Scholarlink

<p>As deepfake technology continues to evolve, existing techniques can manipulate specific regions such as facial components, and the small forgery area makes forgery more difficult to detect. Therefore, deepfake detection technology cannot be limited to real and fake face detection methods based on binary classification. To enable fine-grained detection of facial component forgeries, we propose a detection approach leveraging multi-label classification to simultaneously identify deepfake labels across various facial components. We first process the facial components manipulation dataset using a data augmentation method with information deletion to enrich the dataset with more varied samples. The texture feature enhancement module is also used to enhance the forged traces in the shallow features and enhance the network’s capability in identifying manipulation artifacts. We then apply Asymmetric Polynomial Loss (APL) to address the issue of label imbalance within the dataset. Experimental evaluation demonstrates that our method reaches a mAP of 92.08<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation> and a mAUC of 89.23<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\%\)</EquationSource> </InlineEquation>, effectively supporting precise detection of facial forgery traces.</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

A multi-label classification method combined with texture enhancement for deepfake face detection

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Liwei Deng,
  • Boda Wu,
  • Jiandong Wang

摘要

As deepfake technology continues to evolve, existing techniques can manipulate specific regions such as facial components, and the small forgery area makes forgery more difficult to detect. Therefore, deepfake detection technology cannot be limited to real and fake face detection methods based on binary classification. To enable fine-grained detection of facial component forgeries, we propose a detection approach leveraging multi-label classification to simultaneously identify deepfake labels across various facial components. We first process the facial components manipulation dataset using a data augmentation method with information deletion to enrich the dataset with more varied samples. The texture feature enhancement module is also used to enhance the forged traces in the shallow features and enhance the network’s capability in identifying manipulation artifacts. We then apply Asymmetric Polynomial Loss (APL) to address the issue of label imbalance within the dataset. Experimental evaluation demonstrates that our method reaches a mAP of 92.08 \(\%\) and a mAUC of 89.23 \(\%\) , effectively supporting precise detection of facial forgery traces.