Multidimensional and Adaptive Non-Intrusive Anomaly Detection in Network Services
Performance monitoring and anomaly detection is critical for today's complex network services. However, conventional detection mechanisms like thresholding for single parameters do not deal well with distributed services where many different performance indicators might influence the overall performance of the combined system. The authors propose the analysis of signal-based metrics with a combination of wavelet transform and Mahalanobis distance to automatically detect anomalies in network services. In contrast to conventional detection methods like thresholding, their technique adapts automatically to gradual changes in the measured signals and deals well with periodical load patterns. It supports multidimensional analysis to improve reliability and significance of the detection and provides confidence values, which are the base for judging the anomaly.