Capturing the Lighting Inconsistency for Deepfake Detection

Abstract

AbstractThe rapid development and widely spread of deepfake techniques have raised severe societal concerns. Thus detecting such forgery contents has become a hot research topic. Many deepfake detection methods have been proposed in an artifacts-driven manner. They are well-designed to capture subtle artifacts of the face region in different domains. But since the lighting information is usually ignored during the forgery process, which may cause inconsistent lighting between the original face and forged one, we believe that this kind of semantic information can be useful to promote detection accuracy. In this paper, we propose a lighting inconsistency based deepfake detection method. We apply the color constancy technique to each sample and obtain a pre-processed image. Then the unique lighting information of each sample can be obtained by calculating the difference between the processed image and the original one. The lighting information will be used as an assistant channel for better detection accuracy. Extensive experiments show that our method can achieve obvious enhancements compared to the baseline method.KeywordsDeepfake detectionLight inconsistencyColor constancy