Selection of image features for steganalysis based on the Fisher criterion

Abstract

A steganalytic feature selection method based on the Fisher criterion used in pattern recognition is proposed in this paper in order to reduce effectively the high dimensionality of the statistical features used in state-of-the-art steganalysis. First, the separability of each single-dimension feature in the feature space is evaluated using the Fisher criterion, and these features are reordered in descending order of separability. Then, starting from the first dimension of the reordered features, as the dimension increases, the separability of each feature component is analyzed using the Fisher criterion combined with the Euclidean distance. Finally, the feature components with the best separability are selected as the final steganalytic features. Experimental results based on the selection of SPAM (Subtractive Pixel Adjacency Matrix) features in spatial-domain steganalysis and CC-PEV (Cartesian Calibrated feature extracted by PEVný) features in DCT-domain steganalysis show that the proposed method can not only reduce the dimensionality of the features efficiently while maintaining the accuracy of the steganalysis, but also greatly improve the detection efficiency.