UC Irvine

Autonomous systems, particularly autonomous driving (AD), are rapidly developing in our society, driven by significant progress in perception systems with advanced deep neural networks (DNN) and sensors. Self-driving taxi services and robot delivery services are becoming increasingly common in major cities around the world. Despite their high perception capability in various challenging driving scenarios, it does not always mean high robustness in adversarial scenarios. DNN models, in particular, are known to be vulnerable to adversarial input perturbations, which can significantly affect the model output with only minor changes to the input data. Given the critical importance of road safety in AD vehicles, it is crucial to systematically understand the potential security vulnerabilities of DNN models and sensors in AD systems. My dissertation focuses on exploring these vulnerabilities and their safety implications in autonomous systems, specifically targeting two major AD sensors: cameras and LiDARs. Cameras are widely used in AD perception due to their affordability and the reliance on visual information in human driving, such as lane lines and traffic signs. However, camera inputs can be easily perturbed by various attack vectors, including physical stickers and laser or light reflections. LiDAR is one of the most innovative sensors in the past decade and plays a crucial role in higher automation levels such as Level-4 AD due to their ability to capture 3D surrounding information as point clouds. Nevertheless, LiDAR is fundamentally vulnerable to malicious lasers emitted by adversaries. In this dissertation, I present new practical attacks for these two types of sensors, evaluate the impact of these attacks, and explore potential defenses. By systematically analyzing the security implications at the DNN model level and the closed-loop autonomous driving level, my research aims to contribute to the development of more secure and safer autonomous systems.