基于互信息和特征融合的图像隐写分析

基于互信息和特征融合的图像隐写分析 Image Steganalysis Based on Mutual Information and Feature Fusion 张敏情 ZHANG Minqing 1 first-author 时晨曦 SHI Chenxi 1 武警工程大学电子技术系 武警工程大学电子技术系 Department of Electronic Technology,Engineering University of the Chinese Armed Police Force Department of Electronic Technology,Engineering University of the Chinese Armed Police Force 为了提高隐写分析的检测率和效率,本文提出了一种加权融合的联合图像专家小组JPEG（joint photographic experts group）图像通用隐写分析方法.该方法分别计算离散余弦变换DCT（discrete cosine transform）系数块内和块间的水平、垂直和zigzag三向差分数组,采用联合概率密度矩阵来挖掘信息嵌入对DCT系数间相关性的影响,生成块内和块间三向特征.利用特征与分类类别间的互信息对特征权值进行量化,加权融合得到最终的特征向量,并使用支持向量机进行分类.对3种安全性较高的JPEG隐写算法F5、Outguess和MB2,在不同嵌入率下进行隐写分析.实验结果表明,在不同嵌入率的情况下其检测率均高于88.4%,同时特征融合算法使该方案具有更高的检测效率. A new universal steganalysis scheme based on weighting fusion to attack JPEG（joint photographic experts group）steganography is presented to improve detection rates and efficiency.Firstly,difference arrays within and between DCT（discrete cosine transform）blocks along horizontal,vertical,zigzag directions are computed.Then,joint probability density matrix is applied to capture the impacts on the relevance of DCT coefficients caused by embedding process.And features from three directions are constructed.Finally,weights of each feature are quantified using the mutual information between features and classification classes.Final features vectors for steganalysis are derived by fusing the weighted features and are tested by SVM（support vector machine）.The experimental results show that the proposed scheme provides reliable detection rates in attacking three steganographic schemes including F5,Outguess and MB2.The detection rates are higher than 88.4%,and the feature fusion method increases the detection efficiency. 隐写分析 系数相关性 联合概率密度 互信息 特征融合 steganalysis relevance of coefficients joint probability density mutual information feature fusion 10.14188/j.1671-8836.2013.06.001 TP391.41 国家自然科学基金（61379152）;陕西省自然科学基金（2012JM8014）;武警工程大学基础研究基金（WJY201123）资助项目 2013-06-01 2021-04-01 6 59