Scholarlink

<p>Adversarial attacks pose a significant threat to object detectors reliant on Deep Neural Networks (DNNs), often inducing erroneous predictions. Given the scarcity of effective adversarial defense methods for object detection, this paper proposes an innovative defense approach based on adversarial noise filtering using Luma Non-Local Means (LNLM). LNLM leverages the brightness channel of adversarial sample images to guide the color channel processing, effectively removing adversarial noise and restoring samples to their original manifold space. This method is applied during the preprocessing phase and does not require prior knowledge of adversarial attacks or modifications to the object detection model's internal architecture. Comprehensive experiments on four object detectors (Faster R-CNN, SSD, YOLOv3, and YOLOv5) against eight adversarial attacks demonstrate LNLM's effectiveness, with defense success rates exceeding 60% for Faster R-CNN and reaching up to 90.1% for YOLOv3 against specific attacks. LNLM exhibits minimal interference with normal sample detection outcomes and is readily migratable across different object detectors. Furthermore, integrating Grad-CAM provides insights into the nature of adversarial attacks and the effectiveness of the LNLM defense mechanism.</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

Enhancing object detection robustness through adversarial noise filtering with luma non-local means

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Heng Zhang,
  • Jiasong Ding,
  • Qin Hang,
  • Yonghong Huang,
  • Nongsen Huang

摘要

Adversarial attacks pose a significant threat to object detectors reliant on Deep Neural Networks (DNNs), often inducing erroneous predictions. Given the scarcity of effective adversarial defense methods for object detection, this paper proposes an innovative defense approach based on adversarial noise filtering using Luma Non-Local Means (LNLM). LNLM leverages the brightness channel of adversarial sample images to guide the color channel processing, effectively removing adversarial noise and restoring samples to their original manifold space. This method is applied during the preprocessing phase and does not require prior knowledge of adversarial attacks or modifications to the object detection model's internal architecture. Comprehensive experiments on four object detectors (Faster R-CNN, SSD, YOLOv3, and YOLOv5) against eight adversarial attacks demonstrate LNLM's effectiveness, with defense success rates exceeding 60% for Faster R-CNN and reaching up to 90.1% for YOLOv3 against specific attacks. LNLM exhibits minimal interference with normal sample detection outcomes and is readily migratable across different object detectors. Furthermore, integrating Grad-CAM provides insights into the nature of adversarial attacks and the effectiveness of the LNLM defense mechanism.