Double Watermark for Large Language Models

Abstract

Detecting text generated by large language models (LLMs) with high accuracy is crucial for preventing the spread of fake news and misinformation caused by LLMs. Recently, digital watermark for auto-regressive language models has gained attention as a means of detecting text derived from LLMs. This approach embeds specific token patterns in text as a watermark by increasing token probabilities in a token group selected based on a single key. However, this approach cannot identify the source of text when the single key is leaked. To address this issue, we propose a double watermark which embeds two different watermarks with two corresponding keys in text so that the author of the text can be identified even after the first key is leaked. Our proposed method demonstrated the ability to detect a double watermark with high accuracy without significantly degrading the quality of the text.