IDEAS SoCal AI & Data Science Conference 2018

Modern industrial control systems are highly automated and often controlled by computers and connected devices. In addition to physical failures, cyber attacks present an increasing threat to normal system operations. We propose an unsupervised learning approach for automated detection of adverse cyber attacks. In particular, we will devise the unsupervised learning based approaches for anomaly detection using only datasets collected under normal system conditions. Anomaly detection techniques are extremely useful for real-life system security management because cyber attacks rarely happen in practice even though their impact could be disastrous. We conduct detailed studies on a water treatment plant in which water processes are all connected by computer control systems. We demonstrate our analytics tools can effectively detect new types of cyber attacks and, moreover, can locate potential attacking points and compromised control units or devises