Watermarking Makes Language Models Radioactive

TL;DR

This paper demonstrates that watermarked language models leave detectable traces, allowing reliable identification of training data contamination even at low watermarked data proportions, unlike previous methods.

Contribution

The authors introduce a novel detection method for watermarked language models, providing provable guarantees and linking radioactivity levels to watermark robustness and training conditions.

Findings

Detection confidence exceeds 99.999% with 5% watermarked data

Watermarked models exhibit detectable residual signals

Detection is reliable even with minimal watermarked training data

Abstract

We investigate the radioactivity of text generated by large language models (LLM), i.e. whether it is possible to detect that such synthetic input was used to train a subsequent LLM. Current methods like membership inference or active IP protection either work only in settings where the suspected text is known or do not provide reliable statistical guarantees. We discover that, on the contrary, it is possible to reliably determine if a language model was trained on synthetic data if that data is output by a watermarked LLM. Our new methods, specialized for radioactivity, detects with a provable confidence weak residuals of the watermark signal in the fine-tuned LLM. We link the radioactivity contamination level to the following properties: the watermark robustness, its proportion in the training set, and the fine-tuning process. For instance, if the suspect model is open-weight, we demonstrate that training on watermarked instructions can be detected with high confidence ($p$-value $< 10^{-5}$) even when as little as $5\%$ of training text is watermarked.