Scholarlink

<p>Deepfake technology and its detection methods are in a continuous arms race. Current deepfake detectors, while effective, suffer from generalization and versatility issues. This paper proposes a novel method to enhance deepfakes by utilizing real discrete codebook (RDC) reconstruction to evade common detectors. The approach first identifies and occludes focus areas that affect detectors’ decisions on forged images. It then reconstructs the image using RDC features from real images, aiming to minimize artifacts and deceive detectors. Extensive experiments on the FaceForensics++ DeepFake dataset demonstrate that our method significantly reduces the average detection accuracy of seven commonly used detectors from 95.6 to 17%, without compromising image quality. This work highlights the need for more robust and universal deepfake detection techniques and underscores the importance of research in this area to combat the growing threat of deepfakes. Code is available at <a href="https://gitee.com/kelikeli/RDCReconstruction">https://gitee.com/kelikeli/RDCReconstruction</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 deepfake robustness via real discrete codebook reconstruction

  • 期刊论文

  • 发表日期：

    2025年4月13日

查看全文
  • Ping Chen,
  • Ming Xu,
  • Huanhuan Bao,
  • Lili Xu

摘要

Deepfake technology and its detection methods are in a continuous arms race. Current deepfake detectors, while effective, suffer from generalization and versatility issues. This paper proposes a novel method to enhance deepfakes by utilizing real discrete codebook (RDC) reconstruction to evade common detectors. The approach first identifies and occludes focus areas that affect detectors’ decisions on forged images. It then reconstructs the image using RDC features from real images, aiming to minimize artifacts and deceive detectors. Extensive experiments on the FaceForensics++ DeepFake dataset demonstrate that our method significantly reduces the average detection accuracy of seven commonly used detectors from 95.6 to 17%, without compromising image quality. This work highlights the need for more robust and universal deepfake detection techniques and underscores the importance of research in this area to combat the growing threat of deepfakes. Code is available at https://gitee.com/kelikeli/RDCReconstruction.