Watermarking Matters for Deepfake Detection: A Proactive Method for Detecting Forgeries Under Conventional Attacks

Face deepfake detection is a crucial technique for verifying the authenticity of facial media content and has been a long-standing focus of research in the field of multimedia forensics. However, due to the increasing diversity of forgery techniques and the declining visual quality of forgeries, achieving consistently reliable detection results remains a significant challenge. To tackle this challenge, we propose a proactive approach that integrates watermarking techniques with knowledge distillation to enable robust and accurate deepfake detection. The proposed method is grounded in two key ideas: First, watermark information is embedded into the Quaternion Continuous Orthogonal Moments (QCOMs) space to balance imperceptibility and robustness. Specifically, during the testing phase, the integrity of the watermark remains intact under conventional attacks but is disrupted when deepfakes are introduced. Second, a dual-task learning framework is designed, consisting of a watermark extractor and a deepfake detector. Within this framework, lightweight multiscale adapters are introduced to transfer the forgery-related clues extracted by the watermark extractor to the detector, while preserving the independence of both components. Experimental results on benchmark datasets demonstrate that the proposed method achieves state-of-the-art performance in cross-domain deepfake detection, exhibiting particularly strong robustness against various conventional attacks.