Dynamic Normal Profiling for Anomaly Detection Systems

Our research addresses constructing a dynamic normal profile for anomaly detection systems without requiring timeconsuming retraining. We propose to continuously update normal profiles by keeping the most recently employed patterns whose amount is dynamically determined. Active window adjustment through a simplified concept drift algorithm helps to keep relevant instances without having to contain outdated patterns as well. The ability to dynamically adapt the normal profiles provides a significant foundation for effective anomaly detection.