Date of Award
Summer 8-11-2023
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Autonomous driving systems have gained significant attention in recent years, revolutionizing the transportation industry. However, the increasing complexity and connectivity of these systems introduce new challenges and vulnerabilities. This thesis addresses these challenges by investigating adversarial attacks and defenses in the field of autonomous driving. The thesis commences with a comprehensive literature review, providing insights into the vulnerabilities unique to image recognition and autonomous driving. It examines the cur- rent research lines for defense strategies and explores the ongoing efforts to mitigate these vulnerabilities. In the subsequent sections, the thesis presents four core contributions in adversarial attacks and defenses. Section II focuses on the vulnerabilities of autonomous driving systems and connected autonomous driving fleets. It proposes a fast and differentiable adversarial testing framework for simulated autonomous driving, demonstrating its scalability and effectiveness in identifying vulnerabilities. Additionally, the systemic impact of GPS spoofing attacks on large- scale autonomous vehicle deployment in the context of ride-hailing services is investigated. The research explores innovative approaches to mitigate the risks associated with spoofing devices. Moving forward, Section III delves into defense mechanisms against adversarial attacks. It introduces a novel defense approach to counter adversarial patch attacks in image classification, leveraging contrastive adversarial semantic meaning. Furthermore, the thesis addresses the challenge of maintaining robustness in machine learning-based control systems under adversarial perturbations. It proposes a certified robust control approach that combines robustness certification with control, resulting in a certified robust autonomous driving system. Overall, this thesis contributes to the understanding of adversarial vulnerabilities in autonomous driving systems and provides valuable insights into the development of robust defenses. The findings pave the way for enhancing the security and reliability of autonomous driving technologies, ensuring their safe deployment in real-world scenarios.
Language
English (en)
Chair
Yevgeniy Vorobeychik