Kernel target alignment for feature kernel selection in universal steganographic detection based on multiple kernel SVM

Abstract

Merging multiple features can achieve higher detection rate than single feature in universal steganalysis, however there are drawbacks in most merged feature steganalysis methods: it is only a simple combination of features without analysing the relation of constisten and afoul among these features, besides there is no unified standard for features selection. This paper introduces a quantity to capture the quantitative measure of agreement between different features, which we call alignment, for feature kernel selection in steganographic detection of JPEG images based on Multiple Kernel SVM(MK-SVM). We apply orthogonal feature sets with low alignment value as the "basis kernels on features" in multiple kernel learning model, the linear combination of kernels is optimized using SimpleMKL algorithm. We compare the performance of merged feature method with multiple kernel method to detect six popular steganographic algorithms, result indicates that multiple kernel method outperforms merged feature method for 0.2%-1.7%, and most importantly with theory evidence for features selection.