Selective Ensemble RSM for High Dimensional Steganographic Detection Based on FP-Tree

Abstract

Random subspace method (RSM), which randomly selects low dimensional feature subspace from the original high dimensional feature space to form new training subsets, is an effective ensemble learning method for high dimensional samples. However, RSM also has the drawbacks: Random selection of features does not guarantee that the selected inputs have the necessary discriminant information. If such is the case, poor classifiers are obtained that damage the ensemble. Thus, we put forward a selective ensemble RSM method Based on FP-Tree. The method obtains a refined transaction database and builds a FP-Tree to compact it, next, selects an ensemble size according to the FP-Tree. Since the proposed method only selects part of classifiers to ensemble which can eliminate the poor individual classifiers and obtain better ensemble results than using all the base classifiers. We utilize the proposed method to fight against the newly proposed HUGO steganographic algorithm. Experiment results show that our method has the overall best detection performance.