Unsupervised cycle-consistent adversarial attacks for visual object tracking

Adversarial attacks on visual object tracking have been gaining momentum in the quest to assess and enhance the robustness and security of object tracking models. Most adversarial attack strategies for object tracking rely on fully supervised attacks; that is, they necessitate the use of labels. However, it is not feasible to apply these labeled video object tracking techniques to all attacks. As such, this paper presents an unsupervised attack methodology against visual object tracking models. The approach employs the cycle consistency principle of object tracking models to maximize the inconsistency between forward and backward tracking, thereby providing effective countermeasures. Additionally, this paper introduces a contextual attack method, leveraging the information from the attack object's region and its surrounding contextual regions. This strategy attacks the object region and its surrounding context regions simultaneously, aiming to decrease its response score to the attack. The proposed attack method is assessed across various types of deep learning-based object trackers. Experimental results on multiple benchmarks reveal that the suggested method yields competitive attack outcomes.