Reinforcing mobile device sensor attacks using generative adversarial network

The gamut of sensors inbuilt in mobile and wearable devices help drive a wide range of useful applications, however, they also pose significant privacy concerns. Several studies have proposed the incorporation of noise into sensor measurements in such a way to ward off the privacy threats while keeping the functionality of the user applications acceptable. In this thesis we argue that a committed attacker could overcome such a defense mechanism by reconstructing the sensor signals and executing attacks on user privacy even when noise is incorporated into sensor data. To showcase this kind of attack, we present a GAN design that incorporates a wavelet functionality to clean out the defensive noise. We show, that an attacker using our mechanism is able to attain up to 75% accuracy on a 2-class activity recognition problem even when noise is injected to drive down the recognition accuracy to around random guessing levels. The thesis provides empirical results to support the argument that adversaries equipped with skills in the latest GAN technologies would not be thwarted by noise injection-based defenses.