Scholarlink

<p>Recent advancements in Camouflaged Object Detection (COD) have challenged the robustness of deep learning systems. Instead of treating interpretability as a secondary visualization tool, this study introduces a diagnostic interpretability-guided framework based on Layer Segmented Activation Mapping (LayerSAM) to mechanistically probe and exploit the structural vulnerabilities of COD models. We design a gradient fusion reconstruction strategy that transforms Class Activation Mapping (CAM) into a diagnostic instrument. This allows a fine-grained analysis of how models rely on "decision anchors"—such as boundary continuity and bionic textures—to identify targets. Adversarial examples generated using the ZoomNext network achieve a 13.8% reduction in mIoU while maintaining high visual stealth (PSNR &gt; 30 dB, SSIM &gt;0.8). Our method demonstrates powerful transferability to black-box models like SINetV2 and IdeNet, reducing mIoU by 26.7% and 14.6%, respectively. By isolating perturbations to these interpretability-defined regions, we provide a quantitative evaluation of the model’s internal logic, revealing that the targeted disruption of structural functional regions is more effective than global noise for characterizing and challenging the inherent robustness of deep segmentation architectures. The source code and data are permanently available at <a href="https://github.com/Qiin1104/camouflage-detection-attack">https://github.com/Qiin1104/camouflage-detection-attack</a> and have been archived with the Digital Object Identifier <a href="https://doi.org/10.5281/zenodo.17538737">https://doi.org/10.5281/zenodo.17538737</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

Interpretable adversarial attacks on camouflaged object detection via layer segmented activation mapping

  • 期刊论文

  • 发表日期：

    2026年4月14日

查看全文
  • Qinzi Wang,
  • Chuanzhen Xu,
  • Xiaolin Tan,
  • Sihua Fu

摘要

Recent advancements in Camouflaged Object Detection (COD) have challenged the robustness of deep learning systems. Instead of treating interpretability as a secondary visualization tool, this study introduces a diagnostic interpretability-guided framework based on Layer Segmented Activation Mapping (LayerSAM) to mechanistically probe and exploit the structural vulnerabilities of COD models. We design a gradient fusion reconstruction strategy that transforms Class Activation Mapping (CAM) into a diagnostic instrument. This allows a fine-grained analysis of how models rely on "decision anchors"—such as boundary continuity and bionic textures—to identify targets. Adversarial examples generated using the ZoomNext network achieve a 13.8% reduction in mIoU while maintaining high visual stealth (PSNR > 30 dB, SSIM >0.8). Our method demonstrates powerful transferability to black-box models like SINetV2 and IdeNet, reducing mIoU by 26.7% and 14.6%, respectively. By isolating perturbations to these interpretability-defined regions, we provide a quantitative evaluation of the model’s internal logic, revealing that the targeted disruption of structural functional regions is more effective than global noise for characterizing and challenging the inherent robustness of deep segmentation architectures. The source code and data are permanently available at https://github.com/Qiin1104/camouflage-detection-attack and have been archived with the Digital Object Identifier https://doi.org/10.5281/zenodo.17538737.