Applying temporal feedback to rapid identification of BitTorrent traffic

BitTorrent is one of the dominant traffic generating applications in the Internet. The ability to identify BitTorrent traffic in real-time could allow network operators to manage network traffic more effectively. In this paper we demonstrate that erroneous output of a Machine Learning based classifier is randomly distributed within a flow, allowing the application of temporal feedback to improve the overall classifier performance. We propose and evaluate a number of feedback algorithms. Our results show that we are able to improve classification outcomes (Recall by 2.4% and Precision by 0.1%) whilst both improving classification timeliness from three to two minutes, and improving robustness against future changes to the BitTorrent protocol.