Traffic classification and verification using unsupervised learning of Gaussian Mixture Models

Abstract

This paper presents the use of unsupervised Gaussian Mixture Models (GMMs) for the production of per-application models using their flows' statistics in order to be exploited in two different scenarios: (i) traffic classification, where the goal is to classify traffic flows by application (ii) traffic verification or traffic anomaly detection, where the aim is to confirm whether or not traffic flow generated by the claimed application conforms to its expected model. Unlike the first scenario, the second one is a new research path that has received less attention in the scope of Intrusion Detection System (IDS) research. The term “unsupervised” refers to the method ability to select the optimal number of components automatically without the need of careful initialization. Experiments are carried out using a public dataset collected from a real network. Favorable results indicate the effectiveness of unsupervised GMMs.