Exploiting optimized forgery representation space for general fake face detection

Gaoming Yang, Bang Zuo, Xianjin Fang, Ji Zhang

Abstract: Face forgery has become more realistic with deep learning in computer vision, posing a significant challenge to trustworthy face identification. Existing works have achieved considerable accuracy within the dataset by formulating the detection as a binary classification problem. These methods attempt to amplify the category differences between real and fake faces but ignore the optimization of representation space for learning the specific forgery information within samples, which results in the intra-class distribution collapse and poor generalization in unseen domains. To mitigate this issue, we propose a novel forgery detection framework that combines contrastive learning with supervised learning, named Contrastive Learning Against face Forgery (CLAF). Specifically, a dual branch learning framework is involved in extracting the consistent forgery feature distribution first. Then, we consider the similarity, variance, and covariance constraint term for the representation space, which can better preserve the specific forgery information within each sample for generalization detection. The generalization performance is confirmed on FaceForensics++, Celeb-DF, and DFDC. Extensive experiment results demonstrate the effectiveness of our framework in improving generalization.

External IDs:dblp:journals/paa/YangZFZ25