Online classification of user activities using machine learning on network traffic

The daily deployment of new applications, along with the exponential increase in network traffic, entails a growth in the complexity of network analysis and monitoring. Conversely, the increasing availability and decreasing cost of computational capacity have increased the popularity and usability of machine learning algorithms. In this paper, a system for classifying user activities from network traffic using both supervised and unsupervised learning is proposed. The system uses the behaviour exhibited over the network and classifies the underlying user activity, taking into consideration all of the traffic generated by the user within a given time window. Those windows are characterised with features extracted from the network and transport layer headers in the traffic flows. A three-layer model is proposed to perform the classification task. The first two layers of the model are implemented using K-Means, while the last one uses a Random Forest to obtain the activity labels. An average accuracy of 97.37% is obtained, with values of precision and recall that allow online classification of network traffic for Quality of Service (QoS) and user profiling, outperforming previous proposals.