The integrity of perception data transmitted over in-vehicle networks is important for the safety of autonomous driving. However, legacy protocols like the Controller Area Network (CAN) bus which lacks essential security features make In-Vehicle Networks (IVNs) vulnerable to data tampering attacks. Current research typically focuses on detecting the attack itself but ignores the information recovery from the missing data, leading to an unsafe autonomous driving system. To address the issue, we propose a 3D object recovery framework to recover the missing data caused by the tampering attack that occurred in in-vehicle networks. The proposed framework exploits both temporal and spatial context for the 3D object recovery, where a temporal branch is designed to learn the coordinate offsets of 3D objects based on historical data from previous frames, while a spatial branch employs information from the adjacent views of the attacked objects to locate the recovered objects from the overlapped regions in the current frame. By integrating the temporal and spatial clues, the framework effectively recovers the missing objects from the resting ones, thereby enhancing the immunity of in-vehicle networks for the tampering attack. Extensive experiments on the nuScenes dataset demonstrate that the proposed framework significantly improves 3D object detection performance under the attack when compared to the method without recovery. Additionally, the recovery performance becomes better as the attack intensity increases, highlighting the framework’s robustness in high-risk scenarios. The source will be available upon publication.
Han et al. (Thu,) studied this question.