A Reinforcement Adversarial Framework Targeting Endogenous Functional Safety in ICS: Applied to Tennessee Eastman Process
Xinyi Wu, Yulong Ding, Shuang‐Hua Yang · 2024
Endogenous Safety and Security (ESS) of Industrial Control Systems (ICS) has gained great attention with the advent of Industry 4.0. However, with rising cyber threats, most current research has focused mainly on cybersecurity aspects. Our study aims to fill this research gap by focusing on the endogenous functional safety of ICS, with a particular emphasis on key control parameters “setpoints”. We propose a reinforcement adversarial framework to investigate the functional security issues arising from unexpected operations and malicious tampering against setpoints. In this framework, a deep reinforcement learning(DRL) agent interacts with a custom input rule model, which serves as both a dynamic validator and an adversary, aiming to explore previously unforeseen behaviors. Explored unexpected behaviors are continuously updated to the input rule model, enhancing system adaptability and robustness. Our study employed the Tennessee Eastman Process as a case study, using the proximal policy optimization(PPO) algorithm with Beta and Gaussian distributions. Our approach exhibited significant advantages in exploration efficiency over baseline methods such as random agents and simulated annealing. These findings underscore DRL's important role in augmenting ICS functional safety, thereby enhancing system resilience and security in Industry 4.0.