| With the extensive use of Adaptive multi-rate(AMR)speech in the field of IP phone and mobile communication,steganography methods based on AMR speech emerge one after another.However,like most security techniques,adaptive multi-rate speech-based steganography brings greater security risks to information security if it is abused by criminals.Therefore,steganalysis based on adaptive multi-rate speech has become an important research topic.From the existing researches,there are still many problems to be solved,for example,the feature dimensions in some methods are to o high.Aiming at the problems in existing researches,combined with the principle of AMR speech coding,this paper respectively makes some researches on steganalysis methods in pitch delay,fixed codebook,and linear prediction parameter domain,and the relevant research works are as follows:(1)A steganalysis method based on statistical characteristics of pitch delay is proposed to solve the problems of high feature dimension and incomplete representation of pitch delay characteristics of AMR speech.We obtain low-dimensional but high-efficiency second-order difference statistical features of pitch delay value by screening the existing features,and introduce the parity statistical feature to make up for the insufficient expression of second-order difference statistical feature.A large number of samples are used to evaluate the performance of the proposed method and compared with the existing method with the aid of SVM-based classifier.The experimental results show that the proposed method can afford bet ter detection results than the existing method under different embedding rates.(2)In order to solve the problem of high-dimensional feature in the existing method,an AMR steganalysis method based on XGBoost(e Xtreme Gradient Boosting)is proposed.In this method,XGBoost algorithm is used to select the features based on pulse value to get the low-dimensional features.A large number of samples are used to evaluate the performance of the proposed method and compared with the existing method with the aid o f SVM-based classifier.The results show that the feature dimension of this method(minimum 70 dimension,maximum 289 dimension)is obviously lower than the 498 dimension of the current best method,and its detection performance is better than the suboptimal method,and not inferior to the current best method.(3)In order to realize the efficient detection of steganography in the linear prediction parameter domain,a steganalysis method based on spatial local statistical characteristics is proposed.The p rinciple is to transform the one-dimensional parameter sequence into a two-dimensional "parameter block",and uses the convolutional neural network to model the spatial local statistical characteristics of the speech parameter block for the effective features as a whole.A large number of samples are used to evaluate the performance of the proposed method and compared with the existing method with the aid of SVM-based classifier.The experimental results show that the method is feasible and effective in det ecting linear prediction parameters-based steganography method,and has better detection performance than the existing method. |
