Search papers, labs, and topics across Lattice.
This paper introduces MarkNull, a model-agnostic attack that effectively removes watermarks from AI-generated images through on-manifold latent manipulation, addressing the limitations of existing methods that either target specific models or degrade image quality. By leveraging a statistical dependency between the latent representation and embedded noise, the authors develop the Noise-Latent Alignment Score (NLAS) to optimize watermark removal while maintaining semantic fidelity. Extensive evaluations reveal that MarkNull can reduce watermark accuracy to 53.14% without perceptible visual degradation, and its variant, MarkNull-A, achieves efficient performance with minimal computational overhead.
Achieving near-random watermark removal accuracy without sacrificing image quality, MarkNull challenges the robustness of current digital watermarking techniques.
Digital watermarking has emerged as a critical technique for provenance and copyright attribution in AI-generated imagery, yet its robustness against realistic, model-agnostic removal attacks remains poorly explored. Existing attacks either succeed only against specific generative models or achieve removal at the cost of severe visual degradation. In this paper, we propose MarkNull, a model-agnostic watermark removal attack via on-manifold latent manipulation. MarkNull is grounded in a key observation: watermarked images exhibit a strong statistical dependency between the generated latent representation and the embedded initial noise. To quantify this dependency, we introduce the Noise-Latent Alignment Score (NLAS) and formulate an optimization objective that selectively decorrelates the latent representation from the embedded watermark while preserving semantic fidelity. Extensive evaluations across different categories of watermarking paradigms, including post-hoc, fine-tuning-based, and initial-noise-based schemes, demonstrate that MarkNull reduces average bit accuracy to 53.14%, approaching random-guessing (50%), without perceptible image degradation. To further improve scalability, we propose MarkNull-A, an amortized, optimization-free variant that distills the attack into a single forward pass, achieving 0.50 s/image with modest computational overhead. Notably, our attacks successfully compromise Google's SynthID-Image system while preserving high visual quality and transfer effectively to video watermarking. Finally, we present an attack detection mechanism as a defensive counterpart to MarkNull and MarkNull-A, highlighting the necessity of developing watermark designs resilient to model-agnostic latent-space attacks.
