基于对抗机制的彩色图像隐写分析算法-《郑州大学学报(工学版)》

文章信息/Info

作者:: 张涛¹; 葛育伟²; 韩旭²; 张昊¹; 汪然¹; 1.战略支援部队信息工程大学信息系统工程学院,河南郑州 450001, 2.苏州大学计算机科学与技术学院,江苏苏州 215006

Author(s):: ZHANG Tao¹; GE Yuwei²; HAN Xu²; ZHANG Hao¹; WANG Ran¹; 1.School of Information System Engineering, the University of Strategic Support Forces Information Engineering University, 450001, 2.Zhengzhou, Henan, School of Computer Science and Technology, Suzhou University, Suzhou, Jiangsu 215006

Keywords:: information hidding; steganalysis; deep learning; multiple activation modules; adversarial mechanisms

摘要:: 针对彩色图像的隐写分析问题,引入逐通道卷积、多激活模块以及对抗机制,提出了一种应用于彩色图像隐写分析的深度卷积网络。逐通道卷积能够避免削弱不相关噪声信号,保留更多的隐写嵌入特征;多激活模块利用多种激活函数对卷积结果进行非线性映射,针对嵌入痕迹做出不同反馈,丰富嵌入特征的多样表达;对抗机制能够将内容信息特征和隐写嵌入特征从域类别上严格划分,从而分离出更多的隐写存在性特征。在 PPG-LIRMMCOLOR 数据集上针对多种隐写算法进行了检测实验。结果显示,所提算法比对照方法中性能最好的还要高 1. 83%到 4. 99%。实验结果验证了该彩色图像隐写分析方法的有效性。

Abstract:: Aiming at the steganalysis of color images, a deep convolutional network applied to the steganalysis of color images is proposed by introducing channel-wise convolution, multiple activation module and adversarial mechanism. Channel-wise convolution can avoid weakening irrelevant noise signals and retain additional steganographic embedded features; multiple activation modules use various activation functions to nonlinearly map convolution results and make different feedback for embedded traces to enrich the diverse expressions of embedded features; adversarial mechanisms can divide content information features and steganographic embedding features from domain categories, thereby separating additional steganographic existence features. Experiments are carried out on the PPG-LIRMM-COLOR dataset for various steganographic algorithms. The proposed algorithm is 1. 83% - 4. 99% higher performance than the best performance in the control methods. Results verify the effectiveness of the proposed color image steganalysis method.