A Systematic Framework to Generate Invariants for Anomaly Detection in Industrial Control Systems

Industrial Control Systems (ICS) consisting of integrated hardware and software components designed to monitor and control a variety of industrial processes, are typically deployed in critical infrastructures such as water treatment plants, power grids and gas pipelines. Unlike conventional IT systems, the consequences of deviations from normal operation in ICS have the potential to cause significant physical damage to equipment, the environment and even human life. The active monitoring of invariant rules that define the physical conditions that must be maintained for the normal operation of ICS provides a means to improve the security and dependability of such systems by which early detection of anomalous system states may be achieved, allowing for timely mitigating actions -- such as fault checking, system shutdown -- to be taken. Generally, invariant rules are pre-defined by system engineers during the design phase of a given ICS build. However, this manually intensive process is costly, error-prone and, in typically complex systems, sub-optimal. In this paper we propose a novel framework that is designed to systematically generate invariant rules from information contained within ICS operational data logs, using a combination of several machine learning and data mining techniques. The effectiveness of our approach is demonstrated by experiments on two real world ICS testbeds: a water distribution system and a water treatment plant. We show that sets of invariant rules, far larger than those defined manually, can be successfully derived by our framework and that they may be used to deliver significant improvements in anomaly detection compared with the invariant rules defined by system engineers as well as the commonly used residual error-based anomaly detection model for ICS.