首页|期刊导航|网络与信息安全学报|面向大语言模型的鲁棒多比特水印算法

面向大语言模型的鲁棒多比特水印算法OA

Robust multi-bit watermarking algorithm for large language models

中文摘要英文摘要

随着大语言模型(large language model,LLM)的快速发展,面向模型生成内容嵌入可识别信息的文本水印技术逐渐成为研究热点.现有方法通常依赖前一个词汇构造随机种子,并以伪随机方式确定水印嵌入位置及词汇划分策略,进而实现水印嵌入.然而,这些方法高度依赖局部词汇信息,使得在同义词替换等文本攻击下的鲁棒性受限.为此,提出了一种基于文本语义的鲁棒多比特水印(semantic-oriented robust multi-bit watermarking,SRMW)算法.首先使用预训练的嵌入模型来提取文本的语义表示.在此基础上,不同于现有依赖局部词汇信息构建随机种子的方法,提出了一种水印映射模型(watermark projection model,WPM)来基于文本语义表示构建随机种子,进而伪随机确定水印嵌入位置选择及词汇划分策略.此外,设计了重建损失和语义一致性损失来确保所构建随机种子的泛化性和鲁棒性.在3个LLM和2个数据集上的实验结果验证了所提出的SRMW算法在各种攻击场景(如同义词替换和词汇删除场景)下的鲁棒性.

With the rapid development of large language models(LLM),text watermarking techniques that embed identifiable information into LLM-generated content have gradually become a hot research topic.Existing methods typically rely on the preceding token to generate a random seed,and use a pseudorandom approach to determine watermark embedding positions and partition the vocabulary,enabling watermark injection.Despite the achieved progress,existing methods still exhibit limited robustness against various text attacks,such as synonym substitution attacks.To this end,a semantic-oriented robust multi-bit watermarking(SRMW)algorithm was proposed.Specifi-cally,a pre-trained embedding model was employed to extract semantic representations of text.Then,different from existing methods that rely on local word-level information to generate random seeds,a semantic-driven water-mark projection model(WPM)was designed to construct random seeds based on text semantic representations,thereby pseudo-randomly determining watermark embedding positions and vocabulary partitioning strategies.Fur-thermore,a reconstruction loss and semantic consistency loss were designed to ensure the generalizability and ro-bustness of the constructed random seeds.Extensive experiments on three LLMs and two datasets demonstrated that the proposed SRMW achieved superior robustness compared to existing methods under various attack sce-narios,including synonym substitution and word deletion.

周海林;李博;张琨;洪日昌

合肥工业大学计算机与信息学院,安徽 合肥 230029安徽大学人工智能学院,安徽 合肥 230601合肥工业大学计算机与信息学院,安徽 合肥 230029合肥工业大学计算机与信息学院,安徽 合肥 230029

信息技术与安全科学

多比特水印大语言模型信息隐藏

multi-bit watermarkingLLMinformation hiding

《网络与信息安全学报》 2026 (4)

62-75,14

中央高校基本科研专项资金(No.JZ2025HGTG0289,No.PA2025IISL0114)The Fundamental Research Funds for the Central Universities(No.JZ2025HGTG0289,No.PA2025IISL0114)

10.11959/j.issn.2096-109x.AQ26114

评论