Clustering for online video-conference traffic classification

With video, gaming and multimedia traffic representing almost 80% of total Internet traffic and RTC applications spread increasing each day more, the overall costumers expectation towards quality and performance of such services is also increasing. An emerging metric called Quality of Experience (QoE) has the goal to depict these qualitative aspects linked to users satisfaction and the development of new QoE driven network management frameworks is of key interest in industrial research. However, in order to take proper management actions the first problem to be solved is network traffic classification. Because of the increased sensibility towards privacy issues older methods such as port-based and Deep Packet Inspection are nowadays almost unusable, but recently machine learning techniques relying on statistical analysis of the flows have seen an important development. Anyhow, supervised machine learning approaches usually require a costly and unfeasible labelling of the data, while fully unsupervised solutions could be of difficult interpretation. Case study of this thesis work are video-conferencing systems, particularly Webex Meetings, with the goal of grouping almost in real-time several generated flows into homogeneous groups, each one containing flows of a target category, such as audio, video and screen sharing, by using machine learning unsupervised algorithms. The use of unsupervised algorithms has the potentiality of RTC streams classification increased generalizability to different video-conferencing applications and its actual feasibility also with respect to already existing supervised solutions is explored. Interesting results have been obtained implementing a cascade K-Means algorithm by using as features simple statistics computed per-second on RTP packet flows: first two macro clusters with well separated audio and video flows but not considering screen sharing are identified, then a further clustering allows to separate screen sharing from audio and to distinguish different video qualities. Equal frequency binning on original data have been proved to improve clusters purity and robustness to initial conditions, while a constrained K-Means version exploiting background knowledge was also tried to influence clusters creation in terms of purity and centroids positioning.