Cover signal specific steganalysis: the impact of training on the example of two selected audio steganalysis approaches

The main goals of this paper are to show the impact of the basic assumptions for the cover channel characteristics as well as the impact of different training/testing set generation strategies on the statistical detectability of exemplary chosen audio hiding approaches known from steganography and watermarking. Here we have selected exemplary five steganography algorithms and four watermarking algorithms. The channel characteristics for two different chosen audio cover channels (an application specific exemplary scenario of VoIP steganography and universal audio steganography) are formalised and their impact on decisions in the steganalysis process, especially on the strategies applied for training/ testing set generation, are shown. Following the assumptions on the cover channel characteristics either cover dependent or cover independent training and testing can be performed, using either correlated or non-correlated training and test sets. In comparison to previous work, additional frequency domain features are introduced for steganalysis and the performance (in terms of classification accuracy) of Bayesian classifiers and multinomial logistic regression models is compared with the results of SVM classification. We show that the newly implemented frequency domain features increase the classification accuracy achieved in SVM classification. Furthermore it is shown on the example of VoIP steganalysis that channel character specific evaluation performs better than tests without focus on a specific channel (i.e. universal steganalysis). A comparison of test results for cover dependent and independent training and testing shows that the latter performs better for all nine algorithms evaluated here and the used SVM based classifier.