Steganalysis of adaptive multi-rate speech using statistical characteristics of pulse pairs

In this work, we concentrate on the steganalysis of adaptive multi-rate (AMR) speech, which widely exists in mobile Voice-over-IP services. Compared with the state of the arts, the most significant advantage of this work is that more accurate and more complete steganalysis features are presented. To avoid the impact of the possible interchange of pulse positions in one track, we characterize AMR speech exploiting the statistical properties of pulse pairs. Specifically, we employ the probability distributions of pulse pairs as long-term distribution features, extract Markov transition probabilities of pulse pairs as short-term invariant features, and adopt joint probability matrices of pulse pairs as features based on track-to-track correlations. Moreover, to optimize the feature set and reduce its dimension, we introduce a feature selection mechanism using adaptive boosting technique. Exploiting the well-selected features, we further present a steganalysis of AMR speech based on support-vector-machine. The proposed method is evaluated with a good supply of AMR-encoded speech samples, and compared with the state-of-the-art methods. The experimental results demonstrate that the proposed method can effectively detect the state-of-the-art steganography methods for AMR speech, and significantly outperforms the state-of-the-art steganalysis methods on detection accuracy, false-positive rate and false-negative rate for any given embedding rate or speech samples with any given length. In particular, the proposed method can provide accurate detecting results for the existing steganographic methods in a timely manner, and thereby be applied in the steganalysis scenario for real-time speech streams.