生成式隐写研究  

Research on Generative Steganography

作　　者：周志立 丁淳 李进[2] 彭飞 张新鹏 ZHOU Zhi-Li;DING Chun;LI Jin;PENG Fei;ZHANG Xin-Peng(Engineering Research Center of Digital Forensics,Ministry of Education,Nanjing University of Information Science and Technology,Nanjing 210044;Institute of Artificial Intelligence and Blockchain,Guangzhou University,Guangzhou 510006;School of Computer Science,Fudan University,Shanghai 201203)

机构地区：[1]南京信息工程大学数字取证教育部工程研究中心,南京210044 [2]广州大学人工智能与区块链研究院,广州510006 [3]复旦大学计算机科学技术学院,上海201203

出　　处：《计算机学报》2023年第9期1855-1887,共33页Chinese Journal of Computers

基　　金：国家重点研发计划基金资助项目(No.2022YFB3103100);国家自然科学基金(No.61972205,No.U1936218,No.62122032);江苏省大气环境与装备技术协同创新中心资助项目(CICAEET)资助。

摘　　要：隐写术通常将秘密信息以不可见的形式隐藏到载体中,从而通过传递含密载体实现隐蔽通信.嵌入式隐写方案通过修改载体将秘密信息嵌入其中,但会不可避免地改变载体的统计特性,因此难以抵抗各类隐写分析工具的检测.为了解决此问题,生成式隐写方案以秘密信息为驱动直接生成含密载体.相比于嵌入式隐写方案,生成式隐写方案针对现有基于统计特征的隐写分析方法具有较好的抗检测性能,因此逐渐成为信息隐藏领域的研究热点.本文首先对四类生成式隐写方案进行详细地描述和分析,包括:(1)图像生成式隐写方案;(2)文本生成式隐写方案;(3)音频生成式隐写方案;(4)社交网络行为生成式隐写方案;其次,通过实验详细分析和对比了各种图像和文本生成式隐写方法的性能;最后,本文分析了当前生成式隐写方案仍然存在的问题,并提出相应的解决方案和展望未来的发展方向.Steganography is a technology that generally hides secret information into a cover invisibly to obtain the stego,so that covert communication can be realized by transmitting the stego.Compared to the technology of encryption,steganography can not only secure the content of secret information but also protect the behavior of covert communication itself.In the past two decades,many multimedia steganographic schemes have been proposed,in which the information embedding-based steganographic methods are the most popular ones.The information embedding-based steganographic schemes usually embed the secret information into the cover with slight modification.However,the modification will inevitably change the statistical features of the cover,which makes it difficult to resist the detection of various steganalysis tools.To address this issue,a new kind of steganographic schemes,i.e.,generative steganographic schemes,has been proposed.The generative steganographic schemes usually generate a new multimedia cover as the stego driven by the secret information.Compared with the embedding-based steganographic schemes,the generative steganographic schemes achieve promising anti-detectability to steganalysis,and thus they have become one of hottest research topics in the field of information hiding recently.In this paper,according to the categories of the generated stegos,four kinds of generative steganographic schemes are described,which are:1)image generative steganography,2)text generative steganography,3)audio generative steganography,and 4)social network behavior generative steganography,and the advantages and disadvantages of the four kinds of generative steganographic schemes are also analyzed in detail;Then,by extensive experiments,the performances of recent image and text generative steganographic schemes are analyzed and compared in the aspects of secret information extraction,the robustness against common image attacks,hiding capacity.We also experimentally demonstrate that the stability of stegos generated by the exis

关 键 词：生成式隐写 信息隐藏 数字水印 隐蔽通信 隐写分析 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]