Identification of Network Applications with Co-training

Recently, A new generation of network applications, such as P2P, has started to consume a large amount of network resources, thus produce several problems to network operators, which make the identification of network applications become a new and difficult challenge for both network operators and the network measurement community. Traditional identification techniques that rely on the well-known ports registered by the IANA are no longer valid because of the inaccuracy of its classification results. This situation has motivated us to study how to use statistical features of the flow or characteristics patterns in the payload of the packets to solve the problem of application identification in the network traffic. In this paper, we present an effective supervised machine learning technique based on the characteristics patterns and the behaviors pattern to accurately identify the network traffic. We evaluate our method with an existing passive network monitoring system in our campus network and achieve an overall accuracy greater than 89% with a little training samples.