This document, developed within Work Package 6, Task 3 of the GN5-2 project, presents an approach for detecting anomalies in perfSONAR network performance data using autoencoders. The goal is to improve upon traditional threshold-based monitoring, which often misses subtle or gradually developing issues.
The proposed method learns normal latency patterns and identifies deviations by analysing the full distribution of measurements rather than relying on simple metrics. Two data representations were evaluated, with histogram-based input proving more effective, as it preserves the structure of latency distribution. Results on real and artificially generated synthetic data show that this model can detect both sudden spikes and longer-term changes in network behaviour, including cases where traditional metrics remain stable.
This method achieves high precision with few false positives, although smaller anomalies may sometimes be missed.
Overall, the results demonstrate that the proposed approach is effective in detecting complex changes in network performance and provides a solid foundation for enhancing anomaly detection in monitoring systems.
Read more by following this link: https://zenodo.org/records/21496465
