Anomaly Detection in Cyber-Physical Water Systems

Problem

Water treatment plants are cyber-physical systems. Dozens of sensors and actuators are driven by industrial controllers, and all of them can be attacked. A compromised pump or a spoofed sensor reading can quietly push a plant into an unsafe state. My master's thesis asked whether we can detect those anomalies in real time from the multivariate sensor streams alone, and do it reliably enough that an operator would trust the alerts.

Approach

I built on Graph Deviation Networks (GDN, AAAI 2021), which learn a graph of sensor relationships and flag deviations from expected behavior. Reproducing the paper surfaced the real engineering problems.

• Training instability. Stock GDN failed on about 40 percent of random seeds. I traced the problem to embedding initialization and fixed it with orthogonal initialization, which eliminated the failures entirely.
• Single model brittleness. Individual models disagree about which sensors matter, so I built a diverse ensemble that varies seeds, top-k values, and learning rates, and averages the per-sensor anomaly scores.
• A single signal is not enough. I added a frozen reconstruction decoder (GDN-RD) so that prediction error and reconstruction error vote together.

Architecture & Tech

The PyTorch training pipeline covers three testbeds: SWaT with 51 sensors, WADI with 127, and ACWA with 7, which serves as an extreme low data benchmark. It includes a config driven experiment system, five seed evaluation across Oracle F1, PA-F1, LDH F1, and AUC-ROC, graph based stage localization for root cause analysis, an LLM pipeline that turns raw detections into explanations an operator can read, and an interactive dashboard for exploring alerts.

Results & Impact

• The ensemble reached an Oracle F1 of 0.854 on SWaT 2015, compared with 0.810 for published GDN and 0.767 for our GDN reproduction, with a PA-F1 of 0.905 and an AUC-ROC of 0.892.
• It detected 18 of 35 attack segments on SWaT, and the evaluation on WADI and ACWA is reported honestly, including where limited data hurts.
• The stability fix turned a research prototype into a system that trains reliably, which matters for anything meant to be deployed.

What I'd do next

I would stream the pipeline against live testbed data, quantify detection latency, and harden the dashboard into an operational tool with alert triage and feedback capture for continual evaluation.