DRIFT method reliably removes hidden watermarks from AI-generated images
DRIFT: Removing Diffusion Watermarks by Deflecting the Generative Trajectory
Cryptography and SecurityMachine Learning
Summary
Watermarks are hidden signals placed inside AI-generated images to prove they are authentic. Existing ways to remove these watermarks either guess the same generation path or tweak each image one at a time, which is not very effective. The researchers behind DRIFT found that these watermarks rely on specific steps in the image-making process. Their method partially adds noise and then reverses the process in random ways to confuse the watermark detection, working well without knowing secret keys or needing complicated adjustments for each image. This approach successfully removes a wide variety of watermarks while keeping the images looking good.
diffusion watermarkinggenerative trajectoryforward diffusionreverse resamplingblack-box attacklatent structureimage fidelityWasserstein boundsadaptive refinement
Authors
Rui Bao, Zheng Gao, Xiaoyu Li, Xiaoyan Feng, Yang Song, Jiaojiao Jiang
Abstract
Diffusion watermarking embeds verifiable signals into the generative process and commonly verifies them by recovering trajectory-dependent evidence, making the marks robust to conventional pixel-space distortions. Existing removal attacks either regenerate along deterministic trajectories, which often preserve the watermark-bearing latent structure, or optimize every image separately. We identify the reliance on a recoverable generative trajectory as a common attack surface among the schemes we study. Based on this observation, we propose DRIFT, a black-box attack that combines partial forward diffusion with stochastic reverse resampling. Forward re-noising limits source information available to a fixed-depth recovery pipeline, while stochastic reversal supplies alternative noise-driven paths whose removal benefit we isolate through matched sampler comparisons. Adaptive DRIFT searches a selected ladder for each image's first verifier-rejected rung and refines fidelity while retaining only updates rejected by the same verifier. At fixed depth, we derive information-theoretic and Wasserstein source-dependence bounds; under realized-ladder monotonicity, the first rejected rung is least distorted among rejected rungs on that ladder, and verifier-gated refinement preserves rejection. Across nine watermarks spanning three paradigms, DRIFT achieves 98-100% attack success and the best image quality among the compared attacks, without secret keys, verifier internals, or per-image gradient optimization.