Measuring Complexity and Predictability in Networks with Multiscale Entropy Analysis

We propose to use multiscale entropy analysis in characterisation of network traffic and spectrum usage. We show that with such analysis one can quantify complexity and predictability of measured traces in widely varying timescales. We also explicitly compare the results from entropy analysis to classical characterisations of scaling and self-similarity in time series by means of fractal dimension and the Hurst parameter. Our results show that the used entropy analysis indeed complements these measures, being able to uncover new information from traffic traces and time series models. We illustrate the application of these techniques both on time series models and on measured traffic traces of different types. As potential applications of entropy analysis in the networking area, we highlight and discuss anomaly detection and validation of traffic models. In particular, we show that anomalous network traffic can have significantly lower complexity than ordinary traffic, and that commonly used traffic and time series models have different entropy structures compared to the studied traffic traces. We also show that the entropy metrics can be applied to the analysis of wireless communication and networks. We point out that entropy metrics can improve the understanding of how spectrum usage changes over time and can be used to enhance the efficiency of dynamic spectrum access networks.