Scholarlink

<p>Recent studies have highlighted the vulnerability of deep learning-based models to adversarial attacks. As a fundamental computer vision task, visual object tracking, increasingly powered by deep neural networks, also demands in-depth investigation in this context. However, the impact of diverse tracking scenarios on such vulnerabilities remains underexplored. Specifically, is the cost of misleading trackers consistent across targets with different visual characteristics? If not, which attributes modulate the adversarial strength required? To address these questions, we propose a novel Scenario-Dependent Spatio-Temporal Sparse adversarial Attack framework (SD-STSAtt) that adaptively controls perturbation sparsity. Our analysis reveals that the required perturbation magnitude is strongly influenced by certain visual attributes of the target, such as its visual complexity and region count. Furthermore, unlike frame-wise or pixel-wise adversarial generation schemes, the proposed method perturbs only a sparse set of key pixels in the first frame of a video. This approach induces significant degradation in tracking performance while substantially reducing the overall perturbation magnitude across the video. We believe this work offers not only a sparse adversarial attack method for visual object tracking, but also provides new insights into the vulnerability of deep neural networks to adversarial examples. Codes are available on <a href="https://github.com/DjangoChaogh/STSAtt-main">https://github.com/DjangoChaogh/STSAtt-main</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Attack Intensity is Target-Related: Exploration of Sparse Adversarial Attack against Visual Object Trackers

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • Shaochuan Zhao,
  • Tianyang Xu,
  • Yong Zhou,
  • Zhangyong Tang,
  • Xiao-Jun Wu,
  • Josef Kittler

摘要

Recent studies have highlighted the vulnerability of deep learning-based models to adversarial attacks. As a fundamental computer vision task, visual object tracking, increasingly powered by deep neural networks, also demands in-depth investigation in this context. However, the impact of diverse tracking scenarios on such vulnerabilities remains underexplored. Specifically, is the cost of misleading trackers consistent across targets with different visual characteristics? If not, which attributes modulate the adversarial strength required? To address these questions, we propose a novel Scenario-Dependent Spatio-Temporal Sparse adversarial Attack framework (SD-STSAtt) that adaptively controls perturbation sparsity. Our analysis reveals that the required perturbation magnitude is strongly influenced by certain visual attributes of the target, such as its visual complexity and region count. Furthermore, unlike frame-wise or pixel-wise adversarial generation schemes, the proposed method perturbs only a sparse set of key pixels in the first frame of a video. This approach induces significant degradation in tracking performance while substantially reducing the overall perturbation magnitude across the video. We believe this work offers not only a sparse adversarial attack method for visual object tracking, but also provides new insights into the vulnerability of deep neural networks to adversarial examples. Codes are available on https://github.com/DjangoChaogh/STSAtt-main.