arXiv reaDer
TrackPGD: A White-box Attack using Binary Masks against Robust Transformer Trackers
Object trackers with transformer backbones have achieved robust performance on visual object tracking datasets. However, the adversarial robustness of these trackers has not been well studied in the literature. Due to the backbone differences, the adversarial white-box attacks proposed for object tracking are not transferable to all types of trackers. For instance, transformer trackers such as MixFormerM still function well after black-box attacks, especially in predicting the object binary masks. We are proposing a novel white-box attack named TrackPGD, which relies on the predicted object binary mask to attack the robust transformer trackers. That new attack focuses on annotation masks by adapting the well-known SegPGD segmentation attack, allowing to successfully conduct the white-box attack on trackers relying on transformer backbones. The experimental results indicate that the TrackPGD is able to effectively attack transformer-based trackers such as MixFormerM, OSTrackSTS, and TransT-SEG on several tracking datasets.
updated: Thu Jul 04 2024 14:02:12 GMT+0000 (UTC)
published: Thu Jul 04 2024 14:02:12 GMT+0000 (UTC)
参考文献 (このサイトで利用可能なもの) / References (only if available on this site)
被参照文献 (このサイトで利用可能なものを新しい順に) / Citations (only if available on this site, in order of most recent)アソシエイト