基于信息最大化生成对抗网络的图像隐写方法  

Image Steganography Based on Information Maximizing Generative Adversarial Networks

作　　者：刘义铭[1,2] 过小宇[1,2] 牛一如 LIU Yiming;GUO Xiaoyu;NIU Yiru(Science and Technology on Communication Security Laboratory,Chengdu 610041,China;The 30th Research Institute of CETC,Chengdu 610041,China)

机构地区：[1]保密通信重点实验室,四川成都610041 [2]中国电子科技集团公司第三十研究所,四川成都610041

出　　处：《无线电工程》2021年第11期1214-1219,共6页Radio Engineering

基　　金：国家自然科学基金企业创新发展联合基金-人工智能领域(U20B200063);四川省科技计划项目(2021JDR274)。

摘　　要：针对现有图像隐写方法载体实用性、隐藏信息安全性以及信息嵌入与输出准确性较低等问题,提出了一种基于信息最大化生成对抗网络(Information Maximizing Generative Adversarial Networks,InfoGAN)的图像隐写方法。该方法使用潜码(Latent Code)来表示数据的标签,作为生成模型的一个输入与噪声一起生成含密图像,并使用标签判别器分离出标签,以达到解密的效果。与传统信息隐藏方法不同,该方法抛弃了传统信息隐藏方法中信息嵌入的步骤,而使用生成对抗网络生成含密图像,并且将秘密信息“直接”与噪声一起生成含密图像。实验表明,基于InfoGAN的图像隐写方法相较于其他隐写方法具有更好的安全性,并且加密、解密方式简单,易于操作。Considering the low carrier practicality,hidden information security and accuracy of information embedding and output in existing image steganography methods,an image steganography method based on Information Maximizing Generative Adversarial Networks(InfoGAN)is proposed,in which latent code is used to represent the label of data as an input of the generative model to generate a secret image together with noise,and a label discriminator is used to separate the labels to achieve the decryption effect.Different from the traditional information hiding method,this method abandons the information embedding step in the traditional information hiding method,and uses a generative adversarial network to generate the secret image,and the secret information is used together with the noise to“directly”generate the secret image.Experiments show that the image steganography method based on InfoGAN has better security than other steganography methods,and the encryption and decryption methods are simple and easy to operate.

关 键 词：图像隐写 生成对抗网络 信息最大化生成对抗网络 互信息 码表 

分 类 号：TP309.7[自动化与计算机技术—计算机系统结构]