摘要
深度学习促进了自然语言处理技术的发展,基于文本生成的信息隐藏方法表现出了巨大的潜力.为此,提出了一种基于神经机器翻译的文本信息隐藏方法,在翻译文本生成过程中进行信息嵌入.神经机器翻译模型使用集束搜索(Beam Search)解码器,在翻译过程中通过Beam Search得到目标语言序列各位置上的候选单词集合,并将候选单词依据概率排序进行编码;然后在解码输出目标语言文本的过程中,根据秘密信息的二进制比特流选择对应编码的候选单词,实现以单词为单位的信息嵌入.实验结果表明,与已有的基于机器翻译的文本信息隐藏方法相比,该方法在隐藏容量方面明显提升,并且具有良好的抗隐写检测性和安全性.
Deep learning has promoted the development of natural language processing technology,and the information hiding methods based on text generation show great potential in this area.This paper proposes a text information hiding method based on neural machine translation,which embeds information in the process of generating translated text.The neural machine translation model uses a Beam Search decoder,which is used to obtain the candidate words in the sentence sequence in the translation process,and to encode the candidate words according to the probability ranking.Then,in the process of decoding and outputting target language texts,corresponding encoded candidate words are selected according to the binary bitstream of secret information,so as to realize information embedding at word level.Experimental results show that compared with the existing text information hiding methods based on machine translation,this method significantly improves the rate of information embedding,and shows good capability and security in anti-steganography performance.
作者
尉爽生
杨忠良
江旻宇
黄永峰
YU Shuangsheng;YANG Zhongliang;JIANG Minyu;HUANG Yongfeng(Department of Electronic Engineering,Tsinghua University,Beijing 100084,China;Department of Computer Science,Rice University,Texas 77005,United States)
出处
《应用科学学报》
CAS
CSCD
北大核心
2020年第6期976-985,共10页
Journal of Applied Sciences
基金
国家重点研发计划基金(No.SQ2018YGX210002)
国家自然科学基金(No.U1536201,No.U1636113,No.U1705261)资助。
关键词
文本隐写
信息隐藏
神经网络
机器翻译
text steganography
information hiding
neural network
machine translation
