RAID: Resilient AI Defense Model for Water Utility Cyber–Physical Attacks

Abstract

The increasing digitalization of water utility infrastructures has made them highly susceptible to sophisticated adversarial attacks that inject misleading features into operational data and to sensor anomalies that introduce redundant or noisy information. At the same time, cyber threats continuously evolve, introducing previously unseen attack behaviors that cannot be reliably recognized by conventional models trained on fixed attack distributions. Such disruptions can falsify readings, trigger incorrect valve or pump operations, and lead to water contamination or service interruptions, posing serious risks to public safety and infrastructure reliability. To address these challenges, this article presents the resilient AI defense (RAID) system, a novel integrated framework for adversarially robust cyberattack detection in industrial water infrastructures. RAID combines two complementary components: 1) a fuzzy–Mahalanobis ant colony optimization (FM-ACO) module that constructs compact and adversarially robust feature representations; and 2) an adaptive few-shot Learning module that rapidly generalizes to previously unseen cyberattack scenarios using only a small number of labeled samples. At the analytical core of RAID, FM-ACO employs fuzzy reasoning to evaluate feature importance under uncertainty while Mahalanobis-distance (MD) consistency suppresses adversarially distorted feature representations during optimization. This enables the proposed optimization to simultaneously improve feature compactness, stability, and robustness, thereby providing reliable feature representations for downstream learning. The adaptive few-shot learning module subsequently exploits these robust feature representations to maintain reliable detection performance under both adversarial perturbations and unseen attack scenarios without requiring extensive retraining. The robustness of RAID is validated under multiple adversarial perturbation strategies, including fast gradient sign method (FGSM), projected gradient descent (PGD), and momentum iterative method (MIM). Experimental results demonstrate that RAID significantly reduces attack success rates (ASRs) by 47.5%, improves feature stability by 30%, and lowers computational overhead by 18%, while maintaining 85.4% classification accuracy under adversarial conditions. Comprehensive experiments further demonstrate that FM-ACO consistently outperforms conventional optimization algorithms, while the complete RAID framework surpasses both its ablated variant and state-of-the-art adversarial defense methods, including adversarial training with latent Gaussian perturbation (ATLGP), federated gradient-guided learning (FedG2L), and SmsNet.