Open Collections

Koppisetti, Naga Raghavendra Surya Vara Prasad

University of British Columbia

Radio frequency (RF) wireless systems generate large amounts of data every day on the signal content and the received signal strength (RSS) information. However, limited efforts have been put into analyzing the RF data for security applications, such as the detection of unauthorized transmissions from an intruder. To bridge this gap, the Ph.D. thesis presents machine learning algorithms which detect and localize the intruder devices from their RF transmissions. First, we study the problem of detecting the intruder RF signals in wideband RF traces. An example setup is considered, with the intruder transmitting Wi-Fi signals and there exists interference from Bluetooth and microwave oven signals. We show that the conventional energy-thresholding algorithms are sensitive to noise variations and they require handcrafted parameters for each RF trace. To address this concern, we develop a deep learning solution, which employs convolutional neural networks to perform the signal de- tection. Experiments on both synthetic and real RF traces confirm the superior performance of the proposed solution in terms of the achieved mean average precision. Second, we study the problem of locating the intruder device from the RSS measured passively in the system. We work with the difference of RSS (DRSS), calculated with respect to a reference sensor, in order to handle the unknown transmission power and device heterogeneity of the intruder. The localization problem is formulated as a Gaussian process regression (GP) task to obtain the location estimates and the associated confidence intervals in closed-form. We propose two GP methods which take the stochastic nature of the test DRSS into account and provide more accurate confidence intervals on the test locations than the conventional GP method. Third, to improve the localization accuracy of the proposed GP methods, we present data reconstruction techniques which exploit the low-dimensionality exhibited by the DRSS vectors. Fourth, we study intruder localization for the case when the rate of signal strength decay with distance, also called the path loss exponent, is Gaussian distributed in the area. We propose two low-cost linear least squares estimators for the device location, which employ multilateration on the maximum-likelihood estimates of the distances to the sensors.

Text

2021-09-10

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/79700

Electrical and Computer Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/