Steganalysis of Low Embedding Rates LSB Speech Based on Histogram Moments in Frequency Domain

Abstract

LSB speech steganography with low embedding rate is an effective method to confront speech steganalysis. It is still a big challenging issue to detect LSB speech steganography with low embedding rate. Based on wavelet packet transform focusing micro change of the signal, this study proposes a statistical analysis method to extract the high order histogram moments in frequency domain which are extremely sensitive to LSB speech steganography, and then train the Support Vector Machine (SVM) classifier, which is used to comprehensively analyze in depth the LSB matching steganography with different lower embedding rates, such as 5% or 10% and so on. Experimental results show that statistical moments of histogram and moments in frequency domain can be used to detect the LSB matching steganography, the detection performance of moments in frequency domain and combined moments is superior to that of statistical moments of histogram; The detection accuracy of the histogram features by Wavelet Packet Decomposition (WPD) is higher than that of the corresponding features by Wavelet Decomposition (WD). The moments in frequency domain features by WPD are particularly prominent in detecting LSB speech steganography with low embedding rates, and the accuracy rate can achieve 60.8% when the embedding rate is only 3%.