Distributed steganalysis of compressed speech

In this paper, we present a distributed steganalysis scheme for compressed speech in voice-over-IP scenarios to provide fast and precise detection results. In this scheme, each speech parameter available for concealing information is designed to be detected independently exploiting the corresponding optimal detection feature. To achieve this purpose, we introduce four detection features, including histogram distribution, differential histogram distribution, Markov transition matrix and differential Markov transition matrix. These features stem from both long-time distribution characteristics and short-time invariance characteristics of speech signals. We evaluate their performance for steganalysis based on support vector machines with a large number of steganographic G.729a speech samples at different embedding rates or with various sample lengths and compare them with some existing algorithms. The experimental results demonstrate that the presented algorithms can offer excellent steganalysis performance for all speech parameters in any case and outperform the previous ones. Moreover, it is proved that the four features have diverse performance for steganalysis of different speech parameters, which suggests that it is feasible to achieve the distributed steganalysis employing the optimal feature to detect the corresponding parameter in a faster and more efficient manner.