Reproduce “Stable Signature is Unstable: Removing Image Watermark from Diffusion Models” 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
Yuepeng Hu, Zhengyuan Jiang, Moyang Guo, Neil Gong
Researchers from Duke University developed a model-targeted attack that effectively removes in-generation watermarks from Stable Signature-protected diffusion models, achieving over 94% evasion rate with low FID scores (below 27.5), thereby challenging the robustness claims of this watermarking framework while maintaining high image quality.
Key insights
- In-generation watermarking schemes, despite claims of robustness and deep integration, are vulnerable to strategic fine-tuning attacks that do not compromise the utility (visual quality) of generated images.
- A two-stage optimization approach for estimating denoised latent vectors is crucial for accurate reconstruction and maintaining visual quality, especially in scenarios where the attacker lacks full model access.
- The combination of MSE, perceptual, and adversarial losses in fine-tuning is vital for achieving both high watermark evasion and excellent image quality, outperforming methods that prioritize only one aspect.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.