Feature selection for image steganalysis using hybrid genetic algorithm

Abstract

Learning-based methodology has been demonstrated to be an effective approach to dispose the steganalysis difficulties due to the variety of image texture. A crucial process of the learning-based steganalysis is to construct a low-dimensional feature set. In this study, a feature selection method based on Hybrid Genetic Algorithm (HGA) is presented to select feature subsets which not only contain fewer features, but also provide better detection performance for steganalysis. First, the general framework about utilizing Genetic Algorithm (GA) to do feature selection for steganalysis is presented. Then, we analyze similarity among individuals (SI) in each generation and the Transformation of Generations (TG) to determine whether the GA has converged into a local area. Next, according to the SI and TG, the restarting operation is incorporated into the HGA to allow the algorithm to escape from the unsatisfactory local area. In the experiments, three feature subsets are formed from a universal feature set for three typical steganography methods, respectively. The experimental results show that the classifiers using the feature subsets gain better detection accuracy and higher speed than those using the universal set. © 2009 Asian Network for Scientific Information.