Reproduce “Attack-Resistant Watermarking for AIGC Image Forensics via Diffusion-based Semantic Deflection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qingyu Liu, Yitao Zhang, Zhongjie Ba, Chao Shuai, Peng Cheng, Tianhang Zheng, Zhibo Wang
PAI is a training-free, diffusion-native watermarking framework that links AIGC images to user keys while supporting ownership verification, attack detection, and semantic-tampering localization. Across common degradations, removal and spoofing attacks, and generative edits, it reports up to 100% verification or attack-detection accuracy, with a tampering-localization F1 score of 77.05% and only about 56 ms of additional generation time over standard DDIM sampling.
Key insights
- Coupling the watermark to both the initial noise and early denoising trajectory makes the signal dependent on the generated image’s semantic formation rather than on a standalone post-processing stage.
- Removal and spoofing attacks can have similar scalar watermark magnitudes but different directions in bias space, allowing multidimensional statistical modeling to separate them while retaining ownership attribution.
- The framework treats provenance as more than a yes/no ownership decision by combining identity verification, attack classification, and localization of semantically rewritten regions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.