多层次特征融合的对抗网络图像隐写  

Adversarial Network Image Steganography Based on Multi-level Feature Fusion

作　　者：张震 王真 ZHANG Zhen;WANG Zhen(College of Computer Science and Technology,Shanghai electric power University,Shanghai 200000,China)

机构地区：[1]上海电力大学计算机科学与技术学院,上海200000

出　　处：《计算机仿真》2023年第4期323-329,共7页Computer Simulation

基　　金：国家自然科学基金(61772327,61432017);迁安新工程国家大数据协作安全技术开放项目(QAX-201803)。

摘　　要：针对基于神经网络的图像隐写方法无法同时捕捉载体图像的纹理信息和语义特征,导致存在大量的信息丢失的问题,提出一种基于多层次特征融合的对抗网络图像隐写方法。通过在生成网络中添加多尺度卷积与池化操作对图像进行特征提取,使用跳跃连接融合多层次特征信息,利用隐写分析性能更先进的判别网络与生成网络进行对抗学习,生成嵌入修改图并模拟生成隐写图像。实验结果表明,在高维特征隐写分析与深度学习隐写分析检测下,该方法均具有更高的隐写安全性。与目前所提出的基于深度学习的隐写方法相比,综合安全性能提升3.8%。Aiming at the problem that image steganography based on neural network can not capture the texture information and semantic features of the cover image at the same time,which leads to a large amount of information loss,this paper proposes a method of anti network image steganography based on multi-level feature fusion.Firstly,multi-scale convolution and pooling operations were added to the generative network to extract the features of the images.Then,the multi-level feature information was fused by jumping join.The discriminant network with more advanced performance of stegage analysis was used for adversation learning with the generative network to generate embedded modified graphs and simulate the generation of stego images.Experimental results show that this method has higher steganography security under both high-dimensional feature steganography analysis and deep learning steganography analysis detection.Compared with the proposed stegography method based on deep learning,the comprehensive security performance is improved by 3.8%.

关 键 词：图像隐写 生成对抗网络 多层次特征融合 多尺度卷积 信息丢失 安全性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]