Abstract: Face forensic detection is to distinguish manipulated from pristine face images. The main drawback of existing face forensics detection methods is their limited generalization ability due to differences in domains. Furthermore, artifacts such as imaging variations or face attributes do not persistently exist among all generated results for a single generation method.Therefore, in this paper, we propose a novel framework to address the domain gap induced by multiple deep fake datasets. To this end, the maximum mean discrepancy (MMD) loss is incorporated to align the different feature distributions. The center and triplet losses are added to enhance generalization. This addition ensures that the learned features are shared by multiple domains and provides better generalization abilities to unseen deep fake samples. Evaluations on various deep fake benchmarks (DFTIMIT, UADFV, Celeb-DF, and FaceForensics++) show that the proposed method achieves the best overall performance. An ablation study is performed to investigate the effect of the different components and style transfer losses.