Scholarlink

<p>Face forgery techniques have advanced significantly, posing serious threats to digital media authenticity. This paper introduces a novel forgery-aware multi-scale cross-attention network (FMCA-Net) for robust face forgery detection. FMCA-Net employs a dual-branch architecture, integrating spatial and frequency domain features through cross-modal attention mechanisms. A forgery-aware token multi-scale attention module (FTMA) enhances discriminability, a spatial–frequency cross module (SFCM) enables bidirectional feature interaction, and a learnable spatial–frequency concatenation module (LSFC) ensures stable feature interaction. Extensive experiments on four public datasets demonstrate that FMCA-Net achieves superior performance, reaching an accuracy of 97.47% on FaceForensics++, 98.74% on Celeb-DF-v2, and 86.30% on WildDeepfake, and attaining an average AUC of 80.85% in cross-dataset evaluations, maintaining strong robustness under severe compression and unseen manipulations. Compared with 17 state-of-the-art methods, our findings contribute to the development of reliable face forgery detection methods in diverse conditions. The code is available at&#xa0;<a href="https://github.com/LiJin-sdu/FMCA-Net">https://github.com/LiJin-sdu/FMCA-Net</a> and is permanently archived with a DOI at&#xa0;<a href="https://doi.org/10.5281/zenodo.17596909">https://doi.org/10.5281/zenodo.17596909</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

Multi-scale cross-attention network for enhanced face forgery detection

  • 期刊论文

  • 发表日期：

    2026年2月27日

查看全文
  • Jin Li,
  • Chengyou Wang,
  • Xiao Zhou,
  • Yupeng Zhang

摘要

Face forgery techniques have advanced significantly, posing serious threats to digital media authenticity. This paper introduces a novel forgery-aware multi-scale cross-attention network (FMCA-Net) for robust face forgery detection. FMCA-Net employs a dual-branch architecture, integrating spatial and frequency domain features through cross-modal attention mechanisms. A forgery-aware token multi-scale attention module (FTMA) enhances discriminability, a spatial–frequency cross module (SFCM) enables bidirectional feature interaction, and a learnable spatial–frequency concatenation module (LSFC) ensures stable feature interaction. Extensive experiments on four public datasets demonstrate that FMCA-Net achieves superior performance, reaching an accuracy of 97.47% on FaceForensics++, 98.74% on Celeb-DF-v2, and 86.30% on WildDeepfake, and attaining an average AUC of 80.85% in cross-dataset evaluations, maintaining strong robustness under severe compression and unseen manipulations. Compared with 17 state-of-the-art methods, our findings contribute to the development of reliable face forgery detection methods in diverse conditions. The code is available at https://github.com/LiJin-sdu/FMCA-Net and is permanently archived with a DOI at https://doi.org/10.5281/zenodo.17596909.