Optimization of rich model based on fisher criterion for image steganalysis

Abstract

The rich model features used for steganalysis of modern steganography usually have redundant features, which would influence the detection accuracy and performance. To solve this problem, this paper proposes a feature optimization method for rich model features based on the improved Fisher criterion. Based on the principle that “The within-class variance should be smaller and between-class variance should be larger”, the improved Fisher criterion is used in this paper to evaluate the separability of feature component, sub-model features and feature vector, respectively. Then, two strategies are presented to optimize the rich model features. In the experimental analysis, the proposed method is applied to optimize the typical rich model SRM that can be used to detect typical modern steganography HUGO, and the result is compared with existing feature selection methods. Experimental results show that, in addition to significantly reducing the dimensionality and shortening the detection time, the proposed rich model feature optimization method has a better performance on maintaining the detection accuracy of the original rich model features.