Fast Steganalysis Techniques For Voice-over-IP Streams

As the fast development of Voice over IP(Vo IP)services in instant message softwares,the research community starts paying attention to employing Vo IP streams as steganographic carriers.Although Vo IP-based steganography can provide secure communication,the illegal use of it will pose threats to the public safety.To confront this challenge,steganalysis,the countermeasure of steganography,has attracted increasing attention.However,there are some shortcomings with efficiency for the existing steganalysis methods.For example,the low detection accuracy(especially at low embedding rates)and high-dimensional steganalysis features will make it difficult to obtain accurate detection results fast.To address the above issues,this paper aims at exploring fast and effective steganalysis techniques for some typical steganographic algorithms in Vo IP scenarios in term of features,classifiers and deep learning,which are listed below:Firstly,steganography in inactive Vo IP frames has large steganographic capacity and can maintain acceptable speech quality,which lacks effective detection method.After analyzing the influences induced by steganographic operations,we propose two low-feature-dimensional detection methods.The former utilizes the poker test results of the encoding parameters in inactive frames as the steganalysis features.The latter employs the statistics of zero crossing count and Mel-frequency cepstral coefficients as the steganalysis features.The experimental results demonstrate both methods can effectively detect the steganography behaviors in inactive Vo IP frames.Though the feature dimension of the latter method is slightly higher than the former's,the latter method can provide better detection performance.Moreover,the latter method can effectively detect the steganographic behaviour with several inactive frames.Secondly,to solve the problem that the feature dimension of the existing steganalysis method in the fixed codebook of adaptive multi-rate speech streams is high,we proposed a set of steganalysis features based on the pulse correlation.This feature set is 100-dimensional,which is only 1/5 of that of the existing method.Using support vector machine as the classifier,the detection performance outperforms the existing methods.To further improve the detection performance,we design a novel multiple classifiers combination model.We utilize different steganalysis features to evaluate the proposed multiple classifiers combination model.The experimental results show that the best method based on the proposed model has better performance than the support vetor machine and can reach more than 90% accuracy when the sample length is 10 s at the embedding rate of 30%,which is about 10% higher than the state of the art.Thirdly,steganography based on bitrate modulation neither modifies Vo IP-based protocols nor Vo IP-based payload,which can resist traditional steganalysis methods.The existing steganalysis features based on bitrate interval probability are not complete and the feature dimension is proportion to the length of speech sample.To fill these gaps,we propose a novel end-to-end steganalysis method via deep learning.The proposed method employs recurrent neural network to learn the long-term bitrate correlation and one-dimensional convolution neural network to learn the short-term bitrate correlation.Then,the learned features are fused to obtain the final detection result.The experimental results show that the proposed method is effective and the detection accuracy can be 10% higher than the existing method with the sample length of 10 s and low embedding rates.