Towards Efficient Flow Sampling Technique for Anomaly Detection

With increasing amount of network traffic, sampling techniques have become widely employed allowing monitoring and analysis of high-speed network links. Despite of all benefits, sampling methods negatively influence the accuracy of anomaly detection techniques and other subsequent processing. In this paper, we present an adaptive, feature-aware sampling technique that reduces the loss of information bounded with the sampling process, thus minimizing the decrease of anomaly detection efficiency. To verify the optimality of our proposed technique, we build a model of the ideal sampling algorithm and define general metrics allowing us to compute the distortion of traffic feature distribution for various types of sampling algorithms. We compare our technique with random flow sampling and reveal their impact on several anomaly detection methods by using real network traffic data. The presented ideas can be applied on high-speed network links to refine the input data by suppressing highly-redundant information.