Feature selection for blind steganalysis using localized generalization error model

Abstract

Steganalysis is a technique to fight against steganography. Different kinds of feature extraction methods have been proposed for blind steganalysis. They have their own advantages when attacking different kinds of steganography. Making a combination of different feature sets will improve the performance of the steganalysis system. However, it will increase the dimensionality of features largely at the same time. Meanwhile, it may have many irrelevant features in the system. A proper feature selection method could decrease the computational complexity and also enhance the performance of the steganalysis. In this paper, we proposed a feature selection method based on the Localized Generalization Error Model (L-GEM) to selection the most relevant feature subset for steganalysis system. The proposed method is compared with two other off-the-shelf feature selection methods. The experimental results show that the proposed method outperforms the other two feature selection methods. The steganalysis with the proposed feature selection method yields a higher average testing accuracy than that of using full set of features.