Detecting facial manipulated images via one-class domain generalization

Nowadays, numerous synthesized images and videos generated by facial manipulated techniques have become an emerging problem, which promotes facial manipulation detection to be a significant topic. Much concern about the use of synthesized facial digital contents in society is rising due to their deceptive nature and widespread. To detect such manipulated facial digital contents, many methods have been proposed. Most detection methods focus on specific datasets. It is hard for them to detect facial images or videos manipulated by unknown face synthesis algorithms. In this paper, we propose a method to improve the generalization ability of the facial manipulation detection model using one-class domain generalization. We shape the problem into domain generalization. We divide the dataset into several domains according to different manipulation algorithms. We also try to process the images from the perspective of frequency domain. We utilize two-dimensional wavelet transform to preprocess the images to ensure the effect on compressed images. The results of experiments implemented on FaceForensics++ dataset exceed the baselines and recent works. The feature visualization analyses intuitively show that our method can learn robust feature representation that can be generalized to unseen domains.