基于选择性集成分类器的通用隐写分析

面对高维度的特征集和大规模的样本集,隐写分析技术对分类器的要求越来越高.在集成分类器的基础上提出了一种面向通用隐写分析的选择性集成分类器.首先基于随机森林生成若干个基分类器,然后利用基于遗传算法的选择性集成算法剔除掉个别影响整体性能的基分类器,最后根据遗传优化得到的最优权值向量赋予剩余的基分类器不同权值以用来加权投票集成.实验表明,提出的选择性集成分类器测试性能优于现有分类器,特别在基分类器数量较大、特征维数较高时与现有集成分类器相比,有效降低了检测错误率.

Universal Steganalysis Based on Selective Ensemble Classifier

With massive feature set and high-dimensional sample set, steganalysis has a increasingly demanding for classifiers.Based on ensemble classifier,we proposed a kind of selective ensemble classifier for universal steganalysis.At first, we generated some base learners based on the random forest and then wept out some of them using GASEN(Genetic Algorithm based Selective Ensemble) algorithm.At last,remaining base classifiers were given different weights according to the optimal weight vector from genetic optimization to get used to the weighted vote integration.Experiments showed that selective ensemble classifier proposed performed better than existing single classifier.Compared with the existing ensemble classifier;especially in the case of larger base classifiers or higher number of features,the computational complexity slightly increased,but it reduced the error rate effectively.