Optimizing flow sampling for network anomaly detection

Sampling techniques are widely employed in high-speed network traffic monitoring to allow the analysis of high traffic volumes with limited resources. Sampling has measurable negative impact on the accuracy of network anomaly detection methods. In our work, we build an integrated model which puts the sampling into the context of the anomaly detection used in the subsequent processing. Using this model, we show that it is possible to perform very efficient sampling with limited impact on traffic feature distributions, thus minimizing the decrease of anomaly detection efficiency. Specifically, we propose an adaptive, feature-aware statistical sampling technique and compare it both formally and empirically with other known sampling techniques - random flow sampling and selective sampling. We study the impact of these sampling techniques on particular anomaly detection methods used in a network behavior analysis system.