Asymmetry Vulnerability and Physical Attacks on Online Map Construction for Autonomous Driving

MapTR is the primary model used in our experiments. We extend our attacks to two state-of-the-art models: MapQR and GeMap. MapQR enhances map query capabilities while GeMap considers geometric shapes and relations, both achieving state-of-the-art performance. However, our attacks on these models still demonstrate significant effectiveness, achieving up to 39% Unreachable Goal Rate and up to 31% Collision Rate. Moreover, our experiments on GeMap with both LiDAR point cloud and multi-view camera image inputs show that our attack interference targeting only the camera input can still be somewhat effective against fusion systems, achieving up to 27% unreachable goal rate and 18% collision rate.

End-to-end (E2E) autonomous driving models are increasingly popular, often incorporating online map construction as a key module. We extend our experiments by launching Road Straightening and Early Turn attacks on VAD, a widely used E2E model, across 100 asymmetric scenes. Results demonstrate that our proposed attacks not only compromise dedicated online map construction models but also significantly degrade both map perception and planning performance in E2E autonomous driving systems.