Watermarking for User Identification in Large Language Models

Zihao Fu, Chris Russell

Keywords: Large Language model; Watermark

TL;DR: We propose a watermarking method that simultaneously detects AI-generated text and identifies the LLM user without increasing false positives as the number of users grows, deriving theoretical bounds and discussing ethical issues.

Abstract:

We identify a new task for watermarking -- namely the simultaneous identification of text as being automatically generated alongside the identification of the LLM user.
We show that a naïve approach that treats a text as artificially generated if a user is correctly identified is prone to problems of false positives arising from multiple hypothesis comparison.
We propose a novel approach (Our code is submitted with the supplementary material. We will also open it on Github after the anonymity period.) that retains almost similar rates as the number of users increase. We derive theoretical bounds that support our experimental approach.

Primary Area: foundation or frontier models, including LLMs

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.

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: 2274