An Unsupervised Based Approach to Detecting Anomalies in Hazard Monitoring Networks.
John Hooft Toomey
Committee: Ram Durairajan
Honors Bachelors Thesis(Jun 2024)
Keywords: Anomaly Detection, Machine Learning, Network Measurements

Our society relies heavily on various critical infrastructures (e.g., hazard monitoring networks in the state of Oregon) dedicated to monitoring natural hazards (e.g., wildfires). The hazard monitoring networks comprise sensors and cameras interconnected by wide-area backbones, where failures can result in significant societal and economic loss. Consequently, monitoring of state of health of hazard monitoring networks is paramount.

While Machine Learning (ML) stands as one of the most groundbreaking advancements in Computer Science, its application to detecting and predicting anomalies in hazard monitoring networks is fraught with challenges. For one, each data transfer within hazard monitoring networks represents a distinct relationship, the creation of labeled training datasets for each connection is not feasible. Second, in the absence of labeled data, developing ML models to detect anomalies is impractical.

The main objective of this work is to enhance the robustness of hazard monitoring networks by addressing existing limitations in ML-driven anomaly detection. To this end, we propose the Anomaly Detection Tool (ADT) framework which uses weakly supervised learning techniques (such as heuristics and lightweight ML algorithms) to capture distinct relationships, culminating in weak labels. These weak labels can then be mapped to one stronger label, thus creating a dataset containing labeled anomalies which can be used to notify network operators of potential network failures. Our research stands to benefit the reliability and security for Critical Infrastructures.

Note: The source code of the ADT framework is openly available to the community at: https://github.com/johnhooft/Anomaly-Detection-Tool