Deep visual and audio watermarking for AI video editing forensics

Deep visual and audio watermarking for AI video editing forensics

This paper proposes V2A-Mark, a method to embed invisible visual and audio watermarks into videos to enable precise tamper localization and copyright protection. It addresses limitations of current forensic methods like poor generalizability and singular functionality. V2A-Mark combines fragile video steganography with robust watermarking, using temporal alignment and cross-modal extraction between video and audio. Experiments show it achieves superior localization precision and copyright accuracy compared to other methods, without needing retraining for new tamper types.