ADVERSARIAL DEFENCE MECHANISMS IN NEURAL NETWORKS FOR ICS FAULT TOLERANCE: A COMPARATIVE ANALYSIS

Abstract

This systematic review explores the landscape of adversarial defense mechanisms in neural networks designed for fault-tolerant operations within Industrial Control Systems (ICS). With the increasing integration of artificial intelligence into critical infrastructure, ICS are now vulnerable to adversarial attacks that exploit the fragility of deep learning models, potentially leading to unsafe control actions and operational disruptions. The study systematically reviewed and synthesized findings from 126 peer-reviewed articles published between 2013 and 2024, covering model-centric, preprocessing, and postprocessing defense strategies, as well as their sector-specific implementations in smart grids, chemical processing plants, water treatment systems, and robotic manufacturing. The review was conducted following the PRISMA 2020 guidelines, ensuring a rigorous and transparent methodology. Key findings reveal that adversarial training remains the most widely used and effective model-centric approach, while signal filtering and demising methods serve as efficient preprocessing techniques for input sanitization. Post processing strategies, such as uncertainty estimation and confidence-based detection, are shown to enhance robustness when used in layered defense architectures. Furthermore, it highlights the emerging importance of integrating adversarial robustness into cyber security standards like NIST, ENISA, and IEC 62443. By analyzing both theoretical contributions and practical implementations, this study provides a comprehensive framework for understanding the current state of adversarial defenses in ICS environments. The findings emphasize the need for context-aware, scalable, and explainable defense mechanisms that can operate within the constraints of real-time control systems. This review contributes valuable insights for researchers, engineers, and policymakers working at the intersection of machine learning security and industrial automation.