Scholarlink

<p>Object detection models, widely used in areas like vehicle and person detection, face increasing security challenges, including adversarial patch attacks. Existing research lacks effective defense frameworks for person detection and struggles with generalizing to unknown patch patterns. This paper proposes a defense system combining patch detection and semantic repair. It uses a two-stage framework with a fine-tuned YOLOv11s detector and a UNet repairer integrating CBAM(Convolutional Block Attention Modules) attention and dilated convolution. The system accurately locates adversarial patches and restores image structure, mitigating their impact. Experiments on the Inria and PennFudan datasets show significant improvements in mAP for detectors like YOLOv3tiny, YOLOv8s, and YOLOv11x in person detection. For example, against Advyolo and POI(Patch of invisibility) attacks, our system increases YOLOv3tiny’s mAP by 79.0% and 55.3%, respectively, substantially outperforming baseline defenses such as LGS, JPEG, and Zero. These results validate the effectiveness and generalization capability of our approach in mitigating a wide range of adversarial patch threats.</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

An adversarial patch defense system based on patch detection and occlusion-aware image restoration

  • 期刊论文

  • 发表日期：

    2026年7月23日

查看全文
  • Yunfei Wang,
  • Hong Huang,
  • Jie Gao,
  • Yefei Lei

摘要

Object detection models, widely used in areas like vehicle and person detection, face increasing security challenges, including adversarial patch attacks. Existing research lacks effective defense frameworks for person detection and struggles with generalizing to unknown patch patterns. This paper proposes a defense system combining patch detection and semantic repair. It uses a two-stage framework with a fine-tuned YOLOv11s detector and a UNet repairer integrating CBAM(Convolutional Block Attention Modules) attention and dilated convolution. The system accurately locates adversarial patches and restores image structure, mitigating their impact. Experiments on the Inria and PennFudan datasets show significant improvements in mAP for detectors like YOLOv3tiny, YOLOv8s, and YOLOv11x in person detection. For example, against Advyolo and POI(Patch of invisibility) attacks, our system increases YOLOv3tiny’s mAP by 79.0% and 55.3%, respectively, substantially outperforming baseline defenses such as LGS, JPEG, and Zero. These results validate the effectiveness and generalization capability of our approach in mitigating a wide range of adversarial patch threats.