融合频域矩和MFCC特征的语音隐写分析

摘要

以聚焦信号微细变化的小波包变换为基础,采用统计分析方法,设计并提取对信息隐写过程极为敏感的小波包系数高阶直方图频域统计矩,融合反应人耳感知特性的MFCC矢量,以上述各矢量为特征训练SVM分类器,对不同嵌入率的LSB替换语音隐写方法进行分析,重点研究低嵌入率(如3％和5％等)下的LSB替换隐写检测问题.实验结果表明,小波包系数直方图频域统计矩和MFCC标准及差分特征对LSB替换隐写检测性能均优;在低嵌入率时,MFCC标准及差分特征优于直方图频域矩特征的检测性能;在各种嵌入率下,MFCC标准及差分和直方图频域矩二者融合特征的检测性能尤为突出,在嵌入率为3％时,检测准确率达到68.3％.%Based on wavelet packet transform of micro change of the focusing signal,a statistical analysis method was adopted to extract the high order histogram moments in frequency domain of wavelet packet coefficients which were extremely sensitive to information steganography.Mel frequency cepstrum coefficient (MFCC) vectors which reflected human auditory perception properties were exacted to be integrated with histogram moments in frequency domain.Support vector machine (SVM) was trained for classification based on mentioned features.Least significant bit (LSB) replacing speech steganography was analyzed with some different embedding rates,and the steganography detection problem of LSB replacement with low embedding ratio was focused,such as 3％ or 5％ and etc.Experimental results show that histogram moments in frequency domain and MFCC vectors can be used to detect the LSB replacing steganography excellently.The detection performance of MFCC vectors is superior to that of histogram moments in frequency domain with low embedding rates.The detection performance of the two fusion characteristics is particularly prominent,and the accuracy rate can achieve 68.3 ％ when the embedding rate is only 3 ％.