基于主成分分析的泛盲掩密分析方案

A Universal Blind Steganalysis Based on Principle Component Analysis

为了能够对掩密后图像进行有效的检测,首先从分类特征维数、单个特征的有效性和特征相关性等3个方面对Farid泛盲掩密分析算法的缺陷进行了分析,然后提出了采用主成分分析技术对分类特征进行降维处理的方法,并基于RBF(rad ial basis function)网络构造了新的泛盲掩密分析方案。该方案不但大大降低了用于分类的图像特征的维数,而且提高了掩密分析的检测性能。利用该方案和Farid的方案分别对用JSteg等软件掩密后的图像进行的检测比较实验表明,经主成分分析预处理后,该方案的样本集特征矢量维数比Farid方案分别减少了174(Jsteg)维、163(EzStego)维和180(S-Tools)维,而特征数目的减少又大大简化了分类器的设计,而且,该方案能够有效检测嵌入消息占可嵌容量的比例达60%(Jsteg)、80%(EzStego)、50%(S-Tools)以上的掩密图像,并获得了比Farid算法更高的检测率。

A new universal blind steganalysis scheme is presented based on analysing the limitation of Farid＇ s universal blind steganalysis on the dimension of the feature vector, the validity of the single feature, and the correlation of the features. Preprocessing is proposed using principle component analysis on the image statistics features and steganalysi＇s classifier is constructed using RBF network. The scheme not only reduces the dimension of the feature vector enormously, but also improves the performance of the detection. Apply our scheme and Farid＇ s scheme to detecting the stego images produced by JSteg, EZStego, and S-Tools respectively, and the comparison of these simulation results shows that after the preprocessing using principle component analysis the dimension of the feature vector in our scheme decreases 174（Jsteg）, 163 （EzStego）, 180（ S-Tools）, and therefore simplifies the design of the steganalysis classifier. Furthermore, our scheme is quite more efficient because the stego image that the proportion of the embedding message to the maximal embedding capability is more than 60% （Jsteg）, 80% （EzStego）, 50% （S-Tools） can be detected efficiently by our scheme.