Domain-Based Fuzzing for Supervised Learning of Anomaly Detection in Cyber-Physical Systems

Abstract

A novel approach is proposed for constructing models of anomaly detectors using supervised learning from the traces of normal and abnormal operations of an Industrial Control System (ICS). Such detectors are of value in detecting process anomalies in complex critical infrastructure such as power generation and water treatment systems. The traces are obtained by systematically "fuzzing", i.e., manipulating the sensor readings and actuator actions in accordance with the boundaries/partitions that define the system's state. The proposed approach is tested in a Secure Water Treatment (SWaT) testbed -- a replica of a real-world water purification plant, located at the Singapore University of Technology and Design. Multiple supervised classifiers are trained using the traces obtained from SWaT. The efficacy of the proposed approach is demonstrated through empirical evaluation of the supervised classifiers under various performance metrics. Lastly, it is shown that the supervised approach results in significantly lower false positive rates as compared to the unsupervised ones.