Scholarlink

<p>The increasing use of computer vision-based autonomous driving technology in daily life necessitates further evaluation of its safety. Current physical attack techniques, using non-transparent stickers as perturbations, lack stealth and therefore may not be effective in real-world applications. Some studies use translucent patches on camera lenses to attack deep neural networks (DNNs), but accessing a victim’s camera is impractical. Light-based attacks also struggle to achieve robust effects under varying environmental conditions. To address these issues within the domain of applied pattern recognition, we propose Adversarial Translucent Patch (AdvTP). This method utilizes translucent color patches optimized with a differential evolution algorithm to create effective physical perturbations. These patches are applied to target objects for black-box attacks on object detectors, a key area in computer vision and image processing. Extensive experiments validate the method’s effectiveness, stealth, and robustness. The proposed method achieves a 91.04% success rate in digital attacks and a 100% success rate in most physical attack cases. It demonstrates superior stealth compared to baseline methods and achieves an average attack success rate of 94.04% against advanced object detectors. We also analyze the method’s generalization capabilities across different pattern recognition tasks, including attacking image classifiers and vehicle detectors, as well as its performance in transfer attacks and against adversarial defenses. Given the significant security threats posed by this method to vision-based applications, which are critical in various applied domains, we believe this work will attract considerable attention from the pattern recognition community. The code can be found in <a href="https://github.com/kalbinur90/AdvTP.git">https://github.com/kalbinur90/AdvTP.git</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

Adversarial translucent patch: a robust physical attack technique against object detectors

  • 期刊论文

  • 发表日期：

    2025年8月20日

查看全文
  • Kalibinuer Tiliwalidi,
  • Chengyin Hu,
  • Weiwen Shi,
  • Guangxi Lu,
  • Hao Wu

摘要

The increasing use of computer vision-based autonomous driving technology in daily life necessitates further evaluation of its safety. Current physical attack techniques, using non-transparent stickers as perturbations, lack stealth and therefore may not be effective in real-world applications. Some studies use translucent patches on camera lenses to attack deep neural networks (DNNs), but accessing a victim’s camera is impractical. Light-based attacks also struggle to achieve robust effects under varying environmental conditions. To address these issues within the domain of applied pattern recognition, we propose Adversarial Translucent Patch (AdvTP). This method utilizes translucent color patches optimized with a differential evolution algorithm to create effective physical perturbations. These patches are applied to target objects for black-box attacks on object detectors, a key area in computer vision and image processing. Extensive experiments validate the method’s effectiveness, stealth, and robustness. The proposed method achieves a 91.04% success rate in digital attacks and a 100% success rate in most physical attack cases. It demonstrates superior stealth compared to baseline methods and achieves an average attack success rate of 94.04% against advanced object detectors. We also analyze the method’s generalization capabilities across different pattern recognition tasks, including attacking image classifiers and vehicle detectors, as well as its performance in transfer attacks and against adversarial defenses. Given the significant security threats posed by this method to vision-based applications, which are critical in various applied domains, we believe this work will attract considerable attention from the pattern recognition community. The code can be found in https://github.com/kalbinur90/AdvTP.git.