Deepfake Detection and Localisation Based on Illumination Inconsistency

The rapid development of image synthesis technology has encouraged the spread of some fake news, making people gradually lose trust in digital media. The compression in the process of image propagation brings a major challenge to the existing face forgery detection method. In this paper, we propose a multi-task Deepfake detection method according to the motivation of illumination inconsistency between tampered and non-tampered areas. Specifically, we trained a Siamese network as a feature extractor to estimate the illumination, then distinguish the face image and predict the forged region through a U-shaped network. Our method has achieved great accuracy in classification tasks and can still maintain good performance in compressing data. In addition, we can also show the intensity of tampering while locating the forged area.