Privacy-Preserving Publishing of Moving Objects Databases by Roman Yarovoy B.Sc., The University of British Columbia, 2007 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE in The Faculty of Graduate Studies (Computer Science) THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver) April 2009 c© Roman Yarovoy 2009 Abstract Moving Objects Databases (MOD) have gained popularity as a subject for research due to the latest developments in the positioning technologies and mobile networking. Analysis of mobility data can be used to discover and deliver knowledge that can enhance public welfare. For instance, a study of traffic patterns and congestion trends can reveal some information that can be used to improve routing and scheduling of public transit vehicles. To enable analysis of mobility data, a MOD must be published. However, publication of MOD can pose a threat to location privacy of users, whose movement is recorded in the database. A user’s location at one or more time points can be publicly available prior to the publication of MOD. Based on this public knowledge, an attacker can potentially find the user’s entire trajectory and learn his/her positions at other time points, which constitutes privacy breach. This public knowledge is a user’s quasi-identifier (QID), i.e. a set of attributes that can uniquely identify the user’s trajectory in the published database. We argue that unlike in relational microdata, where all tuples have the same set of quasi-identifiers, in mobility data, the concept of quasi-identifier must be modeled subjectively on an individual basis. In this work, we study the problem of privacy preserving publication of MOD. We conjecture that each Moving Object (MOB) may have a distinct QID. We develop a possible attack model on the published MOD given public knowledge of some or all MOBs. We develop k-anonymity model (based on classical k-anonymity), which ensures that every object is indistinguishable (with respect to its QID) from at least k − 1 other objects, and show that this model is impervious to the proposed attack model. We employ space generalization to achieve MOB anonymity. We propose three anonymization algorithms that generate a MOD that satisfies the k-anonymity model, while minimizing the information loss. We conduct several sets of experiments on synthetic and real-world data sets of vehicular traffic to analyze and evaluate our proposed algorithms. ii Table of Contents Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi List of Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . vii Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . viii Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 Location Tracking and Security Risks . . . . . . . . . . . . . 1 1.2 Protecting Location Privacy . . . . . . . . . . . . . . . . . . 7 2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.1 Moving Objects Database: Terminology and Settings . . . . 10 2.1.1 Eliminating Null Entries . . . . . . . . . . . . . . . . 13 2.1.2 Generalized Moving Objects Database . . . . . . . . 15 2.1.3 Information Loss . . . . . . . . . . . . . . . . . . . . . 17 2.1.4 Range Query Distortion . . . . . . . . . . . . . . . . . 19 2.2 Quasi-Identifiers . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.2.1 Possible Ways to Obtain Quasi-Identifiers . . . . . . . 23 2.2.2 Generating Quasi-Identifiers . . . . . . . . . . . . . . 24 2.3 k-Anonymity . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.4 Hilbert Space-Filling Curve . . . . . . . . . . . . . . . . . . . 29 2.5 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . 31 iii Table of Contents 3 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 3.1 Anonymity Groups . . . . . . . . . . . . . . . . . . . . . . . 36 3.1.1 Overview of Top-K Algorithms . . . . . . . . . . . . 38 3.1.2 Adapting Best Position Algorithm . . . . . . . . . . . 40 3.2 Equivalence Classes . . . . . . . . . . . . . . . . . . . . . . . 43 4 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 4.1 Attack Model . . . . . . . . . . . . . . . . . . . . . . . . . . 48 4.2 k-Anonymity Model for Moving Objects . . . . . . . . . . . . 52 4.2.1 Problem Statement . . . . . . . . . . . . . . . . . . . 54 4.3 Extreme Union . . . . . . . . . . . . . . . . . . . . . . . . . . 55 4.4 Symmetric Anonymization . . . . . . . . . . . . . . . . . . . 58 4.5 Restricted Symmetric Anonymization . . . . . . . . . . . . . 61 4.6 Algorithm Complexity Analysis . . . . . . . . . . . . . . . . 64 4.7 Anonymization Algorithms: Summary . . . . . . . . . . . . . 66 5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 5.1 Experimental Settings . . . . . . . . . . . . . . . . . . . . . . 67 5.1.1 Oldenburg Database . . . . . . . . . . . . . . . . . . . 68 5.1.2 Milan Database . . . . . . . . . . . . . . . . . . . . . 69 5.1.3 Database Characteristics . . . . . . . . . . . . . . . . 69 5.2 Description of Experiments . . . . . . . . . . . . . . . . . . . 70 5.3 Results of Experiments . . . . . . . . . . . . . . . . . . . . . 73 5.4 Discussion of Algorithms . . . . . . . . . . . . . . . . . . . . 79 6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 iv List of Tables 1.1 Location Information of 6 Users at 4 Distinct Time Stamps . 4 1.2 Instance of Moving Objects Database (n=6, m=4) . . . . . . 4 1.3 Instance of Input Moving Objects Database (Running Example) 4 2.1 Generalized 2-Anonymous Moving Objects Database . . . . . 17 2.2 Quasi-identifiers of Moving Objects . . . . . . . . . . . . . . . 23 2.3 Instance of Hospital Microdata . . . . . . . . . . . . . . . . . 27 2.4 2-anonymous Hospital Microdata (Generalized by Age) . . . . 28 2.5 3-anonymous Hospital Microdata (Generalized by Age and City) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 3.1 Hilbert Indexes for Running Example . . . . . . . . . . . . . 38 3.2 Example of Overlapping Anonymity Groups . . . . . . . . . . 43 4.1 Application of Extreme Union or Symmetric Anonymization or Restricted Symmetric Anonymization on Running Exam- ple (k = 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 4.2 Application of Extreme Union on Running Example (k = 3) . 57 4.3 Application of Symmetric or Restricted Symmetric Anonymi- zation on Running Example (k = 3) . . . . . . . . . . . . . . 60 4.4 Time Complexity of Each Stage of Anonymization Algorithm 65 5.1 Databases Used in the Experiments . . . . . . . . . . . . . . . 69 5.2 Characteristics of Oldenburg and Milan Databases . . . . . . 69 5.3 Metrics for Algorithm Evaluation . . . . . . . . . . . . . . . . 71 5.4 Settings of Experiments . . . . . . . . . . . . . . . . . . . . . 72 5.5 Experiment 1: Main Metrics . . . . . . . . . . . . . . . . . . . 74 5.6 Experiment 1: Distribution of Anonymity Group Size . . . . 75 5.7 Experiment 1: Distribution of Equivalence Class Size . . . . . 75 5.8 Experiment 2: Main Metrics . . . . . . . . . . . . . . . . . . . 76 5.9 Experiment 2: Distribution of Equivalence Class Size . . . . . 76 5.10 Experiment 3: Main Metrics . . . . . . . . . . . . . . . . . . . 79 v List of Figures 1.1 Trajectories of Moving Objects (Running Example) . . . . . . 6 1.2 2D View of Trajectories from Running Example . . . . . . . . 7 2.1 Example of Range Query . . . . . . . . . . . . . . . . . . . . 21 2.2 Example of QID Graph . . . . . . . . . . . . . . . . . . . . . 25 2.3 Examples of Hilbert Curve . . . . . . . . . . . . . . . . . . . . 30 3.1 Example of Computing Anonymity Group . . . . . . . . . . . 42 4.1 Example of Attack Graphs . . . . . . . . . . . . . . . . . . . . 50 4.2 Example of Symmetric Attack Graphs . . . . . . . . . . . . . 51 5.1 Results of Experiment 1 . . . . . . . . . . . . . . . . . . . . . 77 5.2 Results of Experiment 2 . . . . . . . . . . . . . . . . . . . . . 78 vi List of Algorithms 1 QID Generator: Random Graph . . . . . . . . . . . . . . . . 26 2 GenerAG : Anonymity Group Generating Top-K Algorithm . 41 3 AGtoEC : Adding AG(O) to EC(t) . . . . . . . . . . . . . . . 46 4 lubSpace: Space Generalization . . . . . . . . . . . . . . . . . 47 5 Extreme Union . . . . . . . . . . . . . . . . . . . . . . . . . . 56 6 Symmetric Anonymization . . . . . . . . . . . . . . . . . . . . 59 7 Restricted Symmetric Anonymization . . . . . . . . . . . . . 62 vii Acknowledgments I want to thank my graduate supervisor Professor Laks V.S. Lakshmanan for his support and contributions to my graduate studies. It was a truly great experience working under his supervision. I am forever grateful for this amazing opportunity. I have certainly learned a lot from him. I would like to acknowledge professor Raymond Ng, who has kindly agreed to proofread this thesis. I sincerely thank professor Ng for his help and guidance. In addition, I would like to acknowledge Dr. Wendy Hui Wang and Dr. Francesco Bonchi for their contributions to the MOD anonymization project, which led to my graduate research. I thank them for constant support and constructive criticism of my ideas. viii Dedication I want to dedicate this work to all scientists, who strive so passionately to improve and expand the field of Computer Science. ix Chapter 1 Introduction Moving Objects Databases have become a focus of scientific research in the field of Computer Science among other fields. Moving Objects Database (MOD) is a collection of spatio-temporal data obtained from one or more Moving Objects (MOBs) such as vehicles. The popularity of MOD came with modern advances in wireless communication and mobile devices. Cel- lular phones, personal digital assistants, and portable vehicle navigators are now equipped with Global Positioning System (GPS) receivers that pro- vide PNT (positioning, navigating, and timing) services. While GPS users benefit from variety of location-based services (LBSs), they also help other users by sharing their location information. The collected spatio-temporal information can be studied and analyzed in order to enhance public welfare (e.g. adjusting traffic light scheduling to disperse traffic jam). But while the navigation systems become more widely used in everyday life, the location privacy of the system users can be at risk of being breached. In fact, there have been cases of stalking involving the use of GPS-enabled devices [31, 40]. Hence, sharing of location information raises some privacy concerns and it is crucial to enable security of this information before it can be published. In this thesis, we address the problem of privacy-preserving publishing of a Moving Objects Database. We present techniques for anonymizing a given MOD by means of space generalization. As a result of generalization pro- cess, some location information will be replaced by more general information and hence, query results over a generalized database will be distorted. So in our approach, we strive to protect location privacy of mobile users, while minimizing the information loss and query distortion. 1.1 Location Tracking and Security Risks Global Navigation Satellite System (GNSS) is a satellite navigation system that provides autonomous geo-spatial positioning with global coverage. The Global Positioning System (GPS) is presently the only fully functional GNSS [17]. The system employs several satellites, which work in constellation, to transmit microwave signals that can be acquired by GPS-enabled receivers. 1 1.1. Location Tracking and Security Risks The main and widely spread application of GPS is satellite navigation sys- tem. Satellite navigation systems embedded into electronic devices provide users with PNT services. That is, these systems allow users to determine their geographical location (longitude, latitude, and altitude) within sev- eral meters, their velocity, and the current time. The ability to determine accurately a location can be beneficial in many situations from individual recreation activities (e.g. hiking, golf) to disaster relief (e.g. vehicular evac- uation plan). In addition, positioning technology is used by law enforcement specialists in crime analysis [42]. By 2008, over 150 million GPS-enabled de- vices were sold and the sales are expected to reach 560 million by the year 2012 [26]. Another mobile technology that has evolved in recent years in the field of transportation systems is the vehicular ad-hoc network (VANET). VANET is a self-organizing and decentralized system consisting of mobile nodes that communicate with each other by means of wireless technologies. The nodes in the network are vehicles and stationary roadside equipment. When vehi- cles move on roads within a network, they share environmental information (e.g. location, vehicle speed, wiper blade motion frequency, tire pressure) with other vehicles and location-based servers. Knowledge derived from this information can be used to promote transportation safety and comfortabil- ity, in particular, to avoid collisions, deliver emergency warning messages, and inform about traffic and road conditions. Several security vulnerabil- ities of VANET have been identified and addressed in academic research (e.g. [24, 35]). Despite the privacy concerns associated with VANET, the network provides various useful functionalities and still remains a valuable source of location information. An example of recent development of technology revolving around lo- cation data is Google Latitude, a geographic location sharing and tracking service launched by Google corporation on February 4, 2009 [16]. Latitude allows its users to track locations of their friends and family who agreed to transmit (i.e. share) their positions to Google using a GPS-equipped devices. Then a Latitude user can view the transmitted positions in the Google’s online feature Google Maps. So if two Latitude users have a com- mon agreement to share their locations with each other, then each one of these users can see the last known position of the other user while they are both signed in to Google Latitude. To protect location privacy, a user can hide his/her location from any individual or from all friends. In general, users may wish to contribute to scientific research by sharing their move- ment history. In such case, more sophisticated methods might be required to protect users’ information. 2 1.1. Location Tracking and Security Risks GPS technology gave rise to location-based services (LBSs), which are personalized services delivered to mobile users based on their current lo- cation. LBSs allow users to utilize their geographical positioning in order to determine the nearest business location (e.g. restaurant, hotel), to re- ceive traffic report, and/or to obtain directions to a user-specified address. The submitted users’ locations are stored at location-based database servers. The database servers accumulate a large number of Moving Object trajec- tories, where each trajectory represents a movement pattern of a mobile user. A collection of such trajectories is referred to as Moving Objects Database (MOD). Table 1.2 shows an example of MOD with 6 trajectories (corresponding to 6 mobile users) spanning over 4 time stamps. The information stored in a given Moving Objects Database can be made publicly accessible by publishing this database. Using a static instance of a Moving Objects Database to analyze movement and behavior of objects has certain beneficial applications such as: • Monitoring traffic patterns and congestion trends: To deter- mine what streets/regions are the busiest during rush hour – Air traffic control: To promote safety in aviation • Resource management: To handle logistics and to coordinate cargo delivery • Sustainable transport (also commonly referred to as sustainable mobility): To improve public transportation, to limit the emissions, and to minimize consumption of natural resources Studying Moving Objects Databases and mining spatio-temporal data can reveal useful information about the movements of objects. For instance, given information about the movement of cars in a city center during rush hours, traffic lights can be adjusted accordingly to minimize the possibility of traffic jam. However, publication of a MOD can pose a threat to the privacy of the users, whose movement is recorded in the database. An adversary can find out a path traveled by a user. If a user’s unique identifiers (e.g. user’s name, driver’s license number) and/or a unique combination of identifiers (e.g. date of birth and postal code) are published, the user’s trajectory can be obtained from the database (relatively) easily. Such identifiers should be removed from or replaced in the published MOD. Pseudo-anonymity refers to a process of substituting real identifiers by fake ids. The following example 3 1.1. Location Tracking and Security Risks User name 12:00 PM 01:00 PM 01:45 PM 04:00 PM Alice (0, 0) (1, 4) (2, 7) null Bob null (5, 7) (7, 7) (7, 4) Cate (0, 1) (0, 2) (2, 4) (3, 7) David (4, 4) (3, 2) (3, 1) (5, 0) Erica (6, 3) (4, 6) (7, 7) (6, 3) Frank null null (0, 6) (7, 1) Table 1.1: Location Information of 6 Users at 4 Distinct Time Stamps User name t1 t2 t3 t4 Alice (0, 0) (1, 4) (2, 7) (2, 7) Bob (5, 7) (5, 7) (7, 7) (7, 4) Cate (0, 1) (0, 2) (2, 4) (3, 7) David (4, 4) (3, 2) (3, 1) (5, 0) Erica (6, 3) (4, 6) (7, 7) (6, 3) Frank (0, 6) (0, 6) (0, 6) (7, 1) Table 1.2: Instance of Moving Objects Database (n=6, m=4) User ID t1 t2 t3 t4 O1 (0, 0) (1, 4) (2, 7) (2, 7) O2 (5, 7) (5, 7) (7, 7) (7, 4) O3 (0, 1) (0, 2) (2, 4) (3, 7) O4 (4, 4) (3, 2) (3, 1) (5, 0) O5 (6, 3) (4, 6) (7, 7) (6, 3) O6 (0, 6) (0, 6) (0, 6) (7, 1) Table 1.3: Instance of Input Moving Objects Database (Running Example) 4 1.1. Location Tracking and Security Risks introduces the running example that we will use repeatedly throughout this thesis. Example 1 (Running Example) Consider a Moving Objects Database in Table 1.1, which stores location information of 6 users recorded at 4 different time stamps. The database can contain null values whenever a position of a mobile user was not recorded at a particular time stamp. We can eliminate all null entries using some assumptions about user’s position given his/her known positions. Section 2.1.1 describes the null entry elimi- nation procedure in details. Table 1.2 is obtained by eliminating null entries from Table 1.1. Figure 1.1 illustrates the trajectories of the 6 users from Table 1.2 (0 hours on Time-axis corresponds to 12:00PM time stamp). Fig- ure 1.2 shows a projection of Figure 1.1 on xy-axes. We can drop the Name attribute that contains explicit personal identifiers or replace all the values of the attribute by fake identifiers. Table 1.3 shows a pseudo-anonymized version of a MOD from Table 1.2, where user names were replaced by fake identifiers. For example, the name Alice was replaced by identifier O1. Table 1.3 exemplifies the format of the data set accepted as input in our anonymization algorithms presented in Chapter 4. Pseudo-anonymity does not guarantee that users’ privacy will be pre- served. Privacy breach can occur in the following cases: 1. Prior Knowledge Attack: An adversary may possess a knowledge of partial user movement. For example, an adversary may know that at 1 o’clock in the afternoon Alice arrives at her workplace. That is, the position of user Alice at time stamp t2 (corresponding to 13:00 hours or 01:00 PM) is known to an adversary. If Alice is the only user at this position at time t2 (i.e. the position of Alice at t2 is unique), then the adversary will find out Alice’s entire trajectory published in the MOD 2. Exclusive Value Attack: An attacker may know that a particular value can occur with probability 0 and thus, he/she can exclude this value from consideration. For example, if an attacker know that Bob cannot be located at position (5, 0) at time stamp t4, then the attacker can conclude that any trajectory that passes through (5, 0) at time stamp t4 does not correspond to Bob’s trajectory. Hence in a rare case such that there is only one trajectory that does not pass through position (5, 0) at time stamp t4, the attacker can conclude that this trajectory corresponds to Bob’s movement 5 1.1. Location Tracking and Security Risks 3. Trajectory Tracking Algorithms: It is possible to publish MOD without any identifiers and to mix location samples in each column at random. Then each row of the altered database may or may not corre- spond to a particular mobile user. Nevertheless, such naive anonymiza- tion will not be sufficient to protect users’ location traces. There exist techniques such as multiple target tracking, which can reconstruct such permuted trajectories with high success rate [41]. An adversary can implement and use such techniques in combination with prior knowl- edge to identify trajectories of interest 0 2 4 6 8 0 2 4 6 8 0 0.5 1 1.5 2 2.5 3 3.5 4 X−coordinateY−coordinate T im e (h ou rs ) Alice Bob Cate David Erica Frank Figure 1.1: Trajectories of Moving Objects (Running Example) Indeed, using, storing, and disclosing location information must be done in privacy-preserving manner. location-based service providers must strive to enhance privacy protection of such information. ”Location Privacy Pro- tection Act of 2001” is a bill introduced to the Committee on Commerce, Science, and Transportation of the Senate of the United States of America [33]. The bill proposes to ensure security and integrity of location informa- tion and to prescribe rules under which the information can be collected, retained, and used by LBS providers. Even though the bill has never become 6 1.2. Protecting Location Privacy 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 X−coordinate Y − c o o r di na te Alice Bob Cate David Erica Frank Figure 1.2: 2D View of Trajectories from Running Example a law, it exemplifies an effort to respect privacy of location information and to protect it from misuse. 1.2 Protecting Location Privacy In order to be trustworthy, a location tracking application must enable preservation of location privacy of its users. Over the years, academic re- searchers have addressed the problem of protecting location data. The com- mon goal is to develop methods that will minimize the chances of deriving and abusing sensitive information by malicious users. Section 2.5 presents an overview of several approaches taken to secure the location traces. Moving Objects Database as any other database can be made publicly accessible. The main purpose of publishing some data is to facilitate re- search. Therefore while protecting users’ privacy, it is important to ensure that the published data is not altered to the point where it is no longer useful for research and/or analysis. Location privacy can be preserved by aggregating two or more different positions. As a result of aggregation, the 7 1.2. Protecting Location Privacy quality of data deteriorates as some information might be lost. The goal is to ensure that the quality of a database is kept high for the database to be a valuable asset in research. In our approach, we consider one particular approach for aggregating data, namely, k-anonymity. In this work, we conduct a study of a mechanism that prevents an adver- sary from identifying any individual trajectory with absolute certainty (i.e. with probability 1) in published Moving Objects Databases. We adopt the notion of k-anonymity [37, 38] to ensure that any Moving Object (MOB) is indistinguishable from at least k−1 other MOBs. Instead of using a fixed set of attributes, which can uniquely identify a particular person in a database, we allow each person to have his/her own set of identifiers that comprise publicly available information about this person. That is, we model the public and sensitive information subjectively with respect to each MOB in a given instance of a database. For each MOB, we generate an anonymity group, i.e. we compute a set of k−1 other MOBs such that these MOBs are in close proximity of the subject MOB. Due to peculiar nature of mobility data, it is possible that public information of two MOBs will overlap and their anonymity groups may overlap as well, which in turn can produce un- sound anonymized results. To ensure the soundness of the output database, we define an equivalence relation of two MOBs with respect to a particular time stamp. Using this relation, we divide MOBs into disjoint equivalence classes. We replace positions of all MOBs in an equivalence class by the least bounding rectangle that contains all of these positions. We measure the information that was lost due to this replacement of a precise position by a generalized region. And we aim to minimize this information loss. We approach the design and development of the mechanism for privacy- preserving publishing of spatio-temporal data in five steps. 1. We extend the classical notion of quasi-identifier (QID) used in [37] 2. We model the application of prior knowledge on a published MOD as a bipartite graph and formulate a possible attack strategy based on this graph 3. We define a robust notion of k-anonymity, which ensures that privacy breach based on our attack model cannot occur 4. We derive three algorithms, which anonymize data by means of space generalization while minimizing the information loss introduced as a result of the anonymization process 8 1.2. Protecting Location Privacy 5. We calibrate our algorithms for simulated and real-world Moving Ob- jects Databases to evaluate their performance The remainder of this thesis is organized as follows. In Chapter 2, we discuss the background information and overview work related to protect- ing privacy of MOBs. Preliminary algorithms, which are used as building blocks of our anonymization algorithms are reviewed in Chapter 3. Chap- ter 4 presents the definition of k-anonymity model in MOD settings and in- troduces our anonymization algorithms. In Chapter 5, we review the setup and results of the experiments used compare the performance and to an- alyze the effectiveness of our algorithms with respect to several metrics. Section 6 concludes this thesis and outlines a number of open issues that can be addressed in a future line of research. 9 Chapter 2 Background In this chapter, we review the background notions and discuss some meth- ods and models used in the thesis. We define the terminology associated with our work of anonymizing Moving Objects by means of spatial gen- eralization. We discuss the classical k-anonymity model in the context of relational databases. We overview an indexing approach based on Hilbert space-filling curve. In addition, this chapter summarizes some existing work on privacy preservation of mobility data. There are several assumptions we make throughout our work: 1. We are given a database D of trajectories of Moving Objects (MOBs). Each trajectory belongs to one MOB and each MOB has no more than one trajectory. 2. We are given a set of quasi-identifiers. Each MOB has its own quasi- identifier (QID), which is a temporal pattern describing the MOB’s usual movements that constitute public knowledge and are known be- fore the database publication. 3. A QID can be used to reidentify a MOB in D. 4. We strive to protect the sensitive information, which is any movements that do not constitute public knowledge. 2.1 Moving Objects Database: Terminology and Settings In the following, we formalize the concepts used in this work on Moving Ob- jects Database (MOD) anonymization. We review the input to our privacy protecting algorithms. We define the constitutes of an input MOD. Then, we discuss the format of the generalized version of the input MOD. Essentially, there are three inputs to our privacy-preserving algorithms: (i) minimal anonymity threshold k, (ii) Moving Objects Database D, and (iii) a list of quasi-identifiers Q. The anonymity threshold k specifies the 10 2.1. Moving Objects Database: Terminology and Settings minimal number of objects in an anonymity group (i.e. a group of tuples identical on the corresponding QID). Theoretically, if MOBO is anonymized with k − 1 other MOBs, then an attacker can determine the private infor- mation of O with probability 1/k. This implies that the larger the value of k is, the smaller are the chances of inferring the correct private information. Thus informally, k specifies the level of security in the anonymized data. A Moving Objects Database D is a collection of n Moving Object iden- tifiers D = {O1, O2, · · · , On}, a set T = {t1, t2, · · · , tn} of m distinct time stamps, and a surjective (i.e. onto) function T : D×T→R2 that maps ob- ject O and time t to the object’s (x, y)-position at time t. For a MOB Oi and all time stamps in T, function T generates the trajectory of MOB Oi. Defi- nition 1 formally defines the notion of trajectory in our settings. Figure 1.1 shows trajectories corresponding to our running example in Table 1.3. Definition 1 (Trajectory of Moving Object) A trajectory Ti of MOB Oi is a polyline in 3-dimensional space that consists of a sequence of spatio- temporal coordinates (x1, y1, t1), (x2, y2, t2), · · · , (xm, ym, tm), where m ≥ 1 and t1 < t2 < · · · < tm. During the time interval [ti, ti+1] (where 1 ≤ i ≤ m − 1), the object is assumed to move along a straight line from (xi, yi) to (xi+1, yi+1) with constant velocity. A MOD can contain null entries for some pairs (O, t). That is for some MOB O, we can have a position record at time tj and a null entry at time tj+1 or vice versa. In some cases, we assume that the object is immobile. In other cases, we assume that the object is positioned somewhere within a rectangle, which incapsulates two consecutive non-null positions. We need to eliminate all null entries in order to efficiently anonymize the objects. Section 2.1.1 describes a systematic approach for eliminating null values from all tuples in an input MOD. In our work, we deal with 2-dimensional location samples. In other words, position of a MOB at some time stamp t is given by a point (x, y). It is also possible to record location samples in terms of 2-dimensional regions in order to incorporate the uncertainty introduced by satellite navigation systems such as GPS. Our privacy-preserving approach can be easily ex- tended to work with regions. Every region can be bounded by a circle or a rectangle. When computing distance between MOB O and some other MOB, we can use circle’s center or the intersection of rectangle’s diagonals as the approximated position of MOB O. When anonymizing the input database, we replace some location samples by a general region. Then in a published database, a generalized region will contain all points or regions of 11 2.1. Moving Objects Database: Terminology and Settings a particular anonymity group. Hence, since our algorithms are not heavily dependent on the the input format (i.e. points or regions), our approach can work with uncertainty-aware Moving Objects Databases. The third input to our algorithms is the list of quasi-identifiers Q = {Q1, Q2, · · · , Qn}. A quasi-identifier Qj is a set of time stamps, i.e. Qj = {t | t ∈ T}. We denote a quasi-identifier (QID) of a MOB O by QID(O). Section 2.2 presents a more elaborated explanation of the concept of QID. The concept of quasi-identifier is modeled on an individual basis; different objects are likely to have different QIDs. Each quasi-identifier corresponds to public information of a particular mobile user. Definition 2 (Public Information of Moving Object) Given a quasi- identifier Qj = QID(Oj), the public information of Oj is the set of (x, y)- positions PI(Oj) = {T (Oj , t) | t ∈ Qj}. Let PI = {PI(O) | O ∈ {O1, O2, · · · , On}} denote the public information of all MOBs in a given MOD. In other words, the public information of a particular user corresponds to user’s positions at all time stamps specified by the user’s quasi-identifier. The set PI is considered to be public knowledge, which is known prior to database publication. We assume that set PI is exactly the background knowledge possessed by one or more malicious users. So our goal is to generalize all the positions in set PI such that a user is indistinguishable from at least k − 1 other users on all the positions corresponding to the user’s QID. This will ensure protection of the sensitive information, which is defined below. Definition 3 (Sensitive Information of Moving Object) Given a quasi-identifier Qj = QID(Oj), the sensitive information of Oj is the set of (x, y)-positions SI(Oj) = {T (Oj , t) | t ∈ T−Qj}. Let SI = {SI(O) | O ∈ {O1, O2, · · · , On}} denote the public information of all MOBs in a given MOD. That is, given trajectory Tj of MOB Oj , the sensitive information of Oj is the set complement of the public information of Oj , i.e. SI(Oj) = Tj − PI(Oj). In traditional k-anonymity model, the same set of attributes comprises public information of all objects (see Section 2.3). However, in our settings the public information is modeled subjectively based on indi- vidual quasi-identifier. In addition, in our settings, not only different QIDs, but also anonymity groups can overlap, which presents some challenges in 12 2.1. Moving Objects Database: Terminology and Settings adopting k-anonymity model. As a consequence of overlapping anonymity groups, not only public information, but some sensitive information will be generalized as well. This will further deteriorate the quality of the data as both public and sensitive information will be altered in the generaliza- tion process. Information loss associated with space generalization in our algorithms is described in Section 2.1.3. 2.1.1 Eliminating Null Entries In this section we discuss the procedure taken to eliminate null entries from a given instance of a MOD. At some time stamp t, a null value can appear in a trajectory of a mobile user, if the user did not transmit his/her location information at this time, or the transmission failed because of device mal- function. But in the synthetic data produced by Brinkhoff’s framework [7] and used in our experiments, a MOB appears in the monitoring network in its start position and then disappears from the monitoring network once the MOB reaches its final destination. So in these synthetic data, if MOB Oi starts moving at time ti and reaches the final destination at some later time tj (for i < j < m), then there will be null entries for positions of MOB Oi at times {t1, · · · , ti−1} and {tj+1, · · · , tm}. We classify all null entries into 3 categories and eliminate null values in each category separately: • Leading null entries: There exists time stamp ta such that T (Oi, ta) 6= null and for every t < ta, T (Oi, t) = null. – To eliminate a leading null entry, we set T (Oi, t) ← T (Oi, ta) for all t < ta. • Trailing null entries: There exists time stamp tb such that T (Oi, tb) 6= null and for every t > tb, T (Oi, t) = null. – To eliminate a trailing null entry, we set T (Oi, t) ← T (Oi, tb) for all t > tb. • Gap null entries: There exist time stamps ta and tb (i.e. gap bound- aries) such that T (Oi, ta) 6= null and T (Oi, tb) 6= null, and for every ta < t < tb, T (Oi, t) = null. – To eliminate a gap null entry, first, we find the smallest rectangle Rgap that encapsulates the points T (Oi, ta) and T (Oi, tb). Then, we set the position T (Oi, t) to a point, which is chosen randomly 13 2.1. Moving Objects Database: Terminology and Settings with uniform distribution and is inside or on the boundary of rect- angle Rgap. We repeat these two steps using the same rectangle Rgap for all t such that ta < t < tb. The approach taken to eliminate leading and trailing null values creates an appearance that the object did not move during the time interval [ta, t] (or [t, tb]), where T (Oi, t) is a leading or trailing null entry. In synthetic data, when a MOB disappears after reaching its final destination, we assume that the object stops and stays at the final destination. In real world, this can occur if the object’s vehicle is parked in a parking lot. To eliminate gap null values, we assumed that an attacker may know the gap boundary positions of a MOB, namely T (Oi, ta) and T (Oi, tb). It is very likely that the MOB travels somewhere inside the rectangular region Rgap that encapsulates these gap boundaries. Therefore, we replace a null entry in the gap by some position inside this rectangular region Rgap as we assume that any position in the rectangle is equally likely (i.e the positions are uniformly distributed). Next, we show the application of the null entry eliminating procedure to our running example. Example 2 Consider the MOD shown in Table 1.1. The position of Alice at time t4 is unknown, which is represented by a null value. This is a trailing null entry. To eliminate this null value, we set the position of Alice at t4 to be equal to her position at time t3 (i.e. T (Alice, t4)← T (Alice, t3)). Bob has a null value at time t1 and there does not exist previous known location of Bob. So this is a leading null entry. Therefore, we set the position of Bob at time t1 to his position at time t2 (i.e. T (Bob, t1) ← T (Bob, t2)). There are two null values in the trajectory of Frank and there does not exist any previous known position, which is another example of leading null entries. To eliminate the null entry at t1, we find the first non-null position at a time stamp greater than t1. So we set T (Frank, t1) ← T (Frank, t3). Even though now the position of Frank at t1 is not null, T (Frank, t2) will still be treated as a leading null entry and not a gap null entry. Hence, we set T (Frank, t2) ← T (Frank, t3). Note that the order of processing the time stamps (i.e. t1 and then t2, or vice versa) does not matter as T (Frank, t1) = T (Frank, t2) = T (Frank, t3) at the end of the procedure. After eliminating all the null entries, we obtain the database shown in Table 1.2. Once the null entries have been eliminated, we can index the positions of MOBs in the the input MOD using Hilbert space-filling curve (see Sec- 14 2.1. Moving Objects Database: Terminology and Settings tion 2.4) and then compute anonymity groups corresponding to each MOB’s quasi-identifier as discussed in Section 3.1.2. 2.1.2 Generalized Moving Objects Database Our objective in this thesis is to create a distorted version of input database D, denoted D′, while (i) minimizing the information loss and distortion caused by anonymization process, and (ii) ensuring that a malicious adver- sary cannot infer sensitive information. We anonymize the input database using space generalization. We replace every (x, y)-position in the public information of MOB Oi by a rectangu- lar region such that the region contains at least k − 1 other positions. By containment we mean that a point (x, y) is either strictly inside or on the boundary of the region. The symbol v will denote the containment of a spatio-temporal point in a spatio-temporal region. Since the considered re- gion is a rectangle, it can be unambiguously represented by two coordinates (xlower, ylower) and (xupper, yupper), where (xlower, ylower) are the coordinates of the lower left corner of the rectangle and (xupper, yupper) are the coordi- nates of the upper right corner of the rectangle. Hence, xlower ≤ xupper and ylower ≤ yupper. At the end of the anonymization process, the trajectory of MOB Oi is replaced by a generalized trajectory if QID(Oi) 6= ∅ and we obtain generalized database D′. A generalized Moving Objects Database D′ is a collection of n Moving Object identifiers D = {O1, O2, · · · , On}, a set T = {t1, t2, · · · , tn} of m distinct time stamps, and a surjective (i.e. onto) function Tg : D × T → R2×R2 that maps object O and time t to the object’s generalized position, represented by a rectangle [(xl, yl), (xu, yu)], at time t 1. For a MOB Oi and all time stamps in T, function Tg generates the generalized trajectory of MOB Oi. Next, we define the notion of generalized trajectory in this context. Definition 4 (Generalized Trajectory of Moving Object) We define generalized trajectory Ti ′ of MOB Oi as a sequence of location-time pairs {(L1, t1), (L2, t2), · · · , (Lm, tm)}, where m ≥ 1, t1 < t2 < · · · < tm, and each location Lj is a rectangle [(xl, yl), (xu, yu)]. In geometric space hierarchy, a rectangle is a more general entity than a point. Definition 4 shows that the core of space generalization approach is 1Point (x, y) can be represented by rectangle [(x, y), (x, y)] with diagonal of length 0 15 2.1. Moving Objects Database: Terminology and Settings to distort a MOD by substituting some spatio-temporal points with spatio- temporal regions. Hence, we create a generalized version of the input database. We say that Ti ′ is a generalization of Ti if and only if: 1. There exist a surjective function f : Ti → Ti′ such that ∀(x, y, t) ∈ Ti either (x, y, t) = f((x, y, t)) or (x, y, t) v f((x, y, t)). 2. The chronological order is preserved. That is, if (xi, yi, ti) precedes (xj, yj , tj) in time, then f((xi, yi, ti)) must precede f((xj , yj, tj). Definition 5 (Trajectory Matching Quasi-Identifier) We say that a generalized trajectory Ti ′ matches the public information of MOB Oi (i.e. PI(Oi)) corresponding to QID(Oi) if and only if ∀(x, y) ∈ PI(Oi), there exists position L ∈ Ti′ such that (x, y) v L at time t ∈ QID(Oi). We let match(Ti ′, QID(Oi)) denote a generalized trajectory matching the public information, which corresponds to the quasi-identifier of MOB Oi. Intuitively, to infer the sensitive information of MOB O, SI(O), an attacker can identify the entire trajectory T , using the public information of O, PI(O). Then SI(O) = T −PI(O). Retrieving an object’s trajectory given the object’s quasi-identifier can be viewed as the following spatio- temporal range query: SELECT T’ FROM D’ WHERE match(T’, QID(O)) If the above query returns a single trajectory, then the attacker can determine a user’s sensitive information with probability 1, which constitutes a privacy breach. In order to protect sensitive information, we adopt k- anonymity, a well-known model for protecting privacy in microdata. We say that a database D′ is a k-anonymous version of the database D if the above range query returns a result set of cardinality at least k for any MOB O ∈ D such that QID(O) 6= ∅. However as we will show in Section 4.1 additional conditions must be imposed in order to preserve location privacy. Example 3 discusses the generalized database in the context of our running example. Example 3 Table 2.1 shows a generalized version of Table 1.3, where the QIDs of each MOB are given in Table 2.2. Table 2.1 is 2-anonymous. That is, each MOB is ”hidden” (i.e. anonymized) with at least one other MOB 16 2.1. Moving Objects Database: Terminology and Settings User t1 t2 t3 t4 ID O1 (0, 0) [(0, 2), (1, 4)] [(2, 4), (2, 7)] (2, 7) O2 [(5, 3), (6,7)] [(4, 6), (5, 7)] (7, 7) (7, 4) O3 (0, 1) [(0, 2), (1, 4)] [(2, 4), (2, 7)] (3, 7) O4 [(0, 4), (4, 6)] (3, 2) [(0, 1), (3, 6)] [(5, 0), (7, 1)] O5 [(5, 3), (6,7)] [(4, 6), (5, 7)] (7, 7) (6, 3) O6 [(0, 4), (4, 6)] (0, 6) [(0, 1), (3, 6)] [(5, 0), (7, 1)] Table 2.1: Generalized 2-Anonymous Moving Objects Database with respect to its quasi-identifier. QID(O1) = {t2} implies that the posi- tion of O1 at time t2 must be hidden and therefore, T (O1, t2) = (1, 4) was generalized to Tg(O1, t2) = [(0, 2), (1, 4)] such that T (O1, t2) v Tg(O1, t2). So as the result of the anonymization, the position of O1 and O3 at time t2 are indistinguishable, i.e. Tg(O1, t2) ≡ Tg(O3, t2). Notice that even though t3 /∈ QID(O1), the position of O1 at t3 was generalized anyway. This gen- eralization was required in order to ensure soundness of the generalized database (see Section 3.2). 2.1.3 Information Loss As mentioned above, we employ space generalization to anonymize an in- put database. Whenever we replace MOB’s (x, y)-position by a rectangular region, we lose some information and the quality of data degrades. We con- sider the probability of finding a particular position in a database. So for each MOB O at time stamp t, we measure the information loss as the reduc- tion in the probability of accurately determining the position of O at time t in D and D′. We define function F(O, t) to be the difference of probability of finding position of object O at time t in input database D and probability of finding position of object O at time t in generalized database D′. The probability of accurately determining a position of a MOB at a given time depends on the type of the position: point, generalized region, or null entry. • Point: In the input database, if a position of MOB O at time t is given by (x, y)-coordinate, then the probability of determining the position of O at t in D equals 1. Similarly, if a MOB’s position was not generalized, then the probability of finding this position in the generalized database is also 1. 17 2.1. Moving Objects Database: Terminology and Settings • Generalized region: In the generalized database, if a position of MOB O at time t is given by rectangle [(xl, yl), (xu, yu)], then the prob- ability of determining the position of O at t in D′ equals 1 area(Tg(O,t)) , where area(Tg(O, t)) denotes the area of rectangle [(xl, yl), (xu, yu)]. But whenever position Tg(O, t) in a generalized database corresponds to a point, we explicitly set 1/area(Tg(O, t)) = 1. • null entry: In the input database, a position of a MOB can be given by a null entry. In Section 2.1.1, we have identified 3 categories of null entries, namely, leading, trailing, and gap null entries. – Leading or trailing null entry: In this case, we assume that the object is immobile and that an adversary may know the clos- est non-null position of the MOB. Therefore in this case, we define the probability of determining the position of O at t in D to be equal to 1. – Gap null entry: Suppose T (O, t) is a gap null entry and T (O, ta) and T (O,tb) are the closest (in time) non-null positions of O (i.e. gap boundaries). As we assumed when we eliminate null entries (see Section 2.1.1), object O at time t is likely to be somewhere inside a rectangle that engulfs the gap boundaries. Therefore, we define the probability of determining the position of O at t in D to be equal to 1/area(T (O, ta),T (O,tb)), where [T (O, ta),T (O,tb)] corresponds to a bounding rectangle of gap boundaries. Equation 2.1 formally defines the function F(O, t), which represents the difference of probabilities of determining a position of MOB O at time t. F(O, t) = ∣∣∣∣ 1area(T (O,ta),T (O,tb)) − 1 area(Tg(Oi,tj)) ∣∣∣∣ , if gap null entry 1− 1 area(Tg(Oi,tj)) , otherwise (2.1) Given input database D and the generalized database D′, we define the information loss resulting in the anonymization of D as follows: IL(D,D′) = n∑ i=1 m∑ j=1 (F(Oi, tj)) (2.2) 18 2.1. Moving Objects Database: Terminology and Settings The above definition of information loss presents a sum of probability differences for all time stamps of every MOB. This metric may not be informative enough as this sum is in the interval [0,∞). We normalize this measure of information loss by dividing it by n ·m, where n is the number of MOBs and m is the number of time stamps in a given database. We refer to this normalized metric (defined in Equation 2.3) as average information loss. avg(IL(D,D′)) = ∑n i=1 ∑m j=1 (F(Oi, tj)) n ·m (2.3) We know that probability difference is in the interval [0, 1], i.e., 0 ≤ F(Oi, tj) ≤ 1 Hence, the numerator in Equation 2.3 is in the interval [0, n · m], which implies that avg(IL(D,D′)) ∈ [0, 1]. Trivially, avg(IL(D,D)) = 0, which says that the closer average information loss to 0, the more similar the generalized data to the original data. When creating a distorted version of the input database, we aim to minimize the corresponding information loss so that the generalized data would still be useful in studies and/or analyses. Therefore, we want the average information loss to be as small as possible. 2.1.4 Range Query Distortion In most cases, the sole purpose of publishing a database is to analyze it. The analysis is usually done by issuing queries to retrieve some data. In case of mobility data, range queries are a natural way to explore the data and to discover knowledge. Given a region R and a time interval [ts, tf ] such that ts ≤ tf , range queries compute the number of MOBs, which are located inside region R at one or more time stamps in the interval [ts, tf ]. Since in the anonymization process a position of a MOB may have been replaced by a region, results of a range query on a generalized database can be distorted compared to the results of the same query on an original database. So we define range query distortion as a measure of the quality of anonymization. In other words, to measure utility of our approach, we compare the results of issuing the same range query on the original input database D and the generalized database D′. In our experiments, we tried to generate regions R and time intervals [ts, tf ] randomly. However in many cases, we obtained biased results, be- cause the random generator produced unreasonably large time intervals for relatively small regions. Therefore, we decided to experiment with range 19 2.1. Moving Objects Database: Terminology and Settings queries that involve a region R and a single time stamp t. Now, we define two types of range queries and the corresponding distortion measures. Let p(D̃, R, t) be a function that returns the number of distinct MOBs in database D̃ that overlap with region R at time stamp t. So p(D̃, R, t) counts the number of exact positions that are inside or on the boundary of region R and the number of generalized positions that overlap with region R. Let d(D̃, R, t) be a function that returns the number of distinct MOBs in database D̃ that are strictly inside region R at time stamp t. Function d(D̃, R, t) counts the number of exact positions that are inside or on the boundary of region R and the number of generalized positions that are completely contained in region R. Given input database D and its k-anonymous version D′, we define two measures of range query distortion using functions p(D̃, R, t) and d(D̃, R, t). Definition 6 (Possibly Inside and Definitely Inside Distortion) For database D and its generalized version D′ given region R and time stamp t, we define Possibly Inside distortion as follows: Possibly Inside = |p(D, R, t)− p(D′, R, t)| p(D′, R, t) (2.4) and we define Definitely Inside distortion as follows: Definitely Inside = |d(D, R, t) − d(D′, R, t)| d(D, R, t) (2.5) The two types of distortion are defined, provided p(D′, R, t) and d(D, R, t) are not equal to 0. After generalization, some MOBs (that were outside of region R at time t) may overlap with the region. Hence, p(D, R, t) ≤ p(D′, R, t), which means 0 ≤ p(D′, R, t)− p(D, R, t) ≤ p(D′, R, t) Therefore by dividing the above inequality by p(D′, R, t), we obtain that Possibly Inside distortion is always in the interval [0, 1]. Similarly, after the generalization process, some MOBs (that were definitely inside region R at time t) may no longer be strictly contained inside region R. Hence, d(D, R, t) ≥ d(D′, R, t), which means 0 ≤ d(D, R, t)− d(D′, R, t) ≤ d(D, R, t) 20 2.1. Moving Objects Database: Terminology and Settings - 6 x y 2 4 6 8 2 4 6 8 uO1 u O2 uO3 u O4 u O5 u O6 R Figure 2.1: Example of Range Query And by dividing the above inequality by d(D, R, t), we obtain that Definitely Inside distortion has to be in the interval [0, 1]. Distortion equals 0, when a range query returns the same set of results after applying the query to both original and generalized databases. This means the results were not distorted by the anonymization process. Therefore, the closer the range query distortion (i.e Possibly Inside and Definitely Inside) to 0, the higher is the quality of the anonymization. Example 4 shows a simple computation of range query distortion using data from our running example. Example 4 Suppose databases D and D′ are given by Table 1.3 and Ta- ble 2.1, respectively. In Figure 2.1, the xy-positions of the MOBs at time t1 are shown using circles, while the generalized positions at time t1 are shown with dash-line rectangles. Let R be the rectangular region [(0, 1), (7, 5)], where (0, 1) is the lower left coordinate and (7, 5) is the upper right coordi- nate of region R. We issue a range query with region R and time stamp t1. Then, p(D, R, t1) = d(D, R, t1) = 3, since MOBs O3, O4, O5 are inside or on 21 2.2. Quasi-Identifiers the boundary of region R at time t1. We also have that p(D′, R, t1) = 5, since O3 is still inside R, while the generalized regions of MOBs O2, O4, O5, O6 overlap with region R. We can see that d(D′, R, t1) = 1, because after the generalization the generalized regions of MOBs O4 and O5 are not strictly contained in region R and only MOB O3 is still contained inside region R. Now we can compute the distortion with respect to D, D′, R, and t1 as described in Definition 6. Possibly Inside = |3 − 5|/5 = 0.4 and Definitely Inside = |3− 1|/3 = 0.6. 2.2 Quasi-Identifiers In our settings, we define a quasi-identifier (QID) as a temporal pattern that corresponds to time stamps at which object’s positions are known pub- licly. Hence, a QID of MOB Oi, denoted by QID(Oi), is a set of time stamps {ti1 , ti2 , · · · , tis}, where s ≤ m, tij ∈ T, and ti1 < ti2 < · · · < tis . And for each t ∈ QID(Oi), T (Oi, t) corresponds to a location that Oi fre- quently visits and/or that is known even before the trajectory of Oi is pub- lished. In other terms, as stated in Definition 2, the public information of Oi, PI(Oi, tij ), corresponds to positions of Oi at the time stamps in QID(Oi). We observed that due to the nature of mobility data we cannot assume that a certain set of locations or time stamps constitutes a quasi-identifier for all individuals in the database. In practice, different mobile users are likely to live in different parts of a city and take different routes when driving from home to work/school. Therefore, it is very likely for various MOBs to have different QIDs, some of which may be disjoint while others may overlap. Thus, the key idea in our work is to model the quasi-identifiers subjectively based on individual movement patterns. So we assume that objects in our settings do not have to share the same quasi-identifier and this is a striking difference between the notions of quasi-identifier in mobility settings and relational microdata; in the latter, each tuple’s quasi-identifier consists of the same set of attributes. We say that QID(Oi) and QID(Oj) overlap, if QID(Oi)∩QID(Oj) 6= ∅. Otherwise, QID(Oi) andQID(Oj) are disjoint. If QID(Oi)∩QID(Oj) = QID(Oi) = QID(Oj), then QID(Oi) and QID(Oj) are identical and we denote it by QID(Oi) ≡ QID(Oj). Example 5 Suppose the quasi-identifiers of the MOBs from our running example (see Example 1 and Table 1.3) are given by Table 2.2. For instance, the quasi-identifier of Alice is the set {t2}, i.e. QID(O1) = {t2}. We say that two QIDs overlap if they share one or more common time stamps. QID(O1) overlaps with QID(O2) on time stamp t2, and QID(O1) and QID(O4) are 22 2.2. Quasi-Identifiers User Name User ID QID Alice O1 t2 Bob O2 t1, t2 Cate O3 t2, t3 David O4 t1, t3, t4 Erica O5 t2 Frank O6 ∅ Table 2.2: Quasi-identifiers of Moving Objects disjoint, since they do not share any common time stamps. It is also possible that a quasi-identifier of a MOB is an empty set, as exemplified by QID(O6). An absolutely natural question to ask is where these quasi-identifiers come from. In Section 2.2.1, we discuss some feasible approaches of acquiring QID lists in real-world situations. 2.2.1 Possible Ways to Obtain Quasi-Identifiers Recent years have witnessed the popular use of location-aware devices, for in- stance, GPS-enabled cell phones, location sensors and active radio-frequency identification (RFID) tags. There exist many ways to collect location traces. However, to the best of our knowledge, no one has tried to collect informa- tion about the quasi-identifiers of mobile users. There is no known algorithm to systematically differentiate between public and sensitive information of a given mobile user. We look at two real-life examples of collecting loca- tion traces and we suggest some possible and plausible ways of obtaining quasi-identifiers in the case of these examples. As the first example, many citizens of the city of Milan, Italy have agreed to participate in the GeoPKDD project [14], provided they would receive a certain discount on the mandatory car insurance. GPS devices were in- stalled in their cars and their movement around the city was recorded. Po- tentially, the organizers of the project could have asked each participant to identify the time stamps at which he/she would want his/her positions to be anonymized. In general, people using some location-based services may construct their own quasi-identifiers and to supply the QIDs to the service provider if they agree to the terms that their location information can be published to be used in some studies aimed to enhance public welfare. That is, users’ involvement may be required to produce QIDs. In our second example, we consider the company Octopus Cards Ltd. 23 2.2. Quasi-Identifiers The company enables its customers (residents of Hong Kong, Macau and Shenzhen) to pay for services such as public transportation, retail, leisure facilities using Octopus RFID smart card [32]. Each Octopus card has a unique ID number associated with it. Over 18 million cards are presently in circulation and over 10 million transactions per day are processed involv- ing Octopus cards. Each point of transaction is essentially a position of an Octopus card user at a particular time. The Octopus point-of-transaction data could be published in order to analyze the movement and behavioral patterns of city residents. As [39] observed, when a user pays using his/her Octopus card at several fast-food restaurants belonging to the same chain, the company (i.e. the chain of fast-food restaurants) accumulates the user’s transaction history. This history corresponds to a subset of the user’s move- ment/trajectory. If this subset of positions can uniquely identify a user in the published data, then this is exactly the public information that must be anonymized. In other words, this subset will correspond to the user’s quasi-identifier. Hence, a QID of every user can be obtained by determin- ing one or more companies, where the user frequently uses his/her Octopus card. Alternatively, when obtaining users’ consent to publish their point- of-transaction data, the data owners may request each user to specify the transaction locations of companies that the user would desire to anonymize in the published data. The above two examples show that some personalized approach needs to be taken to obtain QIDs for each MOD. Since we did not have a real- world QID data, we have design a QID generator to produce some random quasi-identifiers of various lengths. Section 2.2.2 discusses this generator in detail. 2.2.2 Generating Quasi-Identifiers In general, it is very difficult to predict what a quasi-identifier of a given MOB would consist of. Given only an individual trajectory without any additional background information, it is not clear what positions must be anonymized as they constitute public knowledge. Producing realistic QIDs corresponding to a given set of trajectories remains an open issue. We assume that a list of quasi-identifiers is given to us together with the Moving Objects Database to be anonymized. Unfortunately, we did not obtain lists of QIDs for the databases we used in our experiments. That is the reason why we have implemented a QID generator to create a list of QIDs for any input MOD. Our generator is based on the notion of QID graph, which is defined as follows. 24 2.2. Quasi-Identifiers ff Q1, Q5 ff Q2 ff Q3 ff Q4 ff Q6 @ @ @ @ @ @ @ Figure 2.2: Example of QID Graph Definition 7 (QID Graph) QID graph G is an undirected graph with ver- tex set V = {Q1, Q2, · · · , Qp} and edge set E = {e1, e2, · · · , er}, where p ≤ n. Each vertex Qi ∈ V is a distinct quasi-identifier (QID) of one or more Moving Objects (MOBs). There is an edge (Qi, Qj) ∈ E if and only if the quasi-identifiers Qi and Qj overlap. Example 6 Let Qi denote QID(Oi) for i ∈ {1, 2, · · · , 6}. The QID graph in Figure 2.2 is a graphical representation of the quasi-identifiers from Ta- ble 2.2. Notice that since Q1 ≡ Q5, there is one vertex corresponding to the two QIDs. So if two MOBs have identical QIDs, then there will be only one vertex corresponding to this QID in the graph. If QID(O1) ≡ QID(O2) ≡ · · · ≡ QID(Oi), we say the the MOBs O1, O2, · · · , Oi form a block of size i. We generate QIDs on per block basis. The QID generator accepts the following three parameters as inputs: (i) min qid, (ii) max qid, and (iii) block size. • min qid is the minimal number of time stamps per QID • max qid is the maximal number of time stamps per QID • block size is the number of MOBs that have the same QID Since there are n MOBs in the input database, the generator will cre- ate ⌈ n block size ⌉ blocks and one quasi-identifier for each block. Typically, we consider only those values of block size such that block size n. The gen- erated QIDs comprise some QID graph. Our generator can enforce several graph structures. Here are some examples of the supported graph structure: • Random graph: Any randomly generated graph. Algorithm 1 shows the pseudo-code for generating a random QID graph. 25 2.3. k-Anonymity Algorithm 1 QID Generator: Random Graph Input: min qid, max qid, block size Output: A list of quasi-identifiers Q 1: Q← ∅ 2: b← dn/block sizee 3: for j = 1 to b do 4: size← Randomly choose an integer from interval [min qid,max qid] 5: Qj ← Randomly generate a quasi-identifier such that Qj = {t | t ∈ T} and |Qj| = size 6: Q← Q ∪Qj 7: end for • Disjoint graph: A graph without any edges, i.e. V 6= ∅ and E = ∅. This means that QID(Oi)∩QID(Oj) = ∅ for all i and j in the graph. • Tree: A connected acyclic graph. • Cycles: A graph consisted of one or more disjoint cycles. Suppose QID(Oa) = {t1, t2}, QID(Ob) = {t2, t3, t5}, QID(Oc) = {t1, t5, t7}, then QID(Oa), QID(Ob) and QID(Oc) create a cycle. In summary, our motivation for generating QIDs is the lack of real-world examples of such data. Any given list of QIDs can be represented as a QID graph as defined above. In our opinion, random graph configuration is the most likely candidate to capture the QID distribution for a particular set of MOBs as we expect a lot of variability in the QIDs. The other graph configurations were explored merely out of academic curiosity. 2.3 k-Anonymity An important issue that accompanies sharing and publishing of a database, is the security of the sensitive information in the database. Sensitive infor- mation refers to data objects that individuals would want to keep private. Personal explicit identifiers such as full name, driver’s license number and combinations of identifiers such as {postal code, gender, date of birth} can be used by adversaries to derive sensitive information. Dalenius [9] has coined the term quasi-identifier. In the context of relational microdata, quasi-identifier refers to a set of attributes that can be linked with external 26 2.3. k-Anonymity Name Age City Disease Salary ($) Arthur 35 Vancouver Bronchitis 32000 Betty 20 Whistler Hepatitis 50000 Charlie 49 Seattle Bronchitis 50000 Diane 36 Tacoma Diabetes 24000 Edward 21 Vancouver Hepatitis 60000 Fiona 45 Seattle Diabetes 30000 Table 2.3: Instance of Hospital Microdata information (e.g. attacker’s prior knowledge) in order to uniquely identify an individual. The notion of k-anonymity was initially introduced by Samarati and Sweeney in [37] and then further expanded in [38]. k-anonymity is essen- tially a model for protecting data privacy in the publicly released database. The model requires that every individual in a database is indistinguishable from at least k− 1 other individuals (whose information also appears in the database) with respect to a particular quasi-identifier. The set of attributes comprising a quasi-identifier is chosen by a database owner/administrator and is the same for every individual in a database. Two ways were pro- posed to to achieve k-anonymity: (i) generalization and (ii) suppression. The former approach requires to replace a value with more general, but se- mantically consistent value. In the latter approach, a value is deleted from published data. As discussed in Section 2.5, both generalization and sup- pression were previously applied to Moving Objects Databases to anonymize MOBs [1, 39]. However in [1], the generalization was applied to an entire tra- jectory, whereas we generalize each trajectory with respect to each MOB’s quasi-identifier. A set of k or more individuals, that are indistinguishable from each other with respect to a quasi-identifier, form an anonymity group. All anonymity groups must be disjoint on all the attributes in a given quasi-identifier. It has been proved that finding an optimal k-anonymization is NP-hard [3, 27]. Example 7 illustrates the application of k-anonymity to relational microdata. Example 7 Consider a collection of patient records shown in Table 2.3. Attributes Disease and Salary constitute sensitive information that must not be discovered by adversaries when the records are published. Hence, the explicit identifiers such as Name have to be suppressed from the microdata. However, if an adversary possesses some background knowledge about an individual (i.e. one or more non-sensitive values), then this individual’s 27 2.3. k-Anonymity Age City Disease Salary ($) [20, 29] Whistler Hepatitis 50000 [20, 29] Vancouver Hepatitis 60000 [30, 39] Tacoma Diabetes 24000 [30, 39] Vancouver Bronchitis 32000 [40, 49] Seattle Diabetes 30000 [40, 49] Seattle Bronchitis 50000 Table 2.4: 2-anonymous Hospital Microdata (Generalized by Age) Age City Disease Salary ($) ≤ 35 British Columbia Bronchitis 32000 ≤ 35 British Columbia Hepatitis 50000 ≤ 35 British Columbia Hepatitis 60000 > 35 Washington State Diabetes 24000 > 35 Washington State Diabetes 30000 > 35 Washington State Bronchitis 50000 Table 2.5: 3-anonymous Hospital Microdata (Generalized by Age and City) record can be re-identified. For instance, suppose an attacker knows that Arthur is 35 years old, then Arthur’s disease and annual salary can be discovered even after deleting all the names from Table 2.3. To ensure Arthur’s privacy, it is possible to generalize by Age. Table 2.4 shows a 2- anonymous version of Table 2.3, where the quasi-identifier was set to {Age}. If the adversary only knows the Arthur’s age, then he/she will only find out from Table 2.4 that Arthur has either Diabetes or Bronchitis. Similarly, Arthur’s real salary cannot be determined with probability 1. Two common types of privacy breach are identity disclosure and sen- sitive information disclosure. The former occurs whenever an individual is linked to a particular record in a database. Whereas the latter occurs when a new piece of information is revealed that was not known prior to database publication. Identity disclosure leads to sensitive information dis- closure; the converse of this statement is not necessarily true. k-anonymity can protect data against identity disclosure. However, sensitive information disclosure can occur in k-anonymous database when all the tuples in a cer- tain anonymity group are identical on one or more sensitive attributes. That is, k-anonymity model is flawed as it may leak information due to lack of diversity in sensitive attributes [25]. For example, knowing that Betty is 20 28 2.4. Hilbert Space-Filling Curve years old (or that she belongs to [20, 29] age group) will reveal that Betty has hepatitis, since the two individuals in [20, 29] age group in Table 2.4 suf- fer from the same disease. Moreover, privacy can be breached if adversaries know more than one non-sensitive (i.e. public) values for some records. Example 8 Suppose one attacker knows that Arthur is 35 years old and another attacker know that Arthur lives in Vancouver, British Columbia. If the two attackers combine their prior knowledge, they can re-identify Arthur’s record in Table 2.4. Therefore, the hospital microdata must be generalized on both Age and City non-sensitive attributes. Table 2.5 illus- trates microdata that satisfies 3-anonymity. In the 3-anonymous version of hospital microdata, there is some diversity in the sensitive attributes and hence, there is no information leak as in the case of 2-anonymous microdata of Table 2.4. Despite the aforementioned flaws and limitations, k-anonymity remains a fundamental privacy-protecting model with practical application. [8, 36, 44] present an extensive survey of anonymization methods, where a lot of the techniques are based on k-anonymity model. We adopt the concept of k- anonymity to our settings and require that every MOB is indistinguishable from at least k−1 other MOBs with respect to its quasi-identifier (QID in the following). We achieve anonymity by employing space generalization, where a location sample (i.e. (x, y)-coordinate) is replaced by a rectangular region [(x1, y1), (x2, y2)], where (x1, y1) is the coordinate of the lower-left corner of the rectangle and (x2, y2) is the coordinate of the upper-right corner of the rectangle. In our settings, Moving Objects do not require to have the same QID. Due to the fact that QIDs are not necessarily identical, the classical k-anonymity approach used in relational databases is not sufficient to ensure privacy preservation of mobility data. In Chapter 3, we discuss the challenges of adapting k-anonymity framework that arise due to the nature of quasi-identifier values. 2.4 Hilbert Space-Filling Curve In database management, many prominent techniques such as Grid files, KD trees, and R-trees were developed to index spatial data. Space-filling curve is another such technique, which is a continuous surjective function whose domain is the interval [0, 1] and the range is the N -dimensional unit hypercube [0, 1]N . Informally, it is a curve that passes through all point ofN - dimensional unit hypercube. David Hilbert, a German mathematician, has 29 2.4. Hilbert Space-Filling Curve 0 1 2 3 0 1 23 4 5 6 7 8 9 10 11 1213 14 15 0 1 2 3 21 42 62 63 (a) (b) (c) Figure 2.3: Examples of Hilbert Curve: (a) First order (P = 1), (b) Second order (P = 2), (c) Third order (P = 3) introduced a continuous fractal space-filling curve [21] that was later named Hilbert curve. The P th order Hilbert curve (P ≥ 1) can index integers in the interval [0, 2P −1]. That is, the key property of Hilbert curve is that the curve traverses every point with integer coordinates in N -dimensional space. Hence, any N -dimensional point can be mapped to linear space according to the order of Hilbert curve traversal. Figure 2.3 shows the first three orders of Hilbert curve in 2-dimensional space. Hilbert index of an N -dimensional point (x1, x2, · · · , xN ) is an integer corresponding to the linear order in which Hilbert curve traverses the point. For example, if points A and B have Hilbert indexes 0 and 8, then A and B are the first and the ninth points, respectively, visited by the curve. We denote a Hilbert index of MOB Oi at time stamp t by Ht(Oi). Similarly, H(x, y) denotes the Hilbert index of point (x, y). It has been shown that Hilbert curve preserves spatial locality [30]. We use the locality-preserving property of Hilbert curve to estimate spatial proximity of objects. We as- sume that whenever |Ht(O1)−Ht(O2)| < |Ht(O1)−Ht(O3)| the MOB O2 is closer to MOB O1 than MOB O3 at time t. Unfortunately, this assumption is not true in all cases. For instance, in the third order Hilbert curve for 2-dimensional space (see Figure 2.3), 6 = |0− 6| = |H(0, 0) −H(3, 1)| < |H(0, 0) −H(0, 2)| = |0− 14| = 14 but point (3, 1) is farther away from point (0, 0) than point (0, 2). We use Hilbert space-filling curve to index positions of MOBs at differ- ent time stamps. Hamilton and Rau-Chaplin designed an efficient algorithm 30 2.5. Related Work to compute Hilbert indexes of a hypercube whose sides have different car- dinality [20]. As discussed above, we consider 2-dimensional spatial data, where each position is given by two floating-point positive numbers. When computing a Hilbert index, we do not need to round these floating-point numbers to the nearest integer. Instead, we find a bit representation of the numbers using the IEEE Standard for Floating-Point Arithmetic (IEEE 754-2008). Given a static instance of a Moving Objects Database, the space in which all objects move can be bound by a rectangle, whose sides are not necessarily of equal length. We implement the algorithm presented in [20] to compute a Hilbert index of (x, y)-coordinate of a given MOB. As men- tioned in Section 2.3, computing an optimal k-anonymization of microdata is NP-hard. By adapting k-anonymity model, we inherit the NP-hardness of finding an optimal solution. We approximate our solution by using the above assumption about absolute difference of Hilbert indexes, which we formulate as the following heuristic. Space Proximity heuristic. Due to locality-preserving prop- erty of Hilbert space-filling curve, the smaller the absolute differ- ence of Hilbert indexes of two MOBs at time t, the closer these MOBs are to each other in space That is, we approximate the distance between two MOBs at time t by an absolute difference of their corresponding Hilbert indexes. We generate an anonymization group of each MOB by computing k Nearest Neighbors us- ing a Top-k algorithm, which employs Hilbert indexes as distance measure. Section 3.1.1 presents a detailed overview the Top-k algorithm. 2.5 Related Work A large part of the existing work on anonymity of spatio-temporal moving points has been mainly conducted in the context of location-based service (LBS)). In this context a trusted server is set to handle users’ requests and to pass them on to the third party service providers, and the general goal is to provide the service on-the-fly without threatening the anonymity of the user requiring the service. That is, this work targets real-time applications. The trusted server uses space generalization to make the request sent from a user indistinguishable from k − 1 other requests [18]. This approach is called spatial-cloaking. Based on this idea, a system that supports different queries and various anonymization requirements was developed in [28, 29]. Spatial-cloaking has also been applied in a distributed environment, where 31 2.5. Related Work users dynamically form and maintain the k-anonymized spatial regions [15]. In [13] instead of using the same k for all messages, the system allows each message to specify an independent anonymity value and the maximum spa- tial and temporal tolerance resolutions it can tolerate based on its privacy requirements. The proposed algorithm tries to identify the smallest spa- tial area and time interval for each message, such that there exist at least k−1 other messages from distinct users, with the same spatial and temporal dimensions. The work described in [23] proposes a privacy system that takes into account only the spatial dimension: the area in which location anonymity is evaluated is divided into several regions and position data is delimited by the region. Anonymity is achieved by satisfying the following two crite- ria: (i) ubiquity requires that a user passes through at least k regions; (ii) congestion requires that the number of users in each region is at least k. High ubiquity guarantees the location anonymity of every user, while high congestion guarantees location anonymity of local users in a specified region. Beresford et al. [5] have introduced the concept of mix zones. A mix zone is an area, where the location based service providers cannot trace users’ movements. When a user enters a mix zone, the service provider does not receive the real identity of the user, but a pseudonym that changes whenever the user enters a new mix zone. This approach ensures that the identities of users entering a mix zone in the same time period are mixed. The mix zone method used only predefined fixed areas. Ouyang et al. [34] have designed a dynamic mix zone method, which creates virtual zones according to user movement and perturbs location information to preserve privacy. This is a greedy method, which finds a circle (of predefined radius) that covers the maximum number of MOB positions at time stamp t. Users’ positions inside a circle are replaced by a randomly chosen positions (withing the same circle). The circles are generated dynamically when two or more MOBs are in each other’s vicinity and destroyed when the MOBs are moving away from each other. Similarly to mix zone, the work in [19] presents sensitivity map, an approach that classifies areas. Each location is classifies as either sensitive or insensitive. Three algorithms were designed in order to hide users’ positions in sensitive areas. Contrary to the notions of mixed zones and sensitivity maps, the ap- proach introduced by Bettini et al. in [6] is geared on the concept of lo- cation based quasi-identifier (lbqid), i.e. a spatio-temporal pattern that can uniquely identify one individual. More specifically, a location based quasi-identifier (lbqid) is a sequence of spatio-temporal constraints plus a recurrence formula. This approach aims to ensure that no sensitive data are 32 2.5. Related Work released from the trusted server receiving the requests, to a third party ser- vice provider, when the data can be personally identified through a lbqid. In addition, [6] has introduced the concept of historical k-anonymity. Given the set of requests issued by a certain user (i.e. its trajectory in our ter- minology), it satisfies historical k-anonymity if there exist k − 1 personal history locations (i.e. trajectories in our terminology) belonging to k − 1 different users such that they are location-time consistent (i.e. indistin- guishable). Hence, what the framework in [6] aims at, is to make sure that if a set of requests from a user matches a location based quasi-identifier, then it satisfies historical k-anonymity. The idea is that if a service provider could successfully track the requests of a user through all the constraints of a lbqid, then there should be at least k − 1 other users whose personal history of locations is consistent with these requests, and thus any of these users may have issued those requests. In other words, a set of trajectories that satisfies historical k-anonymity can be considered safe for the purpose of data publishing. Although defined in a completely different context, and with slightly different concepts of quasi-identifiers, this work has some similarities to our proposal. The main difference with our context, is the fact that Bettini et al. consider data points (requests) continuously arriving, and thus they provide online anonymity. The anonymity group is chosen once and for all, at the time this is needed. This means that the k − 1 MOBs passing closer to the point of the request are selected to form the anonymity group regardless of what they will do later, as this information is not yet available. In our context anonymity is provided offline. Our aim is to anonymize a static database of Moving Objects, assuring that the quality of the data is kept high (i.e. minimizing the information loss). To this end, we must select anonymity groups considering the MOBs as the location information is given priori. Thus the solution proposed in [6] is not suitable for our purposes. Processing MOBs online or offline is the main difference between our work and all the work developed for lbs. In this thesis, we face the problem from the perspective of privacy aware data publishing, i.e. the same context as in the case of classical k-anonymity. In our setting, we have a static database of moving objects. We want to pub- lish it (for instance, for analysis purposes) in such a way that the anonymity of the individuals is preserved, but also the quality of the data is kept high. On the contrary, in the LBS context the aim is to provide the service without learning user’s exact position. Consider, for instance, the aforementioned concept of mix zones. Mix zone approach enhances anonymity of LBS users while still allowing to provide the service. However, it is not applicable to 33 2.5. Related Work data publishing, since the quality of the data is completely destroyed. Indeed, a considerable amount of work has been done on anonymity in the context of location-based services, but most of this work cannot be adapted to our settings to ensure secure data publication. There also exist approaches that focus on enforcing MOD privacy offline, but which could also be extended to online applications. An example of such work is the path confusion algorithm developed by Hoh and Gruteser [22]. This algo- rithm finds a pair of trajectory segments, which do not intersect. Then the algorithm perturbs the location samples in these segments such that the two trajectories intersect while minimizing the error imposed on the altered database. The authors model the privacy problem as a constraint optimiza- tion problem and their algorithm has exponential running time with respect to the number of MOBs. Thus, this algorithm is not extensible to be applied to real-world databases. The following two papers have tackled the problem of privacy preserv- ing data publishing of MOD based on the idea of k-anonymity. First, Abul et al. [1] propose a novel concept of k-anonymity based on co-localization that exploits the inherent uncertainty of the acquired positions of Mov- ing Objects. Due to sampling and positioning systems (e.g., gps) impreci- sion, the trajectory of a moving object is no longer a polyline in a three- dimensional space, instead it is a cylindrical volume, where its radius δ represents the possible location imprecision: we know that the trajectory of the moving object is within this cylinder, but we do not know exactly where. If another object moves within the same cylinder they are indistinguishable from each other. This leads to the definition of (k, δ)-anonymity for MODs. A method for obtaining (k, δ)-anonymity is introduced in [1] and named NWA (N ever Walk Alone). The method is based on trajectory clustering and spatial translation (i.e., moving in space the necessary points, in order to make a trajectory lie within the anonymity cylinder). The main difference of this work with our proposal is that it does not account for any notion of quasi-identifiers: the trajectories are clustered together considering their similarity all along their lifetime. Moreover, the same considerations made above about the possibility of using classical k-anonymity methods in our context, hold also for NWA. In particular the fact that NWA produces a complete partitioning of the trajectories is an overkill with respect to data quality given that our problem definition does not require partitioning. Second, Terrovitis and Mamoulis study the problem of protecting privacy in the publication of trajectory databases in [39]. The basic assumption of this work is that different adversaries own different, disjoint parts of the trajectories, and more importantly, the data publisher is aware of the 34 2.5. Related Work adversarial knowledge. The goal of this work is to ensure that any adversary cannot learn anything more than what he/she already knows from the data publication. The typical scenario is when the data publisher is a credit card company, and the adversaries are the various companies where the customers make some transactions at point-of-sale services using the credit card. The data publisher owns the whole trajectories, while the adversaries own only few points for each trajectory, corresponding to the transactions in their point-of-sale. Anonymization is obtained by means of suppression of the dangerous observations from each trajectory. As generalization is not used, it is not even needed to consider the spatial proximity of the locations, and many of the challenges (e.g. overlapping anonymity groups) that arise when we employ space generalization, are not an issue in this setting. 35 Chapter 3 Preliminaries This chapter presents problems associated with adapting k-anonymity model from classical microdata to MOD settings. In particular, we address the challenges of computing anonymity groups efficiently and ensuring sound- ness of the generalized database, which could be undermined due to overlap- ping anonymity groups. To address the former challenge, we employ Hilbert space-filling curve (see Section 2.4) and adapt an efficient top-K algorithm presented in Section 3.1. In Section 3.2, we show that there could be some inconsistencies whenever AG(Oi) overlaps with AG(Oj) and QID(Oi) over- laps with QID(Oj). To address this challenge, we define an equivalence relation of MOBs with respect to a time stamp and divide MOBs into dis- joint equivalence classes. The solutions to the challenges presented in this chapter provide the core functionality of our anonymization algorithms dis- cussed in Chapter 4. 3.1 Anonymity Groups We adopt k-anonymity model from relational microdata to Moving Object settings and aim to compute k-anonymous version of the input Moving Ob- jects Database. So for every Moving Object whose quasi-identifier is not an empty set (we refer to this object as subject MOB), we want to compute a set of k − 1 other MOBs and make these MOBs and the subject MOB indistinguishable from each other on all time stamps in the subject’s QID. We call this set of k MOBs an anonymity group. To make all the MOBs in a particular anonymity group indistinguishable from each other, we re- place their positions at a particular time stamp by the smallest rectangle that contains all of these positions. There are 2 key steps in achieving our goal: (i) To compute an anonymity group (denoted AG(O)) for every sub- ject MOB O, (ii) To add AG(O) to every equivalence class EC(t) such that t ∈ QID(O). The latter step is discussed in Section 3.2. In that section, we show that when two anonymity groups AG(Oi) and AG(Oj) have a MOB in common and at least one time stamp t is in both QID(Oi) and QID(Oj), then all the MOBs in the union of these two anonymity groups will be in 36 3.1. Anonymity Groups the same equivalence class with respect to time t. In the former step, given MOB O, we seek a set of k − 1 other MOBs such that when these k − 1 MOBs are generalized with the subject MOB O, the resulting information loss will be minimized. The task of computing anonymity groups presents two computational difficulties. First, calculating the distance from a subject MOB to all other MOBs in a database is ex- pensive, and hence, will affect the performance of our algorithms. Second, we need to find k − 1 MOBs that are close in space to the subject MOB at all the time stamps in the quasi-identifier of the subject. Thus, we need to consider distance from a subject MOB to some other MOB aggregated over all time stamp in the subject’s QID. So to compute an anonymity group, we need to find k− 1 nearest neighbors with respect to a particular subject MOB and its QID. If we use Euclidean distance to measure space proximity between two MOBs Oi and Oj at time t, then we will have to compute the distance for every pair of MOBs at every time stamp. There are n(n − 1) pairs of MOBs and m time stamp, hence the Euclidean distance computation can be performed in O(n2m) time. This can result in prohibitively large running time for large values of n. Therefore, instead of Euclidean distance, we make use of Hilbert space-filling curve. As illustrated in Section 2.4, Hilbert curve preserves locality of multi- dimensional points and the distance between two MOBs Oi and Oj at time t can be approximated by a difference of their corresponding Hilbert in- dexes. We refer to this approximation as space proximity heuristic. To take advantage of this heuristic, we preprocess the location information in an in- put MOD. More specifically, we compute Hilbert indexes of positions of all MOBs at each time stamp t ∈ T. For each time stamp t, we generate a list Lt of pairs of MOB identifiers and their corresponding Hilbert indexes. Each entry in list Lt is of the form (O,Ht(O)). The list is sorted in ascending order by Hilbert index. We refer to list Lt as Hilbert list at time t. Example 9 We index every position in the MOD from our running example using third order Hilbert curve. Table 3.1 shows the Hilbert indexes of the positions from Table 1.3. To construct a Hilbert list at a given time stamp, we insert MOBs and their Hilbert indexes into a list and sort the list in the increasing order by Hilbert index. For instance, the Hilbert list at t1 will be equal to Lt1 = {(O1, 0), (O3, 1), (O6, 20), (O4, 32), (O2 , 38), (O5, 51)}. Similarly, Lt4 = {(O1, 25), (O3, 26), (O2, 47), (O5, 51), (O4, 59), (O6 , 62)}. Now, given subject MOB O, the k − 1 nearest MOBs at time t can be approximated by the MOBs with the smallest Hilbert index deviation from 37 3.1. Anonymity Groups User t1 t2 t3 t4 ID O1 0 17 25 25 O2 38 38 42 47 O3 1 14 30 26 O4 32 9 6 59 O5 51 36 42 51 O6 20 20 20 62 Table 3.1: Hilbert Indexes for Running Example O at t. The main advantage of using Hilbert lists is that we can find the k−1 MOBs with the least Hilbert index deviation without necessarily computing the deviation from the subject MOB for all MOBs in the database. We define the local score of MOB Oi with respect to subject MOB O in list Lt as the absolute difference of the Hilbert indexes of the two MOBs at time t, i.e. scoret(Oi) = |Ht(O) − Ht(Oi)|. Our goal is to determine a set of k − 1 MOBs whose aggregate distance from the subject MOB is as small as possible. We can use any monotonically increasing function as an aggregate distance function. But in our implementation, we have used sum as the aggregate distance function, that is, we add up all local scores to obtain a global overall score. Then the problem of finding k − 1 nearest neighbors is reduced to determining top-K MOBs with the lowest overall score, where K = k−1. In Section 3.1.1, we discuss an efficient algorithm to answer top- K queries and in Section 3.1.2 show how to adopt the algorithm in order to apply it on Hilbert lists. 3.1.1 Overview of Top-K Algorithms The problem of computing the K items with the top overall score, where the data items are stored in lists sorted by the local score, has been extensively studied in the field of computer science. The most common settings of the problem can be described as follows. We are given M lists of N items, where each item has a local score. Each list is sorted in descending order by local score. Given an aggregate function, the overall score of an item is computed by applying the function on its local scores in all of the M lists. The aim is to find K items with the largest overall score. The brute- force approach is to compute overall score of all items, to sort the scores in descending order, and then to select the first K items. This naive approach has O(M ·N+N · log(N)) time complexity. This approach maybe infeasible 38 3.1. Anonymity Groups for large values of M and N . Fagin and colleagues developed a family of efficient algorithms that use a constant-size buffer and may stop early without having to compute overall score of every item in a list [10]. One landmark algorithm introduced in [10] is Threshold Algorithm (TA). TA performs two types of data accesses: (i) sorted access, and (ii) random access. In sorted access, the algorithm pro- ceeds through each list sequentially by retrieving the local score of the next item in a currently visited list. The random access is the mode, by which the algorithm access a specified item in a specified list. The algorithms per- forms sorted access into each list in parallel. When object R is encountered under sorted access in one of the lists, the algorithm does random access to M − 1 other lists to retrieve the local scores of R. If the overall score of R is higher than the smallest overall score of an item in the top-K set of items found so far, then R will replace that item in the top-K set. The thresh- old is computed by applying the aggregate function to the last seen local score in each list. If the smallest overall score of an item in the top-K set is greater or equal to the threshold, then the algorithm halts as the top-K items have been found. It has been proven that if the aggregate function f is a monotonically increasing function (i.e. x ≤ y → f(x) ≤ f(y)), then the algorithm finds the correct top-K items. The distinguishing property of TA is the fact that it is an instance optimal algorithm. As defined in [10], given an algorithm class A and a set of inputs D, algorithm A ∈ A is instance optimal over any other algorithm B ∈ A for every input D ∈ D, if there exist constants c1 and c2 such that cost(A,D) ≤ c1 · cost(B,D) + c2 Essentially, instance optimality guarantees that, on any valid input, algo- rithm B can compute the answer in the number of steps, which is smaller than the number of steps taken by algorithm A only by constant factor, provided algorithm B does not make any wild lucky guesses. Constant c1 is called optimality ratio. Akbarinia et al. noticed that TA may perform some redundant accesses on some input instances and proposed two Best Position Algorithms, BPA and BPA2, which are more efficient than TA [4]. Akbarinia et al. proved that similar to TA, if the aggregate function is monotonically increasing, then BPA and BPA2 find the correct top-K answers to a query. In addition, they proved that both BPA and BPA2 are instance optimal and the optimality ratios of BPA and BPA2 are smaller or equal to the optimality ratio of TA. It was shown in [4], that the execution cost of BPA can be (M − 1) times 39 3.1. Anonymity Groups smaller than the execution cost of TA, where M is the number of lists. We adapt BPA2 algorithm because it was shown that BPA performs at most as many accesses as TA and BPA2 performs at most as many accesses as BPA. TA and BPA may access the same item several times in different lists. For instance, an item can be accessed in list Li under random access and later it can be accessed again in the same list under sorted access. BPA2 optimizes the performance by accessing the same item at most once. Let best position bp in list Li denote the greatest position such that for any position j ∈ [1, bp], the item at position j in list Li has been seen under sorted or random access. BPA2 stores best positions in each list. In order to avoid redundant accesses, instead of doing sorted access, BPA2 does direct access to position bp+ 1 when processing the lists of items in parallel. The threshold is computed by applying the aggregate function to the local scores from the best positions of all lists. As in TA, BPA2 stores the set of K items with the highest overall scores and halts when the smallest overall score of an item in the top-K set is greater or equal to the threshold. 3.1.2 Adapting Best Position Algorithm Given a subject MOB, our goal is to compute its corresponding anonymity group. In other words, for each MOB whose QID is not an empty set, we need to compute a set of k−1 other MOBs such that the aggregate distance of these MOBs from the subject MOB is the lowest among all MOBs. As discussed in Section 3.1.1, we adapt BPA2 algorithms developed in [4]. So for subject MOB O, we are given q lists with n items each and a monotonically increasing aggregate function f , where q = |QID(O)| and n is the number of distinct MOBs in the input database. Each entry in the list is a pair of MOB identifier and Hilbert index. Each list is sorted in ascending order by the Hilbert index. The local score of MOB Oa with respect to subject MOB O is given by |Ht(O) − Ht(Oa)|. We denote the position of subject MOB O in Hilbert list Lj by SPj . We access items in positions that are smaller (i.e. below) than SPj and in positions that are larger (i.e. above) than SPj . Let BPbelow denote the smallest position such that for any position p ∈ [BPbelow, SPj ], the item in position p has been seen under direct or random access. Similarly, let BPabove denote the largest position such that for any position p ∈ [SPj , BPabove], the item in position p has been seen under direct or random access. SupposeMOBOa is at position BPbelow − 1 and MOB Ob is at position BPabove + 1 with respect to subject MOB O. We compute score below and score above as |Ht(Oi) − Ht(Oa)| and |Ht(Oi)−Ht(Ob)|, respectively. If score below < score above, then we 40 3.1. Anonymity Groups Algorithm 2 GenerAG : Anonymity Group Generating Top-K Algorithm Input: K, lists L1, L2, · · · , Lq, function f Output: A set of K MOBs 1: For each Hilbert list Lj, initialize SPj to the position of subject MOB O in Lj. Initially, set BPbelow and BPabove to be equal to SPj . 2: Do direct access to best position BP in each of the q Hilbert lists in parallel. As object Oi is seen under direct access in some list, do random access to the other q − 1 lists to compute the local score of Oi. Then compute the overall score of Oi using function f . Maintain set H of MOBs with the lowest overall score among the objects seen so far (ties are broken arbitrarily). 3: After a direct or random access to a list, update BPbelow and BPabove and propagate the changes to BP in this list. Let scorej(BP ) be the local score of an object at best position BP in list Lj. 4: Compute threshold λ = f(score1(BP ), score2(BP ), · · · , scoreq(BP )). If set H contains K objects such that the largest overall score of an object in set H is less or equal to threshold λ, then halt. Otherwise, go to 1. 5: Return set H. set the best position BP = BPbelow−1. Otherwise, we set the best position BP = BPabove + 1. We maintain pointers to SPj , BPbelow and BPabove for every Hilbert list. Algorithm 2 shows the pseudo-code for generating one anonymity group for subject MOB O, given the lists of Hilbert indexes corresponding to the time stamps in QID(O), anonymity group size K, and aggregate function f as inputs. Example 10 We continue with our running example, where Table 1.3 gives the input database and Table 3.1 given the indexing of objects positions using Hilbert curve. Suppose k = 3 and we want to compute the anonymity group for MOB O3, given that QID(O3) = {t2, t3}. So we need to find 2 other MOBs such that their aggregated distance from O3 is the smallest among all MOBs. Lists L1 and L2 correspond to Hilbert lists at time stamps t2 and t3, respectively. Figure 3.1 shows the Hilbert lists L1 and L2 and the positions of the pointers (maintained for each list) at the time the algorithm returns. In L1, we initialize SP1, BPbelow, and BPabove to be equal to 2. Whereas in L2, we initialize SP2, BPbelow, and BPabove to be equal to 4. We have |Ht2(O3) −Ht2(O1)| < |Ht2(O3) −Ht2(O4)|, i.e. score(O1) = 3 in L1 with 41 3.1. Anonymity Groups Position L1 L2 MOB Index MOB Index 1 2 3 4 5 6 O4 O3 O1 O6 O5 O2 9 14 17 20 36 38 O4 O6 O1 O3 O2 O5 6 20 25 30 42 42 ff ff ff ffSP1, BPbelow BPabove SP2, BPabove BPbelow Figure 3.1: Example of Computing Anonymity Group respect to subject MOB O3 is less than score(O4) = 5 in the same list. Therefore, the best position in L1 equals 3. We set BPabove in L1 to 3, while BPbelow still equals to SP1. We do a random access to L2 in order to get the local score of O1 in L2. We add O1 to the top-K set of MOBs. We update the BPbelow in L2 to be equal to 3, while BPabove still equals SP2. The current threshold equals score1(BP ) + score2(BP ) = 3 + 5 = 8. At the next iteration, we do a direct access to the best position in list L2. MOB O6 is at the best position in L2 and score2(O6) = 10. We do a random access to L1 to get score1(O6). We update BPabove in L1 to be equal to 4, while BPbelow still equals to SP1. We add O6 to the top-K set. We compute the current threshold: λ = f(score1(BP ), score2(BP )) = 6 + 10 = 16. Since score(O1) ≤ score(O6) ≤ λ, the algorithm halts. MOBs O1 and O6 are returns. That is, we obtain AG(O3) = {O1, O3, O6}. We refer to Algorithm 2 as GenerAG and use it to compute anonymity group of each MOB in all of our anonymization algorithms presented in Chapter 4. GenerAG algorithm is instance optimal. The instance optimality is inherited from the Best Position algorithm. That is, for all MOBs in a given Hilbert list, we can compute the absolute Hilbert index differences from the subject MOB and store it in a list. We can sort this list in the 42 3.2. Equivalence Classes ID t1 t2 t3 QID(O) AG(O) O1 (0, 1) (2, 2) (3, 2) t1 O1, O3 O2 (1, 3) (2, 3) (4, 2) t1, t2 O2, O1 O3 (0, 0) (1, 0) (3, 1) t3 O3, O1 Table 3.2: Example of Overlapping Anonymity Groups ascending order and repeat this process for all other Hilbert lists. Then we can apply BPA2 algorithm on such lists to compute the top-K objects with the smallest absolute Hilbert index difference. 3.2 Equivalence Classes As we will show in this section, providing a robust and sound definition of k-anonymity in the case of MOD is not a straightforward task. The difficulty in establishing such definition stems from the fact that different MOBs may have different QIDs, and their corresponding anonymity groups may overlap. In classical k-anonymity model, the anonymity groups must be disjoint, otherwise the anonymized database will not be sound. The following example illustrates the problem with soundness of a generalized MOD due to overlapping anonymity groups. Example 11 Let us look at the MOD in Table 3.2. Suppose we are trying to create 2-anonymous generalized database (i.e. k = 2). QID(O1) = {t1} and AG(O1) = {O1, O3} because O3 is the closest MOB to O1 at time t1. Similarly, QID(O2) = {t1, t2} and AG(O2) = {O2, O1} because O1 is the closest MOB to O2 at times t1 and t2. So at time stamp t1, the object O1 will be first generalized with O3 and then it will be generalized with O2 or the other way around. In order to create a sound generalized MOD, each object must be generalized with the same set of k − 1 other MOBs with respect to this object’s QID. Using any generalization order, the anonymization of O1 will be inconsistent, i.e. the resulting generalized database will not be sound. In general, for any two MOBs Oi and Oj , if AG(Oi) ∩AG(Oj) 6= ∅ and QID(Oi)∩QID(Oj) 6= ∅, then there can be inconsistencies in the generalized database. By inconsistency we mean a case where a MOB is generalized with two different groups of MOBs with respect to the same time stamp. To resolve the inconsistencies, we can merge the overlapping anonymity 43 3.2. Equivalence Classes groups with respect to time stamps t ∈ QID(Oi)∩QID(Oj). We define the following equivalence relation of MOBs with respect to time t. Definition 8 (Equivalence Relation of Moving Objects) Given any two MOBs Oa and Ob, we say that Oa is equivalent to Ob (i.e. Oa ∼ Ob) with respect to time stamp t if and only if there exist MOBs Oi and Oj such that: • Oa ∈ AG(Oi) and Ob ∈ AG(Oj) • AG(Oi) ∩AG(Oj) 6= ∅ • t ∈ QID(Oi) ∩QID(Oj) From Definition 8, we can see that Oa is equivalent to every MOB in AG(Oi)∪AG(Oj) with respect to every time stamp in QID(Oi)∩QID(Oj). Our definition of equivalence relation satisfies the three properties of equiv- alence relations for any MOBs Oa, Ob, Oc and time stamp t as shown below. • Reflexivity: Oa ∼ Oa. Trivially, if Oa ∈ AG(Oi), then AG(Oi) 6= ∅ andAG(Oi)∩AG(Oi) 6= ∅. And if t ∈ QID(Oi), then t ∈ QID(Oi) ∩QID(Oi). • Symmetry: If Oa ∼ Ob, then Ob ∼ Oa. If Oa ∼ Ob and Oa ∈ AG(Oi) and Ob ∈ AG(Oj), then AG(Oi) ∩ AG(Oj) = AG(Oj) ∩ AG(Oi) 6= ∅ and t ∈ QID(Oi) ∩ QID(Oj) = QID(Oj) ∩QID(Oi) • Transitivity: If Oa ∼ Ob and Ob ∼ Oc, then Oa ∼ Oc. Suppose we have that Oa ∈ AG(Oi), Ob ∈ AG(Oj), Oc ∈ AG(Oh) and t ∈ QID(Oi) ∩ QID(Oj) ∩ QID(Oh). We have that Oa is equiv- alent to every MOB in AG(Oj) with respect to t. By symmetry, Oc ∼ Ob and therefore, Oc is equivalent to every MOB in AG(Oj) with respect to t. So we can say that Oa ∈ AG(Oi) ∪ AG(Oj) and Oc ∈ AG(Oh) ∪ AG(Oj). Then we know that (AG(Oi) ∪ AG(Oj)) ∩ (AG(Oh) ∪ AG(Oj)) 6= ∅ and t ∈ QID(Oi) ∩ QID(Oh). Therefore, Oa ∼ Oc. We use the equivalence relation of MOBs in order to merge overlapping anonymity groups into disjoint equivalence classes. The equivalence class (EC) corresponding to equivalence relation (from Definition 8) is defined as follows. 44 3.2. Equivalence Classes Definition 9 (Equivalence Class at time t) An equivalence class with respect to time stamp t, denoted by EC(t), is a set of MOBs such that Oi ∈ EC(t) and Oj ∈ EC(t) if and only if Oi ∼ Oj with respect to time t. Equally, if AG(Oi)∩AG(Oj) 6= ∅ and t ∈ QID(Oi)∩QID(Oj), then EC(t) contains all the MOBs in AG(Oi) ∪AG(Oj). There can be more than one equivalence class with respect to any given time stamp. If two equivalence classes overlap, then they are merged into one equivalence class in order to preserve the property that any two equivalence classes (with respect to the same time t) are either disjoint or equal. We implement equivalence classes as disjoint set forests using Union-Find algo- rithm with union-by-rank and path compression. The original Union-Find algorithm was introduced in [12]. Algorithm 3 presents the pseudo-code for adding a set of MOBs to an equivalence class with respect to a specific time stamp. We refer to Algorithm 3 as AGtoEC and employ it in all of our anonymization algorithm in order to transform each anonymity group to a collection of equivalence classes with respect to a particular time stamp set. AGtoEC algorithm work as follows. After the initialization, we find the root of the equivalence class that contains a particular MOB Oi (Step 4). If MOB Oi is not in any equivalence class with respect to time stamp t, then we create a new equivalence class that contains only MOB Oi (Step 6). All the MOBs in AG(O) have to be in the same equivalence class, which does not overlap with any other equivalence class with respect to time stamp t. So if two MOBs in AG(O) are in different equivalence classes, then these equivalence classes must be merged into one equivalence class (Step 9). At the end, the algorithm return the root of the equivalence class the contains all the MOBs of the input set AG(O). Example 12 In Table 3.2, the anonymity groups of O1 and O2 overlap (i.e. AG(O1) ∩AG(O2) = {O1} 6= ∅). In addition, QID(O1) ∩QID(O2) = {t1}. Hence, O1 ∼ O2 with respect to time stamp t1, and since O3 ∈ AG(O1), we have that O3 ∼ O2 with respect to t1. Thus, EC(t1) = [O1, O2, O3], which implies that all MOBs in EC(t1) must be anonymized together at time stamp t1. We require that all MOBs in a given equivalence class are generalized to the same region. This is analogous to generalization of a disjoint anonymity group in case of classical k-anonymity model in microdata. The difference is that in MOD settings the anonymity groups may not be disjoint and to achieve this disjointness, we transform anonymity groups into equivalence 45 3.2. Equivalence Classes Algorithm 3 AGtoEC : Adding AG(O) to EC(t) Input: AG(O) = a set of MOBs, time stamp t Output: All MOBs in AG(O) are added to the appropriate equivalence class with respect to time stamp t. The pointer to the root of the equivalence class is returned 1: currentRoot← null 2: lastRoot← null 3: for all Oi ∈ AG(O) do 4: currentRoot← FIND(Oi) 5: if currentRoot = null then 6: currentRoot← Make a new node with rank 0 7: end if 8: if lastRoot 6= null and currentRoot 6= lastRoot then 9: lastRoot← UNION(currentRoot, lastRoot) 10: else 11: lastRoot← currentRoot 12: end if 13: end for 14: return lastRoot classes with respect to time stamps t ∈ T. The use of equivalence classes ensures that there are no inconsistencies in the generalized database, which in turn, provides a sound definition of k-anonymity in our settings. However, the disadvantage of using equivalence classes lies in the fact that size of any equivalence class is in the interval [k, n]. So an object is generalized with all other objects in the same equivalence class, which can contain more than k− 1 other objects. On one hand, larger equivalence classes (in the number of MOBs) provide a higher degree of anonymity, but on the other hand, they result in higher information loss. One of the open issues in our work is to try to find a schedule for anonymizing MOBs such that the resulting equivalence classes are as small as possible. If a MOB is in some equivalence class at time stamp t, then the position of this MOB will be generalized to a region, which engulfs all the position of the MOBs in this equivalence class. Algorithm 4 shows the pseudo-code of the generalization procedure. When we say we generalized to the least upper bound (Step 1 of Algorithm 4), we mean replacing all the position in the equivalence class EC(t) by the smallest rectangle that contains these positions. 46 3.2. Equivalence Classes Algorithm 4 lubSpace: Space Generalization Input: EC(t) = a set of MOBs, time stamp t Output: All MOBs at time t are assigned either the original or generalized position 1: Rlub ← Generalized EC(t) to the least upper bound 2: for all i ∈ {1, 2, · · · , n} do 3: if Oi ∈ EC(t) then 4: Tg(Oi, t)← Rlub 5: else 6: Tg(Oi, t)← T (Oi, t) 7: end if 8: end for Algorithms GenerAG, AGtoEC, and lubSpace presented in this chapter are used by each one of our proposed anonymization algorithms, which we discuss in the following chapter. 47 Chapter 4 Algorithms In this chapter, we present our anonymization algorithms to compute k- anonymous version of the input database, given this database, anonymity threshold, and a list of quasi-identifiers as inputs. But we begin with a discussion of an attack model on a published generalized database. We il- lustrate some pitfalls of simple adaptation of k-anonymity to MOD settings, which can lead to privacy breach in our proposed attack model. In this chapter, we formally define the notion of k-anonymity for Moving Objects Databases, which guarantees that privacy breach cannot occur using our attack model. We then present three algorithms that satisfy this definition of k-anonymity in MOD settings and prove their correctness. The three al- gorithm we developed are Extreme Union (EU), Symmetric Anonymization (SA), and Restricted Symmetric Anonymization (RSA). Algorithms EU and SA (presented in Section 4.3 and 4.4, respectively) were first published in proceedings of EDBT conference [43]. RSA is a modified version of SA as explained in Section 4.5. 4.1 Attack Model Our ultimate goal in this work is to create k-anonymous version of the input MOD for some user-specified anonymity threshold k. A basic building block of constructing a notion of anonymity is indistinguishability. We achieve anonymity by ensuring that a MOB is indistinguishable from at least k − 1 other MOBs with respect to its QID. In other terms, we substitute public information of a MOB by more general information such that every position in the public information is replaced by a region. So if an attacker knows the exact position (x, y) of MOB O at time stamp t, then in the published database, the attacker will find that the position (x, y) is inside some region R and there are at least k MOBs, which are located in region R at time t. However, only one of the object in region R corresponds to MOB O. We refer to this object at anonymized counterpart of MOB O. Essentially, anonymized counterpart A is MOB O after generalizing O with at least k−1 other MOBs. It should be noted that some or all of the sensitive information 48 4.1. Attack Model of O could be generalized in the anonymization process as illustrated by Example 3. Definition 10 (Anonymized Counterpart of Moving Object) As a result of anonymization process, every MOB O in input MOD D is trans- formed into MOB A in generalized MOD D′ such that the sensitive informa- tion of A is either identical to or a generalization of the sensitive information of O and the public information of A is the generalized version of the public information of O. We call A the anonymized counterpart of O. We model the correspondence between the public information in D and D′ as a bipartite graph, which we call the attack graph. Intuitively, given MODs D and D′, the attack graph shows the information that can be derived from D′ by applying the public information of all MOBs (i.e. PI) in D to the generalized trajectories in D′. Definition 11 (Attack Graph) Attack graph G is a bipartite graph with vertex set V = {VO, VA} and edge set E. Every vertex O ∈ VO corresponds to MOB O ∈ D. Every vertex A ∈ VA corresponds to MOB A ∈ D′. There exists edge (O,A) ∈ E if and only if T (O, t) v Tg(A, t) for all t ∈ QID(O). From Definition 11 it follows that there always exists an edge in attack graph G between a MOB and its anonymized counterpart. The assignment of individuals to anonymized MOBs is consistent if there exist a perfect matching in the attack graph G. Figure 4.1 illustrates two possible attack graphs for some hypothetical databases. Every vertex Oi ∈ VO in Figure 4.1 has degree 2 or higher, which implies that every MOB in indistinguishable from at least one other MOB with respect to its QID. Privacy of a mobile user will be breached if an attacker can identify this user’s anonymized counterpart in the published database D′ using the pub- lic information (as described in Definition 2) of this user. In terms of the attack graph, this means that an attacker can find an edge e such that e must participate in every perfect matching. Equivalently, the attacker can prune any edge that does not participate in any perfect matching as this edge cannot be an edge between a MOB and its anonymized counterpart. So edge (Oi, Aj) that participates in every perfect matching must be the real mapping between MOB Oi and its anonymized counterpart Aj . If the attacker can identify the real anonymized counterpart of MOB Oi, then the attacker can discover the sensitive information of Oi (as described in Def- inition 3). We consider the identification of Aj as privacy breach despite 49 4.1. Attack Model u u u u u u PPPPPPPPP PPPPPPPPP O1 O2 O3 A1 A2 A3 u u u u u u u u u u PPPPPPPPP PPPPPPPPP @ @ @ @ @ @ @ @@ O1 O2 O3 O4 O5 A1 A2 A3 A4 A5 (a) (b) Figure 4.1: Example of Attack Graphs: Privacy breached at the following edges: (a) edge (O1, A1), (b) edge (O5, A5) the fact that some sensitive information of Oi could have been generalized in the anonymization process as mentioned above. The following defini- tion outlines an attack procedure that can be used to identify one or more anonymized counterparts. Definition 12 (Attack Model) Given published database D′, which cor- responds to a Moving Objects Database D and PI of D corresponding to the quasi-identifiers of the MOBs in D, the attacker can procede in the following manner: 1. Construct attack graph G(V,E) 2. Identify edge e that cannot participate in any perfect matching in bi- partite graph G. Prune edge e to obtain attack graph G̃(V, Ẽ), where V = VO ∪ VA and Ẽ = E − {e} 3. Repeat Step 2 for graph G̃(V, Ẽ) 4. After pruning all such edges, if there exists edge (Oi, Aj) such that the degree of vertex Aj is 1, then Aj is the anonymized counterpart of MOB Oi. So privacy breach occurs when the attacker can identify at least one vertex A ∈ VA such that deg(A) = 1. The attack graphs shown in Figure 4.1 are susceptible to privacy breach. In Figure 4.1(a) edge (O1, A1) must participate in every perfect matching 50 4.1. Attack Model u u u u u u PPPPPPPPP PPPPPPPPP O1 O2 O3 A1 A2 A3 u u u u u u u u u u PPPPPPPPP PPPPPPPPP @ @ @ @ @ @ @ @@ O1 O2 O3 O4 O5 A1 A2 A3 A4 A5 (a) (b) P P P P P Figure 4.2: Example of Symmetric Attack Graphs and hence, it can be deduced that the generalized trajectory of A1 cor- responds to MOB O1, which constitutes privacy breach. If we apply our attack model (from Definition 12) to Figure 4.1(b), then we can prune edge (O2, A5) as it does not participate in any possible perfect matching in this graph. Then vertex A5 has degree 1 and hence, the attacker can conclude that (O5, A5) is the real mapping of MOB O5 to its anonymized counterpart. The attack graphs in Figure 4.1 provide some insights for defining a robust notion of k-anonymity in MOD settings. For instance, from Fig- ure 4.1(a), we can conclude that it is not sufficient for every vertex O ∈ VO to have degree k as some vertices in VA can have degree less than k. The example in Figure 4.1(b) illustrates that privacy breach can occur even if all vertices in a given attack graph have degree k or higher. To provide a robust notion of k-anonymity, we define the notion of symmetric attack graph. We base our k-anonymity model for MOD settings on this notion (see Section 4.2) and prove that this model will not admit privacy breach by applying the attack model described in Definition 12. Definition 13 (Symmetric Attack Graph) Attack graph G with vertex set V = {VO, VA} and edge set E is symmetric if and only if the following condition holds for all edges in E: If edge (Oi, Aj) ∈ E, then there exists edge (Oa, Ab) ∈ E such that: • Ab is the anonymized counterpart of Oi • Aj is the anonymized counterpart of Oa 51 4.2. k-Anonymity Model for Moving Objects Note that any edge between a MOB and its anonymized counterpart automatically satisfies the condition in Definition 13. We can transform the attack graphs from Figure 4.1 into symmetric attack graphs by adding several edges. For instance, in Figure 4.1(a), we need to add edge (O2, A1) to make the attack graph symmetric. Whereas in Figure 4.1(b), we need to add edges (O4, A5) and (O5, A2). We show these symmetric attack graphs in Figure 4.2, where dashed lines correspond to the aforementioned edges that we added. In Section 4.2, given the definition of attack model and symmetric attack graph, we formally define the notion of k-anonymity in the case of Moving Objects Databases. Using the definition of k-anonymity, we form our prob- lem statement and present three algorithms in the forthcoming sections, which solve the problem we address. 4.2 k-Anonymity Model for Moving Objects As we observed in Section 4.1, privacy breach can occur in a non-symmetric attack graph, where every vertex has degree k or higher. In this section, we formally define the notion of k-anonymity of Moving Objects, which guarantees there exists a perfect matching for any arbitrary edge in a given attack graph. The fact that every edge participates in some perfect matching implies that no edge can be pruned from the attack graph according to the attack model presented in Definition 12. Next, we define the k-anonymity model in MOD settings and prove that it does not admit privacy breach based on our proposed attack model by showing how to construct a perfect matching for any arbitrary edge in the associated attack graph. Definition 14 (k-Anonymity of Moving Objects) Let D be an input Moving Objects Database and D′ be a generalized version of D. Given a list of QIDs for MOBs in D, let G(V,E) be the attack graph corresponding to D and D′, where V = {VO, VA}. Then D′ satisfies k-anonymity (i.e. D′ is a k-anonymous version of D) if and only if the following two conditions are both true: • G is symmetric (see Definition 13) • Every vertex A ∈ VA has degree k or higher (i.e. ∀A ∈ VA,deg(A) ≥ k) It should be noted that if an attack graph satisfies the two conditions in Definition 14, then every vertex O ∈ VO has degree k or higher (i.e. ∀O ∈ VO, deg(O) ≥ k). We now show that privacy breach cannot occur 52 4.2. k-Anonymity Model for Moving Objects using our definition of k-anonymity and our attack model. In our attack model an edge in an attack graph can be pruned if and only if the edge does not participate in any perfect matching. We will show that in our k-anonymity model, no edge can be pruned. Theorem 1 (Correctness of k-Anonymity) Let D′ be a k-anonymous version of MOD D. Using the attack model from Definition 12, privacy of MOBs in D′ will not be breached. Proof : Let G(V,E) be a symmetric attack graph corresponding to the public information of D and the published database D′. There are at least k vertices adjacent to every vertex A ∈ VA. In other words, for every vertex Aj ∈ VA, there are at least k edges incident to vertex Aj. If neither of these edges can be pruned using our attack model, then an attacker can only guess with probability 1/k what MOB has Aj as its anonymized counterpart. Hence, no privacy breach can occur if no edge can be eliminated from G. To prove that no edge can be pruned from G, we show that for any arbitrary edge (Oi, Aj) ∈ E, there exists a perfect matching such that (Oi, Aj) participates in this matching. There are two cases to consider. • Case 1 : Aj is the anonymized counterpart of Oi. In this case, we map every MOB O ∈ VO to its anonymized counterpart A ∈ VA. Since every object has its anonymized counterpart in the attack graph, then there always exists a perfect matching between O and A. • Case 2 : Aj is not the anonymized counterpart of Oi. MOB Oi must have its anonymized counterpart in G. Let Ab denote the anonymized counterpart of Oi. Also, Aj has to be an anonymized counterpart of some MOB. Let Oa denote the MOB such that Aj is the anonymized counterpart of Oa. If (Oi, Aj) ∈ E, then there exists edge (Oa, Ab) ∈ E because G is symmetric. We can map Oi to Aj , and Oa to Ab. For all other vertex O ∈ VO such that O 6= Oa and O 6= Oi, we map O to its anonymized counterpart A. Note that A 6= Ab and A 6= Aj . This mapping is a perfect matching. This completes the proof. 2 Theorem 1 states that our anonymity model is impervious to the attack strategy proposed in Definition 12. This attack strategy can breach loca- tion privacy when applied on non-symmetric attack graphs. In the future line of research, it would be interesting to consider more aggressive attack models. For instance, we have investigated the following attack strategy on 53 4.2. k-Anonymity Model for Moving Objects a symmetric attack graph. An attacker can construct all possible perfect matchings in a given symmetric attack graph. Let popularity of edge e de- note the number of distinct perfect matchings that edge e is included in. After constructing all perfect matchings, the attacker can compute popular- ity of each edge. The attacker can then proceed using the following greedy approach. The attacker selects edge (Oi, Aj) with the maximal (or minimal) popularity (ties are broken arbitrarily) and assumes that edge (Oi, Aj) spec- ifies the true mapping of MOB Oi to its anonymized counterpart Aj . The attacker removes all edges incident with vertices Oi and Aj and repeats the greedy approach of selecting an edge with maximal (or minimal) popularity. We found that in some cases the most popular edge corresponds to a true mapping, whereas in other cases the least popular edge was in fact mapping a MOB to its anonymized counterpart. In other words, we were able to show that this greedy attack model allows an adversary to guess that an edge is a true edge, but the model does not lead consistently to privacy breach. By presenting the aforementioned attack models, we hope to ignite research, which would explore more aggressive attack strategies in order to provide stronger security and anonymity of mobile users. 4.2.1 Problem Statement Now that we have formally defined the notion of k-anonymity of MOBs (Definition 14) and information loss (Equation 2.2), we state the problem that we address and solve in this thesis. Problem Statement: We are given the following three inputs: (i) anonymity threshold k, where k ∈ N, (ii) a Moving Ob- jects Database D, and (iii) a list Q of quasi-identifiers for Mov- ing Objects in database D. We create a k-anonymous version of database D using space generalization, while minimizing the corresponding information loss. In Section 3.1.2, we have introduced GenerAG algorithm (see Algo- rithm 2), which computes an anonymity group for a subject MOB such that the MOBs in the anonymity group have the smallest aggregate distance from the subject MOB, where distance is approximated by an absolute difference of Hilbert indexes. If we simply generalize a MOB with all the MOBs in its anonymity group, then the resulting attack graph may not necessarily satisfy the conditions of k-anonymity. Therefore, we need a systematic approach of extending the anonymity groups produced by GenerAG in order to en- sure that the conditions of k-anonymity are true. The algorithms presented 54 4.3. Extreme Union in the remainder of this chapter employ two different approaches to ensure k-anonymity conditions hold. Extreme Union algorithm takes union of all QIDs of the MOBs in a particular anonymity group. Whereas Symmetric Anonymization and Restricted Symmetric Anonymization algorithms en- force symmetry among different anonymity group such that if Oj ∈ AG(Oi), then Oi ∈ AG(Oj). Thus, these algorithms solve the problem identified in the above problem statement. 4.3 Extreme Union Our first anonymization algorithm, which uses space generalization to com- pute the k-anonymous version of the input database, is Extreme Union (EU) algorithm. Extreme Union algorithm essentially proceeds in three stages. First, for each MOB O, we compute its anonymity group AG(O) using Gen- erAG algorithm (Algorithm 2). Second, we take the union of QIDs of all MOBs in AG(O). Third, we add AG(O) to equivalence classes with respect to all time stamps in this union using AGtoEC algorithm (Algorithm 3). Finally, we use lubSpace algorithm (Algorithm 4) to generalize positions of MOBs in a given equivalence class to the region that encapsulated all positions of these MOBs. Algorithm 5 shows the pseudo-code for Extreme Union algorithm. Next, we prove Theorem 2, which shows the correctness of Extreme Union algorithm. Theorem 2 (Correctness of Extreme Union) Given MOD D as input, the generalized database D′ computed by Extreme Union algorithm is a k- anonymous version of D. Proof : Let AG(Oi) and AG(Oj) be the anonymity groups of MOBs Oi and Oj , respectively. Suppose Oj ∈ AG(Oi), then AG(Oi) ∩ AG(Oj) 6= ∅. AG(Oi) is generalized with respect to time stamp union TUi, where TUi =⋃ O∈AG(Oi) QID(O). Similarly, AG(Oj) is generalized with respect to time stamp union TUj, where TUj = ⋃ O∈AG(Oj) QID(O). Then, QID(Oi) ∪ QID(Oj) is a subset of TUi ∩ TUj. So Oi and Oj will be in the same equivalence classes with respect to each time stamp t ∈ QID(Oi)∪QID(Oj). Hence, Oi and Oj will be generalized to the same region with respect to all time stamps t ∈ QID(Oi)∪QID(Oj). So, if Ai and Aj are the anonymized counterparts of Oi and Oj respectively, then there will be edges (Oi, Aj) and (Oj , Ai) in the attack graph corresponding to D and D′ (i.e. the resulting attack graph is symmetric). Every MOB Oj ∈ AG(Oi) contributes one 55 4.3. Extreme Union Algorithm 5 Extreme Union Input: D,Q, and anonymity threshold k Output: D′ = k-anonymous version of D 1: D← {Oi | Oi ∈ D, QID(Oi) 6= ∅} 2: T← {t | t ∈ QID(O), QID(O) ∈ Q} 3: for all Oi ∈ D do 4: AG(Oi) ← Compute anonymity group (of size k) of Oi with respect to QID(Oi) using GenerAG algorithm 5: T imeStampUnion← ∅ 6: for all Oj ∈ AG(Oi) do 7: T imeStampUnion← T imeStampUnion ∪QID(Oj) 8: end for 9: Add AG(Oi) to equivalence classes with respect to all time stamps in T imeStampUnion using AGtoEC algorithm 10: end for 11: for all t ∈ T do 12: Generalize equivalence class EC(t) using lubSpace algorithm 13: end for edge (Oj , Ai) incident to the anonymized counterpart of Oi. And since |AG(Oi)| ≥ k, there will be at least k edges incident to vertex Ai. 2 While Extreme Union algorithm does in fact produce a k-anonymous generalized database, the algorithm can result in substantial information loss. This disadvantage of the algorithm is due to generalization with re- spect to union of all QIDs in a given anonymity group. For any MOB O, we compute AG(O) by selecting k−1 other MOBs that are in close proximity of MOB O and thus, heuristically minimize the generalized area with respect to QID(O). However, AG(O) is not generalized with respect to QID(O), but with respect to time stamp union TU , where TU = ⋃ Oi∈AG(O) QID(Oi). For all time stamps t ∈ TU−QID(O), any MOB Oi ∈ AG(O) (Oi 6= O) can be arbitrarily far away from O. Generalizing O and Oi together can lead to a large region, which in turn results in considerable information loss. In Sec- tion 4.4, we present Symmetric Anonymization algorithm, which uses a less aggressive approach for generalizing MOBs and achieving k-anonymity of the generalized database. We finish this section by showing the application of Extreme Union algorithm to our running example. Example 13 Recall that Extreme Union algorithm requires three inputs: D, Q, and k. Database D is given by Table 1.3 and QID list Q is given 56 4.3. Extreme Union ID t1 t2 t3 t4 O1 (0, 0) [(0, 2), (1, 4)] [(2, 4), (2, 7)] (2, 7) O2 [(5, 3), (6,7)] [(4, 6), (5, 7)] (7, 7) (7, 4) O3 (0, 1) [(0, 2), (1, 4)] [(2, 4), (2, 7)] (3, 7) O4 [(0, 4), (4, 6)] (3, 2) [(0, 1), (3, 6)] [(5, 0), (7, 1)] O5 [(5, 3), (6,7)] [(4, 6), (5, 7)] (7, 7) (6, 3) O6 [(0, 4), (4, 6)] (0, 6) [(0, 1), (3, 6)] [(5, 0), (7, 1)] EC(ti) [O2, O5] [O1, O3] [O1, O3] [O4, O6] [O4, O6] [O2, O5] [O4, O6] Table 4.1: Application of Extreme Union or Symmetric Anonymization or Restricted Symmetric Anonymization on Running Example (k = 2) ID t1 t2 t3 t4 O1 (0, 0) [(0, 2), (5, 7)] [(0, 1), (7, 7)] (2, 7) O2 [(0, 3), (6,7)] [(0, 2), (5, 7)] [(0, 1), (7, 7)] [(5, 0), (7, 4)] O3 (0, 1) [(0, 2), (5, 7)] [(0, 1), (7, 7)] (3, 7) O4 [(0, 3), (6, 7)] [(0, 2), (5, 7)] [(0, 1), (7, 7)] [(5, 0), (7, 4)] O5 [(0, 3), (6,7)] [(0, 2), (5, 7)] [(0, 1), (7, 7)] [(5, 0), (7, 4)] O6 [(0, 3), (6, 7)] [(0, 2), (5, 7)] [(0, 1), (7, 7)] [(5, 0), (7, 4)] EC(ti) [O2, O4, O5, O6] [O1, O2, O3, [O1, O2, O3, [O2, O4, O5, O6] O4, O5, O6] O4, O5, O6] Table 4.2: Application of Extreme Union on Running Example (k = 3) by Table 2.2. First, we consider k = 2. Table 4.1 shows the 2-anonymous version of input database D and the equivalence classes for each time stamp in D. The average information loss in this case is 0.29652778. Second, we run Extreme Union algorithm on the running example for k = 3. In case of k = 3, the average information loss is 0.78960317. Table 4.2 shows the 3-anonymous generalized database after applying Extreme Union algorithm on input database D. Table 4.2 also shows the equivalence classes we obtain at the end of the algorithm for each time stamp in D. 57 4.4. Symmetric Anonymization 4.4 Symmetric Anonymization One of the two conditions in our definition of k-anonymity is that the induced attack graph must be symmetric. Extreme Union algorithm achieves this at the expense of generalizing each anonymity group with respect to time stamp union of all QIDs of the MOBs in the anonymity group. Alternatively, we can ensure that each anonymity group AG(O) is generalized with respect to a fixed set of time stamps, which equals to QID(O). But then we must carefully control the composition of anonymity group in order to ensure the resulting attack graph is symmetric. We compute anonymity group using GenerAG algorithm (see Algorithm 2), which only finds k−1 nearest MOBs without taking the induced attack graph into consideration. In Symmetric Anonymization (SA) algorithm, we enforce symmetry in between a subject MOB and k − 1 nearest MOBs in its anonymity group. We achieve sym- metry by inserting the subject MOB into anonymity groups of the subject’s nearest neighbors. In other words, after we compute the anonymity group for some subject MOB Oi, we enforce that Oi is in anonymity group of MOB Oj , for each Oj ∈ AG(Oi). Algorithm 6 presents the pseudo-code of Symmetric Anonymization algorithm. In Theorem 3, we prove that the at- tack graph induced by Symmetric Anonymization algorithm satisfies the two conditions of k-anonymity; the attack graph is symmetric and every vertex corresponding to some anonymized counterpart has degree k or higher. Theorem 3 (Correctness of Symmetric Anonymization) Given MOD D as input, the generalized database D′ computed by Symmetric Anonymization algorithm is a k-anonymous version of D. Proof : Let AG(Oi) and AG(Oj) be the anonymity groups of MOBs Oi and Oj , respectively. Let Ai and Aj be the anonymized counterparts of Oi and Oj respectively. Each anonymity group contains at least k distinct MOBs, i.e. |AG(Oi)| ≥ k for all Oi such that QID(Oi) 6= ∅. Hence, Oi is anonymized with at least k − 1 MOBs. That is, there are at least k edges incident with Ai in the attack graph, i.e. deg(Ai) ≥ k. Now, we show the other condition of k-anonymity is satisfied. In particular, we show that the induced attack graph is symmetric. Suppose Oj ∈ AG(Oi), then there will be edge (Oi, Aj) in the attack graph, since Oi is generalized with Oj (and all other MOBs in AG(Oi)) with respect to all time stamps in QID(Oi). If Oj ∈ AG(Oi), then Symmetric Anonymization will force Oi to be in AG(Oj). Therefore, Oj will be anonymized with Oi with respect to all time stamps in QID(Oj), which implies there will be edge (Oj , Ai) in the attack 58 4.4. Symmetric Anonymization Algorithm 6 Symmetric Anonymization Input: D,Q, and anonymity threshold k Output: D′ = k-anonymous version of D 1: D← {Oi | Oi ∈ D, QID(Oi) 6= ∅} 2: T← {t | t ∈ QID(O), QID(O) ∈ Q} 3: for all Oi ∈ D do 4: AG(Oi)← {Oi} // Anonymity group initialization 5: end for 6: for all Oi ∈ D do 7: k̃ ← k − |AG(Oi)| 8: if k̃ > 0 then 9: HS(Oi) ← Compute k̃ nearest neighbors of Oi with respect to QID(Oi) using GenerAG algorithm 10: AG(Oi)← AG(Oi) ∪HS(Oi) 11: for all Oj ∈ AG(Oi) do 12: AG(Oj)← AG(Oj) ∪ {Oi} // Enforce symmetry 13: end for 14: end if 15: end for 16: for all Oi ∈ D do 17: Add AG(Oi) to equivalence classes with respect to QID(Oi) using AGtoEC algorithm 18: end for 19: for all t ∈ T do 20: Generalize equivalence class EC(t) using lubSpace algorithm 21: end for graph. So for every edge (Oi, Aj), there is edge (Oj , Ai) in the attack graph. Hence, by Definition 13, the induced attack graph is symmetric. 2 Example 14 This example discusses Symmetric Anonymization algorithm and its application on the MOD from our running example. So given MOD D from Table 1.3 and a list of QIDs Q from Table 2.2, we can compute the k-anonymization of D, where 1 ≤ k ≤ n. For k = 2, Symmetric Anonymization algorithm produces the same collection of equivalence classes as Extreme Union algorithm. The resulting generalized database and the collection of corresponding equivalence classes are given by Table 4.1. In case of k = 3, applying Symmetric Anonymization algorithm on the in- put database D results in average information loss of 0.71247024, which is 59 4.4. Symmetric Anonymization ID t1 t2 t3 t4 O1 (0, 0) [(0, 2), (5, 7)] [(0, 1), (7, 7)] (2, 7) O2 [(0, 3), (6,7)] [(0, 2), (5, 7)] [(0, 1), (7, 7)] [(5, 0), (7, 4)] O3 (0, 1) [(0, 2), (5, 7)] [(0, 1), (7, 7)] (3, 7) O4 [(0, 3), (6, 7)] [(0, 2), (5, 7)] [(0, 1), (7, 7)] [(5, 0), (7, 4)] O5 [(0, 3), (6,7)] [(0, 2), (5, 7)] (7, 7) (6, 3) O6 [(0, 3), (6, 7)] [(0, 2), (5, 7)] [(0, 1), (7, 7)] [(5, 0), (7, 4)] EC(ti) [O2, O4, O5, O6] [O1, O2, O3, [O1, O2, O3, [O2, O4, O6] O4, O5, O6] O4, O6] Table 4.3: Application of Symmetric or Restricted Symmetric Anonymi- zation on Running Example (k = 3) smaller than the average information loss corresponding to applying Ex- treme Union algorithm on the same database for the same value of k. The reason for decrease in the (average) information loss is that the equivalence classes at time stamps t3 and t4 produced by Symmetric Anonymization algorithm are smaller (in terms of the number of MOBs) than the equiv- alence classes produced by Extreme Union algorithm. Both algorithms compute AG(O2) = {O2, O4, O5} and AG(O4) = {O2, O4, O6}. In Sym- metric Anonymization algorithm, AG(O2) is generalized with respect to QID(O2) = {t1, t2}. Whereas in Extreme Union algorithm, AG(O2) is gen- eralized with respect to QID(O2) ∪ QID(O4) ∪ QID(O5) = {t1, t2, t3, t4}. Similarly, AG(O4) is generalized with respect to QID(O4) = {t1, t3, t4} in Symmetric Anonymization and with respect to to QID(O2) ∪ QID(O4) ∪ QID(O6) = {t1, t2, t3, t4} in Extreme Union. Then, in case of Symmet- ric Anonymization, AG(O2) is not generalized with respect to time stamp t4, and so O5 is not in any equivalence class at time t4. However in Ex- treme Union algorithm, AG(O2) and AG(O4) are merged into the same equivalence class with respect to time stamp t4, which produces a larger equivalence class. Similarly, the equivalence class at time stamp t3 created by Extreme Union algorithm is larger than the equivalence class created by Symmetric Anonymization algorithm. It should be noted that the size of an anonymity group in Extreme Union algorithm is fixed, i.e. |AG(O)| = k for all MOBs. However in Symmetric Anonymization algorithm, the size of an anonymity group is in the interval [k, n]. For instance, suppose Symmetric Anonymization algo- rithm anonymizes MOBs in the following order: O1, O2, · · · , Os−1, Os, where 60 4.5. Restricted Symmetric Anonymization k < s ≤ n. It is possible that Os ∈ AG(Oi) for all i ∈ {1, 2, · · · , s−1} and by symmetry, Oi will be included in AG(Os). So at the time the algorithm picks Os to be anonymized, the anonymity group of Os contains all MOBs that in- clude Os in their anonymity groups, i.e. AG(Os) = {O1, O2, · · · , Os−1, Os} and |AG(Os)| = s > k. In the worst case, s = n k. To counteract the drawback of Symmetric Anonymization algorithm associated with large anonymity groups, we propose Restricted Symmetric Anonymization algo- rithm presented in Section 4.5. In Restricted Symmetric Anonymization algorithm, we ensure that when |AG(Oi)| ≥ k, no MOB Oj is allowed to include Oi in AG(Oj), provided there are enough other MOBs to construct an anonymty group of size at least k. 4.5 Restricted Symmetric Anonymization This section discusses a modified version of SA algorithm, which we call Re- stricted Symmetric Anonymization (RSA). As we observed in Section 4.4, Symmetric Anonymization algorithm can result in anonymity groups whose size is much larger than the anonymity threshold k. Restricted Symmet- ric Anonymization algorithm is designed to control the sizes of anonymity groups. RSA algorithm computes anonymity groups of size k and enforces symmetry by requiring that if Oi ∈ AG(Oj), then Oj must be in the anonymity group of Oi. So RSA follows the same strategy as SA to ensure that the induced attack graph is symmetric. In addition, Restricted Sym- metric Anonymization algorithm keeps track of MOBs that already have at least k − 1 other MOBs to be anonymized with. That is, the algorithm stores set processedMobSet of MOBs O such that |AG(O)| ≥ k. To com- pute anonymity groups, RSA algorithm applies GenerAG algorithm only on those MOBs that are not on processedMobSet. In other words, RSA disal- lows MOBs to include any MOB Oi ∈ processedMobSet in their anonymity groups. Algorithm 7 presents the pseudo-code of Restricted Symmetric Anonymization algorithm. Theorem 4 (Correctness of Restricted Symmetric Anonymization) Given MOD D as input, the generalized database D′ computed by Restricted Symmetric Anonymization algorithm is a k-anonymous version of D. Proof : Restricted Symmetric Anonymization algorithm adds MOB O to the processed set if and only if |AG(O)| ≥ k. Therefore in the generalized database, every anonymized counterpart is indistinguishable from at least k − 1 MOBs, which means that in the induced attack graph, every vertex 61 4.5. Restricted Symmetric Anonymization Algorithm 7 Restricted Symmetric Anonymization Input: D,Q, and anonymity threshold k Output: D′ = k-anonymous version of D 1: D← {Oi | Oi ∈ D, QID(Oi) 6= ∅} 2: T← {t | t ∈ QID(O), QID(O) ∈ Q} 3: processedMobSet← ∅ 4: for all Oi ∈ D do 5: AG(Oi)← {Oi} // Anonymity group initialization 6: end for 7: for all Oi ∈ D do 8: k̃ ← k − |AG(Oi)| 9: if k̃ > 0 then 10: if (|D| − |processedMobSet|) < k then 11: processedMobSet← ∅ // Clear the set of processed MOBs 12: end if 13: D̃← D− processedMobSet 14: HS(Oi) ← Compute k̃ nearest neighbors of Oi with respect to QID(Oi) using GenerAG algorithm choosing only MOBs from set D̃ as neighbors 15: AG(Oi)← AG(Oi) ∪HS(Oi) 16: for all Oj ∈ AG(Oi) do 17: AG(Oj)← AG(Oj) ∪ {Oi} // Enforce symmetry 18: if |AG(Oj)| ≥ k then 19: processedMobSet← processedMobSet ∪ {Oj} 20: end if 21: end for 22: end if 23: end for 24: for all Oi ∈ D do 25: Add AG(Oi) to equivalence classes with respect to QID(Oi) using AGtoEC algorithm 26: end for 27: for all t ∈ T do 28: Generalize equivalence class EC(t) using lubSpace algorithm 29: end for corresponding to an anonymized counterpart has at least k edges incident with it. After computing anonymity group of MOB Oi, if Oj ∈ AG(Oi), then RSA algorithm adds Oi to AG(Oj). So even if Oi is added to the 62 4.5. Restricted Symmetric Anonymization processed set right after computing AG(Oi), Oi is already in AG(Oj), and hence, there will be edges (Oi, Aj) and (Oj , Ai) in the induced attack graph, where Ai and Aj are the anonymized counterparts of Oi and Oj, respectively. This shows that the attack graph associated with the generalized database produced by Restricted Symmetric Anonymization algorithm satisfies the conditions of k-anonymity of Moving Objects. 2 We proved Theorem 4 to show that RSA algorithm satisfies the k- anonymity model defined in Section 4.2. It was expected that Restricted Symmetric Anonymization (RSA) will generate a k-anonymous version of an input database, since RSA is a variation of SA algorithm and we showed in Theorem 3 that SA produces a k-anonymous generalized database. The following example reviews the application of RSA algorithm on the running example. Example 15 If we apply Restricted Symmetric Anonymization algorithm on our running example with the input MOD and the QID list given by Table 1.3 and Table 2.2, respectively, then we obtain the same collection of equivalence classes as by applying Symmetric Anonymization algorithm on the example. That is, the generalized databases and the associated collec- tions of equivalence classes corresponding to application of RSA algorithm on the running example for k = 2 and k = 3 are shown in Table 4.1 and Table 4.3, respectively. RSA algorithm produces anonymity groups, whose size is less or equal to the size of anonymity groups produced by SA algorithm. Nevertheless, it is not true that RSA would always result in smaller information loss than SA. For example, we suppose Oj is anonymized after Oi, |AG(Oi)| ≥ k and Oj /∈ AG(Oi). So Oi is added to the processed set. It is possible that Oi is relatively close to Oj (in space) at all time stamp in QID(Oj). However in RSA algorithm, Oj will not be able to include Oi in AG(Oj) and so Oj might have to find another neighbor MOB Oh to ensure that |AG(Oj)| ≥ k. But, the distance between Oj and Oh can be much greater than the distance between Oj and Oi. Thus, even though RSA ensures that the size of the resulting anonymity group of Oi is kept relatively small, the anonymity groups of Oj can lead to significant information loss. Section 4.6 discusses the complexity of the anonymization algorithms. We set up and performed several experiments to evaluate our anonymization algorithms with respect to speed and information loss among other measures. Chapter 5 discusses these experiments in detail. 63 4.6. Algorithm Complexity Analysis 4.6 Algorithm Complexity Analysis Our approach to computing anonymization of a MOD proceeds in three stages: 1. Anonymity group computation: For all MOBs in a given database, we compute an anonymity group of size at least k. In this stage we use GenerAG algorithm. 2. Transformation of anonymity group to equivalence classes: For all MOBs in a given database, we add each MOB’s anonymity group to equivalence classes with respect to a particular time stamp set. In this stage we use AGtoEC algorithm. 3. Generalization of certain positions: For all time stamps and all MOBs in a given database, we set some positions to generalized re- gions, while other positions remain unmodified. In this stage we use lubSpace algorithm. Let T (GenerAG), T (AGtoEC), and T (lubSpace) denote the time com- plexities of algorithms GenerAG, AGtoEC, and lubSpace, respectively. Sup- pose A is one of the anonymization algorithms Extreme Union, Symmet- ric Anonymization, or Restricted Symmetric Anonymization, then the time complexity of A is given by T (A): T (A) = n · T (GenerAG) + n · T (AGtoEC) +m · T (lubSpace) where n is the number of distinct MOBs andm is the number of distinct time stamps in a given MOD. As already discussed in Section 3.1.2, GenerAG algorithm (Algorithm 2) is instance optimal with respect to the number of accesses on any input instance of Hilbert lists. However in the worst possible scenario, the algorithm may access every data item in each list, and so the worst case time complexity of GenerAG algorithm is in O(M ·N), where M is the number of lists and N is the number of distinct data items in each list. We note that the upper bound on the time complexity of GenerAG algorithm is very pessimistic. In majority of cases, the algorithm accesses only a small portion of all data items. The complexity of AGtoEC algorithm (Algorithm 3) is directly related to the data structure used to implement equivalence classes. As mentioned in Section 3.2, we employ the disjoint set forests data structure based on Union-Find algorithm with union-by-rank and path compression to repre- sent equivalence classes. Let s denote the size of anonymity group AG(O) for 64 4.6. Algorithm Complexity Analysis Algorithm Worst case time complexity GenerAG O(m · n) AGtoEC O(s · α(n)) lubSpace O(n) Table 4.4: Time Complexity of Each Stage of Anonymization Algorithm some MOB O, i.e. s = |AG(O)|. In the worst case, AGtoEC algorithm per- forms s FIND operations and s−1 UNION operations. Using union-by-rank and path compressions techniques together, the amortized time complexity of each FIND and UNION operation is in O(α(n)), which has been shown to be the optimal complexity per operation on average [11]. α(n) denotes the inverse Ackermann function [2] that grows very slowly such that α(n) < 5 for most feasible inputs of size n. Therefore, T (AGtoEC) ∈ O(s · α(n)). In lubSpace algorithm (Algorithm 4), we compute the least upper bound for a given equivalence class. In the worst case, an equivalence class at time stamp t is of size n. To carry out the space generalization, we traverse the positions of all MOBs in a particular equivalence class twice. First, we find the minimal and maximal x and y coordinates of the encapsulating region Rlub. Second, we replace all the MOB positions in a given equivalence class by Rlub. This can be done in O(n) time. Hence, T (lubSpace) ∈ O(n). Table 4.4 summarizes the worst case time complexities of the three algo- rithms that our anonymization algorithms rely on. The analysis shows that the time complexity of each of the anonymization algorithms presented in this chapter is given by: T (A) = n · T (GenerAG) + n · T (AGtoEC) +m · T (lubSpace) T (A) = O(n2 ·m) +O(n · s · α(n)) +O(m · n) We defined s = |AG(O)| for some MOB O. In EU algorithm, the largest size of an anonymity group equals k. So in case of EU algorithm, s = k. However in case of SA algorithm, s ∈ [k, n], i.e. theoretically in the worst case, s = n. Whereas in RSA algorithm, s can be larger than k, but is expected to be relatively close to k. In our experiments, we observed that for SA algorithm, s ∈ [k, β · k], where β is a small integer, and for RSA algorithm, s ∈ [k, k + 1]. The next section presents a summary of the anonymization algorithms introduced in this chapter. 65 4.7. Anonymization Algorithms: Summary 4.7 Anonymization Algorithms: Summary In this chapter, we introduced a possible attack model, which can compro- mise privacy of one or more objects in the published MOD. We developed a robust definition of k-anonymity model for MOBs based on the notion of at- tack graph and proved that the model is impervious to the proposed attack strategy. We designed anonymization algorithms (Extreme Union (EU), Symmetric Anonymization (SA), and Restricted Symmetric Anonymization (RSA)) that satisfy the conditions of the k-anonymity model and hence, ensure that the generalized database produced by these algorithms does not admit privacy breach based on our attack model. Extreme Union algorithm achieves k-anonymity by creating anonymity groups of size k and generalizing each group with respect to time stamp union of QIDs of all MOBs in a particular anonymity group. The union of QIDs can lead to large (in the number of MOBs) equivalence classes. Every MOB O has to be generalized with respect to all time stamp in QID(O), and hence, QID(O) is the smallest time stamp set, at which O must be generalized. So we designed Symmetric Anonymization algorithm to achieve k-anonymity, while generalizing each anonymity group with re- spect to the subject MOB’s QID. SA algorithm ensures that the resulting attack graph is symmetric by enforcing the following condition among the anonymity groups: If Oj ∈ AG(Oi), then Oi ∈ AG(Oj). However, the en- forcement of this condition can result in anonymity groups, whose size is in the interval [k, n], where n is the number of distinct MOBs in the input database. The larger an anonymity groups is, the more information will be lost in the anonymization process. To eliminate this disadvantage of produc- ing large anonymity groups, we modified SA algorithm to obtain Restricted Symmetric Anonymization algorithm, which controls the size of generated anonymity groups. When |AG(O)| ≥ k, RSA algorithm adds MOB O to the set of processed MOBs and disallows any unprocessed MOB (i.e. MOB whose anonymity group is of size less than k) to include O in its anonymity group. We ran an extensive set of experiments to demonstrate the effectiveness of the aforementioned algorithms and to compare their relative performance with respect to each other. Chapter 5 discusses the experimental settings and the experiment results. 66 Chapter 5 Experiments In this chapter, we discuss the experimental settings and present a set of ex- periments to test our algorithms on large real-world and synthetic databases. 5.1 Experimental Settings We implemented all of our anonymization algorithms using Java 6.0. We ran our experiments on a machine with Intel Pentium 4 processor (3.00GHz), 3 gigabytes (GB) of main memory, and GNU/Linux - openSUSE 10.3 oper- ating system. Although the size of main memory was 3 GB, Java Virtual Machine could only allocate 2686 megabytes (MB) of virtual memory. We experimented with two Moving Objects Databases, one of which was synthetically generated, while the other was acquired in real-world settings. Section 5.1.1 discusses the synthetic data set, whereas Section 5.1.2 provides a detailed overview of the real-world data set. The expected format of an input data set is an ASCII file containing four columns separated by TAB. The four columns correspond to: 1. Moving Object ID (integer) 2. Time stamp (integer) 3. x-coordinate (float) 4. y-coordinate (float) Entries in a data set do not have to appear in any particular order. We assume that all MOBs in a data set move for the duration of time interval [tmin, tmax], where tmin and tmax are the minimal and the maximal time stamps in the given data set, respectively. Therefore, after eliminating null entries from a data set, we obtain another set, where every trajectory is of size m, where m is the number of distinct time stamps in the data set. We did not have lists of QIDs for the MOBs in our experimental data sets. So we have implemented a QID generator as discussed in Section 2.2.2. For each database, we constructed a list of QID instances of different sizes. 67 5.1. Experimental Settings We varied the size of the largest QID in each set. We formed the QIDs according to the random graph approach (see Algorithm 1), where each distinct QID contains a randomly chosen set of time stamps and corresponds to a vertex in a graph. There is an edge between two vertices if and only if the corresponding QIDs overlap on one or more time stamps. To compute range query distortion, we randomly chose 100 different time stamps. For each time stamp, we generated 100 different rectangular regions. Each region was inside the xy-space, where the MOBs move. So we averaged the range query distortion over 100 · 100 = 10000 values. To compute anonymity groups, we must index the MOB positions using Hilbert space-filling curve. For each database, we need to generate the lists of Hilbert indexes only once. This is considered as a preprocessing step. 5.1.1 Oldenburg Database In our experiments, we used two data sets. The first data set is synthetic and it was generated using Brinkhoff’s network-based generator of Moving Objects [7]. In the absence of real-world data, the Brinkhoff’s network-based generator is an excellent software for obtaining trajectories of objects that mimic the behavior of real MOBs. The Brinkhoff’s generator combines real data (given by the road network of a real city) with user-specified properties of the generated spatio-temporal database. In fact, the generator tries to create realistic settings, where each object has a starting and a final destina- tion positions, and computes an adequate route between the two positions. In addition, the generator simulates road and traffic conditions in such a way that speed and path of a MOB is influenced by other MOBs in the network. Even though we were able to obtain real-world data (see Section 5.1.2), we still chose to use the Brinkhoff’s generator to create synthetic data set in order to investigate the performance of our anonymization algorithms on a large set of trajectories. The data set, we created with the generator, contains 150, 000 trajectories with 400 distinct time stamps over the road network of the city of Oldenburg (Germany). We will refer to this database as Oldenburg. The generation of Hilbert index lists for Oldenburg set takes approxi- mately 60 minutes and requires about 480 megabytes of memory. 68 5.1. Experimental Settings D n m x-range y-range t-range Oldenburg 150000 400 [281, 23854] [3935, 30851] [0, 399] Milan 45639 2009 [4.537E7, [9050000, [7776300, 4.556E7] 9279999] 8380500] Table 5.1: Databases Used in the Experiments D Non-null radius(D) Maximal Average positions length step Oldenburg 16199628 17889.6 400 125.7 Milan 4729613 149164.0 287 2541.9 Table 5.2: Characteristics of Oldenburg and Milan Databases 5.1.2 Milan Database The second Moving Objects Database used in our experiments is a real- world set that contains trajectories of cars moving in the city of Milan (Italy). Comune di Milano (Municipality of Milan) provided these data to be studied in the realm of GeoPKDD project [14]. These data is not freely available and was obtained by us in the collaboration with GeoPKDD project when we worked on [43]. From this point on, we will refer to this database as Milan. After preprocessing, the data set contains over 45 thousand trajectories and a total of approximately 500 thousand non-null spatio-temporal points. These trajectories were obtained by GPS-equipped cars moving in the city of Milan starting from January 2007. During the pre-processing, the data collectors enforced a constant sampling rate of one point every 5 minutes. The collectors cut trajectories when two consecutive observations were too far in time or in space. Moreover, the trajectories were cut everyday at 3 AM, to focus on a single day. At the end, the data set contains 2009 different time stamps. Finally, the collectors removed completely noisy trajectories such as trajectories with too small temporal or spatial extension. The generation of Hilbert index lists for Milan set takes approximately 85 minutes and requires about 915 megabytes of memory. 5.1.3 Database Characteristics This section presents some metadata about the two databases used in our experiments. Table 5.1 shows the following parameters associated with Old- enburg and Milan databases. n and m are the number of distinct MOB 69 5.2. Description of Experiments IDs and the number of distinct time stamps in an input database, respec- tively. The area in which the objects move is specified by x and y-ranges. All the objects are assumed to move in the time interval given by t-range. In Table 5.2, we report the following characteristics of the two experimental databases. Non-null positions column shows the total number of spatio- temporal points that are not null in the original input database. radius(D) represents the half of the length of the diagonal of the minimal bounding box of the area, where the objects move. That is, if the x-range is given by the interval [xmin, xmax] and y-range is given by the interval [ymin, ymax], then: radius(D) = 1 2 √ (xmax − xmin)2 + (ymax − ymin)2 (5.1) Maximal length corresponds to the maximum number of spatio-temporal points in one trajectory, whereas average step presents the average distance between two points at two consecutive time stamps. In addition to the empirical differences of the two databases shown in Tables 5.1 and 5.2, we noticed a structural difference between them. In Sec- tion 2.1.1, we identified three categories of null entries: leading, trailing, and gap null entries. Both Oldenburg and Milan contain leading and trailing null entries. However, we observed that unlike Milan set, Olden- burg data set does not contain gap null entries. So in Oldenburg, once an object appears in the road network, its position is known at every time stamp until the object reaches its final destination or the maximum time elapses. 5.2 Description of Experiments In the section, we describe three sets of experiments that we performed to assess the efficiency and quality of our anonymization algorithms. We outline the parameters and their values involved in each set of experiments. We explain what parameters will be held constant and what parameter will vary in each experiment. We compare our algorithms both in terms of efficiency and efficacy. We base our empirical evaluation of the algorithms on the set of metrics de- scribed in Table 5.3. The runtime and memory peak metrics assess the efficiency of the algo- rithms. Whereas average information loss and range query distortion are measures of the quality of the resulting anonymized data. In addition, we consider the following distributions in order to better understand the quality of anonymization. First, we consider the distribution of sizes of anonymity 70 5.2. Description of Experiments Metric Description Runtime Runtime metric represents the running time of the anonymization algorithm (in minutes) Memory Peak This metric shows the largest amount of main mem- ory (in megabytes) allocated by an algorithm Average This metric is defined by Equation 2.3 in Information Loss Section 2.1.3 Range Query The distortion is measured to assess the usability Distortion of the published data. It is defined by Equation 2.4 and Equation 2.5 in Section 2.1.4 Table 5.3: Metrics for Algorithm Evaluation groups. In case of Extreme Union algorithm, every anonymity group is of size exactly k. Therefore, this distribution does not convey any interesting information about the EU algorithm. However, this distribution can pro- vided some insights into the differences of Symmetric Anonymization and Restricted Symmetric Anonymization algorithms. The second distribution we consider is the distribution of the sizes of equivalence classes. This dis- tribution can shed some light on the differences between the algorithms and the results they produce. For each distribution we record the minimal and maximal values, mean and median. And for the distribution of equivalence class sizes, we also look at the coverage. We define coverage as follows, where k denotes the anonymity threshold. Definition 15 (Coverage of Equivalence Classes) Let Ns denote the number of equivalence classes, whose size is in the interval [k, 2k − 1]. Let Nt denote the total number of equivalence classes. Then we define: Coverage = Ns Nt In relational microdata, if an anonymity group has size of at least 2k, then it could be divided further into two or more groups, which will yield less information loss. However in MOD settings, an equivalence class of size 2k or more cannot be split into subclasses as the resulting database may not be k-anonymous according to our definition. Thus, it is interesting to compute the percentage of equivalence classes that have size strictly less than 2k. Coverage · 100% gives us this percentage. Intuitively, the closer the coverage of equivalence classes is to 1, the better the anonymization. 71 5.2. Description of Experiments Experiment Settings Parameter Value 1 Constant Database Milan Max |QID| 200 Variable k {2, 4, 8, 16, 32} 2 Constant Database Milan k 16 Variable Max |QID| {50, 100, 150, 200} 3 Constant Database Oldenburg k 16 Max |QID| 40 Variable n {25K, 50K, 100K, 150K} Table 5.4: Settings of Experiments Next, we review the setup of the experiments we performed to evaluate our anonymization algorithms: EU, SA, and RSA. Table 5.4 summarizes the settings of each set of experiments we performed. Experiment 1 In our first set of experiments, we use the entire Milan data set. We gener- ate a list of QIDs for all MOBs in Milan. The largest QID contains 200 dis- tinct time stamps, i.e. maximal QID size equals 200, which is approximately 10% of all time stamps. In this set of experiments, we wanted to study the effect of anonymity threshold k on the set of metrics outlined in Table 5.3. In this set of experiments, we test the algorithms for k ∈ {2, 4, 8, 16, 32}. Experiment 2 In this experiment set, we look at the influence of the maximal QID size on the efficiency and quality of the anonymization algorithm. We ran the experiments on the entire Milan data set. We set the anonymity threshold k = 16 and held it constant. We generated several instances of QID sets and varied the maximal QID size in these lists between 50 and 200 time stamps. Experiment 3 For this set of experiments, we took the Oldenburg data set with 150, 000 distinct MOBs and created subsets of 25, 000, 50, 000, and 100, 000 MOBs. To create a subset of size ñ, we simply extracted the first ñ trajectories from 72 5.3. Results of Experiments Oldenburg. The number of distinct time stamps was 400 in Oldenburg and its subsets. For each of the four data sets (i.e. Oldenburg and its three subsets), we generated an instance of QID list with maximal QID size of 40 time stamps. We set k = 16. We applied our algorithms to the four data set for this value of k using the corresponding instances of QID sets. Our goal is to analyze the effect of database size (i.e. number of distinct MOBs) on the proposed metrics (see Table 5.3). 5.3 Results of Experiments In this section, we present the results of the sets of experiments described in Section 5.2. For each set of experiments, we present empirical values of the considered metrics. For Experiment 1, we present the distributions of anonymity group and equivalence class sizes. The distribution of anonymity group sizes follows the same pattern in the other sets of experiments, and does not convey any additional information. For Experiment 2, we included the distribution of the sizes of equivalence classes. For Experiment 3, we report the values of the main metrics from Table 5.3. Section 5.4 compares and contrasts the performance of the algorithms in details. Experiment 1: Results In Table 5.5, we show the empirical results of this set of experiments. Fig- ure 5.1 shows graphs corresponding to the aforementioned results from Table 5.5. Tables 5.6 and 5.7 show the statistics of the distributions of sizes of anonymity groups and equivalence classes, respectively. Extreme Union (EU) algorithm produces anonymity groups of fixed size equal to k. Therefore, the distribution of anonymity group sizes for EU algorithm has all statistical values from Table 5.6 equal to k. As expected, all considered metrics increase as k increases. It should be noted that in this set of experiments, k increases exponentially. From Figure 5.1, we can see that the main metrics increase according to f(x) = a √ x distribution, where a ∈ N. This implies that the metrics from Table 5.3 increase linearly as k increases linearly. In terms of quality (i.e. information loss and range query distortion), RSA algorithm outperforms the other two algorithms, namely, Extreme Union (EU) and Symmetric Anonymization (SA), although SA is very close to RSA with respect to possibly inside range query distortion. For k = 8, the memory peak of EU algorithm is higher than the memory peaks for k > 8. This occurs because for k > 8, EU may run out of memory, and therefore, 73 5.3. Results of Experiments Algo- k Run- Memory Average Possibly Definitely rithm time Peak Info. Inside Inside (min) (MB) Loss Distortion Distortion EU 2 943 1785 0.107934 0.219966 0.079098 4 1135 2116 0.289367 0.623322 0.306324 8 1380 2645 0.547444 0.733918 0.601555 16 2224 2549 0.715077 0.761985 0.801663 32 2389 2573 0.760656 0.768059 0.861070 SA 2 676 1675 0.092055 0.168581 0.059784 4 917 1838 0.171128 0.387498 0.159872 8 1121 2104 0.293229 0.616752 0.313656 16 1289 2304 0.450169 0.710679 0.494947 32 1504 2652 0.602970 0.745192 0.672137 RSA 2 521 1702 0.079231 0.136118 0.047526 4 756 1794 0.145121 0.343187 0.131846 8 1083 1969 0.249257 0.587681 0.265895 16 1447 2242 0.388484 0.689527 0.426159 32 1767 2478 0.533165 0.732049 0.590570 Table 5.5: Experiment 1: Main Metrics we need to save the equivalence classes from main memory to secondary storage. In this case, the memory peak occurs when we store all equivalence classes in main memory. Whereas for k ≤ 8, the memory peak occurs when we store all equivalence classes in memory and also generate the generalized positions, which are also stored in main memory. Experiment 2: Results In this set of experiments, we varied the maximal number of distinct time stamps per QID. The results of the experiments are shown in Tables 5.8 and 5.9. The former table presents the values of the metrics from Ta- ble 5.3. Whereas, the latter table shows the distribution of sizes of equiva- lence classes. The plots in Figure 5.2 are based on the experimental results from Table 5.8. In Figures 5.2(a) and 5.2(b) for maximal QID size of 50 and 100 time stamps, we can see a common trade-off between running time and memory. EU algorithm takes more than twice the time of SA and RSA algorithms to complete the anonymization task. Yet, EU algorithm requires less memory as equivalence classes are written onto hard disk. 74 5.3. Results of Experiments Algo- k Min Max Median Mean rithm Value Value SA 2 2 10 2 2.39 4 4 20 4 4.91 8 8 30 9 9.92 16 16 54 18 19.89 32 32 107 37 39.83 RSA 2 2 3 2 2.00 4 4 5 4 4.00 8 8 9 8 8.00 16 16 17 16 16.00 32 32 33 32 32.00 Table 5.6: Experiment 1: Distribution of Anonymity Group Size Algo- k Min Max Median Mean Coverage rithm Value Value EU 2 2 32 3 3.06 0.742145 4 4 35687 7 52.83 0.517593 8 8 44496 15 7960.17 0.512603 16 16 45591 44060 41813.57 0.011789 32 42377 45639 45564 45292.66 0 SA 2 2 15 2 2.44 0.907658 4 4 16401 5 7.06 0.761555 8 8 37882 10 57.82 0.721696 16 16 44145 19 1209.16 0.754199 32 32 45569 33158 23153.10 0.334968 RSA 2 2 3 2 2.00 1 4 4 861 4 5.20 0.906942 8 8 35073 8 38.88 0.831855 16 16 42872 16 715.49 0.824413 32 32 45331 52 12187.51 0.542073 Table 5.7: Experiment 1: Distribution of Equivalence Class Size 75 5.3. Results of Experiments Algo- Max Run- Memory Average Possibly Definitely rithm QID time Peak Info. Inside Inside size (min) (MB) Loss Distortion Distortion EU 50 539 1308 0.365735 0.399888 0.401627 100 1027 1324 0.529808 0.620337 0.641757 150 2053 2358 0.666969 0.724339 0.740056 200 2224 2549 0.715077 0.761985 0.801663 SA 50 226 1426 0.159555 0.349773 0.169128 100 505 1795 0.259747 0.562097 0.332788 150 1122 2151 0.389801 0.662193 0.417038 200 1289 2304 0.450169 0.710679 0.494947 RSA 50 203 1380 0.132335 0.330887 0.140364 100 421 1734 0.219215 0.537296 0.281970 150 839 2123 0.330105 0.635040 0.350911 200 1447 2242 0.388484 0.689527 0.426159 Table 5.8: Experiment 2: Main Metrics Algo- Max Min Max Median Mean Coverage rithm |QID| Value Value EU 50 33849 45458 38344 39921.01 0 100 16 45566 38050 40344.18 0.009047 150 30795 45590 42755 41475.67 0 200 16 45591 44060 41813.57 0.011789 SA 50 16 36620 19 429.79 0.794998 100 16 40498 18 537.49 0.788560 150 16 43735 19 829.00 0.758994 200 16 44145 19 1209.16 0.754199 RSA 50 16 31910 16 135.91 0.801143 100 16 38538 16 263.76 0.816348 150 16 42316 16 423.22 0.805704 200 16 42872 16 715.49 0.824413 Table 5.9: Experiment 2: Distribution of Equivalence Class Size 76 5.3. Results of Experiments 5 10 15 20 25 30 1000 1500 2000 k R un tim e (m in) EU SA RSA (a) Runtime 5 10 15 20 25 30 1800 2000 2200 2400 2600 k M em or y Pe ak (M B) EU SA RSA (b) Memory Peak 5 10 15 20 25 30 0.1 0.2 0.3 0.4 0.5 0.6 0.7 k A ve ra ge In fo rm at io n Lo ss EU SA RSA (c) Information Loss 5 10 15 20 25 30 0.2 0.3 0.4 0.5 0.6 0.7 k Po ss ib ly In sid e D ist or tio n EU SA RSA (d) Possibly Inside 5 10 15 20 25 30 0.2 0.4 0.6 0.8 k D ef in ite ly In sid e D ist or tio n EU SA RSA (e) Definitely Inside Figure 5.1: Results of Experiment 1 (k ∈ {2, 4, 8, 16, 32}) 77 5.3. Results of Experiments 50 100 150 200 500 1000 1500 2000 Maximal QID Size R un tim e (m in) EU SA RSA (a) Runtime 50 100 150 200 1400 1600 1800 2000 2200 2400 Maximal QID Size M em or y Pe ak (M B) EU SA RSA (b) Memory Peak 50 100 150 200 0.2 0.3 0.4 0.5 0.6 0.7 Maximal QID Size A ve ra ge In fo rm at io n Lo ss EU SA RSA (c) Information Loss 50 100 150 200 0.4 0.5 0.6 0.7 Maximal QID Size Po ss ib ly In sid e D ist or tio n EU SA RSA (d) Possibly Inside 50 100 150 200 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Maximal QID Size D ef in ite ly In sid e D ist or tio n EU SA RSA (e) Definitely Inside Figure 5.2: Results of Experiment 2 (|QID| ∈ {50, 100, 150, 200}) 78 5.4. Discussion of Algorithms Algo- n Run- Memory Average Possibly Definitely rithm (103) time Peak Info. Inside Inside (min) (MB) Loss Distortion Distortion EU 25 57 366 0.862822 0.674977 0.778467 50 256 875 0.952506 0.780603 0.922211 100 1214 1805 0.980141 0.792022 0.950513 150 3281 2558 0.985430 0.794133 0.958346 SA 25 19 323 0.546975 0.596153 0.475374 50 63 581 0.591333 0.710145 0.584412 100 319 1222 0.626259 0.735036 0.599374 150 334 1984 0.633931 0.742934 0.612997 RSA 25 17 354 0.494436 0.563435 0.425671 50 58 527 0.531178 0.686755 0.522739 100 235 1147 0.564743 0.717647 0.539318 150 551 1684 0.564018 0.727360 0.544735 Table 5.10: Experiment 3: Main Metrics Experiment 3: Results In this set of experiments, we varied the number of distinct MOBs per data set. The results of this set of experiments are shown in Tables 5.10. This table presents the values of the metrics from Table 5.3 after applying our algorithms on Oldenburg database and its subsets. In this set of experiments, we observed the disadvantage of EU approach as the quality of the resulting data is very low even for small data set size of 25000 MOBs. Again, RSA algorithm had the best performance among the three algorithm in terms of memory requirement and quality of the output database. 5.4 Discussion of Algorithms In this section, we discuss the performance of the three anonymization al- gorithms presented in Chapter 4. The anonymization process takes a lot of time; sometimes more than one day (i.e. 1440 minutes). Computation of anonymity groups takes approximately 90 − 95% of the time of the entire anonymization process. To generate an anonymity group, we compute the exact top-K results (see Section 3.1.2) in order to find MOBs, which are the nearest neighbors of the subject MOB. Alternatively, we could compute the approximate top-K nearest neighbors, which is likely to improve the 79 5.4. Discussion of Algorithms runtime of an anonymization algorithm. However, we could lose more infor- mation, since the approximate top-K neighbors can be further away from a subject MOB than the exact top-K neighbors. Using an approximate top- K algorithm instead of an exact top-K algorithm is one potential research direction we consider to explore in the future. When transforming an anonymity group into equivalence classes with respect to a particular time stamp set, we need to have access to all equiva- lence classes created so far, because the anonymity group can add MOBs to any of these equivalence classes. So we store the equivalence classes in main memory when running AGtoEC algorithm (Algorithm 3). After transform- ing all anonymity groups and storing all equivalence classes in main memory, we start the generalization process, which is the point where the memory peak occurs. This is also the point where our algorithm implementation can run out of memory. In our experiments, Extreme Union algorithm was running out of memory for k > 8. Therefore in Extreme Union algorithm, we wrote the equivalence classes onto secondary storage (i.e. hard drive), and then read the equivalence classes one by one during the generalization process. This approach allows Extreme Union algorithm to finish, although it drastically affects the performance of the algorithm as writing equivalence classes on a disk can take over 9 hours. We did not have to write equivalence classes onto a disk in case of SA and RSA algorithms, which explains the notable difference of the running time of the two algorithms compared with the running time of EU algorithm. Another reason why SA and RSA algorithms have smaller running time than EU algorithm is because of the approach they follow to enforce sym- metry of the induced attack graph. In SA (and RSA), if we process MOBs O1, · · · , Oi, Oj in some order, then it is possible that by the time we pick Oj as the subject MOB, the anonymity group of Oj contains k or more MOBs, i.e. |AG(Oj)| ≥ k. And so we do not have to run GenerAG top-K algorithm to compute the anonymity group of Oj, which clearly saves time. The experiments we conducted clearly show that the Extreme Union ap- proach performs much worse than the Symmetric Anonymization and Re- stricted Symmetric Anonymization approaches, both in terms of efficiency and efficacy. Extreme Union is always slower and in the majority of cases, it has larger memory requirements. The main difference of the three al- gorithms is manifested in the quality tests, where Restricted Symmetric Anonymization performs better both in terms of information loss and range query distortion than SA, which in turn, performs better than EU on these evaluation metrics. That is, RSA approach yields smaller information loss and results in less distortion than the other two approaches. In Table 5.5 for 80 5.4. Discussion of Algorithms k = 4 and 8, EU algorithm results in information loss that is approximately two times larger then the corresponding information loss associated with RSA algorithm. An analysis of all three anonymization algorithms on the real-world database is summarized in Tables 5.5 - 5.9, where the effects of (i) the anonymity threshold k and (ii) the QID size are studied. The first ob- servation is that the median of the distribution of equivalence class sizes increases quickly with k, while it seems not to be influenced by the size of the largest QID. Second, the information loss grows with larger k and larger QIDs. We have expected such results as larger k and larger QIDs clearly make for a more difficult anonymization task. Range query distortion and coverage of equivalence class sizes confirm these trends. This observation is also reflected in runtime and memory consumption. However, we can ob- serve that the maximal size of a quasi-identifier has a larger impact on the runtime performances. While running time grows linearly with k, it grows super-linearly with the increase of the maximal size of QID. We can see in Figures 5.2(a) and 5.2(b) that running time and memory peak increase super-linearly with the maximal QID size. Whereas Figures 5.2(c)–5.2(e) shows that the quality of the generalized data degrades sub-linearly as QID size increases. We experimentally discovered that for SA, the quality of the data main- tained after anonymization degrades very slow with the increase of the size of data sets. In other words, SA approach benefits from the larger variety of MOBs available. Similar trend was observed for RSA algorithm. In fact in larger data set, an object Oi has more candidate objects to be anonymized with. So it is more likely that anonymity groups will not overlap among themselves, and consequently, SA will create small-sized equivalence classes. In addition, it is more likely that two or more objects move close to each other and/or follow the same path when the data set contains a larger num- ber of distinct trajectories. That is, in a larger data set, anonymity groups are likely to include MOBs that are densely clustered in space, which would result in smaller information loss. Theoretically, a size of an equivalence class should not have a direct effect on information loss. An equivalence class can contain a lot of MOBs that are densely clustered together, or it can contain a very few MOBs that are dispersed far away from each other. However, our experiments on real-wold Milan database show that smaller information loss and smaller range query distortion correspond to smaller equivalence classes. Tables 5.7 and 5.9 show the statistics of the distribution of equivalence class sizes in Experiments 1 and 2, respectively. RSA algorithm results in smaller equivalence classes. 81 5.4. Discussion of Algorithms That is, the mean and median of the distribution are smaller for RSA algo- rithm than for SA algorithm, which in turn has smaller equivalence classes than EU algorithm. And it was observed that RSA outperforms the other two algorithms with respect to information loss and range query distortion. Thus, we conclude that the size of equivalence classes does have an effect on quality of the anonymization and the desirable property of an anonymization algorithm is to minimize the sizes of the resulting equivalence classes. Of course, it is a challenging and very hard problem to find the order of MOB anonymization that would result in an optimal partitioning of equivalence classes. Two factors affect the sizes of resulting equivalence classes: (i) the sizes of anonymity groups, and (ii) the time stamp sets, with respect to which we add anonymity groups to equivalence classes. The smallest size of an anonymity group equals k (i.e. |AG(Oi)| ≥ k). And each anonymity group must be anonymized with respect to a time stamp set, which is a superset of the QID of the subject MOB (i.e. QID(Oi) ⊆ T imeSet). Ultimately in order to minimize information loss, we want every anonymity group to be of size exactly k and to be anonymized with respect to subject MOB’s quasi- identifier. The experimental results shown in Tables 5.7 and 5.9 support this statement as both SA and RSA add anonymity groups to equivalence classes with respect to QID of a subject MOB, but RSA results in less information loss because its anonymity groups have smaller sizes. As discussed in Sec- tion 4.7, our anonymization algorithms control either the sizes of anonymity groups (i.e. EU algorithm) or the time sets used to transform anonymity groups into disjoint equivalence classes (i.e. SA and RSA algorithms). De- signing an anonymization algorithm that controls both anonymity group sizes and time stamp sets, with respect to which the anonymity groups are generalized, remains an open issue. 82 Chapter 6 Conclusion In this thesis, we considered the problem of preserving location privacy when publishing a Moving Objects Database (MOD). The key assumption in our work is that every Moving Object (MOB) can have a distinct quasi- identifier (QID) and that adversarial knowledge is modeled by this QID. We proposed a possible attack model on the published database. We focused on providing a form of k-anonymity based on space generalization. In this work, we showed the challenges of adapting k-anonymity model from relational settings to MOD settings. We developed a robust notion of k-anonymity and showed that this notion ensures privacy breach (based on the proposed attack model) cannot occur in an anonymized MOD. We have devised a novel approach to computing nearest neighbors using a Top-K algorithm and Hilbert space-filling curve, which is used in the anonymization process to determine an anonymity group of an individual MOB. Based on our notion of k-anonymity in MOD settings, we designed three anonymization algorithms and proved that each of these algorithms satisfies k-anonymity. The three algorithms we developed are Extreme Union (EU), Symmetric Anonymization (SA), and Restricted Symmetric Anonymization (RSA). We have implemented and tested our algorithms on large synthetic and real-world databases. EU algorithm was found to per- form worse than the other two algorithms with respect to the efficiency and efficacy metrics that we considered. Our experiments showed that RSA algo- rithm outperforms both EU and SA algorithms with respect to the quality of the published data. However in a few cases, the Symmetric Anonymization approach had the smallest running time out of the three algorithms. The experimental results have revealed that the performance of the al- gorithms is strongly influenced by the distribution of QIDs. As a future research direction, we intend to investigate the sizes and configurations of QIDs in real world. To better understand the distribution of QIDs, it is possible to conduct a user study by administering a questionnaire. In ad- dition, data mining techniques can be employed to uncover sub-trajectories that can uniquely identify a MOB. Such sub-trajectories can constitute quasi-identifiers of objects. 83 Chapter 6. Conclusion The anonymization algorithms, presented in this thesis, work on a static instance of a database. If one or more trajectories are added to the database, the entire database needs to be anonymized from scratch. We plan to design anonymization techniques that provide efficient support for database up- dates. Another possible issue that can be considered in the future research is discovering stronger and more aggressive attack models and developing privacy-preserving mechanisms against such attack models. Finally, we be- lieve that the presence of outliers in the input MOD deteriorates the overall data quality of the generalized database. In future work, we want to explore ways of detecting and removing outliers from Moving Objects Databases. 84 Bibliography [1] Osman Abul, Francesco Bonchi, and Mirco Nanni. Never walk alone: Uncertainty for anonymity in moving objects databases. In Proceedings of the 24th International Conference on Data Engineering, pages 376– 385, 2008. [2] Wilhelm Ackermann. Zum hilbertschen aufbau der reellen zahlen. Mathematische Annalen, 99(1):118–133, 1928. [3] Gagan Aggarwal, Tomás Feder, Krishnaram Kenthapadi, Rajeev Mot- wani, Rina Panigrahy, Dilys Thomas, and An Zhu. Anonymizing tables. In Proceedings of the 10th International Conference on Database Theory (ICDT’05), pages 246–258, 2005. [4] Reza Akbarinia, Esther Pacitti, and Patrick Valduriez. Best position algorithms for top-k queries. In VLDB ’07: Proceedings of the 33rd international conference on Very large data bases, pages 495–506, 2007. [5] Alastair R. Beresford and Frank Stajano. Mix zones: User privacy in location-aware services. In Proceedings of the 2nd IEEE Annual Conference on Pervasive Computing and Communications Workshops (PERCOMW04), pages 127–131, 2004. [6] Claudio Bettini, X. Sean Wang, and Sushil Jajodia. Protecting privacy against location-based personal identification. In Proceedings of the 2nd VLDB Workshop on Secure Data Management, pages 185–199, 2005. [7] Thomas Brinkhoff. A framework for generating network-based moving objects. Geoinformatica, 6(2):153–180, 2002. [8] Valentina Ciriani, Sabrina De Capitani di Vimercati, Sara Foresti, and Pierangela Samarati. k-anonymity. In Secure Data Management in Decentralized Systems, volume 33 of Advances in Information Security, pages 323–353. 2007. [9] T. Dalenius. Finding a needle in a haystack - or identifying anonymous census record. Journal of Official Statistics, 2(3):329–336, 1986. 85 [10] Ronald Fagin, Amnon Lotem, and Moni Naor. Optimal aggregation algorithms for middleware. Journal of Computer and System Sciences, 66(4):614–656, 2003. [11] M. Fredman and M. Saks. The cell probe complexity of dynamic data structures. In STOC ’89: Proceedings of the twenty-first annual ACM symposium on Theory of computing, pages 345–354, 1989. [12] Benrard A. Galler and Michael J. Fisher. An improved equivalence algorithm. Communications of the ACM, 7(5):301–303, 1964. [13] Bugra Gedik and Ling Liu. Location privacy in mobile systems: A personalized anonymization model. In ICDCS ’05: Proceedings of the 25th IEEE International Conference on Distributed Computing Sys- tems, pages 620–629, 2005. [14] GeoPKDD. Geographic Privacy-aware Knowledge Discovery and De- livery. http://www.geopkdd.eu. [15] Gabriel Ghinita, Panos Kalnis, and Spiros Skiadopoulos. Prive: anony- mous location-based queries in distributed mobile systems. In WWW ’07: Proceedings of the 16th international conference on World Wide Web, pages 371–380, 2007. [16] Google. Google Latitude. http://www.google.com/latitude. [17] GPS. Global Positioning System. http://www.gps.gov. [18] Marco Gruteser and Dirk Grunwald. Anonymous usage of location- based services through spatial and temporal cloaking. In MobiSys ’03: Proceedings of the 1st international conference on Mobile systems, ap- plications and services, pages 31–42, 2003. [19] Marco Gruteser and Xuan Liu. Protecting privacy in continuous location-tracking applications. IEEE Security and Privacy, 2(2):28–34, 2004. [20] Chris H. Hamilton and Andrew Rau-Chaplin. Compact Hilbert indices: Space-filling curves for domains with unequal side lengths. Information Processing Letters, 105(5):155–163, 2008. [21] D. Hilbert. Über die stetige Abbildung einer Linie auf ein Flachenstück. Mathematische Annalen, 38:459–460, 1891. 86 [22] Baik Hoh and Marco Gruteser. Protecting location privacy through path confusion. In SECURECOMM ’05: Proceedings of the First In- ternational Conference on Security and Privacy for Emerging Areas in Communications Networks, pages 194–205, 2005. [23] Hidetoshi Kido, Yutaka Yanagisawa, and Tetsuji Satoh. Protection of location privacy using dummies for location-based services. In ICDEW ’05: Proceedings of the 21st International Conference on Data Engi- neering Workshops, page 1248, 2005. [24] Chun-Ta Li, Min-Shiang Hwang, and Yen-Ping Chu. A secure and effi- cient communication scheme with authenticated key establishment and privacy preserving for vehicular ad hoc networks. Computer Commu- nications, 31(12):2803–2814, 2008. [25] Ashwin Machanavajjhala, Johannes Gehrke, Daniel Kifer, and Muthu- ramakrishnan Venkitasubramaniam. l-diversity: Privacy beyond k- anonymity. In Proceedings of the 22nd International Conference on Data Engineering, ICDE 2006, page 24, 2006. [26] André Malm. GPS and Mobile Handsets. Berg Insight Report, January 2008. [27] Adam Meyerson and Ryan Williams. On the complexity of optimal k-anonymity. In PODS ’04: Proceedings of the twenty-third ACM SIGMOD-SIGACT-SIGART symposium on Principles of database sys- tems, pages 223–228, 2004. [28] Mohamed F. Mokbel, Chi-Yin Chow, and Walid G. Aref. The new casper: query processing for location services without compromising privacy. In VLDB ’06: Proceedings of the 32nd international conference on Very large data bases, pages 763–774, 2006. [29] Mohamed F. Mokbel, Chi-Yin Chow, and Walid G. Aref. The new casper: A privacy-aware location-based database server. In Proceedings of the 23rd International Conference on Data Engineering, pages 1499– 1500, 2007. [30] Bongki Moon, H. V. Jagadish, Christos Faloutsos, and Joel H. Saltz. Analysis of the clustering properties of the Hilbert space-filling curve. IEEE Transactions on Knowledge and Data Engineering, 13(1):124– 141, 2001. 87 [31] FOX News. Man Accused of Stalking Ex-Girlfriend With GPS, Septem- ber 2004. [32] Octopus. Octopus Cards Limited. http://www.octopuscards.com/. [33] Location Privacy Protection Act of 2001. The Library Of Congress. http://www.loc.gov/index.html, 2001. [34] Yi Ouyang, Yurong Xu, Zhengyi Le, Guanling Chen, and Fillia Make- don. Providing location privacy in assisted living environments. In PETRA ’08: Proceedings of the 1st international conference on Perva- sive Technologies Related to Assistive Environments, pages 1–8, 2008. [35] Maxim Raya and Jean-Pierre Hubaux. Securing vehicular ad hoc net- works. Journal of Computer Security, 15(1):39–68, 2007. [36] Daniele Riboni, Linda Pareschi, and Claudio Bettini. Privacy in georef- erenced context-aware services: A survey. In Proceedings of the 1st In- ternational Workshop on Privacy in Location-Based Applications, vol- ume 397 of CEUR Workshop Proceedings, 2008. [37] P. Samarati and L. Sweeney. Protecting privacy when disclosing infor- mation: k-anonymity and its enforcement through generalization and suppression. Technical report, SRI Computer Science Laboratory, 1998. [38] Latanya Sweeney. k-anonymity: a model for protecting privacy. In- ternational Journal on Uncertainty, Fuzziness and Knowledge-Based Systems, 10(5):557–570, 2002. [39] Manolis Terrovitis and Nikos Mamoulis. Privacy preservation in the publication of trajectories. In MDM ’08: Proceedings of the The Ninth International Conference on Mobile Data Management (mdm 2008), pages 65–72, 2008. [40] John Voelcker. Stalked by Satellite. IEEE Spectrum, July 2006. [41] Nikolay Vyahhi, Spiridon Bakiras, Panos Kalnis, and Gabriel Ghinita. Tracking moving objects in anonymized trajectories. In DEXA ’08: Proceedings of the 19th international conference on Database and Expert Systems Applications, pages 158–171, 2008. [42] Fahui Wang. Geographic Information Systems and Crime Analysis. Idea Group Inc (IGI), Hershey, Pennsylvania, 2005. 88 [43] Roman Yarovoy, Francesco Bonchi, Laks V.S. Lakshmanan, and Wendy Hui Wang. Anonymizing Moving Objects: How to Hide a MOB in a Crowd? In Proceedings of the 12th International Conference on Extending Database Technology (EDBT 2009), pages 72–83, 2009. [44] Bin Zhou, Jian Pei, and Wo-Shun Luk. A brief survey on anonymization techniques for privacy preserving publishing of social network data. ACM SIGKDD Explorations, 10(2):12–22, 2008. 89
- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- Privacy-preserving publishing of moving objects databases
Open Collections
UBC Theses and Dissertations
Featured Collection
UBC Theses and Dissertations
Privacy-preserving publishing of moving objects databases Yarovoy, Roman 2009
pdf
Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.
Page Metadata
Item Metadata
Title | Privacy-preserving publishing of moving objects databases |
Creator |
Yarovoy, Roman |
Publisher | University of British Columbia |
Date Issued | 2009 |
Description | Moving Objects Databases (MOD) have gained popularity as a subject for research due to the latest developments in the positioning technologies and mobile networking. Analysis of mobility data can be used to discover and deliver knowledge that can enhance public welfare. For instance, a study of traffic patterns and congestion trends can reveal some information that can be used to improve routing and scheduling of public transit vehicles. To enable analysis of mobility data, a MOD must be published. However, publication of MOD can pose a threat to location privacy of users, whose movement is recorded in the database. A user's location at one or more time points can be publicly available prior to the publication of MOD. Based on this public knowledge, an attacker can potentially find the user's entire trajectory and learn his/her positions at other time points, which constitutes privacy breach. This public knowledge is a user's quasi-identifier (QID), i.e. a set of attributes that can uniquely identify the user's trajectory in the published database. We argue that unlike in relational microdata, where all tuples have the same set of quasi-identifiers, in mobility data, the concept of quasi-identifier must be modeled subjectively on an individual basis. In this work, we study the problem of privacy preserving publication of MOD. We conjecture that each Moving Object (MOB) may have a distinct QID. We develop a possible attack model on the published MOD given public knowledge of some or all MOBs. We develop k-anonymity model (based on classical k-anonymity), which ensures that every object is indistinguishable (with respect to its QID) from at least k-1 other objects, and show that this model is impervious to the proposed attack model. We employ space generalization to achieve MOB anonymity. We propose three anonymization algorithms that generate a MOD that satisfies the k-anonymity model, while minimizing the information loss. We conduct several sets of experiments on synthetic and real-world data sets of vehicular traffic to analyze and evaluate our proposed algorithms. |
Extent | 603320 bytes |
Genre |
Thesis/Dissertation |
Type |
Text |
FileFormat | application/pdf |
Language | eng |
Date Available | 2009-04-15 |
Provider | Vancouver : University of British Columbia Library |
Rights | Attribution-NonCommercial-NoDerivatives 4.0 International |
DOI | 10.14288/1.0051481 |
URI | http://hdl.handle.net/2429/7154 |
Degree |
Master of Science - MSc |
Program |
Computer Science |
Affiliation |
Science, Faculty of Computer Science, Department of |
Degree Grantor | University of British Columbia |
GraduationDate | 2009-05 |
Campus |
UBCV |
Scholarly Level | Graduate |
Rights URI | http://creativecommons.org/licenses/by-nc-nd/4.0/ |
AggregatedSourceRepository | DSpace |
Download
- Media
- 24-ubc_2009_spring_yarovoy_roman.pdf [ 589.18kB ]
- Metadata
- JSON: 24-1.0051481.json
- JSON-LD: 24-1.0051481-ld.json
- RDF/XML (Pretty): 24-1.0051481-rdf.xml
- RDF/JSON: 24-1.0051481-rdf.json
- Turtle: 24-1.0051481-turtle.txt
- N-Triples: 24-1.0051481-rdf-ntriples.txt
- Original Record: 24-1.0051481-source.json
- Full Text
- 24-1.0051481-fulltext.txt
- Citation
- 24-1.0051481.ris
Full Text
Cite
Citation Scheme:
Usage Statistics
Share
Embed
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"
src="{[{embed.src}]}"
data-item="{[{embed.item}]}"
data-collection="{[{embed.collection}]}"
data-metadata="{[{embed.showMetadata}]}"
data-width="{[{embed.width}]}"
data-media="{[{embed.selectedMedia}]}"
async >
</script>
</div>
Our image viewer uses the IIIF 2.0 standard.
To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0051481/manifest