Scholarlink

<p>Against the backdrop of increasing risks of adversarial attacks on face recognition systems, a defense framework based on generative adversarial networks (GANs) called AdvFaceDefGAN is proposed to enhance the robustness of the system against dual-identity impersonation attacks. This method does not require modification of the original recognition model. Instead, it constructs a generative network composed of a perturbation generator, a disturbance eliminator, a discriminator, and an auxiliary identity classifier. By employing a joint mechanism of feature preservation and similarity suppression, the network effectively purifies the perturbations in the input images. Experiments were conducted on typical face recognition models such as IR50, ArcFace, and FaceNet, along with various adversarial samples, including FGSM, MI-FGSM, and AdvFaceGAN, to validate the performance. The results demonstrate that the purified images retain high similarity to the original identity features while significantly reducing the similarity to the target identity. This leads to a notable reduction in the success rate of both traditional impersonation and dual-identity impersonation attacks. The method also shows excellent model compatibility and defense stability in both black-box and white-box environments, proving its practical deployment value in real-world engineering applications. The source code and datasets used in our experiments are available at: <a href="https://github.com/yuwantang/AdvFaceDefGAN">https://github.com/yuwantang/AdvFaceDefGAN</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

Enhancing face recognition robustness: AdvFaceDefGAN against dual-identity attacks

  • 期刊论文

  • 发表日期：

    2026年6月27日

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

摘要

Against the backdrop of increasing risks of adversarial attacks on face recognition systems, a defense framework based on generative adversarial networks (GANs) called AdvFaceDefGAN is proposed to enhance the robustness of the system against dual-identity impersonation attacks. This method does not require modification of the original recognition model. Instead, it constructs a generative network composed of a perturbation generator, a disturbance eliminator, a discriminator, and an auxiliary identity classifier. By employing a joint mechanism of feature preservation and similarity suppression, the network effectively purifies the perturbations in the input images. Experiments were conducted on typical face recognition models such as IR50, ArcFace, and FaceNet, along with various adversarial samples, including FGSM, MI-FGSM, and AdvFaceGAN, to validate the performance. The results demonstrate that the purified images retain high similarity to the original identity features while significantly reducing the similarity to the target identity. This leads to a notable reduction in the success rate of both traditional impersonation and dual-identity impersonation attacks. The method also shows excellent model compatibility and defense stability in both black-box and white-box environments, proving its practical deployment value in real-world engineering applications. The source code and datasets used in our experiments are available at: https://github.com/yuwantang/AdvFaceDefGAN.