一种结合旋转森林变换与多分类器集成的隐写分析算法

Novel Steganalysis Algorithm Combine Rotating Forest Transformation with Multiple Classifiers Ensemble

针对图像隐写分析中,集成分类器的基分类器精度较低、分类器种类单一的缺点,提出一种结合旋转森林变换与多分类器集成的隐写分析算法.首先随机生成若干特征子空间并与训练样本组成不同的样本子集,然后对每个样本子集使用旋转森林算法训练费歇尔线性分类器、极限学习机与支持向量机三种分类器,并通过加权投票的方式将三种分类器集成为一个基分类器.最后将各基分类器分类结果使用简单投票法进行集成.实验结果表明,在不同的隐写算法与嵌入率的条件下,与传统集成分类器和集成极限学习机分类器相比,该算法降低了3.2%与1.1%的误检率,能够有效提升集成分类器的检测精度.

Concerning the problems that the accuracy of basic classifier is low and the kind of basic classifier is single in ensemble classifiers for steganalysis,a novel algorithm combine rotating forest transformation with multiple classifiers ensemble was proposed. First, some feature subsets generated randomly combine with training samples to generate different sample subsets. Second, every sample subset was transformed by rotating forest algorithm and trained fisher linear discriminate, extreme learning machine and support vector machine,then the weighted voting method was used to ensemble three classifiers as a basic classifier. At last,the majority voting method was used to integrate the decisions of basic classifiers. Experimental results show that by different steganography approaches and in different embedding rate conditions, the prediction error rate of proposed method decreased by 3.2% and 1.1% in compared with the typical ensemble classifiers and ensemble classifiers of extreme learning machines, therefore demonstrating the proposed method could improve the detection accuracy of ensemble classifiers.