中文摘要：

在偏序Markov的基础上,以相邻DCT系数间关系对为统计对象,提出一种新的JPEG偏序隐写分析模型.在模型的无环有向图中,将相邻系数关系对作为终点集合,而系数本身作为始点集合.与原偏序分析模型相比,此终点集合的统计空间对DCT系数关系的分布规律有更直接的描述,并且始点集缩小,每个始点对应的终点数增加,使特征的信息熵上升,有助于分类效果.本文综合了系数间相关性较强的两个方向的统计值,采用像素裁剪重压缩进行图像校准,把待测图像与校准图像的统计概率之差作为特征.对3种代表性的DCT域隐写方法F5、MB1和Steghide进行隐写分析测试,实验结果表明：改进后的特征比原模型特征更有效,针对这3种隐写算法的检测效果,本文特征优于现有的单一模型低维特征.

英文摘要：

A novel JPEG steganalysis model which is based on partially ordered Markov and with the relationships of adjacent DCT（discrete cosine transform）coefficients for statistical object is proposed.In this model＇s directed acyclic graph,the relationships of adjacent coefficients is the set connected to the head of directed edge,and the coefficient itself as the set of tail.Compared with the original steganalysis model of partially ordered,the statistics space of this head set is more effective to describe the relationship between DCT coefficients.Besides,the number of head vertexes connected with each tail vertex increase with the set of tail shrinking,which makes information entropy of feature rise,improving results of classification.This paper have integrated the statistical values of two directions with stronger correlation between coefficients,setting the image feature by the difference between statistical probability of the test image and the calibration image which is got using pixel cropping and re-compression.The experiment results of three representative DCT domain steganographic methods F5,MB1 and Steghide show that the improved features are more effective than the original ones,and superior than the existing low dimensional features of single-model in the detection of these three steganographies.