Estimation and detection of network traffic

The classification of Internet traffic is of interest in areas like differentiated services and network security. Such classification is usually done using the packet header field of ´port number´. However, recent developments in networking techniques have rendered the port numbers unreliable for this purpose. Our scheme of classification uses the distribution of packet sizes in a buffer or collected during a short time interval at a switch or router. We demonstrate that applications can be classified by these distributions and that estimation of the amount of each application is possible. We compare two methods for estimation of the traffic in various applications: MMSE estimation; POCS and neural networks. It is observed that a highly accurate detection and estimation is obtained using the artificial neural networks.