Text Zero-Watermarking Algorithm Based on Semantic Undirected Weighted Graph

在生成式语言模型兴起的今天，人工智能为文本创作和传播带来了前所未有的变革，但是生成式语言模型的广泛应用也带来了版权保护的问题。本研究基于文本的语义特征，提出了一种创新的文本零水印算法，通过语义相似度编码模型将文本的基础粒度编码为高维向量，接着利用文本粒度的高维语义嵌入向量的方向各异性，构建文本语义特征图，对文本特征进行相关性分析实现相似度的评估。经实验证明，本文所提出的零水印算法，在误判率方面的表现较好；在鲁棒性上，对同义改写和文本添加攻击具有良好的抵抗力，对文本的删除攻击具有一定的鲁棒性。With the rise of generative language models, artificial intelligence has brought unprecedented changes to text creation and dissemination, but the widespread application of generative language models has also brought the problem of copyright protection. Based on the semantic features of the text, this study proposes an innovative text zero watermark algorithm, which encodes the basic granularity of the text into high-dimensional vectors through the semantic similarity coding model, and then uses the directional heterogeneity of the high-dimensional semantic embedding vectors of the text granularity to construct a text semantic feature map, and analyzes the relevance of the text features to achieve similarity evaluation. Experiments show that the zero-watermark algorithm proposed in this paper has a better performance in terms of false positive rate. In terms of robustness, it has good resistance to synonymous rewriting and text addition attacks, and has a certain robustness to text deletion attacks.