中文摘要：

自然语言与图像、音频信号的性质截然不同，图像水印等的鲁棒性分析方法不适用于自然语言水印，但是直到目前还没有专门针对自然语言水印鲁棒性的研究和评估工作．文中针对自然语言的特点，提出适合自然语言水印的敌手模型．然后将现有的自然语言水印分类，并总结各类的一般算法模型．利用文本提出的敌手模型分析自然语言水印编码算法的鲁棒性，并通过实验验证鲁棒性的理论模型．本项工作为对比、评估自然语言水印算法的鲁棒性提供了理论依据．

英文摘要：

The nature of natural language are quite different from the signal of ：images and sound, and the methods of robust analysis on image watermarks are not to be able to apply to NLW. However, the study of robustness of NLW is still absent. In this paper, based on the nature of natural language, we propose an adversary model that suit for NLW. Then, we classify the exist- ing NLW methods and propose common algorithms. Third, we analyze the robustness of NLW algorithms using the adversary model that we proposed, and verify the theoretical model by ex- perimental results. Our theoretical robust models are useful for comparison and evaluation ro- bustness of different NLW algorithms.