Adaptive feature selection for image steganalysis based on classification metrics

Abstract

Feature selection can remove redundant and useless features, which is an essential way to improve steganalysis efficiency. However, with the diversity of steganalysis features, feature selection has run into bottlenecks of high time cost, poor universality, and experience depending on parameter setting. To this end, an adaptive steganalytic feature selection based classification metrics is proposed. First, the categories to which the features belong are redefined and classified. Secondly, three metrics are proposed for three different categories of features, to make the metric more precise. Then, to reduce the computational cost and optimize parameters, two adaptive threshold models are designed, which achieve the purpose of fast and effective feature selection without relying on the time-consuming classification results. Experimental results on 11 typical steganalytic features demonstrate that compared with classic and state-of-the-art feature selection methods, the proposed method achieves competitive performance on detection accuracy, calculation cost, storage cost, and universality.