A novel traffic classification algorithm using machine learning

Internet traffic classification is of prime importance to the areas of network management and security monitoring, network planning, and QoS provision. But the Traditional Classifications depend on certain header fields (take port numbers for instance). These port-based and payload-based approaches will be out of action when a lot of applications like P2P use dynamic port numbers. Masquerading techniques and payload encryption requires a high amount of resource of computing and is easily infeasible in the protocol that unknown or encrypted. This paper describes a different level in network traffic-analysis using an unsupervised machine learning technique. In this approach flows are automatically classified by exploiting the different statistics characteristics of flow. We implement and estimate the efficiency and feasibility of our approach using data at different location of Internet. A new attribute selection method is put forward to determine optimal attribute set and evaluate the influence.