基于贝叶斯分类器的图像隐写分析

集成分类器是目前用于图像隐写分析的主流分类器。为提高集成分类器的检测精度,针对集成分类器基分类器组合方法过于简单,无法体现基分类器之间的内在联系,不能从整体上对结果进行判定的缺点,依据图像特征在集成分类器分类超平面上的投影值服从多维正态分布这一特性,提出了一种基于贝叶斯分类器的图像隐写分析算法。首先基于随机森林算法生成若干基分类器,然后计算类条件概率密度函数与先验概率并训练贝叶斯分类器,最后使用经过训练的贝叶斯分类器代替简单投票方法进行分类判决。算法的检测错误率比以往算法平均降低了1.6%,ROC曲线比简单投票方法更接近于左上角,即具有更高的检测率,AUC值平均增长约2.12%,并且训练时间仅有少量提高,最大提高约2.610s。可以有效提高集成分类器的检测精度。

Image steganalysis based on bayesian classifier

Recently, ensemble classifier is predominantly used for steganalysis of digital media. For increasing the detection accuracy of ensemble classifier and focused on the defects that the fusion method of ensemble classifier is too simple to reflect the correlation of base learners and can't give an overall result, an algorithm based on Bayesian Classifier is proposed according to a characteristic that the projection of feature on the hyper planes of ensemble classifier obeys the framework of the multivariate Gaussian distribution. At first, some base learners were generated based on the random forest, then the probability density function and prior probability were calculated for training a Bayesian classifier. At last, the algorithm uses a Bayesian Classifier instead of the majority voting rule to fuse the decisions of base learners. The error detection rate lower than before by an average of 1.6%, ROC curve is closer to the top left corner than simple voting method, which has higher detection rate, the growth of AUC value is about 2.12%, and the training time increase about 2.610 s. The proposed method is able to increase steganalysis performance of ensemble classifier.