摘要： 论证了通用图像隐写分析是一个类间很聚合、类内很分散的2类模式识别的困难分类问题。提出一种基于JPEG图像量化DCT域的块内和块间2个马尔可夫链获得高维特征，给出2种高维特征的分类器，即改进贝叶斯分类器和CNPCA分类器，后者简单而性能略低，但仍略优于SVM分类器。针对4种公认的JPEG隐藏数据方法，即F5, Outguess, MB1和MB2进行隐写分析，在CorelDraw图像库上做实验，取得了较好的效果。

关键词: 隐写分析, JPEG图像, DCT系数, 马尔可夫模型, 改进贝叶斯分类器, CNPCA分类器

Abstract: This paper proves that the universal steganalysis is a difficult two-class recognition problem, of which the between-class distribution is quite close and the within-class distribution is very scattered. This paper proposes the high-dimension feature based on the two Markov models of inner-block and inter-blocks in DCT domain of JPEG image. The paper also proposes two types of classifiers for high-dimension classification. One is the improved Bayesian classifier, and the other is the Class-wise Non-Principal Components Analysis(CNPCA)classifier. The latter is simple and slightly lower performance, but is still superior to SVM classifier. Experiments are taken out in CorelDraw image database, and the result shows that the scheme outperforms the existing steganalysis technique in attacking modern JPEG steganographic schemes F5, Outguess, MB1 and MB2.
【Key words】steganalysis; JPEG image; DCT coefficient; Markov model; improved Bayesian classifier; Class-wise Non-Principal Components

Key words: steganalysis, JPEG image, DCT coefficient, Markov model, improved Bayesian classifier, Class-wise Non-Principal Components Analysis(CNPCA) classifier

中图分类号: 

• TP391.41