Towards Robust Dual-Trigger Physical Backdoor Attacks against Multi-Object Tracking

In recent years, backdoor attacks have posed a significant threat to the security of deep models. Attackers can induce erroneous behavior in victim models through carefully designed triggers. However, existing backdoor attacks primarily originate from image classification tasks and are designed for the digital world, while research on physical backdoor attacks against multi-object tracking (MOT) is scarce. Moreover, existing methods typically implant a single trigger in the victim model, limiting the diversity of attacks. To fill this gap, we propose a robust dual-trigger physical backdoor attack framework against MOT under a poison-only scenario. Specifically, we do not require knowledge of the training components of the model. We only need to implant two triggers with different backdoor effects (disappearance and resizing) in a few video frames. To enhance attack effectiveness and prevent confusion between these two backdoor effects, we introduce contrastive loss to optimize the triggers and apply various transformations to the triggers to simulate physical world scenarios, thereby further improving the robustness of the attacks. Extensive experiments in both digital and physical worlds demonstrate that our attacks significantly degrade the performance of state-of-the-art MOT trackers. Additionally, we discuss the transferability and imperceptibility of our attacks and their robustness against various potential backdoor defenses.