基于仿射传播聚类的富模型降维方法  

Dimension Reduction Method of Rich Modelusing Affinity Propagation Clustering Algorithm

作　　者：张敏情[1] 马占山[1] 刘佳[1] 李军[1] 

机构地区：[1]武警工程大学电子技术系网络与信息安全武警部队重点实验室,陕西西安710086

出　　处：《四川大学学报（工程科学版）》2016年第2期91-96,共6页Journal of Sichuan University (Engineering Science Edition)

基　　金：国家自然科学基金资助项目(61379152)

摘　　要：为了解决隐写分析中富模型特征维数高,冗余大,容易造成维数灾难问题,提出了一种基于仿射传播聚类的特征降维方法;该方法通过分析富模型特征结构,利用非线性距离定义子模型特征间距离,使用仿射聚类方法和图的谱理论,确定特征的聚类中心,并将聚类中心中所出现频率较高的子模型特征作为最优聚类中心,并使用Fisher集成分类器进行隐写分析。实验结果表明,经过该降维方法处理的空域富模型（spatial rich model,SRM）在特征维数降低到5 525维时,对S-UNIWARD、WOW、HUGO隐写算法隐写分析时检测错误率EOOB比原始特征下降1%～2%,因此,本文方法可以实现提升特征降维与隐写分析效果的目的。A feature dimension reduction means based on affine clustering was worked out to solve the problem of the features of rich model＇s high-dimension and dimension disaster led by much redundancy. This method analyzed the structure of rich model and defined sub model feature distance with nonlinear distance definition,using the affine clustering algorithm and the image spectrum theory to determine the clustering center of features. The sub model feature which has a high occurrence probability in the clustering center was considered as the optimal clustering center,and conducted steganalysis by Fisher ensemble classifier. The experiment showed that the testing error rate EOOBcan be decreased by 1% ~ 2% compared to original feature for the steganalysis of S-UNIWARD,WOW,HUGO stegography by the dimension reduction means when the dimensionality of SRM（ spatial rich model） is down to 5 525. Therefore,the method in this paper will fulfill the improvement in feature dimension reduction and the effect of steganalysis.

关 键 词：隐写分析 仿射聚类 富模型 降维 

分 类 号：TP309.1[自动化与计算机技术—计算机系统结构]