MACHINE LEARNING FOR WIRELESS CYBER-PHYSICAL SYSTEMS SECURITY

Abstract

Wireless cyber-physical systems (CPS) have been progressively adopted in many applications such as smart industrial control systems, intelligent vehicular transportation, unmanned aerial vehicles (UAV), etc. Despite the CPS huge potential benefits in paving the path to develop new applications, the open and broadcast nature of the wireless communication medium has made these systems vulnerable to cyber attacks. In this thesis, we propose a family of novel online machine learning algorithms which can be employed to defend against jamming attacks in wireless CPS, and wireless communication systems in general. In addition, we study the problem of fast detection and identification of intruding consumer UAVs and propose a new method which exploits wireless network traffic information and utilizes machine learning techniques to identify the UAVs in a timely manner. More specifically, in this thesis, we discuss four research projects which briefly are summarized as follows. 1) We study security of remote state estimation in wireless CPS where a sensor sends its measurements to the remote state estimator over a multi-channel wireless link in presence of a jamming attacker. We propose a novel online learning-based policy which can be employed by the sensor to jointly choose the transmission channel and power to defend against the attack. We theoretically prove that the proposed algorithm achieves a sublinear order-optimal learning regret bound in time. 2) We focus on the security of multi-channel wireless communication systems with a scenario in which the jammer always successfully attacks on the acknowledgment link and the transmitter loses throughput due to dynamic channel switching latency. We model this problem as self-unaware bandits with arm switching costs problem and propose two novel online learning algorithms with theoretical performance guarantees. We prove a sublinear regret upper bound for both algorithms and bound the switching costs such that it can improve the regret bound. 3) We study the security of cooperative adaptive cruise control (CACC) system under jamming attacks. We propose a novel time domain approach to analyze the mean string stability and impact of the jammer's location on the string stability. We derive a condition for the packet successful delivery probability which indicates that the jammer has a higher probability to destabilize the string when it is closer to the first vehicle following the lead vehicle. As a defense strategy for the setting of multi-channel wireless communication among the vehicles, we derive the mean string stability condition with respect to the minimum packet loss probability and number of channels, when both the vehicles and jammer employ online learning-based channel access policies for data transmission and attack, respectively. 4) Finally, we study detecting and identifying intruding consumer UAVs as an urgent need for both invasion detection and forensics purposes. We propose a machine learning-based framework for fast UAV identification over encrypted Wi-Fi traffic. The framework jointly optimizes feature selection and prediction performance in a unified objective function. Furthermore, we identify the UAVs' operation mode through data traffic analysis which implies that there is a strong correlation or coupling between cyber information (data traffic) and physical information (operation mode) of UAVs. This finding is expected to motivate new cyber-physical defense and forensics mechanisms that leverage this cyber-physical coupling. We believe the proposed methodology can be applied to other CPS and motivate more in-depth study on cyber-physical attack co-detection or co-defense for many Internet-of-Things (IoT) applications, such as smart home, smart healthcare, and smart manufacturing.