In this paper we extend the state-of-the-art steganalysis tool developed by Kodovsky and Fridrich: the Kodovsky's ensem-ble classifiers. We propose to boost the weak classifiers com-posing the Kodovsky classifier. For this, we minimize the probability of error thanks to a regression approach of low complexity. We also propose a post-selection of features, achieved after the learning step of all the weak classifiers. For each weak classifier, we identify a subset of features reducing the probability of error. Both proposals are of neg-ligeable complexity compared to the complexity of the Kodovsky classifier. Moreover, these two proposals significantly increase the performance of classification.