An Efficient and Flexible Black-Box Watermarking Framework for Large Language Models

To prevent misuse of text generated by large language models (LLMs), watermarking technology offers a straightforward and effective means of verification. Existing approaches, however, suffer from several limitations, including a high dependency on the underlying logic of LLMs, inadequate performance against sophisticated watermarking attacks, and poor stealth, making them less suitable for complex and dynamic real-world applications. To address these challenges, we propose an efficient and flexible black-box watermarking framework for large language models. This framework enhances robustness and stealth by embedding watermarks in key areas of the text through the identification and sampling of high-entropy words. Furthermore, we introduce an encryption-based encoding algorithm that allows users to verify watermarks with their personal keys, supporting the customization of watermarking algorithms. Comparative experiments on four datasets with five baseline methods demonstrate that the proposed method excels in accuracy while significantly improving robustness and concealability with minimal semantic changes.