Deepfake Detection Using Graph Representation with Multi-dimensional Features
Jia Chen, Weiguo Lin, Junfeng Xu · 2023
The proliferation of fake video poses a significant threat to the authority and authenticity of news across multiple domains. The most existing methods of deepfake detection primarily focus on identifying the face as a whole in a video, ignoring the correlation between the facial components. However, our investigation indicates that constituent potions of a face have different effects in deepfake detection. To address this issue, we divided the face in a video frame into several regions and explored the relationship between these regions. Our approach involves constructing a feature graph of this correlation, aiming to make use of the relationship and temporal characteristics between regions of a face in a deepfake video. To begin with, the features of each facial region are extracted through CNN. Subsequently, the feature graph of the entire video is constructed with these features being the vertices and the correlation being the edge. A graph neural network is finally utilized to determine whether the video has been tampered with. Our experiments on several publicly accessible datasets demonstrate that the proposed approach outperforms other state-of-the-art deepfake detection techniques in most scenarios.