Publicly-Detectable Watermarking for Language Models
Machine Learning
2025-01-07 v4 Computation and Language
Cryptography and Security
Abstract
We present a publicly-detectable watermarking scheme for LMs: the detection algorithm contains no secret information, and it is executable by anyone. We embed a publicly-verifiable cryptographic signature into LM output using rejection sampling and prove that this produces unforgeable and distortion-free (i.e., undetectable without access to the public key) text output. We make use of error-correction to overcome periods of low entropy, a barrier for all prior watermarking schemes. We implement our scheme and find that our formal claims are met in practice.
Keywords
Cite
@article{arxiv.2310.18491,
title = {Publicly-Detectable Watermarking for Language Models},
author = {Jaiden Fairoze and Sanjam Garg and Somesh Jha and Saeed Mahloujifar and Mohammad Mahmoody and Mingyuan Wang},
journal= {arXiv preprint arXiv:2310.18491},
year = {2025}
}