A Diffusion-Based Approach for Restoring Face-swapped Images

The misuse of deepfake technology underscores the critical importance of deepfake forensics. Current research in this domain primarily employs binary classification models that are limited to detecting the authenticity of facial images, and are incapable of tracing the original face replaced in the face-swapping process. Consequently, this limitation results in the inability to provide substantial judicial evidence.To address this issue, this paper introduces a novel forensic methodology, which aims to reconstruct the original face from the forged face. Specifically, we propose a novel framework that reformulates the restoration process from the forged face to the original face as a special image editing task guided by a series of conditions. Our research demonstrates that forged face images after face-swapping still retain some features of the original face, and these features can be used to restore the original face.Furthermore, we also construct the first dataset and set up evaluation metrics for this important research problem. Extensive experiments demonstrate the qualitative and quantitative effectiveness of our proposed framework.