Learning Mixed Traffic Signatures in Shared Networks.

On shared, wide-area networks (WAN), it can be difficult to characterise the current traffic. There can be different protocols in use, by multiple data streams, producing a mix of different traffic signatures. Furthermore, bottlenecks and protocols can change dynamically. Yet, if it were possible to determine the protocols (e.g., congestion control algorithms (CCAs)) or the applications in use by the background traffic, appropriate optimisations for the foreground traffic might be taken by operating systems, users, or administrators. We extend previous work in predicting network protocols via signatures based on a time-series of round-trip times (RTT). Gathering RTTs is minimally intrusive and does not require administrative privilege. Although there have been successes in using machine learning (ML) to classify protocols, the use cases have been relatively simple or have focused on the foreground traffic. We show that both k-nearest-neighbour (K-NN) with dynamic time warp (DTW), and multi-layer perceptrons (MLP), can classify (with useful accuracy) background traffic signatures with a range of bottleneck bandwidths.