Image Steganalysis using Feature Selection based on Mutual Information and Adaptive Particle Swarm Optimization

Abstract

In recent years, Steganalysis has been an area of active research. Feature Selection is a necessary phase of steganalysis in order to achieve high detection accuracy. Steganalytic feature selection methods based on Minimal Information (MI) and Adaptive Weight based Particle Swarm Optimization (APSO) are proposed in this work in order to effectively reduce the high space dimensionality of the sta tistical features used in state-of-the-art steganalysis. First the differentiability of each feature dimension is calculated using MI parameter and further weight based sorting is carried out for the whole feature dimensions. The features sorted in descending order of differentiability among another features. So first few feature dimensions are selected in this phase. In order to further optimize the feature reduction and to reduce inseparable features, weight adaptive PSO in used which further reduces the feature vector space to a threshold value decided. The fitness function used to get the best particles in the PSO is AUC measure of a classifier which is calculated for each particle and the particle having high AUC is chosen as final optimal selected features. The effectiveness of the proposed feature selection is carried out by classifying the combined dataset of stego and cover images taken from BOSSbase dataset using three different machine learning techniques named as decision tress, kNN and SVM. The statistical features used for evaluation are SPAM and CC-PEV feature extraction methods which have 686 and 548 feature space dimensionality. The final feature reduction is tested at 80 and 100 selected features. Experimental results shows that high classification accuracy is achieved with the proposed features reduction methods.