A hybrid feature selection approach based on improved PSO and filter approaches for image steganalysis

This paper proposes a novel feature selection approach to improve the classification accuracy and reduce the computational complexity in image steganalysis. It is a hybrid filter-wrapper approach based on improved Particle Swarm Optimization (PSO). It consists of two phases: the first phase is composed of two filter techniques namely t test and multiple-regression which selects the features based on their ability to discriminate images as stego or cover. The second phase further reduces the number of features by working on the significant features selected during the first phase using an improved PSO. This approach overcomes the disadvantages of global best PSO by integrating it with local best PSO and dynamically changing the population size (Hope/Rehope). The proposed approach is tested on two sets of features extracted from spatial domain (SPAM-Subtractive Adjacency Matrix) and transform domain (CCPEV-Cartesian Calibrated features extracted by Pevny) for four embedding algorithms nsF5, Outguess, Perturbed Quantization and Steghide using SVM (Support Vector Machine) classifier. Experimental results demonstrate that this approach significantly improves the classification accuracy and drastically reduces dimensionality as compared to results produced by some well-known feature selection algorithms.