《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (4): 1292-1299.DOI: 10.11772/j.issn.1001-9081.2025040461

• 多媒体计算与计算机仿真 • 上一篇    

基于三分支特征提取的人脸伪造检测方法

许盛伟, 王健波(), 韩季杰, 白怡婕   

  1. 北京电子科技学院,北京 100070
  • 收稿日期:2025-04-27 修回日期:2025-07-30 接受日期:2025-08-04 发布日期:2025-08-15 出版日期:2026-04-10
  • 通讯作者: 王健波
  • 作者简介:许盛伟(1976—),男,江西吉安人,教授,博士,主要研究方向:大数据安全、人工智能与密码应用
    韩季杰(2001—),男,江西上饶人,硕士研究生,主要研究方向:网络流量分类、密码应用、大数据安全
    白怡婕(2001—),女,山西临汾人,硕士研究生,主要研究方向:密码安全分析、密码应用和评估。
  • 基金资助:
    国家重点研发计划项目(2022YFB3104402)

Face forgery detection method based on tri-branch feature extraction

Shengwei XU, Jianbo WANG(), Jijie HAN, Yijie BAI   

  1. Beijing Electronic Science and Technology Institute,Beijing 100070,China
  • Received:2025-04-27 Revised:2025-07-30 Accepted:2025-08-04 Online:2025-08-15 Published:2026-04-10
  • Contact: Jianbo WANG
  • About author:XU Shengwei, born in 1976, Ph. D., professor. His research interests include big data security, AI-driven cryptographic applications.
    HAN Jijie, born in 2001, M. S. candidate. His research interests include network traffic classification, cryptographic applications, big data security.
    BAI Yijie, born in 2001, M. S. candidate. Her research interests include cryptographic security analysis, cryptographic applications and evaluation.
  • Supported by:
    National Key Research and Development Program of China(2022YFB3104402)

摘要:

针对现有检测方法在应对多样化伪造方式和低质量图像时存在的特征表达不足、鲁棒性差和跨域泛化能力弱等问题,提出一种基于三分支特征提取的人脸伪造检测方法Tri-BranchNet(Tri-Branch feature extraction Network),以实现多类型特征的互补与融合,并提升伪造痕迹的表征能力和模型的检测性能。具体架构为:1)利用ViT(Vision Transformer)捕获全局语义表征;2)引入可逆神经网络(INN)增强局部纹理特征的建模能力;3)设计边缘特征提取分支解决传统模型对边界伪造区域特征提取不足的问题。在多个公开数据集上的实验结果表明,所提方法在FaceForensics++(C23)数据集上的准确率达98.75%,相较于F3-Net(Frequency in Face Forgery Network)和CORE(COnsistent REpresentation learning)分别提升了1.26%和1.17%;在跨压缩率与跨数据集测试中,所提方法的曲线下面积(AUC)值分别达到85.26%(C40)和81.09%(Celeb-DF),显示出良好的鲁棒性与泛化性能。可见,所提三分支融合机制在复杂伪造场景下能显著提升检测准确率,为伪造图像的多维度特征建模提供了一种新思路。

关键词: 人脸伪造检测, 多分支网络, Vision Transformer, 可逆神经网络, 边缘特征

Abstract:

To address the problems of insufficient feature representation, poor robustness, and weak cross-domain generalization in handling diverse forgery types and low-quality images, a face forgery detection method based on tri-branch feature extraction, Tri-BranchNet (Tri-Branch feature extraction Network), was proposed to achieve the complementarity and integration of multiple types of features, and enhance forgery trace representation and model’s detection performance. The architecture was designed as: 1)global semantic representation were captured by using Vision Transformer (ViT); 2)local texture feature modeling ability was improved by introducing Invertible Neural Network (INN); 3)an edge feature extraction branch was designed to solve the problem that traditional models inadequately extracted features from boundary forgery regions. Experimental results on multiple public datasets show that the proposed method achieves 98.75% accuracy on FaceForensics++ (C23) dataset, outperforming F3-Net (Frequency in Face Forgery Network) and CORE (COnsistent REpresentation learning) by 1.26% and 1.17%, respectively. In cross-compression and cross-dataset tests, the proposed method has the Area Under Curve (AUC) scores reached 85.26% and 81.09% on C40 and Celeb-DF, respectively, demonstrating strong robustness and generalization. It can be seen that the proposed tri-branch fusion mechanism enhances detection accuracy in complex forgery environments significantly and provides a novel idea for multi-dimensional feature modeling of forgery images.

Key words: face forgery detection, multi-branch network, Vision Transformer (ViT), Invertible Neural Network (INN), edge feature

中图分类号: