Move Statistics-Based Traffic Classifiers Online

A number of recent works have proposed using data mining and machine learning techniques to classify traffic flows based on statistical flow characteristics. Most of these classifiers work offline, since full-flow statistics are not available until a flow is finished. Therefore, it is usually too late to take actions for online deployment. In this paper, we propose a simple and effective technique to make these classifiers workable online. The idea is that different applications will show distinct traffic patterns from the very beginning of the flow, and so using statistics extracted from the first few packets can distinguish applications. Preliminary result shows that our method can achieve high flow accuracy, just a bit lower than using full-flow statistics. Using fewer packets per flow is promising, since it not only enables early classification and can easily apply to most of the existing classifiers, but also saves memory and computing power.