Impact of Feature Selection in Classification for Hidden Channel Detection on the Example of Audio Data Hiding

The classification accuracy achieved in applied classification problems depends strongly on the choice of classifiers and, if a model based approach is chosen, on the quality of the model. In this paper, for a selected classification problem from the area of determination of the existence of hidden channels in audio data, the relevance of single features for model generation in a support vector machine based classification procedure is determined. Here we consider nine audio data hiding algorithms as well as an existing audio steganalysis approach. The goal is to sharpen the model used for classification by algorithm specific generation of the feature set used and thereby reducing its dimensionality while keeping the same degree of classification accuracy for hidden channel detection on audio data. We show that for a multi-genre audio test set the impact of feature space reduction is less severe than for a set containing only speech. The fractions of the feature space considered significant in the performed multi-genre and speech evaluations (best results for the percentage of the available feature space considered significant in the tests performed here: 37.4% and 54.5% respectively) are determined for different thresholds of considering a feature significant in single feature classification. A first evaluation on embedding domain distinction is performed, distinguishing between time- and firequency/wavelet-domain. The results for application specific steganalysis achieved here are compared to the results achieved in current image steganalysis schemes.