Representation learning based deepfake detection with facial region features

As the Deepfake phenomena have become famous today, a more stabilize and versatile Generative Adversarial Network (GAN) architect was purposed every single year. Recently, there is some state-of-art GAN architect that able to face-swap, expression-swap, voice- swap, or crafting a new face on the video by providing just a source image and a target video. It has become a real threat for celebrities and world leaders to fear that the high-realistic Deepfake video may use as a way to defame them in one day. The goal of this study is to propose an improved representation learning framework based Deepfake detector to allow the model to discover the representations need for feature detection automatically. In this work, it reviews the latest GAN architect, face manipulation type, and related work of face manipulation detector. Then addressed the fact that overfitting and less generalization happened in most of the face manipulation detectors. Overall, the study will benefit the Deepfake detector in generalizability and detecting the unseen GAN generated images.