Abstract: To address the problems of insufficient feature representation, low robustness, and weak generalization in handling diverse forgery types and low-quality images, a face forgery detection method based on tri-branch feature extraction (Tri-Branch Feature Extraction Network, Tri-BranchNet) was proposed to enhance forgery trace representation and detection performance. The architecture was designed with three branches: 1) global semantics were captured by using Vision Transformer (ViT); 2) local texture modeling was improved by Invertible Neural Network (INN); 3) an edge-aware branch was constructed to extract boundary-level features. Experimental results show that the proposed method achieves 98.75% accuracy on FaceForensics++ (C23), outperforming F3-Net (Frequency in Face Forgery Network) and CORE (COnsistent REpresentation learning) by 1.23% and 1.14%, respectively. In cross-compression and cross-dataset evaluations, AUC (Area Under Curve) scores reach 85.26% and 81.09% on C40 and Celeb-DF, demonstrating strong robustness and generalization. The proposed tri-branch fusion mechanism significantly enhances detection accuracy and provides a novel approach to multi-dimensional feature modeling.

Key words: face forgery detection, multi-branch network, Vision Transformer (ViT), invertible neural network, edge feature

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)，以实现多类型特征互补与融合，提升伪造痕迹的表征能力和模型的检测性能。该方法具体架构为：1)利用Vision Transformer(ViT)捕获全局语义表征；2)引入可逆神经网络(INN)以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1.23%和1.14%；在跨压缩率与跨数据集测试中，曲线下面积(AUC)值分别达到85.26%(C23→C40)、81.09%(Celeb-DF)，显示出良好的鲁棒性与泛化性能。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 (ViT), 可逆神经网络, 边缘特征