TrackSpear: Attention-Guided Adversarial Patches for Probing Security in Visual Tracking

Le résumé fourni par la source

In traffic cyber-physical systems, accurately tracking the movements of surrounding objects is crucial for the safety and reliability of autonomous vehicles. In these safetycritical scenarios, physical adversarial attacks present a significant threat. Existing attack methods primarily rely on global perturbations, which are difficult to implement in real-world scenarios, while effective attacks in practice often require local patches. However, current patch-based attacks are less effective against Transformer-based visual tracking models, as these models capture long-range dependencies and demonstrate strong adversarial robustness. To address this, TrackSpear is proposed, an attention-guided network designed to generate adversarial patches and evaluate the security of visual tracking models. TrackSpear offers several key features: (i) it dynamically generates adversarial perturbations for video frames based on the tracker's attention maps; (ii) it employs a simple endto-end network, along with three innovative loss functions, to carry out the attack. Our experiments based on several real datasets, including OTB100, VOT2018, and GOT10K show that TrackSpear significantly degrades the performance of visual tracking models, successfully disrupting target tracking and misleading the tracker into focusing on incorrect locations, ultimately causing misjudgments or loss of the target object.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.