基于改进生成对抗网络的图像自适应隐写模型  

Image adaptive steganography model based on improved generative adversarial network

作　　者：刘荣 李冠 贾斌 LIU Rong;LI Guan;JIA Bin(College of Computer Science and Engineering,Shandong University of Science and Technology,Qingdao 266590,China)

机构地区：[1]山东科技大学计算机科学与工程学院,山东青岛266590

出　　处：《计算机工程与设计》2021年第6期1551-1561,共11页Computer Engineering and Design

基　　金：国家自然科学基金项目(71772107);青岛市社会科学规划研究基金项目(QDSKL1901125、QDSKL2001142)。

摘　　要：针对现有图像隐写模型存在网络训练不易收敛、梯度爆炸且生成样本质量差等问题,提出一种基于改进生成对抗网络的图像自适应隐写模型SWGAN-GP。将生成图像作为载体,使用HUGO自适应隐写算法进行信息隐藏;在损失函数中加入梯度惩罚,在网络结构中引入注意力机制,设置双判别器与生成器进行对抗训练。实验结果表明,该方法生成图像的IS值、PSNR值等均有提高,判别器分类效果明显改善。该模型可以提高收敛速度,使网络训练更稳定,载密图像更具安全性,有效抵御隐写分析算法的检测。Aiming at the problems of the existing image steganography models,such as difficulty in network convergence,gra-dient explosion and poor quality of generated samples,an image adaptive steganography model SWGAN-GP based on improved generative adversarial network was proposed.The generated image samples were used as cover images,and the adaptive stega-nography algorithm HUGO(highly undetectable steganography)was used to generate the dense images for information hiding.The gradient penalty was added to the loss function,and an attention mechanism was introduced into the network structure.The discriminator D and steganographic discriminator S were used to conduct adversarial training with generator G.Experimental results show that the IS value,PSNR value,etc.of the generated images are improved,and the classification effects of the discriminators are also significantly improved.This model can improve the convergence speed and make the network training more stable.And it makes the generated dense images safer,so they can resist the detection of steganalysis algorithms effectively.

关 键 词：生成对抗网络 梯度惩罚 自适应隐写 高度不可检测隐写 信息隐藏 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]