A physical ID-transfer attack on multi-object tracking based on adversarial trajectories

The researchers propose AdvTraj, a physical ID-transfer attack against multi-object tracking (MOT). The authors describe it as the first online physical ID-manipulation attack against tracking-by-detection MOT: the attacker uses adversarial trajectories to transfer its own ID to a target object in order to confuse the tracking system, without attacking the object detection module. CARLA simulations show that the white-box attack on SORT achieves a 100% success rate across multiple scenarios, and the transfer attack on SOTA MOT algorithms achieves a success rate of up to 93%. The authors identify two general adversarial maneuvers that can be executed by pedestrians or drivers in everyday scenarios, and state that the work reveals an understudied weakness in the object association stage of SOTA MOT systems and provides ideas for enhancing the robustness of such systems.