Attack-resistant watermarking for AIGC image forensics: diffusion-based semantic deflection

The researchers propose a training-free intrinsic watermarking framework, PAI, for AIGC image copyright protection, which can be plug-and-play for diffusion-based AIGC services. To address the difficulty existing watermarking methods have in balancing anti-forgery and anti-removal attacks and their lack of support for semantic-level tampering localization, PAI simultaneously provides robust ownership verification, attack detection and semantic-level tampering localization. Its key-conditioned deflection mechanism guides the denoising trajectory of the diffusion model according to the user key, enhancing the semantic entanglement of identity and content through trajectory-level coupling, and it is theoretically proven that only a valid key can pass verification. In experiments with 12 attack methods, PAI achieves a verification accuracy of 98.43%, an average improvement of 37.25% over SOTA methods, and maintains strong tampering localization performance against advanced AIGC editing. Code is open-sourced at https://github.com/QingyuLiu/PAI。