Steganalysis of AMR Speech Based on Multiple Classifiers Combination

In this paper, we focus on steganalysis in adaptive multi-rate (AMR) speech streams, whose goal is to detect covert communication behaviors effectively to prevent illegal uses of AMR steganography. Differing from the existing methods for speech steganalysis, we present a novel steganalysis scheme based on multiple classifiers combination. Specifically, we design a new combined classifier model, whose main idea is as follows. First of all, steganalysis features are respectively inputted into two different classifier sets to obtain the first and second types of prediction results. Then, the second type of prediction results are viewed as a special type of features and fed into another classifier set to obtain the third type of prediction results. Finally, all the three types of prediction results are fused to achieve a final detection result. The presented steganalysis scheme is comprehensively evaluated using two types of state-of-the-art features for fixed codebook (FCB) parameters, namely, pulse-pair characteristics-based features and pulse-correlation based features, and compared with the existing methods based on support vector machines (SVM). The experimental results demonstrate that the proposed scheme is feasible and effective, and significantly outperforms the previous SVM-based methods for detecting steganography in FCB domains in every case.