Machine Learning for User Traffic Classification in Wireless Systems

With the advent of Internet of Things telecommunications will play a crucial role in every day life. The rapidly growing demand for radio services by millions of user all over the world will make the radio spectrum an increasingly valuable resource. The modern standards of communications provide a static utilization of the radio spectrum resources, which results in its under-utilization. Therefore let us imagine a dynamic sharing of the radio resources, where every device can use a portion of such resources if and only if they are not utilized yet. In this regard, the Federal Communication Commission (FCC), the authority that regulatizes specturm sharing in the USA, has decided to free some portions of the radio spectrum in order to allow its dynamic usage. From this perspective, devices will have to probe the RF scene in time, space and frequency domain to ensure that a well defined portion of the spectrum is free, making multidimensional spectrum analysis mandatory. On large scale infrastructures indeed, the classification of trasmissions, the spatial localization of the events and the search for spectrum holes, might be done with an extensive use of machine learning algorithms. Traffic classification allows to automatically recognize the user-level application that has generated a given stream of packets from direct observation of the packets or from the spectrum occupancy. An in-depth knowledge of the composition of traffic, as well as the identification of trends in application usage, may help operators improving network design and provisioning. Moreover, traffic classification represents the first step in the direction of activities such as anomaly detection for the identification of malicious use of network resources, and for security operation such as firewalling and filtering of unwanted traffic. This work proposes a machine learning approach for wireless traffic classification in common bands, such as WiFi, with low-cost measurement devices.