Improved ensemble growing method for steganalysis of digital media

Authors

Ramin Toosi, Sadaf Salehkalaibar, Mohammad Ali Akhaee

Abstract

Ensemble methods provide significantly a low training complexity which enables a steganalyst to work with high dimensional cover models. In this paper, we propose an improved method for growing an ensemble of classifiers. Each classifier of the ensemble is trained with a subspace of the features. The features are selected based on a predefined probability distribution. This distribution is proportional to the information that each feature carries about the class label. To make the final decision, the outputs of a subset of classifiers are fused using the majority voting. An information-theoretic classifier selection method is employed to find an appropriate subset of classifiers for the fusion phase. Using these modifications, we pursue different improvements such as reduction of memory needed to store the model, reduction of the prediction time, and enhancing the generalization ability of the model. The proposed method is compared with some state of the art methods. Results show that the number of the classifiers and the number of features in each subspace are significantly reduced. Also, there exists a small reduction in the error rate. On the other hand, simulations show a small increment in the training time while having a significant reduction in the prediction time of the ensemble.