ADoPT: LiDAR Spoofing Attack Detection based on Point-Level Temporal Consistency

Minkyoung Cho, Yulong Cao, Zhuoqing Mao

Abstract:

Deep neural networks (DNNs) are increasingly integrated into LiDAR (Light Detection and Ranging)-based perception systems for autonomous vehicles (AVs), requiring robust performance under adversarial conditions.
One pressing concern is the challenge posed by LiDAR spoofing attacks, where attackers inject fake objects into LiDAR data, leading AVs to misinterpret their surroundings and make faulty decisions. Many current defense algorithms predominantly depend on perception outputs, such as bounding boxes. However, these outputs are intrinsically limited as they are generated by imperfect perception models that process a restricted set of points, acquired from the ego vehicle's specific viewpoint. The reliance on bounding boxes is a manifestation of this fundamental constraint.
To overcome these limitations, we propose a novel framework, named ADoPT, which quantitatively measures temporal consistency across consecutive frames and identifies abnormal objects based on the coherency of point clusters. In our evaluation using the nuScenes dataset, our algorithm effectively counters various LiDAR spoofing attacks, achieving a low ($<$ 10\%) false positive ratio and high ($>$ 85\%) true positive ratio, outperforming existing state-of-the-art defense methods, CARLO and 3D-TC2. Moreover, ADoPT shows promising potential for accurate defense in diverse road environments.