.png){kind=icon}
Return
Box-spoof attack against single object tracking
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-023-05264-2.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Recent research has revealed that single object tracking (SOT) is susceptible to adversarial examples, with even small perturbations added to video frames leading to tracking failure. However, most of the existing methods are online attack methods. These methods generate video-specific or target-specific perturbation to deceive the trackers through a frame-by-frame online paradigm. This method is inherently inefficient in the production of adversarial examples, rendering it less than ideal for real-time application scenarios. To address this, we propose a novel offline attack method that can achieve high-frequency attack strength by using only one frame of video, called box-spoof attack (BS-Attack). In contrast to prior methods perturb feature maps or confidence maps in a frame-by-frame manner. Our BS-Attack directs its focus towards the proposal selection stage, a relatively late stage within the tracking pipeline. Notably, our method generates perturbations by utilizing the information from the initial frame, which can achieve rapid disturbance generation and high-frequency attack strength. In addition to confidence loss, we incorporate joint intersection-over-union and shrink loss to keep the predicted bounding box away from the ground truth and shrink it. BS-Attack requires only one video to generate a perturbation that can deceive the entire dataset. The UAV123, OTB100, VOT2018, and VOT2019 datasets are used to demonstrate the effectiveness of this method in attacking popular trackers such as SiamRPN, SiamRPN++, and SiamMask. The results show that BS-Attack outperforms state-of-the-art online and offline attack methods.
Keywords:
Single object tracking
Adversarial examples
Offline attack
Box-spoof attack
Journal
IF:
3.5
Papers:
7.6K
Citations:
1.7W
Organization
Cited Papers
ShareSaveShareSaveFine-tuning more stable neural text classifiers for defending word level adversarial attacks.png)
APPLIED INTELLIGENCE
APPLIED INTELLIGENCEIF3.5
PREAI
ShareSaveShareSaveSiamese network with transformer and saliency encoder for object trackingAPPLIED INTELLIGENCE
APPLIED INTELLIGENCEIF3.5
PREAI
ShareSave
View more