基于语义的鲁棒文本水印算法  

Semantic-based robust text watermarking algorithm

作　　者：张琨 李博[2] 陈希 杨晓依 吴乐[1,3] 洪日昌 ZHANG Kun;LI Bo;CHEN Xi;YANG Xiaoyi;WU Le;HONG Richang(School of Computer and Information,Hefei University of Technology,Hefei 230029,China;School of Artificial Intelligence,Anhui University,Hefei 230601,China;Institute of Dataspace,Hefei Comprehensive National Science Center,Hefei 230088,China)

机构地区：[1]合肥工业大学计算机与信息学院,安徽合肥230029 [2]安徽大学人工智能学院,安徽合肥230601 [3]合肥综合性国家科学中心数据空间研究院,安徽合肥230088

出　　处：《大数据》2024年第6期49-61,共13页Big Data Research

基　　金：国家自然科学基金资助项目(No.62436003)。

摘　　要：文本水印算法能够确定文本数据的版权归属,进而促进数据安全流通和共享。现有文本水印算法通常预先对原始文本中的词汇进行标记并采用词汇替换的方法来注入水印。然而,这些算法仅基于原始文本词汇的前一个词汇的哈希值来标记当前词汇,限制了水印算法的鲁棒性。为了解决这一问题,提出了SRTW算法。具体而言,SRTW算法首先利用现有的嵌入模型获取文本语义嵌入;其次,通过训练的词汇标记模型将这些文本语义嵌入转换为词汇标记(-1或1);最后,选择标记为1的词汇替换原词汇来注入水印。与现有的较先进的基准方法相比,提出的SRTW算法在3种不同攻击场景下AUC指标分别提高了2.08%、5.17%和3.09%,充分证明了SRTW算法的有效性。Text watermarking can determine the copyright ownership of text data,facilitating secure circulation and sharing of data.Existing text watermarking algorithms typically pre-mark words and employ word substitution methods to embed watermarks.However,these algorithms only mark candidate words based on the hash value of the previous word,limiting the robustness of the watermarking algorithm.To address this issue,SRTW algorithm was proposed.Specifically,semantic embeddings of the text were obtained using existing embedding models.Then,these embeddings were converted into word markers(-1 or 1)through a trained word marking model.Finally,words marked as 1 were selected to replace the original words to construct the watermark.Compared with existing more advanced benchmark algorithms,the proposed SRTW algorithm improves the AUC metric by 2.08%,5.17%,and 3.09%in three different attack scenarios,respectively,demonstrating the effectiveness of the SRTW algorithm.

关 键 词：文本水印 数据确权 数据安全 数据流通 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]