Speech Steganalysis of Least Significant Bits Based on the Percentage of Equal Adjacent Samples

The popularity of audio formats usually attracts the attention of intruders and criminals to use this medium as a cover for establishing their secret communications. The extensive use of this formats, along with  various modern techniques, designed for audio steganography, can cause the cyber spaces to be insecure environments. In order to deal with threats, some audio steganalysis techniques have been presented that statistically analyze various audio formats, such as music, MP3, and VoIP, efficiently. Among the presented approaches, combining the techniques of signal processing and machine learning has made possible the creation of steganalyzers that are highly accurate. However, since the statistical properties of audio files differ from purely speech ones, the current steganalysis methods cannot detect speech stego files,            accurately. Another issue is the large number of analysis dimensions which increase the implementation cost, significantly. As response to these issues, this paper proposes the percentage of equal adjacent  samples (PEAS) feature, as a one-dimensional feature for speech steganalysis. Using a classifier, based on the Gaussian membership function, on stego instances with 50% embedding ratio, the evaluation results for the designed steganalyzer, show a sensitivity of 99.82%. Additionally, it can efficiently estimate the length of a hidden message with the desirable accuracy. Also, the PEAS steganalysis was evaluated on a database, containing classic music instances, and the results show an 81.2% efficient performance.