Threshold feature selection (TFS), an improvement over forward selection in steganalysis feature selection

High-dimensional feature sets proposed for steganalysis are able to model a lot of dependencies between pixels. Although these dependencies can reveal the changes induced by steganography in pixels or JPEG coefficients values, for some known steganography methods, some features may not change significantly. The significance of feature selection in steganalysis relies on two facets. First, by keeping particular features constant in steganography, a more resistant embedding scheme can be obtained; in other words, the most prominent features for detecting a particular steganography method can reveal the weaknesses of the method. Second, extraction of high-dimensional feature sets in steganalysis is a time-consuming process and this issue prevents steganalysis applicability in the real-word problems, whereas in some cases feature selection might lead to reduction of feature extraction time. In this paper, a novel and simple method is suggested in order to select the most prominent features from the feature sets. The aim of the proposed method is not to increase the classification accuracy, but it aims at decreasing the negative effect of removing weak-discriminant features on classification accuracy and therefore at decreasing the classification complexity along with minimum degradation. Also another goal is to find the features that should be preserved in steganography process in order to avoid detection of steganography. The proposed method begins with sorting the features based on a selective ranking function, and then by considering a value as threshold, only the features that can increase classification accuracy more than the threshold value are selected. Due to its similarities to Forward selection algorithm, it is compared with this selection. The comparison results showed improvement in terms of the selected features.