Improved classification of known and unknown network traffic flows using semi-supervised machine learning

Abstract

Modern network traffic classification approaches apply machine learning techniques to statistical flow properties, allowing accurate classification even when traditional approaches fail. We base our approach to the task on a stateof- the-art semi-supervised classifier to identify known and unknown flows with little labelled training data. We propose a new algorithm for mapping clusters to classes to target classes that were previously difficult to classify. We also apply alternative statistical features. We find our approach has an accuracy of 95.10 %, over 17 % above the technique on which it is based. Additionally, our approach improves the classification performance on every class.