Recently, the rapid progress in facial image manipulation raises significant concerns on the implication on society. These facial manipulations such as face swapping and face merging could cause serious social, political, or commercial problems. There is an urgent need for effective methods to expose fake face images. While these facial image manipulation methods show stunning results, there are artifacts could be exposed. We find CNN-based networks can effectively distinguish the fake face images from the real ones.
In this presentation, we will present two effective methods and describe how low-level and high-level features are used to automatically detect facial manipulation frames. Firstly, we use a simple yet effective CNN architecture with several convolutional layers to build a powerful Deepfakes detector showing a prediction accuracy of 99% for Deepfakes and 94% for face merging. Secondly, we find a FaceNet based method is an effective binary classifier. FaceNet is one of the state-of-the-art convolutional neural networks designed for face recognition, producing embeddings catching face features. We use these features to train an SVM classifier reaching an accuracy rate of 94% for Deepfakes and 76% for face merging among our test.