Recent results
(K. Fukuda, K. Cho, H. Esaki, R. Fontugne from the NII, IJJ and Univ. Tokyo (Japan), CNRS-JSPS program) Signal processing is a great asset to study the communications over the Internet network. Accurate host-level traffic classification is made possible by relying on statistical features describing traffic of a host, e.g. from the Multi-Scale Gamma Model, or from traffic patterns reminiscent of traffic graphlets. Leveraging on previous works, using sketches and multi-resolution analysis, we prove that long memory is a robust property in traffic, as shown on seven years of collected traffic. We re-investigated the relationship between long memory (modeled as self-similarity) and heavy-tailness of flows, theoretically (Taqqu’s theorem), also questioning the respective roles of the flow and session levels, and experimentally on a grid, proving that long memory is a stable feature of Internet traffic. Finally, lacking methods to characterize and benchmark anomaly detectors, we used graph analyses to compare them, and we validated that by annotating the anomalies in the MAWI traffic database.