HCGAN:一种基于GAN的高容量信息隐藏算法  

HCGAN:A High Capacity Information Hiding Algorithm Based on GAN

作　　者：张克君 李旭 于新颖 冯丽雯 秦昊聪 张健毅 ZHANG Kejun;LI Xu;YU Xinying†;FENG Liwen;QIN Haocong;ZHANG Jianyi(Department of Cyberspace Security,Beijing Electronic Science and Technology Institute,Beijing 100071,China;School of Cyberspace Security,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]北京电子科技学院网络空间安全系,北京市100071 [2]北京邮电大学网络空间安全学院,北京市100876

出　　处：《湖南大学学报（自然科学版）》2022年第4期35-46,共12页Journal of Hunan University:Natural Sciences

基　　金：北京高校高精尖学科建设项目(20210086Z0401);国家重点研发计划网络空间安全重大专项课题资助(2018YFB0803601)。

摘　　要：针对现有信息隐藏算法存在隐写容量低、信息提取困难以及安全性差等问题,本文提出了一种基于生成对抗网络的高容量信息隐藏算法(High Capacity Information Hiding Al⁃gorithm Based on GAN,HCGAN).在秘密信息嵌入方面,使用基于Im-Residual结构的编码器将秘密信息嵌入载体图像中,避免了秘密信息嵌入时由卷积层提取特征导致的信息损失.在秘密信息提取方面,使用基于稠密结构的解码器从含秘图像中提取出秘密信息,利用特征复用来增加秘密信息的提取率.在抗隐写分析方面,利用基于隐写分析的鉴别器与基于Im Residual结构的编码器进行对抗训练,以提高含秘图像的抗隐写分析能力.实验表明,经过对抗训练后,HCGAN在2 bpp嵌入率下比WOW和S-UNIWARD在0.4 bpp嵌入率下具有更低的隐写分析检测率.Aiming at the problems of low steganographic capacity,difficult information extraction,and poor secu⁃rity in existing information hiding algorithms,this paper proposes a high capacity information hiding algorithm based on GAN(HCGAN).For secret information embedding,an Im-Residual structure-based encoder is applied to embed the secret information into the carrier image,avoiding the information loss caused by the feature extraction of the con⁃volution layer.For secret information extraction,a dense structure-based decoder is utilized to extract secret informa⁃tion from the secret image,and feature reuse is used to increase the extraction rate of secret information.In terms of anti-steganalysis,the discriminator based on steganalysis and the encoder based on Im-Residual structure are used for adversarial training to improve the anti-steganalysis ability of the secret image.Experiments show that after adver⁃sarial training,HCGAN has a lower steganalysis detection rate at an embedding rate of 2bpp than the WOW and SUNIWARD algorithms at an embedding rate of 0.4bpp.

关 键 词：信息隐藏 深度学习 生成对抗网络 自编码器 卷积神经网络 

分 类 号：TN915.08[电子电信—通信与信息系统]