Securing Autonomous Vehicles with Robust Physical Invariants

Automation is an exciting and developing field that, today, is growing at the speed of human innovation. Autonomous Vehicles (AVs) in particular are being implemented in pioneering ways across many industries—medical, education, even law enforcement—with arguably its most noticeable implications on the transportation and logistics industry. AVs are not simply limited to ground vehicles, but encompass aerial, sea, and underwater vehicles. While this shift toward automation could improve the global quality of life, it comes with a difficult problem to solve: security. Currently, AVs’ method of interacting with their surrounding environment is by information gathered through their sensors. While these sensors have multiple purposes and vary between vehicle types, they all have the common problem of being susceptible to sensor targeted attacks. These types of attacks focus on tampering with the physical signals observed by sensors. AVs intrinsically use this sensor information to navigate their environments without any form of verification, and current conventional security mechanisms do not address this new type of attacks. To solve this issue, this dissertation proposes SAVIOR, a physics-based anomaly detector that leverages the physical invariants of AVs and evaluates the veracity of sensor input before acting upon it in real time. This is achieved via the usage of non-linear models that describe the behavior of AVs and the implementation of an Extended Kalman Filter (EKF) algorithm for state prediction in real time. This work also introduces a type of stealthy attack that is able to bypass current anomaly detection approaches by injecting a small amount of deviation over time. Our proposed anomaly detector is able to identify these types of attacks due to the implementation of a cumulative sum (CUSUM) algorithm that keeps track of accumulated discrepancies over time and raises an alarm once any sensor has passed a predefined threshold. This work has been implemented on both aerial and ground AVs and presents an anomaly detection framework that improves on the current state-of-the-art with low run-time overhead. Developing further our work on SAVIOR, we increased our focus on ground AVs to enhance their functionality and anomaly detection capabilities. Chapter 3 introduces Securing Platooning Vehicles with Shared Reality which demonstrates improvements to the self-driving algorithm (vehicle platooning), the camera sensor, implementation of LiDAR (Light Detection And Ranging) sensor, and the development of a new shared reality module. We aim at improving our anomaly detector by implementing sensor redundancy via the usage of a LiDAR sensor and our shared reality module that will ensure that both sensors are experiencing the same physical reality in real time. We also introduce a new type of slow attack that is not detectable by SAVIOR and show how sensor redundancy and shared reality can aid in identifying subtle consistent attacks against sensors. We implemented our design on a popular robotic operating system platform and show that platooning vehicles remain secure while performing several autonomous tasks with low overhead.