Internet Traffic Classification using multi-classifier systems

Abstract

The field of Internet traffic classification has been growing fast recently. This growth is based on the increasing number of Internet users, and varieties of data intensive applications being used, such as video streaming and file sharing services. Traffic classification is a method for assigning classes to network flows based on features passively observed in the traffic. It is used by ISPs to implement QOS schemes and fine-tune their network configurations for different kinds of traffic. Different classifiers work better on different traffic classes, so a system combining different methods would potentially achieve higher accuracy than any single one, and the multi-classifier system would be more robust to changes in the distribution of applications in the traffic data. We have extended an already implemented classification platform (TIE) to support ensemble classifiers and tried different configurations of classifiers, traffic features, and decision combiners in order to enhance both classification accuracy and early classification of traffic flows, which is important in real-time implementations. We have gathered anonymous traffic data with its corresponding ground truth to be used for the training and testing of the proposed methods. We will explain the extended platform and its experimental results.