MarkNull: A New Model-Agnostic Watermark Removal Attack for AI Images
A recent study published on arXiv (2608.10166) presents MarkNull, a model-independent method for removing watermarks from images created by AI. This method takes advantage of a statistical link between the latent representation of watermarked images and the original embedded noise. The researchers introduce a Noise-Latent Alignment Score (NLAS) to measure this connection and establish an optimization goal that effectively decorrelates the watermark from the latent representation while maintaining semantic integrity. Comprehensive assessments across various watermarking types confirm the attack's success. The findings underscore the susceptibility of existing watermarking methods to realistic removal strategies, raising significant concerns regarding provenance and copyright in AI-generated visuals.
Key facts
- Paper arXiv:2608.10166 introduces MarkNull, a model-agnostic watermark removal attack.
- MarkNull uses on-manifold latent manipulation to remove watermarks.
- The attack is based on the statistical dependency between latent representation and initial noise.
- A Noise-Latent Alignment Score (NLAS) is introduced to quantify this dependency.
- The optimization objective decorrelates latent representation from watermark while preserving semantic fidelity.
- Extensive evaluations across different watermarking categories are reported.
- The paper addresses the gap in robustness of watermarking against model-agnostic removal attacks.
- Existing attacks are either model-specific or cause severe visual degradation.
Entities
Institutions
- arXiv