SUMMARY

The prevalence of Machine Learning (ML) technology in industrial control systems (ICS) and operational technology (OT) environments is on the rise. Autonomous control systems (ACS) based on ML digital twins (DTs) are being proposed as a new way to control nuclear power plant (NPP) subsystems using data-driven analytics and stochastic modeling. However, the implementation of ACS in advanced reactor controls must be assessed for its resilience against adversarial action, given the increasing incidence of cyber-attacks against critical infrastructure and OT environments. This thesis delves into the implementation techniques and cyber-resilience of ML-based ACS designs for advanced reactors. Its recommendations provide security and safeguards for physical and digital systems to protect against cyber-attacks on ML algorithms. Two implementations of ACS were considered: one as a virtual testbed using the International Atomic Energy Agency (IAEA) Asherah pressurized water reactor (PWR) nuclear power plant simulator, and one as a cyber-physical testbed using the Western Services Corporation (WSC) Generic Pressurized Water Reactor (GPWR) and real-time programmable logic controller (PLC) data. The purpose was to determine the impact of cyber-attacks on ACS in both a medium and high-fidelity environment. Cyber-attacks targeting ACS’ ML models’ training data, real-time data, and architecture were conducted to encapsulate potential targets against ML DTs and determine the impact and associated cyber risk of using ML-based DTs in control environments.