Research Of Speech Steganalysis Method

The development of steganography in networks and communication arises the danger for illegal activities. Stegalysis can be used for watching, analyzing and deciphering of information to protect networks and society. This paper gives some research of stegalysis based on digital audio data, and the detailed contents are as follows.After researching on the statistical discrepancy of audio quality before data embedding and after, two blind speech stegalysis methods: stegalysis based on multiple linear regression and SVM (Support Vector Machine) are proposed. They both measure quality evaluation methods depends on ANOVA (Analysis of Variance), choose some quality metrics which reflect statistical discrepancy of pure speech and stego speech well, then training classifier can be modeled. The former uses multiple linear regression classifier and latter is SVM. Simulation results indicated that they general do well in transform domain steganographic, such as DFT, DCT and DWT domain. Correct detection rate of method using SVM is higher.By analyzing the correlation of different quality metrics, an audio steganalysis based on integrated quality metrics is described. The blind detection simulating model with SVM is created with integrated audio quality metrics, which chosen by multi-dimension feature extraction, and the probably stego-audio can be detected. Simulation results indicated that multi-dimension feature extraction performs higher conect detection rate.After researching principle of echo hiding, a steganalysis for echo hiding based on statistics charcters is proposed.Coefficients of skewness and kurtosis, belonging to complex cepstrum coefficient of pure and stego speech, are compared and trained for thresholds,which used for blind steganalysis of echo hiding. Expermental results show that the presented strategy can effectively detect four kinds of echo hiding methods, including custom, positive and negative, backward and forward kernels etc.Blind stegalysis system has been designed and verified by experiment results.