OptAML: Optimized adversarial machine learning on water treatment and distribution systems

Abstract

This research presents the optimized adversarial machine learning framework, OptAML, which is developed for use in water distribution and treatment systems. In consideration of the physical invariants of these systems, the OptAML generates adversarial samples capable of deceiving a hybrid convolutional neural network-long short-term memory network model. The efficacy of the framework is assessed using the Secure Water Treatment (SWaT) and Water Distribution (WADI) datasets. The findings demonstrate that OptAML is capable of effectively evading rule checkers and significantly reducing the accuracy of anomaly detection frameworks in both systems. Additionally, the study investigates a defense mechanism that demonstrates enhanced robustness against these adversarial attacks and is based on adversarial training. Our results underscore the necessity for robust and flexible protection tactics and highlight the shortcomings of the machine learning-based anomaly detection systems for critical infrastructure that are currently in place.