Abstract: Because it is difficult to identify encrypted VoIP traffic,we proposed a UDP statistical fingerprinting mixture models to enhance the accuracy and stability of VoIP traffic identification． We used the statistical features of UDP flow along with the anomaly score of traffic flow in which the statistical fingerprinting model is used to identify a traffic flow to train a Support Vector Machine(SVM)classification model,and used a hyperplane of high-dimensional space instead of a threshold point to classify the traffic．Because we use both the packet level features and flow level features in our mixture models,the impact of the deviation of traffic features which is caused by the instability of network will be decreased．The results of our experiment show that the precision of VoIP traffic is over 97% in our model,and our model is more stable compare with the statistical fingerprinting model and Support Vector Machine(SVM)．

Key words: Statistical fingerpinting,Voice over internet protocol,Traffic classification,Support vector machine,Internet