ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2021, Vol. 58 ›› Issue (3): 548-568.doi: 10.7544/issn1000-1239.2021.20200360

• 信息安全 • 上一篇    下一篇

基于深度学习的图像隐写研究进展

付章杰1,2,李恩露1,程旭1,黄永峰3,胡雨婷3   

  1. 1(南京信息工程大学计算机与软件学院 南京 210044);2(鹏城实验室 广东深圳 518066);3(清华大学电子工程系 北京 100084) (fzj@nuist.edu.cn)
  • 出版日期: 2021-03-01
  • 基金资助: 
    国家自然科学基金项目(U1836110, 61602253, 61802058);江苏省自然科学基金(BK20200039)

Recent Advances in Image Steganography Based on Deep Learning

Fu Zhangjie1,2, Li Enlu1, Cheng Xu1, Huang Yongfeng3, Hu Yuting3   

  1. 1(School of Computer and Software, Nanjing University of Information Science & Technology, Nanjing 210044);2(Peng Cheng Laboratory, Shenzhen, Guangdong 518066);3(Department of Electronic Engineering, Tsinghua University, Beijing 100084)
  • Online: 2021-03-01
  • Supported by: 
    This work was supported by the National Natural Science Foundation of China (U1836110, 61602253, 61802058) and the Natural Science Foundation of Jiangsu Province (BK20200039).

摘要: 图像隐写是信息安全领域的研究热点之一.早期隐写方法通过修改载体图像获得含密图像, 导致图像统计特性发生变化, 因此难以抵抗基于高维统计特征分析的检测.随着深度学习的发展, 研究者们提出了许多基于深度学习的图像隐写方法, 使像素修改更隐蔽、隐写过程更智能.为了更好地研究图像隐写技术, 对基于深度学习的图像隐写方法进行综述.首先根据图像隐写过程, 从3个方面分析了基于深度学习的图像隐写方法:1)从生成对抗网络和对抗样本2个角度介绍载体图像获取方法; 2)分析基于深度学习的隐写失真设计方法; 3)阐述基于编码-解码网络的含密图像生成方法.然后, 分析和总结了无载体图像隐写方法的优缺点, 该类方法无需载体图像即可实现图像隐写, 因此在对抗统计分析方面存在天然优势.最后, 在深入分析与总结基于深度学习的图像隐写与无载体图像隐写2类方法优缺点的基础上, 对图像隐写的发展方向进行了探讨与展望.

关键词: 图像隐写, 深度学习, 生成对抗网络, 对抗样本, 无载体图像隐写

Abstract: Image steganography is one of the research hotspots of information security. In the early stages of image steganography study, pixel modifications change the statistical properties of the image in a non-negligible way. So it is difficult to resist detection based on high-dimensional statistical characterization. With the development of deep learning, researchers have proposed a number of deep learning-based image steganography methods, which make image modification more imperceptible and the steganography process more intelligent. To further study image steganography technology, this paper provides a thorough review of the image steganography methods based on deep learning. First of all, according to the image steganography process, this paper profoundly analyzes the deep learning-based image steganography methods from three aspects. The first is to introduce the cover image acquisition methods from two perspectives: generative adversarial networks and adversarial samples. The second is to illustrate the steganographic distortion design methods. The third is to expound the methods of generating stego images based on the encoding-decoding network. In addition, the advantages and disadvantages of coverless image steganography are also discussed and summarized in this paper. This kind of method can achieve imperceptible steganography without the cover image, so it has inherent superiority in resisting steganalysis. Finally, we conclude the pros and cons of deep learning-based image steganography methods and coverless image steganography. On the basis of the analysis, we discuss and prospect the future research direction of image steganography.

Key words: image steganography, deep learning, generative adversarial networks, adversarial examples, coverless image steganography

中图分类号: