Abstract: Obtaining qualified training data for protocol identification generally requires domain experts to be involved, which is time-consuming and laborious. A novel approach for network protocol identification based on active learning and SVM algorithm was proposed. The experimental evaluations on real-world network traces show this approach can accurately and efficiently classify the target network protocol from mixed Internet traffic, and meanwhile display a significant reduction in the number of labeled samples. Therefore, this approach can be employed as an auxiliary tool for analyzing unknown protocols in real-world environment.