Page-Based Anomaly Detection in Large ScaleWeb Clusters Using Adaptive MapReduce

摘要

While anomaly detection systems typically work on single server, most commercial web sites operate cluster environments, and user queries trigger transactions scattered through multiple servers. For this reason, anomaly detectors in a same server farm should communicate with each other to integrate their partial profile. In this paper, we describe a real-time distributed anomaly detection system that can deal with over one billion transactions per day. In our system, base on Google MapReduce algorithm, an anomaly detector in each node shares profiles of user behaviors and propagates intruder information to reduce false alarms. The web log data, about 250GB in size, contains over one billion transactions recorded in a day.