Timezone: »

A Watermark for Large Language Models
John Kirchenbauer · Jonas Geiping · Yuxin Wen · Jonathan Katz · Ian Miers · Tom Goldstein

Wed Jul 26 07:24 PM -- 07:32 PM (PDT) @ Ballroom C
Event URL: https://github.com/jwkirchenbauer/lm-watermarking »

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.

Author Information

John Kirchenbauer (University of Maryland, College Park)
John Kirchenbauer

Researcher at the Center for Machine Learning (ml.umd.edu, umiacs.umd.edu), advised by Professor Tom Goldstein My current research focuses on understanding various aspects of model behavior including robustness, reliability, and efficiency in the discrete data domains of both natural language and graphs. I am interested in understanding how deep learning models generalize under distribution shift from an empirical and theoretical perspective through dataset curation, architectural changes, and novel analysis techniques. One overarching goal is the development of principled yet practical definitions of what "In" and "Out-of" distribution means in practice for these data domains.

Jonas Geiping (University of Maryland, College Park)
Yuxin Wen (University of Maryland)
Jonathan Katz (University of Maryland)
Ian Miers (Department of Computer Science, University of Maryland, College Park)
Tom Goldstein (University of Maryland)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors