Depth Priors-informed Purification Defense for Car-Borne LiDAR Vehicle Detection

LiDAR-based 3D vehicle detection is a fundamental perception task of autonomous driving systems. However, recent studies show that physical or physically plausible adversarial examples can severely degrade the vehicle detection performance and then affect downstream tasks such as motion planning. However, the existing defense methods do not achieve satisfactory trade-offs between computational efficiency and defense effectiveness. In this paper, we identify two attack-indicative priors in the depth of the perturbed point cloud area and propose a two-stage informed purification algorithm to remove adversarial points while keeping essential benign points for vehicle detection. With low computational overhead, this new input purification achieves defense performance comparable to the state-of-the-art neural network-based methods while remaining highly efficient.
© 2026 Copyright held by the owner/author(s).