Black-Box Text Watermarking Algorithm towards Large Language Models

为应对大语言模型快速发展及其API接口日常生活使用带来的机器生成文本内容检测问题，本文提出了一种面向大语言模型生成文本的黑盒文本水印算法。该算法使用强化学习技术，通过设计的水印嵌入状态、水印嵌入智能体和水印嵌入奖励三个部分进行文本水印嵌入的训练，使得方法能够在水印可检测性、文本语义一致性和抗常规文本水印攻击上均表现出色。与其他算法相比，所提方法在平均准确率达到94.22%的同时，仍能保持更优的文本质量。所提出的算法不仅提升了文本水印的嵌入质量，也为保护模型使用和知识产权保障提供了支持。To address the detection issue of AI-generated text content arising from the rapid development of large language models and the daily usage of their API interfaces, this paper proposes a black-box text watermarking algorithm for text generated from the API of large language model. The algorithm employs reinforcement learning techniques to train text watermark embedding through three parts: watermark embedding statement, watermark embedding agent, and watermark embedding reward, which enables the good performance in watermark detectability, text semantic consistency, and resistance to conventional text watermarking attacks. The proposed algorithm has better text quality than other algorithms with an average detection accuracy of 94.22%. The proposed algorithm not only improves the embedding quality of text watermarking, but also provides significant support for protecting model usage and intellectual property rights.