A Watermark for Large Language Models

doi:10.48550/arXiv.2301.10226

Paper

A Watermark for Large Language Models

Published Jan 24, 2023 · John Kirchenbauer, Jonas Geiping, Yuxin Wen

UNKNOWN SJR score

335

Citations

77

Influential Citations

Full text

Semantic Scholar

Abstract

Potential harms of large language models can be mitigated by watermarking model output, i.e., embedding signals into generated text that are invisible to humans but algorithmically detectable from a short span of tokens. We propose a watermarking framework for proprietary language models. The watermark can be embedded with negligible impact on text quality, and can be detected using an efficient open-source algorithm without access to the language model API or parameters. The watermark works by selecting a randomized set of"green"tokens before a word is generated, and then softly promoting use of green tokens during sampling. We propose a statistical test for detecting the watermark with interpretable p-values, and derive an information-theoretic framework for analyzing the sensitivity of the watermark. We test the watermark using a multi-billion parameter model from the Open Pretrained Transformer (OPT) family, and discuss robustness and security.

Highly Cited

Study Snapshot

Key takeawayThis paper proposes a watermarking framework for proprietary language models, which can be embedded with negligible impact on text quality and detected using an efficient open-source algorithm without access to the language model API or parameters.

PopulationOlder adults (50-71 years)

Sample size24

MethodsObservational

OutcomesBody Mass Index projections

ResultsSocial networks mitigate obesity in older groups.

Sign up to use Study Snapshot

Consensus is limited without an account. Create an account or sign in to get more searches and use the Study Snapshot.

Create an account

Paper

A Watermark for Large Language Models

Sign up to use Study Snapshot

References

Citations

Humanizing the Machine: Proxy Attacks to Mislead LLM Detectors

Our proxy-attack strategy effectively compromises large language models, producing outputs that align with human-written text and misleads detection systems, while maintaining the generation quality of the attacked models.

Provably Robust Watermarks for Open-Source Language Models

Our watermarking scheme for open-source language models is unremovable and can be detected from the model's outputs, making it a valuable tool for identifying AI-generated text.