Predicting network failures using anomaly detection methods

点击翻译按钮获取中文摘要

Networks of all kinds often experience anomalous behaviour that is a sudden and short-lived deviation from its normal operation. Examples include attacks or large data transfers in networks. Some anomalies are deliberately caused by intruders with malicious intent such as cyber-attack, while others may be purely accidental such as an overpass falling in a busy road network. Quick detection is needed to initiate a timely response, especially if the goal of the detector is to identify and/or predict network failures, to detect security issues such as intrusion or denial of services attacks, to detect changing traffic patterns, etc. The purpose of this project is to generate a general tool applicable to different kind of networks. For this reason, it is useful to consider machine learning based techniques that can adapt to each network's specific conditions. Furthermore, given that anomalies in networks can take many different forms that cannot be anticipated ahead of time, anomaly detection algorithms such as one-class learning systems are desirable. Indeed, such approaches will allow to model the normal operation of the network and issue an alarm when a deviation from this normal operation is detected. The networks that will be considered in this project will be massive networks and may be generating stream data. Therefore, we will adapt existing one-class learning techniques or develop new ones that scale well to big data sets and can deal with a fast online data presentation regimen.