LOCAT: Localization-Driven Text Watermarking via Large Language Models

Abstract

The rapid advancement of large language models (LLMs) has raised concerns regarding potential misuse and underscores the importance of verifying text authenticity. Text watermarking, which embeds covert identifiers into generated content, offers a viable means for such verification. Such watermarking can be implemented either by modifying the generation process of an LLM or via post-processing techniques like lexical substitution, with the latter being particularly valuable when access to model parameters is restricted. However, existing lexical substitution-based methods often face a trade-off between maintaining text quality and ensuring robust watermarking. Addressing this limitation, our work focuses on enhancing both the robustness and imperceptibility of text watermarks within the lexical substitution paradigm. We propose a localization-based watermarking method that enhances robustness while maintaining text naturalness. First, a precise localization module identifies optimal substitution targets. Then, we leverage LLMs to generate contextually appropriate synonyms, and the watermark is embedded through binary-encoded substitutions. To address different usage scenarios, we focus on the trade-off between watermark robustness and text quality. Compared to existing methods, our approach significantly enhances watermark robustness while maintaining comparable text quality and achieves similar robustness levels while improving text quality. Even under severe semantic distortions, including word deletion, synonym substitution, polishing, and re-translation, the watermark remains detectable.