Challenges in Machine Learning based approaches for Real-Time Anomaly Detection in Industrial Control Systems

Abstract

Data-centric approaches are becoming increasingly common in the creation of defense mechanisms for critical infrastructure such as the electric power grid and water treatment plants. Such approaches often use well-known methods from machine learning and system identification, i.e., the Multi-Layer Perceptron, Convolutional Neural Network, and Deep Auto Encoders to create process anomaly detectors. Such detectors are then evaluated using data generated from an operational plant or a simulator; rarely is the assessment conducted in real time on a live plant. Regardless of the method to create an anomaly detector, and the data used for performance evaluation, there remain significant challenges that ought to be overcome before such detectors can be deployed with confidence in city-scale plants or large electric power grids. This position paper enumerates such challenges that the authors have faced when creating data-centric anomaly detectors and using them in a live plant.