改进SteGAN的嵌入式图像隐写方案  

Improved SteGAN Embedded Image Steganography Scheme

作　　者：杨忠鹏 李启南[1] YANG Zhong-peng;LI Qi-nan(School of Electronic and Information Engineering,Lanzhou Jiaotong Universityanzhou 730070,China)

机构地区：[1]兰州交通大学电子与信息工程学院,兰州730070

出　　处：《兰州交通大学学报》2022年第4期48-57,共10页Journal of Lanzhou Jiaotong University

基　　金：教育部人文社会科学研究项目(18YJAZH044)。

摘　　要：针对隐写生成对抗网络训练时图像特征信息丢失导致生成载密图像质量较低的问题,提出了一种基于改进隐写生成对抗网络的图像隐写模型.改进的模型采用密集连接方式将浅层网络的特征输送到深层网络结构的每一层,有效地保留图像结构信息;为了使模型更多关注图像复杂纹理区域,在网络中加入卷积注意力模块获取图像的深层次特征,并在判别器中引入谱归一化方法,以提高模型训练的稳定性.仿真实验结果表明:相比已有隐写算法0.4位/像素的嵌入容量,在保证隐写图像质量的前提下,改进的隐写生成对抗网络能够将嵌入容量提升至0.6位/像素的同时还能抵抗隐写分析器的检测.Aiming at the problem that the loss of image feature information during the training of steganographic generative adversarial network will lead to the low quality of the generated secret image,an image steganography model based on improved steganographic generative adversarial network is proposed.The improved new model adopts the dense connection method to transfer the features of the shallow network to each layer of the deep network structure,which can effectively retain the image structure information;in order to make the model pay more attention to the complex texture area of the image,convolution attention is added to the network module to obtain the deep-level features of the image,and a spectral normalization method is introduced in the discriminator to improve the stability of model training.The simulation results show that compared with the 0.4 bit per pixel embedding capacity of the existing steganography algorithm,ISteGAN can increase the embedding capacity to 0.6 bit per pixel while keeping the steganographic image undistorted and resisting the detection of the steganalyzer.

关 键 词：图像隐写 密集连接 卷积注意力 谱归一化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]