Low-complexity fake face detection based on forensic similarity

In recent years, face synthesis and manipulation technology have been developed rapidly, and now it is feasible to synthesize extremely realistic fake face videos, which can easily deceive existing face recognition systems. Due to the high quality of fake videos, allowing fake face videos to propagate on Internet may cause serious ethical, moral and legal problems. Therefore, the effective and reliable detection method is urgently needed to distinguish fake face videos. We notice that existing face forgery methods commonly extract face area of each frame first and perform manipulations only on face areas while background areas remain unchanged. Therefore, the difference between the face area and the background area in a forged face frame is significantly larger than the difference between the face area and the background area in the corresponding unforged frame. In this paper, based on such observation, we propose a new detection method-forensic similarity method-which judges the authenticity of face video frames by detecting the difference in similarity between the face area and the background area. For evaluation, we conduct training and testing on FaceForensics++ dataset, and evaluate the generalization capability on Celeb-DF dataset. From the experimental results, we can find that the proposed method has a better or comparable performance, especially in the term of generalization capability. Compared with Xception, our model can attain 8-12% accuracy gains under Celeb-DF dataset. In addition, our model has lower complexity than Xception.