Robust Machine Learning Models for Cybersecurity in Critical Infrastructures A Systematic Review, Empirical Synthesis, and Framework Proposal

Abstract

ML-based intrusion detection systems for critical infrastructure work well in controlled research settings.The gap between that and reliable operation against adversaries who probe, adapt, and know the detection layer is where this paper focuses.We reviewed 44 primary sources via a PRISMA-adapted protocol covering ensemble learning, GNNs, transformer architectures, deep reinforcement learning, adversarial defences, concept drift adaptation, federated learning, and XAI across three CI sectors: energy and water, healthcare IoMT, and intelligent transportation.The main findings: ensemble hybrids are the most deployable near-term architecture.Certified robustness methods are theoretically principled but fall short of operational security guarantees under realistic CI threat models, as Cullen et al. (2025) demonstrated at ICML.SHAP and LIME measurably improve analyst trust but enable adversaries to reconstruct model decision boundaries with over 90% success; this paper calls that tension the Adversarial XAI Paradox, and no reviewed study resolves it.Transportation sector evaluation is the weakest of the three, with no public V2X-specific benchmark comparable to SWaT or CICIoMT2024.We identify five testable research gaps, construct a five-layer framework aligned to NIST CSF 2.0, and close with specific recommendations.