Waterfall traffic identification: optimizing classification cascades

The Internet transports data generated by programs whch cause various phenomena in IP flows. By means of machine learning techniques, we can automatically discern between flows generated by different traffic sources and gain a more informed view of the Internet. In this paper, we optimize Waterfall, a promising architecture for cascade traffic classification. We present a new heuristic approach to optimal design of cascade classifiers. On the example of Waterfall, we show how to determine the order of modules in a cascade so that the classification speed in maximized, while keeping the number of errors and unlabeled flows at minimum. We validate our method experimentally on 4 real traffic datasets, showing significant improvements over random cascades.