de-vice authentication and encryption by using the PUF challenge/response outputs. Ouranalysis shows that there is strong relationship between the energy consumption of thesensor node and the dynamic range of the TransImpedance Amplifier (TIA) circuit,which is determined by signal strength and noise at the output of TIA. Thus, one ofpossible design choices for configuring the PUF circuit is to use RF = 50MOhm andCF= 10pF to get a dynamic range of 130.3 (7bits) with energy consumption of 0.157 J.Through simulations and measurements, we have shown that design is better in termsof energy and costs requirements than other state-of-the-art security algorithms. More-over, it provides a two-fold secure data transfer and is resilient towards various attacks.This method can be extended to other IoT sensors, if suitable physically varying andunclonable circuit properties are chosen.Energy management in smart cities is an indispensable challenge to address due torapid urbanization. In Chapter 5, we first present an overview of energy managementin smart homes to build a green and sustainable smart city, and then present a unify-ing framework for IoT in building green smart homes. To achieve our goal, a neural152network based Q-learning algorithm is proposed to reduce the peak load demand ofa typical Canadian home while minimizing the user inconvenience and enhancing therobustness of the system. The user convenience level for 5% and 10% load reductionis maintained at and above 80%. Whereas other levels of peak power reduction causesmore discomfort for the users. While Canada Green Building Council is working to-wards finding ways of making buildings greener and community sustainable, a novelmethod has been applied for finding suitable strategies for efficient management ofhome energy demand and reducing the energy consumption during peak period in atypical Canadian condo. In a purely market-driven energy management environment,peak-reduction is mostly valuable for utilities. In order to make the demand side man-agement more user friendly and consumer centric, a reward matrix based self-learningalgorithm has been applied. The energy savings and carbon-footprint reduction is alsoshown to be quite significant. In future, it has been planned to incorporate real timescheduling into the system to schedule and pause appliance operation. Moreover, it isalso proposed to design a system that learns from feedback smart sensors in the envi-ronment to ease the MCCU decision making and reduce user input, yet still maintaininga high enough user convenience.6.2 Future WorkThe approaches presented in this thesis are not exhaustive. In this section, we proposesome possible research directions that can be followed from this thesis.1536.2.1 Highly Efficient, Low-cost, and Small-Form-Factor WirelessEnergy Harvesting SystemThe key challenge in successful large-scale deployment of sensor devices in an IoEinfrastructure is to minimize their impact on users and the environment. Non-intrusivedevices need to be small, be fabricated and deployed at very low cost, and are expectedto operate in a selfsufficient manner for a long time. A WEH unit as an integral partof such devices must comply with such cost and size requirements. Efficiency is an-other crucial factor for a WEH system. High efficiency becomes increasingly relevantconsidering that the transmitted power by the dedicated source is usually limited dueto health issues and interference constraints. Commercial RF harvesting systems cur-rently existing in the market enable single-band RF harvesting at sub-milliwatt powerlevels with efficiencies as high as 50 percent. However, extensive studies are still be-ing carried out to improve the performance of WEH systems at the circuit and systemlevels. Energy beamforming [166], high gain antennas, and multi-band harvesting areamong the other hot topics in the context of WEH systems for IoE.6.2.2 Channel Statistics for IoE SystemsThe scenarios and their respective analysis in our thesis in chapter 2 and Chapter 3 as-sume the channel as static and time invariant. Practically, channel characteristics varydepending on the environment in which the number of interferers and the number ofpaths available from source device to sink. Harvested energy depends on the distancebetween sink and sensor node. In the presence of fading or multipath, the receivedenergy for the purpose of harvesting and the transmitted data are adversely affected. In[167], a compressive sensing based approach is proposed to recover sparse signals frommultiple spatially correlated data transmitted to a fusion center. Recently, in [168], re-154searchers have proposed techniques to reduce the amount of packets to be retransmittedin case of faulty transmission, eventually saving energy.6.2.3 Cross-Layer DesignAlthough in Chapter 3 we have used data-awareness for our design in the physicallayer, the sensor device still has to operate in duty-cycled mode due to limited en-ergy collection from the environment, and dynamically adjust duty cycles to adapt tothe availability of environmental energy. Such dynamic duty cycles pose challengesfor medium access control (MAC) layer protocol design in terms of synchronization,reliability, efficiency of utilizing channel resource and energy, and so on. Therefore,solutions of duty-cycling-aware middleware between MAC and physical layer powermanagement are highly desired. Moreover, dynamic duty cycling also has nontrivialimpact on the end-to-end performance of the network layer, including end-to-end de-lay, throughput, and so on. However, the current routing protocol design for IoT haspaid very little attention to duty cycling. The problem of seamlessly integrating duty-cycle awareness into the multi-path routing scenario has been dealt with in [169] usinga sleep scheduling mechanism; however, it still remains an open question.6.2.4 Security and Privacy ConcernsIn Chapter 4, we have not dealt with profiles of same manufacturer’s photodiode darkcurrents in detail. In the future work, investigating the range of dark current profilefor the large sample of the same type of photodiode can provide a useful backgroundto determine the required range of bias voltage. Then, the existing actual dark currentprofile can be used to evaluate the necessity of a step-up voltage converter which canboost the typical voltage range used for a microcontroller (3.3V 5.0V) to the voltage155level which is large enough for biasing a photodiode. In our prototype circuit, separatediscrete components such as a TIA and an opamp were used with an external powersource. However, these circuits can be more optimally designed and integrated in asingle die to increase the performance of a sensor node such as lower noise and reducedpower consumption. Integrating to a single chip would also increase the portability ofa sensor node, which is one of key requirements for IoT application.6.2.5 Home Energy ManagementReal-time management is a challenging issue for resource constrained sensor networks.In the Chapter 5, the IoE system needs to rely on effcient service gateway to minimizethe amount of data to be sent by constantly receiving the feedback data from users, andintelligent data oriented middleware design to only transmit real time information whena reward matrix is to be calculated. The modelling is done through radial-basis neuralnetwork. In future works, deep learning with boosting can be applied to faster train thedata to achieve better models.Dynamic registration, bootstrap and management will be particularly considered fora large scale deployment with devices coming in and out and changing their character-istics and functionalities. The IoE device management should be suitable to develop anopen and universal ecosystem with sustainable interactions and interoperability amongthings.156Bibliography[1] D. Evans, The Internet of Everything: How More Relevant and Valuable Con-nections Will Change the World. CISCO Internet Business Solutions Group(IBSG), 2012. [Online]. Available: my/assets/ciscoinnovate/pdfs/IoE.pdf[2] A. Zanella, N. Bui, A. Castellani, L. Vangelista, and M. Zorzi, “Internet of thingsfor smart cities,” IEEE Internet of Things Journal, vol. 1, no. 1, pp. 22–32, Feb2014.[3] Z. Sheng, S. Yang, Y. Yu, A. V. Vasilakos, J. A. Mccann, and K. K. Leung, “Asurvey on the ietf protocol suite for the internet of things: standards, challenges,and opportunities,” IEEE Wireless Communications, vol. 20, no. 6, pp. 91–98,December 2013.[4] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and M. Ayyash, “In-ternet of things: A survey on enabling technologies, protocols, and applica-tions,” IEEE Communications Surveys Tutorials, vol. 17, no. 4, pp. 2347–2376,Fourthquarter 2015.[5] L. Atzori, A. Iera, and G. Morabito, “The internet of things: A survey,” Com-puter networks, vol. 54, no. 15, pp. 2787–2805, 2010.[6] S. Sen, J. Koo, and S. Bagchi, “Trifecta: Security, energy efficiency, and com-munication capacity comparison for wireless iot devices,” IEEE Internet Com-puting, vol. 22, no. 1, pp. 74–81, Jan 2018.[7] IEEE-SA-P2413, Standard for an Architectural Framework for the Internet ofThings (IoT). IEEE Standards Association, 2016. [Online]. Available:[8] A. Ahmed and E. Ahmed, “A survey on mobile edge computing,” in 2016 10thInternational Conference on Intelligent Systems and Control (ISCO), Jan 2016,pp. 1–8.157[9] J. Lin, W. Yu, N. Zhang, X. Yang, H. Zhang, and W. Zhao, “A survey on in-ternet of things: Architecture, enabling technologies, security and privacy, andapplications,” IEEE Internet of Things Journal, vol. 4, no. 5, pp. 1125–1142, Oct2017.[10] R. Arshad, S. Zahoor, M. A. Shah, A. Wahid, and H. Yu, “Green iot: An inves-tigation on energy saving practices for 2020 and beyond,” IEEE Access, vol. 5,pp. 15 667–15 681, 2017.[11] J. Huang, Y. Meng, X. Gong, Y. Liu, and Q. Duan, “A novel deployment schemefor green internet of things,” IEEE Internet of Things Journal, vol. 1, no. 2, pp.196–205, April 2014.[12] R. Want, B. N. Schilit, and S. Jenson, “Enabling the internet of things,” Com-puter, vol. 48, no. 1, pp. 28–35, Jan 2015.[13] D. Bol, J. D. Vos, F. Botman, G. de Streel, S. Bernard, D. Flandre, and J. D.Legat, “Green socs for a sustainable internet-of-things,” in 2013 IEEE FaibleTension Faible Consommation, June 2013, pp. 1–4.[14] Y. W. Lim, S. B. Daas, S. J. Hashim, R. M. Sidek, N. A. Kamsani, and F. Z.Rokhani, “Reduced hardware architecture for energy-efficient iot healthcare sen-sor nodes,” in 2015 IEEE International Circuits and Systems Symposium (IC-SyS), Sept 2015, pp. 90–95.[15] M. Sain, Y. J. Kang, and H. J. Lee, “Survey on security in internet of things: Stateof the art and challenges,” in 2017 19th International Conference on AdvancedCommunication Technology (ICACT), Feb 2017, pp. 699–704.[16] A. M. Nia and N. K. Jha, “A comprehensive study of security of internet-of-things,” IEEE Transactions on Emerging Topics in Computing, vol. PP, no. 99,pp. 1–1, 2017.[17] J. Gilbert and F. Balouchi, “Comparison of energy harvesting systems for wire-less sensor networks,” International Journal of Automation and Computing,vol. 5, no. 4, pp. 334–347, 2008.[18] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, “Wireless networks with RFenergy harvesting: A contemporary survey,” CoRR, vol. abs/1406.6470, 2014.[Online]. Available:[19] C. Mahapatra, Z. Sheng, V. Leung, and T. Stouraitis, “A reliable and energy ef-ficient iot data transmission scheme for smart cities based on redundant residuebased error correction coding,” in Sensing, Communication, and Networking -158Workshops (SECON Workshops), 2015 12th Annual IEEE International Confer-ence on, June 2015, pp. 1–6.[20] X. Peng, M. Bessho, N. Koshizuka, and K. Sakamura, “Epdl: Supportingcontext-based energy control policy design in iot-enabled smart buildings: Pro-graming the physical world with epdl,” in Data Science and Data Intensive Sys-tems (DSDIS), 2015 IEEE International Conference on. IEEE, 2015, pp. 297–303.[21] A. Fensel, V. Kumar, and S. D. K. Tomic, “End-user interfaces for energy-efficient semantically enabled smart homes,” Energy Efficiency, vol. 7, no. 4,pp. 655–675, 2014.[22] M. Moreno, B. U´beda, A. F. Skarmeta, and M. A. Zamora, “How can we tackleenergy efficiency in iot basedsmart buildings?” Sensors, vol. 14, no. 6, pp. 9582–9614, 2014.[23] E. Capo-Chichi, H. Guyennet, and J.-M. Friedt, “K-rle: A new data compres-sion algorithm for wireless sensor network,” in Sensor Technologies and Ap-plications, 2009. SENSORCOMM ’09. Third International Conference on, June2009, pp. 502–507.[24] T. Schoellhammer, B. Greenstein, E. Osterweil, M. Wimbrow, and D. Estrin,“Lightweight temporal compression of microclimate datasets [wireless sensornetworks],” in Local Computer Networks, 2004. 29th Annual IEEE InternationalConference on, Nov 2004, pp. 516–524.[25] N. Xu, S. Rangwala, K. K. Chintalapudi, D. Ganesan, A. Broad, R. Govindan,and D. Estrin, “A wireless sensor network for structural monitoring,”in Proceedings of the 2Nd International Conference on Embedded NetworkedSensor Systems, ser. SenSys ’04. New York, NY, USA: ACM, 2004, pp.13–24. [Online]. Available:[26] M. Rubin and T. Camp, “On-mote compressive sampling to reduce power con-sumption for wireless sensors,” in Sensor, Mesh and Ad Hoc Communicationsand Networks (SECON), 2013 10th Annual IEEE Communications Society Con-ference on, June 2013, pp. 291–299.[27] E. Candes and M. Wakin, “An introduction to compressive sampling,” SignalProcessing Magazine, IEEE, vol. 25, no. 2, pp. 21–30, March 2008.[28] N. Kimura and S. Latifi, “A survey on data compression in wireless sensor net-works,” in Information Technology: Coding and Computing, 2005. ITCC 2005.International Conference on, vol. 2, April 2005, pp. 8–13 Vol. 2.159[29] G. Bovet, A. Ridi, and J. Hennebert, “Machine learning with the internet of vir-tual things,” in Protocol Engineering (ICPE) and International Conference onNew Technologies of Distributed Systems (NTDS), 2015 International Confer-ence on, July 2015, pp. 1–8.[30] GeSi better 2030. GeSi, 2018. [Online]. Available:[31] F. P. Kelly, A. K. Maulloo, and D. K. Tan, “Rate control for communicationnetworks: shadow prices, proportional fairness and stability,” Journal of the Op-erational Research society, pp. 237–252, 1998.[32] R. Madan and S. Lall, “Distributed algorithms for maximum lifetime routingin wireless sensor networks,” Wireless Communications, IEEE Transactions on,vol. 5, no. 8, pp. 2185–2193, Aug 2006.[33] S. Ehsan, B. Hamdaoui, and M. Guizani, “Radio and medium access contentionaware routing for lifetime maximization in multichannel sensor networks,” Wire-less Communications, IEEE Transactions on, vol. 11, no. 9, pp. 3058–3067,September 2012.[34] J. Chen, W. Xu, S. He, Y. Sun, P. Thulasiraman, and X. Shen, “Utility-basedasynchronous flow control algorithm for wireless sensor networks,” Selected Ar-eas in Communications, IEEE Journal on, vol. 28, no. 7, pp. 1116–1126, Sep.2010.[35] S. He, J. Chen, D. Yau, and Y. Sun, “Cross-layer optimization of correlated datagathering in wireless sensor networks,” in Sensor Mesh and Ad Hoc Communica-tions and Networks (SECON), 2010 7th Annual IEEE Communications SocietyConference on, June 2010, pp. 1–9.[36] D. S. Lun, M. Me´dard, R. Koetter, and M. Effros, “On coding for reliable com-munication over packet networks,” Physical Communication, vol. 1, no. 1, pp.3–20, 2008.[37] K. Yu, F. Barac, M. Gidlund, and J. Akerberg, “Adaptive forward error correctionfor best effort wireless sensor networks,” in Communications (ICC), 2012 IEEEInternational Conference on, June 2012, pp. 7104–7109.[38] W. Xu, Q. Shi, X. Wei, Z. Ma, X. Zhu, and Y. Wang, “Distributed optimal ratereliability lifetime tradeoff in time varying wireless sensor networks,” WirelessCommunications, IEEE Transactions on, vol. 13, no. 9, pp. 4836–4847, Sept2014.160[39] J. Zou, H. Xiong, C. Li, R. Zhang, and Z. He, “Lifetime and distortion optimiza-tion with joint source/channel rate adaptation and network coding-based errorcontrol in wireless video sensor networks,” Vehicular Technology, IEEE Trans-actions on, vol. 60, no. 3, pp. 1182–1194, March 2011.[40] T. He, K.-W. Chin, and S. Soh, “On wireless power transfer and max flow inrechargeable wireless sensor networks,” IEEE Access, vol. 4, pp. 4155–4167,2016.[41] M. Magno, D. Boyle, D. Brunelli, E. Popovici, and L. Benini, “Ensuring surviv-ability of resource-intensive sensor networks through ultra-low power overlays,”Industrial Informatics, IEEE Transactions on, vol. 10, no. 2, pp. 946–956, May2014.[42] R. Deng, Y. Zhang, S. He, J. Chen, and X. Shen, “Globally optimizing net-work utility with spatiotemporally-coupled constraint in rechargeable sensor net-works,” in Global Communications Conference (GLOBECOM), 2013 IEEE, Dec2013, pp. 4810–4815.[43] P. Kamalinejad, C. Mahapatra, Z. Sheng, S. Mirabbasi, V. Leung, and Y. L.Guan, “Wireless energy harvesting for the internet of things,” CommunicationsMagazine, IEEE, vol. 53, no. 6, pp. 102–108, June 2015.[44] Z. Mao, C. Koksal, and N. Shroff, “Near optimal power and rate control of multi-hop sensor networks with energy replenishment: Basic limitations with finiteenergy and data storage,” Automatic Control, IEEE Transactions on, vol. 57,no. 4, pp. 815–829, April 2012.[45] S. Chen, P. Sinha, N. Shroff, and C. Joo, “Finite-horizon energy allocation androuting scheme in rechargeable sensor networks,” in INFOCOM, 2011 Proceed-ings IEEE, April 2011, pp. 2273–2281.[46] A. Biason and M. Zorzi, “Joint online transmission and energy transfer policiesfor energy harvesting devices with finite batteries,” in European Wireless 2015;21th European Wireless Conference; Proceedings of. VDE, 2015, pp. 1–7.[47] G. Koutitas, “Green network planning of single frequency networks,” Broadcast-ing, IEEE Transactions on, vol. 56, no. 4, pp. 541–550, Dec 2010.[48] M. Naeem, U. Pareek, D. C. Lee, and A. Anpalagan, “Estimation of distributionalgorithm for resource allocation in green cooperative cognitive radio sensor net-works,” Sensors, vol. 13, no. 4, pp. 4884–4905, 2013.161[49] J. N. Al-Karaki and A. E. Kamal, “Routing techniques in wireless sensor net-works: a survey,” Wireless communications, IEEE, vol. 11, no. 6, pp. 6–28,2004.[50] W. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “An application-specific protocol architecture for wireless microsensor networks,” Wireless Com-munications, IEEE Transactions on, vol. 1, no. 4, pp. 660–670, 2002.[51] S. Lindsey and C. Raghavendra, “Pegasis: Power-efficient gathering in sensorinformation systems,” in Aerospace Conference Proceedings, 2002. IEEE, vol. 3,2002, pp. 3–1125–3–1130 vol.3.[52] O. Younis and S. Fahmy, “Heed: a hybrid, energy-efficient, distributed clusteringapproach for ad hoc sensor networks,” Mobile Computing, IEEE Transactionson, vol. 3, no. 4, pp. 366–379, 2004.[53] L. Qing, Q. Zhu, and M. Wang, “Design of a distributed energy-efficient cluster-ing algorithm for heterogeneous wireless sensor networks,” Computer commu-nications, vol. 29, no. 12, pp. 2230–2237, 2006.[54] T. Qureshi, N. Javaid, M. Malik, U. Qasim, and Z. Khan, “On performance eval-uation of variants of deec in wsns,” in Broadband, Wireless Computing, Com-munication and Applications (BWCCA), 2012 Seventh International Conferenceon, Nov 2012, pp. 162–169.[55] K. Zhao and L. Ge, “A survey on the internet of things security,” in 2013Ninth International Conference on Computational Intelligence and Security,Dec 2013, pp. 663–667.[56] J. Granjal, E. Monteiro, and J. S. Silva, “Security for the internet of things: Asurvey of existing protocols and open research issues,” IEEE CommunicationsSurveys Tutorials, vol. 17, no. 3, pp. 1294–1312, thirdquarter 2015.[57] B. Schneier, Applied Cryptography: Protocols, Algorithms, and Source Code inC, John Wiley & Sons, 2007.[58] A. J. Menezes, P. C. Van Oorschot, and S. A. Vanstone, Handbook of AppliedCryptography, CRC press, 1996.[59] C. Karlof, N. Sastry, and D. Wagner, “TinySec: a link layer security architecturefor wireless sensor networks,” in Proceedings of the 2nd international confer-ence on Embedded networked sensor systems. ACM, 2004, pp. 162–175.[60] R. Roman, P. Najera, and J. Lopez, “Securing the Internet of things,” Computer,vol. 44, no. 9, pp. 51–58, 2011.162[61] M. Potkonjak and V. Goudar, “Public Physical Unclonable Functions,” Proceed-ings of the IEEE, vol. 102, no. 8, pp. 1142–1156, Aug. 2014.[62] C. Herder, M.-D. Yu, F. Koushanfar, and S. Devadas, “Physical UnclonableFunctions and Applications: A Tutorial,” Proceedings of the IEEE, vol. 102,no. 8, pp. 1126–1141, Aug. 2014.[63] Y. Gao, D. C. Ranasinghe, S. F. Al-Sarawi, O. Kavehei, and D. Abbott, “Emerg-ing physical unclonable functions with nanotechnology,” IEEE Access, vol. 4,pp. 61–80, 2016.[64] U. Ru¨hrmair, F. Sehnke, J. So¨lter, G. Dror, S. Devadas, and J. Schmidhuber,“Modeling attacks on physical unclonable functions,” in Proceedings of the 17thACM conference on Computer and communications security. ACM, 2010, pp.237–249.[65] J. Guajardo, S. S. Kumar, G. J. Schrijen, and P. Tuyls, “Fpga intrinsic pufs andtheir use for ip protection,” in CHES, vol. 4727. Springer, 2007, pp. 63–80.[66] S. S. Kumar, J. Guajardo, R. Maes, G. J. Schrijen, and P. Tuyls, “Extended ab-stract: The butterfly puf protecting ip on every fpga,” in 2008 IEEE InternationalWorkshop on Hardware-Oriented Security and Trust, June 2008, pp. 67–70.[67] P. Tuyls, G.-J. Schrijen, B. Sˇkoric´, J. Van Geloven, N. Verhaegh, and R. Wolters,“Read-proof hardware from protective coatings,” in International Workshop onCryptographic Hardware and Embedded Systems. Springer, 2006, pp. 369–383.[68] J. Delvaux, D. Gu, D. Schellekens, and I. Verbauwhede, “Secure lightweightentity authentication with strong pufs: Mission impossible?” in InternationalWorkshop on Cryptographic Hardware and Embedded Systems. Springer, 2014,pp. 451–475.[69] U. Ru¨hrmair, J. So¨lter, F. Sehnke, X. Xu, A. Mahmoud, V. Stoyanova, G. Dror,J. Schmidhuber, W. Burleson, and S. Devadas, “Puf modeling attacks on simu-lated and silicon data,” IEEE Transactions on Information Forensics and Secu-rity, vol. 8, no. 11, pp. 1876–1891, 2013.[70] R. Pappu, B. Recht, J. Taylor, and N. Gershenfeld, “Physical one-way functions,”Science, vol. 297, no. 5589, pp. 2026–2030, 2002.[71] B. Gassend, D. Clarke, M. Van Dijk, and S. Devadas, “Silicon physical randomfunctions,” in Proceedings of the 9th ACM conference on Computer and commu-nications security. ACM, 2002, pp. 148–160.163[72] G. E. Suh and S. Devadas, “Physical unclonable functions for device authen-tication and secret key generation,” in Proceedings of the 44th annual designautomation conference. ACM, 2007, pp. 9–14.[73] D. Lim, J. W. Lee, B. Gassend, G. E. Suh, M. van Dijk, and S. Devadas, “Ex-tracting secret keys from integrated circuits,” IEEE Transactions on Very LargeScale Integration (VLSI) Systems, vol. 13, no. 10, pp. 1200–1205, Oct 2005.[74] J.-L. Zhang, G. Qu, Y.-Q. Lv, and Q. Zhou, “A survey on silicon pufs and recentadvances in ring oscillator pufs,” Journal of computer science and technology,vol. 29, no. 4, pp. 664–678, 2014.[75] D. Choi, S. H. Seo, Y. S. Oh, and Y. Kang, “Two-factor fuzzy commitment forunmanned iot devices security,” IEEE Internet of Things Journal, pp. 1–1, 2018.[76] C. Marchand, L. Bossuet, U. Mureddu, N. Bochard, A. Cherkaoui, and V. Fis-cher, “Implementation and characterization of a physical unclonable functionfor iot: A case study with the tero-puf,” IEEE Transactions on Computer-AidedDesign of Integrated Circuits and Systems, vol. 37, no. 1, pp. 97–109, Jan 2018.[77] D. Mukhopadhyay, “Pufs as promising tools for security in internet of things,”IEEE Design Test, vol. 33, no. 3, pp. 103–115, June 2016.[78] B. Zhou, W. Li, K. W. Chan, Y. Cao, Y. Kuang, X. Liu, and X. Wang, “Smarthome energy management systems: Concept, configurations, and schedulingstrategies,” Renewable and Sustainable Energy Reviews, vol. 61, pp. 30–40,2016.[79] D. Dı´az Pardo de Vera, A. Siguenza Izquierdo, J. Bernat Vercher, and L. A.Herna´ndez Go´mez, “A ubiquitous sensor network platform for integrating smartdevices into the semantic sensor web,” Sensors, vol. 14, no. 6, pp. 10 725–10 752,2014. [Online]. Available:[80] H. Farhangi, “The path of the smart grid,” IEEE power and energy magazine,vol. 8, no. 1, 2010.[81] A. Kumar and G. Hancke, “An energy-efficient smart comfort sensing systembased on the IEEE 1451 standard for green buildings,” IEEE Sensors Journal,vol. 14, no. 12, pp. 4245–4252, 2014.[82] V. R. Shen, C.-Y. Yang, and C. H. Chen, “A smart home management sys-tem with hierarchical behavior suggestion and recovery mechanism,” ComputerStandards & Interfaces, vol. 41, pp. 98–111, 2015.164[83] M. Rahman, M. Kuzlu, M. Pipattanasomporn, and S. Rahman, “Architectureof web services interface for a home energy management system,” InnovativeSmart Grid Technologies Conference (ISGT), 2014 IEEE PES, pp. 1–5, 2014.[84] Y.-T. Lee, W.-H. Hsiao, C.-M. Huang, and T. C. Seng-cho, “An integrated cloud-based smart home management system with community hierarchy,” IEEE Trans-actions on Consumer Electronics, vol. 62, no. 1, pp. 1–9, 2016.[85] F. Ciancetta, B. D’Apice, D. Gallo, and C. Landi, “Plug-n-play smart sensorbased on web service,” IEEE Sensors Journal, vol. 7, no. 5, pp. 882–889, 2007.[86] C. Chen, J. Wang, Y. Heo, and S. Kishore, “Mpc-based appliance schedulingfor residential building energy management controller,” IEEE Transactions onSmart Grid, vol. 4, no. 3, pp. 1401–1410, 2013.[87] S. Li, D. Zhang, A. B. Roget, and Z. O’Neill, “Integrating home energy simu-lation and dynamic electricity price for demand response study,” IEEE Transac-tions on Smart Grid, vol. 5, no. 2, pp. 779–788, 2014.[88] Q. Wei, F. L. Lewis, G. Shi, and R. Song, “Error-tolerant iterative adaptive dy-namic programming for optimal renewable home energy scheduling and batterymanagement,” IEEE Transactions on Industrial Electronics, vol. 64, no. 12, pp.9527–9537, Dec 2017.[89] A.-H. Mohsenian-Rad, V. W. Wong, J. Jatskevich, R. Schober, and A. Leon-Garcia, “Autonomous demand-side management based on game-theoretic en-ergy consumption scheduling for the future smart grid,” IEEE transactions onSmart Grid, vol. 1, no. 3, pp. 320–331, 2010.[90] K. Dehghanpour, H. Nehrir, J. Sheppard, and N. Kelly, “Agent-based modelingof retail electrical energy markets with demand response,” IEEE Transactions onSmart Grid, vol. PP, no. 99, pp. 1–1, 2016.[91] M. Magno, T. Polonelli, L. Benini, and E. Popovici, “A low cost, highly scalablewireless sensor network solution to achieve smart led light control for greenbuildings,” IEEE Sensors Journal, vol. 15, no. 5, pp. 2963–2973, 2015.[92] D. O’Neill, M. Levorato, A. Goldsmith, and U. Mitra, “Residential demand re-sponse using reinforcement learning,” 2010 First IEEE International Conferenceon Smart Grid Communications, pp. 409–414, 2010.[93] F. Ruelens, B. J. Claessens, S. Vandael, S. Iacovella, P. Vingerhoets, and R. Bel-mans, “Demand response of a heterogeneous cluster of electric water heatersusing batch reinforcement learning,” 2014 Power Systems Computation Confer-ence, pp. 1–7, 2014.165[94] K. Turitsyn, S. Backhaus, M. Ananyev, and M. Chertkov, “Smart finite statedevices: A modeling framework for demand response technologies,” 2011 50thIEEE Conference on Decision and Control and European Control Conference,pp. 7–14, 2011.[95] E. C. Kara, M. Berges, B. Krogh, and S. Kar, “Using smart devices for system-level management and control in the smart grid: A reinforcement learning frame-work,” 2012 IEEE Third International Conference on Smart Grid Communica-tions (SmartGridComm), pp. 85–90, 2012.[96] G. Deak, K. Curran, and J. Condell, “A survey of active and passive indoorlocalisation systems,” Computer Communications, vol. 35, no. 16, pp. 1939–1954, 2012.[97] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, “Wireless networks withrf energy harvesting: A contemporary survey,” IEEE Communications SurveysTutorials, vol. 17, no. 2, pp. 757–789, Secondquarter 2015.[98] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge universitypress, 2004.[99] H. How, T. Liew, E.-L. Kuan, L.-L. Yang, and L. Hanzo, “A redundant residuenumber system coded burst-by-burst adaptive joint-detection based cdma speechtransceiver,” Vehicular Technology, IEEE Transactions on, vol. 55, no. 1, pp.387–396, Jan 2006.[100] B. Zarei, V. Muthukkumarasay, and X.-W. Wu, “A residual error control schemein single-hop wireless sensor networks,” in Advanced Information Networkingand Applications (AINA), 2013 IEEE 27th International Conference on, March2013, pp. 197–204.[101] M. C. Vuran and I. Akyildiz, “Error control in wireless sensor networks: Across layer analysis,” Networking, IEEE/ACM Transactions on, vol. 17, no. 4,pp. 1186–1199, Aug 2009.[102] J.-D. Sun and H. Krishna, “A coding theory approach to error control in redun-dant residue number systems. ii. multiple error detection and correction,” Cir-cuits and Systems II: Analog and Digital Signal Processing, IEEE Transactionson, vol. 39, no. 1, pp. 18–34, Jan 1992.[103] W. Jenkins and E. Altman, “Self-checking properties of residue number errorcheckers based on mixed radix conversion,” Circuits and Systems, IEEE Trans-actions on, vol. 35, no. 2, pp. 159–167, Feb 1988.166[104] M. Russo, P. Sˇolic´, and M. Stella, “Probabilistic modeling of harvested GSMenergy and its application in extending UHF RFID tags reading range,” Journalof Electromagnetic Waves and Applications, vol. 27, no. 4, pp. 473–484, 2013.[Online]. Available:[105] P. Kamalinejad, K. Keikhosravy, R. Molavi, S. Mirabbasi, and V. Leung, “Ef-ficiency enhancement techniques and a dual-band approach in RF rectifiers forwireless power harvesting,” in IEEE International Symposium on Circuits andSystems (ISCAS),, June 2014, pp. 2049–2052.[106] V. Jelicic, M. Magno, D. Brunelli, V. Bilas, and L. Benini, “Analyticcomparison of wake-up receivers for WSNs and benefits over the wake-on radio scheme,” in Proceedings of the 7th ACM Workshop on PerformanceMonitoring and Measurement of Heterogeneous Wireless and Wired Networks,ser. PM2HW2N ’12. New York, NY, USA: ACM, 2012, pp. 99–106. [Online].Available:[107] V. Jelicic, M. Magno, D. Brunelli, V. Bilas, L. Benini, “Bene-fits of wake-up radio in energy-efficient multimodal surveillance wireless sensornetwork,” IEEE Sensors Journal,, vol. 14, no. 9, pp. 3210–3220, Sept 2014.[108] P. Kamalinejad, K. Keikhosravy, M. Magno, S. Mirabbasi, V. Leung, andL. Benini, “A high-sensitivity fully passive wake-up radio front-end for wire-less sensor nodes,” in IEEE International Conference on Consumer Electronics(ICCE),, Jan 2014, pp. 209–210.[109] M. Bhardwaj and A. Chandrakasan, “Bounding the lifetime of sensor networksvia optimal role assignments,” in INFOCOM 2002. Twenty-First Annual JointConference of the IEEE Computer and Communications Societies. Proceedings.IEEE, vol. 3, 2002, pp. 1587–1596 vol.3.[110] M. A. Razzaque and S. Dobson, “Energy-efficient sensing in wireless sensornetworks using compressed sensing,” Sensors, vol. 14, no. 2, pp. 2822–2859,2014.[111] O. Arnold, F. Richter, G. Fettweis, and O. Blume, “Power consumption modelingof different base station types in heterogeneous cellular networks,” in FutureNetwork and Mobile Summit, 2010, June 2010, pp. 1–8.[112] E. Olivetti, J. Gregory, and R. Kirchain, “Life cycle impacts of alkaline batter-ies with a focus on end-of-life,” Study Conducted for the National ElectricalManufacturers Association, 2011.167[113] D. Kumar, “Performance analysis of energy efficient clustering protocols formaximising lifetime of wireless sensor networks,” Wireless Sensor Systems, IET,vol. 4, no. 1, pp. 9–16, March 2014.[114] J. Yu, Y. Qi, G. Wang, and X. Gu, “A cluster-based routing protocol for wire-less sensor networks with nonuniform node distribution,” {AEU} - InternationalJournal of Electronics and Communications, vol. 66, no. 1, pp. 54 – 61, 2012.[115] M. Hossain and V. Prybutok, “Consumer acceptance of rfid technology: Anexploratory study,” Engineering Management, IEEE Transactions on, vol. 55,no. 2, pp. 316–328, May 2008.[116] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, “Wireless networks with rfenergy harvesting: A contemporary survey,” Communications Surveys Tutorials,IEEE, vol. 17, no. 2, pp. 757–789, 2015.[117] H. Ostaffe, “Rf-based wireless charging and energy harvesting.” [Online].Available: energy harvesting/[118] S. S. Kumar and K. Kashwan, “Research study of energy harvesting in wirelesssensor networks,” International Journal of Renewable Energy Research (IJRER),vol. 3, no. 3, pp. 745–753, 2013.[119] Y. Hori, T. Yoshida, T. Katashita, and A. Satoh, “Quantitative and statisticalperformance evaluation of arbiter physical unclonable functions on fpgas,” inReconfigurable Computing and FPGAs (ReConFig), 2010 International Confer-ence on. IEEE, 2010, pp. 298–303.[120] A. Maiti, V. Gunreddy, and P. Schaumont, “A systematic method to evaluateand compare the performance of physical unclonable functions,” in Embeddedsystems design with FPGAs. Springer, 2013, pp. 245–267.[121] A. Alvarez, W. Zhao, and M. Alioto, “14.3 15fJ/b static physically unclonablefunctions for secure chip identification with <2% native bit instability and 140xInter/Intra PUF hamming distance separation in 65nm,” in IEEE InternationalSolid- State Circuits Conference (ISSCC), Feb. 2015, pp. 256–257.[122] P. Pongle and G. Chavan, “A survey: Attacks on rpl and 6lowpan in iot,” in 2015International Conference on Pervasive Computing (ICPC), Jan 2015, pp. 1–6.[123] C. Mahapatra, P. Kamalinejad, T. Stouraitis, S. Mirabbasi, and V. C. M. Le-ung, “Low-complexity energy-efficient security approach for e-health applica-tions based on physically unclonable functions of sensors,” in 2015 IEEE Inter-national Conference on Electronics, Circuits, and Systems (ICECS), Dec 2015,pp. 531–534.168[124] “AFE4400 Integrated Analog Front-End for Heart Rate Monitors and Low-Cost Pulse Oximeters,”, Ac-cessed: 2015/10/20.[125] C. Mathas, “Infrared sensors evolve from military to iot applications,” ElectronicProducts, NewYork, , Sept 2014.[126] V. Paliouras and T. Stouraitis, “Novel high-radix residue number system archi-tectures,” IEEE Transactions on Circuits and Systems II: Analog and DigitalSignal Processing, vol. 47, no. 10, pp. 1059–1073, Oct 2000.[127] D. Schinianakis and T. Stouraitis, “Multifunction residue architectures for cryp-tography,” IEEE Transactions on Circuits and Systems I: Regular Papers,vol. 61, no. 4, pp. 1156–1169, April 2014.[128] T. Stouraitis, “Efficient convertors for residue and quadratic-residue number sys-tems,” IEE Proceedings-G: Circuits, Devices and Systems, vol. 139, no. 6, pp.626–634, 1992.[129] Y. Chen, J.-S. Chou, and H.-M. Sun, “A novel mutual authentication schemebased on quadratic residues for RFID systems,” Computer Networks, vol. 52,no. 12, pp. 2373–2380, 2008.[130] D. J. Soudris, V. Paliouras, T. Stouraitis, and C. E. Goutis, “A vlsi design method-ology for rns full adder-based inner product architectures,” IEEE Transactionson Circuits and Systems II: Analog and Digital Signal Processing, vol. 44, no. 4,pp. 315–318, Apr 1997.[131] T. ElGamal, “A public key cryptosystem and a signature scheme based on dis-crete logarithms,” IEEE transactions on information theory, vol. 31, no. 4, pp.469–472, 1985.[132] “LMV793/LMV794 88 MHz, Low Noise, 1.8V CMOS Input, DecompensatedOperational Amplifiers,”, Ac-cessed: 2017/01/03.[133] “Noise Analysis of FET Transimpedance Amplifiers,”, Accessed: 2017/01/03.[134] “LMP2231 Single Micropower, 1.6V, Precision Operational Amplifierwith CMOS Inputs,”, Accessed:2017/01/03.169[135] R. Maes and I. Verbauwhede, “Physically unclonable functions: A study on thestate of the art and future research directions,” in Towards Hardware-IntrinsicSecurity. Springer, 2010, pp. 3–37.[136] J. Deepakumara, H. M. Heys, and R. Venkatesan, “Fpga implementation of md5hash algorithm,” in Canadian Conference on Electrical and Computer Engi-neering 2001. Conference Proceedings (Cat. No.01TH8555), vol. 2, 2001, pp.919–924 vol.2.[137] D. Eastlake 3rd and P. Jones, “Us secure hash algorithm 1 (sha1),” Tech. Rep.,2001.[138] H. Krawczyk, R. Canetti, and M. Bellare, “Hmac: Keyed-hashing for messageauthentication,” 1997.[139] G. de Meulenaer, F. Gosset, O.-X. Standaert, and O. Pereira, “On the EnergyCost of Communication and Cryptography in Wireless Sensor Networks,” inIEEE International Conference on Wireless and Mobile Computing, Oct. 2008,pp. 580–585.[140] A. Kaur, “Energy analysis of wireless sensor networks using rsa and ecc encryp-tion method,” International Journal of Scientific & Engineering Research, vol. 4,no. 5, p. 2212, 2013.[141] A. S. Wander, N. Gura, H. Eberle, V. Gupta, and S. C. Shantz, “Energy analysisof public-key cryptography for wireless sensor networks,” in Pervasive Comput-ing and Communications, 2005. PerCom 2005. Third IEEE International Con-ference on. IEEE, 2005, pp. 324–328.[142] A. G. Reddy, A. K. Das, E. J. Yoon, and K. Y. Yoo, “A secure anonymous au-thentication protocol for mobile services on elliptic curve cryptography,” IEEEAccess, vol. 4, pp. 4394–4407, 2016.[143] C. Klein and G. Kaefer, “From smart homes to smart cities: Opportunities andchallenges from an industrial perspective,” in International Conference on NextGeneration Wired/Wireless Networking. Springer, 2008, pp. 260–260.[144] C.-F. Liao and P.-Y. Chen, “Rosa: Resource-oriented service managementschemes for web of things in a smart home,” Sensors, vol. 17, no. 10, p. 2159,2017.[145] T. D. Mendes, R. Godina, E. M. Rodrigues, J. C. Matias, and J. P. Catala˜o,“Smart home communication technologies and applications: Wireless protocolassessment for home area network resources,” Energies, vol. 8, no. 7, pp. 7279–7311, 2015.170[146] W. Ejaz, M. Naeem, A. Shahid, A. Anpalagan, and M. Jo, “Efficient energymanagement for the internet of things in smart cities,” IEEE CommunicationsMagazine, vol. 55, no. 1, pp. 84–91, January 2017.[147] Y.-L. Hsu, P.-H. Chou, H.-C. Chang, S.-L. Lin, S.-C. Yang, H.-Y. Su, C.-C.Chang, Y.-S. Cheng, and Y.-C. Kuo, “Design and implementation of a smarthome system using multisensor data fusion technology,” Sensors, vol. 17, no. 7,p. 1631, 2017.[148] T. Daily, survey of commercial and institutional energy use, 2014. StatisticsCanada, 2016. [Online]. Available:[149] S. DMO, Smart Home Report. Statista Digital Market Out-look, 2016. [Online]. Available:[150] S. A. P. Kani and M. H. Nehrir, “Real-time central demand response for pri-mary frequency regulation in microgrids,” 2013 IEEE PES Innovative SmartGrid Technologies Conference (ISGT), pp. 1–1, 2013.[151] S. Borenstein, M. Jaske, and A. Rosenfeld, “Dynamic pricing, advanced meter-ing, and demand response in electricity markets,” Center for the Study of EnergyMarkets, 2002.[152] P. Palensky and D. Dietrich, “Demand side management: Demand response,intelligent energy systems, and smart loads,” IEEE transactions on industrialinformatics, vol. 7, no. 3, pp. 381–388, 2011.[153] E. M. Rodrigues, R. Godina, M. Shafie-khah, and J. P. Catala˜o, “Experimentalresults on a wireless wattmeter device for the integration in home energy man-agement systems,” Energies, vol. 10, no. 3, p. 398, 2017.[154] U. Hunkeler, H. L. Truong, and A. Stanford-Clark, “MQTT-s : A publish/sub-scribe protocol for wireless sensor networks,” Communication Systems Softwareand Middleware and Workshops, 2008. COMSWARE 2008. 3rd InternationalConference on, pp. 791–798, 2008.[155] C. M. Bishop, Pattern recognition and machine learning. springer, 2006.[156] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel,M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg et al., “Scikit-learn: Ma-chine learning in python,” Journal of Machine Learning Research, vol. 12, no.Oct, pp. 2825–2830, 2011.171[157] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction. MITpress Cambridge, 1998, vol. 1, no. 1.[158] A. M. Andrew, A. Zakaria, S. Mad Saad, and A. Y. Md Shakaff, “Multi-stagefeature selection based intelligent classifier for classification of incipientstage fire in building,” Sensors, vol. 16, no. 1, 2016. [Online]. Available:[159] M. L. Puterman, Markov Decision Processes: Discrete Stochastic Dynamic Pro-gramming, 1st ed. John Wiley & Sons, Inc., 2014.[160] C. J. Watkins and P. Dayan, “Q-learning,” Machine learning, vol. 8, no. 3, pp.279–292, 1992.[161] D. Ernst, P. Geurts, and L. Wehenkel, “Tree-based batch mode reinforcementlearning,” Journal of Machine Learning Research, vol. 6, pp. 503–556, 2005.[162] M. Riedmiller, “Neural fitted q iteration–first experiences with a data effi-cient neural reinforcement learning method,” European Conference on MachineLearning, pp. 317–328, 2005.[163] O. o. E. E. Natural Resources Canada, Energy consumption of majorhousehold appliances shipped in Canada, summary report. Energy Publica-tions, 2012. [Online]. Available:[164] J. Park and I. W. Sandberg, “Universal approximation using radial-basis-functionnetworks,” Neural computation, vol. 3, no. 2, pp. 246–257, 1991.[165] I. E. S. O. (IESO), ontario power stats. canadian energy issues, 2017. [Online].Available:[166] G. Yang, C. K. Ho, and Y. L. Guan, “Dynamic resource allocation for multiple-antenna wireless power transfer,” IEEE Transactions on Signal Processing,vol. 62, no. 14, pp. 3565–3577, July 2014.[167] G. Yang, V. Y. F. Tan, C. K. Ho, S. H. Ting, and Y. L. Guan, “Wireless com-pressive sensing for energy harvesting sensor nodes over fading channels,” in2013 IEEE International Conference on Communications (ICC), June 2013, pp.4962–4967.[168] H. Sharma and P. Balamuralidhar, “A transmission scheme for robust delivery ofurgent/critical data in internet of things,” in 2013 Fifth International Conferenceon Communication Systems and Networks (COMSNETS), Jan 2013, pp. 1–7.172[169] L. Shu, Z. Yuan, T. Hara, L. Wang, and Y. Zhang, “Impacts of duty-cycle on tpgfgeographical multipath routing in wireless sensor networks,” in 2010 IEEE 18thInternational Workshop on Quality of Service (IWQoS), June 2010, pp. 1–2.[170] D. Bertsekas, Nonlinear Programming. Athena Scientific, 1995. [Online].Available: AProof of the Lemma 2We define E¯1 and E¯2 inℜ|N|+|L(i)| as ET X(1+E(T,hi))+ERX(1+E(T,hi)) and EPR(1+E(T,hi)P′)+Ri jESN respectively. If we denote ∞-norm as ‖.‖∞ and q-norm as ‖.‖q,the lifetime objective functions of (2.17) are represented by -‖E¯1r + E¯2R‖∞ and -(1/(ε+1))‖E¯1r+ E¯2R‖ε+1 respectively. Suppose {r∗,R∗} and {r∗ε ,R∗ε} be the optimalsolutions for the two objective functions. Then we have the following inequalities usingapproximation of ‖.‖∞ from [170]‖E¯1r∗ε + E¯2R∗ε‖∞≤ ‖E¯1r∗ε + E¯2R∗ε‖ε+1≤ ‖E¯1r∗+ E¯2R∗‖ε+1≤ |N|1/(ε+1)‖E¯1r∗+ E¯2R∗‖∞(A.1)174The corresponding network lifetimes become Ti=1/‖E¯1r∗+E¯2R∗‖∞ and T εi =1/‖E¯1r∗ε+E¯2R∗ε‖∞. From (39) we have,1|N|1/(ε+1)Ti ≤ Tεi ≤ Ti (A.2)At limε→∞ T εi = Ti, and thus the lemma holds.175Appendix BProof of the Proposition 2From (27), the gradient of the objective function D(λ ,µ) w.r.t λl ,∇λD(λ ,µ) =α∑i∈N∑j∈Ni∇λUi(Ri j,Ps)− (1−α)sεi ·∇λ si≤ α∑i∈N∑j∈Ni∇λUi(Ri j,Ps)≤ αU(B.1)By Definition 1 and Assumption 1 in Chapter 2, we can find the error in the costestimation of the link price λl when iteration c→c+1‖D(λ (c+1))−D(λ (c))‖ ≤ ‖∇λD(λ )T (λ (c+1)−λ (c))‖≤ ‖∇λD(λ )‖ · ‖(λ (c+1)−λ (c))‖≤ L1/2αU‖(λ (c+1)−λ (c))‖(B.2)From the above inequalities, we see that function is Lipschitz. Thus the solution gen-erated with step size ϕc is optimal [170]. Let the update at each iteration c is given by176∆λ (c). Then,|∆(λ (c))|= | ri j(c)α∇λUi(Ri j,Ps)∇λD(λ )| ≤Rα|∇λD(λ )| (B.3)|∇λD(λ )T∆(λ (c))|‖∆(λ (c))‖2 ≤Rα ‖∇λD(λ )‖2( Rα )2‖∇λD(λ )‖2=αR(B.4)According to [170], the step size satisfies 0<ϕc<2L1/2U R. Similarly, the step sizebound can be proven for ψc.177


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items