AI-Driven Anomaly Detection for Securing U.S. Critical Infrastructure Networks

The growing frequency and sophistication of cyberattacks targeting critical infrastructure in the United States necessitate advanced, real-time detection mechanisms beyond traditional security approaches. This study investigates the application of artificial intelligence (AI) and machine learning (ML) techniques for anomaly detection in critical infrastructure networks, focusing on Industrial Control Systems (ICS), Supervisory Control and Data Acquisition (SCADA) systems, and smart city infrastructures. The research evaluates supervised, unsupervised, and deep learning models including Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, autoencoders, and ensemble methods using a mixed-methods approach combining systematic literature review, comparative performance analysis, and case studies across water treatment facilities, power grids, and IoT-enabled municipal systems. Results indicate that hybrid CNN–LSTM models achieve detection accuracy exceeding 98% for known attacks, while autoencoder-based approaches effectively identify zero-day attacks with 94.7% accuracy. Ensemble methods reduce false positive rates by 43% compared to single-model implementations, addressing a critical operational challenge in safety-critical environments. Despite these advances, significant limitations persist, including high false positive rates, concept drift, adversarial vulnerabilities, computational overhead, and explainability constraints. This study contributes a critical-infrastructure-specific taxonomy of AI-driven anomaly detection techniques, empirically validated performance benchmarks, and a practical implementation framework that supports secure, explainable, and operationally feasible deployment of AI-based detection systems.