Facial Deepfake Detection Using Gaussian Processes

Facial deepfake detection involves detecting images and videos with tampered faces. In this paper, we automatically detect four types of deepfakes: Deepfake, Face2Face, FaceSwap and Neural Textures. From a deepfake video, we extract the faces in its image frames to serve as input. Given a facial image, we preprocessed the image by capturing the different compression levels, extracted features using a feature extraction backbone that focuses on mesoscopic properties, and we make use of Gaussian Processes (GPs) for binary classification because GPs inherently provide more accurate uncertainty estimation which leads to lower prediction error and higher accuracy. To the best of our knowledge, we are the first to apply GPs in deepfake detection. The proposed method was compared with state-of-the-art baseline methods; we performed a cross-dataset evaluation and observed no significant accuracy difference between this approach and the baseline. Also, we experimented on varying dataset sizes and the results show that our method has a competitive accuracy on large datasets and outperforms on small datasets. The implementation is available at: https://github.com/harmz123/FacialDeepfakeDetection