Adaptive Multi-Rate Speech Steganalysis Method Based On Pitch Delay

With the widespread application of adaptive multi-rate speech steganography in instant messaging,steganalysis technology,as a countermeasure against steganography,has also become a research hotspot.In this paper,an adaptive multi-rate speech steganalysis method based on pitch delay is proposed by studying the principle of AMR codec.The main work includes the following two parts:(1)In view of the problem of low detection accuracy of the current steganalysis scheme based on pitch delay,a speech steganalysis method based on the correlation between sub-frames of pitch delay is proposed.Firstly,this method improves the existing steganalysis method based on the Markov second-order difference characteristic of pitch delay to reduce the dimension of the detection feature.the existing steganalysis method based on Markov second-order difference characteristics is improved to reduce the dimension of detection features.Then,a new detection feature is extracted by using the ratio of pitch delay between the first and second sub-frames and between the first and third sub-frames,which can break the correlation between sub-frames by steganography.Finally,the above two detection characteristics are combined to form the final detection characteristics,which are input into the support vector machine.Experimental results show that the detection performance of the proposed scheme is obviously superior to the existing steganalysis schemes.(2)Aiming at the poor detection performance of current steganalysis methods based on support vector machine,a speech steganalysis method based on multi-classifier fusion technology is proposed.This method constructs a new multi-classifier model,which contains two classifier sets and each classifier set contains four sub-classifiers.In the process of steganalysis,the steganalysis characteristics of pitch delay are input into the first classifier set,then the output of the first classifier set is input into the second classifier set as a special detection characteristic,and then the prediction results of each classifier set are fused through the trained weights,and the fusion algorithm is the weighted average method.Finally,the final detection result is generated.The experimental results show that when the same detection features are used,the performance of the proposed speech steganalysis scheme based on multi-classifier fusion is significantly better than that based on support vector machine.