The paper addresses issues arising from deepfake face swapping, which despite offering convenience and entertainment, has led to serious privacy issues. Existing models usually face performance issues due to the high-quality synthetic images used in deepfake face swapping. The paper proposes a robust identity perceptual watermarking framework that can effectively detect and trace the source of deepfake face swapping. The watermarks are encoded and recovered by jointly training an encoder-decoder framework. The proposed model demonstrates superior performance in both cross-dataset and cross-manipulation settings.
Publication date: 3 Nov 2023
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2311.01357
