Scholarlink

<p>The adversarial robustness of deep neural networks is crucial for person re-identification tasks. Existing research has indicated that adversarial examples can significantly degrade the recognition accuracy of person re-identification models. Despite various defense methods, adversarial distillation stands out by enhancing the robustness of lightweight student models through the effective extraction of robust features from larger teacher models. However, existing adversarial distillation approaches predominantly rely on a single adversarial teacher model, overlooking the potential benefits of multiple teacher models providing more comprehensive and diverse guidance. Additionally, the widening capability gap between teacher and student models tends to introduce semantic misalignment and reduce the effectiveness of knowledge transfer. To address those challenges, we propose a novel adversarial distillation method named multi-teacher adversarial distillation (MTF-ARD). MTF-ARD dynamically assigns weights to different teacher models based on their predictive confidence and similarity with the student model. These weights are used then to adaptively fuse robust knowledge from multiple teacher models to provide a unified and diverse guidance to the student model. Experimental results demonstrate that MTF-ARD significantly improves the robustness of person re-identification models against state-of-the-art adversarial attacks. Particularly in the Market1501 dataset with ResNet-18 as the student model, MTF-ARD achieves 62.8% robustness under the PGD attack, outperforming AdaAD by 6.6%. Our code is available at <a href="https://github.com/yinghualuo614/MTF-ARD">https://github.com/yinghualuo614/MTF-ARD</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

MTF-ARD: multi-teacher adversarial distillation for robust person re-identification

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Ya Li,
  • Jinbo Ye,
  • Hao Zhang,
  • Qing Wang

摘要

The adversarial robustness of deep neural networks is crucial for person re-identification tasks. Existing research has indicated that adversarial examples can significantly degrade the recognition accuracy of person re-identification models. Despite various defense methods, adversarial distillation stands out by enhancing the robustness of lightweight student models through the effective extraction of robust features from larger teacher models. However, existing adversarial distillation approaches predominantly rely on a single adversarial teacher model, overlooking the potential benefits of multiple teacher models providing more comprehensive and diverse guidance. Additionally, the widening capability gap between teacher and student models tends to introduce semantic misalignment and reduce the effectiveness of knowledge transfer. To address those challenges, we propose a novel adversarial distillation method named multi-teacher adversarial distillation (MTF-ARD). MTF-ARD dynamically assigns weights to different teacher models based on their predictive confidence and similarity with the student model. These weights are used then to adaptively fuse robust knowledge from multiple teacher models to provide a unified and diverse guidance to the student model. Experimental results demonstrate that MTF-ARD significantly improves the robustness of person re-identification models against state-of-the-art adversarial attacks. Particularly in the Market1501 dataset with ResNet-18 as the student model, MTF-ARD achieves 62.8% robustness under the PGD attack, outperforming AdaAD by 6.6%. Our code is available at https://github.com/yinghualuo614/MTF-ARD.