Robust Facial Manipulation Detection via Domain Generalization

Abstract

Face generation and forgery algorithms are available on the Internet, which promotes facial manipulation detection to be an important topic. Recently, many methods have been presented to detect facial manipulation images and videos. Most of which focus on specific datasets and achieve promising results on them. However, it is hard for them to detect the facial images manipulated by unknown face synthesis algorithms. In this paper, we present a method to improve the generalization ability of the detection models using one class domain generalization. Unlike the methods using datasets to train deep neural networks directly, we propose to shape the problem to domain generalization. The images manipulated by different algorithms are regarded as different domains. To obtain domain-invariant features, we take the fake facial images from multiple domains into the domain discriminator for domain adversarial training. The models can discriminate between the real and fake facial images from different domains, even the fake images generated by unknown algorithms. The experiments implemented on FaceForensics++ dataset demonstrate that the proposed method achieves outstanding performance and improves the robustness of the detection models.