Revisiting Adversarial Perception Attacks and Defense Methods on Autonomous Driving Systems
Cheng Chen, Yuhong Wang, Nafis S Munir, Xiangwei Zhou, Xugui Zhou
TL;DR
This work systematically assesses adversarial threats to ADS perception by evaluating a broad set of attacks on traffic sign recognition and relative distance estimation within OpenPilot and YOLOv8. It compares defenses spanning image preprocessing, adversarial training, diffusion-based restoration, and contrastive learning, highlighting that no single method robustly guards against all perturbations while also noting trade-offs in latency and long-range accuracy. Key findings include Auto-PGD being highly effective for distance errors, diffusion-based restoration offering strong gains in some tasks but substantial runtime costs, and mixed adversarial training providing broader but imperfect cross-attack robustness. The study provides practical guidance for building more resilient ADS perception through integrated, task-aware defense strategies and points to future work on efficiency and distance-aware defense design.
Abstract
Autonomous driving systems (ADS) increasingly rely on deep learning-based perception models, which remain vulnerable to adversarial attacks. In this paper, we revisit adversarial attacks and defense methods, focusing on road sign recognition and lead object detection and prediction (e.g., relative distance). Using a Level-2 production ADS, OpenPilot by Comma$.$ai, and the widely adopted YOLO model, we systematically examine the impact of adversarial perturbations and assess defense techniques, including adversarial training, image processing, contrastive learning, and diffusion models. Our experiments highlight both the strengths and limitations of these methods in mitigating complex attacks. Through targeted evaluations of model robustness, we aim to provide deeper insights into the vulnerabilities of ADS perception systems and contribute guidance for developing more resilient defense strategies.