Use of a Cepstral Information Norm for Anomaly Detection in a BGP-inferred Internet

In this paper we use a particular type of mutual information norm — the cepstral information norm — for anomaly detection at the router level in the Internet. We combine the cepstral norm with a state space Kalman filter to define two distance metrics to capture anomalous behaviour. These metrics are implemented using a subspace-based model-free paradigm to aid realtime analysis. We infer a top level Internet topology using Border Gateway Protocol router updates and characterise the structural evolution of the network using a selection of graph metrics. Analysis over one week of non time-homogeneous updates, which includes The SQL Slammer worm event, shows the combined use of the two cepstral distance metrics detects the occurrence and severity of anomalous network events.