Watermark Smoothing Attacks against Language Models

ICLR 2025 Conference Submission13239 Authors

Keywords: LLM Watermark

Abstract: Statistical watermarking is a technique used to embed a hidden signal in the probability distribution of text generated by large language models (LLMs), enabling the attribution of the text to the originating model. We introduce the smoothing attack and show that existing statistical watermarking methods are not robust against minor modifications of text. In particular, with the help of a weaker language model, an adversary can smooth out the distribution perturbation caused by watermarks. The resulting generated text achieves comparable quality to the original (unwatermarked) model while bypassing the watermark detector. Our attack reveals a fundamental limitation of a wide range of watermarking techniques.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 13239