Privacy-First Explainable Federated Learning with Zero-Trust AI Infrastructure for Proactive Healthcare Decision Support

Abstract

Healthcare AI is increasingly deployed in decision support and patient-facing workflows, but scale-up is limited by privacy constraints on protected health information (PHI), expanding attack surfaces in hybrid delivery models, and weak transparency of complex models. This paper articulates ZT-XFL, a reference architecture that integrates federated learning (FL) with differential privacy (DP) and explainable AI (XAI) under a zero trust architecture (ZTA). The control plane enforces explicit verification, least-privilege authorization, and comprehensive audit logging as defined in ZTA guidance [1], while the training plane coordinates institution-local optimization (FedAvg-style aggregation) [3] and optionally applies DP-SGD and privacy budgeting to bound leakage from updates [6]. Secure aggregation is incorporated to prevent the coordinator from learning individual client updates [7]. The governance plane binds post-hoc explanations (LIME/SHAP) to immutable model versions and inference events [12], [13], reflecting the view that clinical explainability requirements are context-dependent and must be assessed alongside validation and system role [16]. We formalize a threat model spanning endpoint compromise, insider misuse, gradient-based inference, supply-chain risks, and poisoning of federated updates; and we map each threat to enforceable controls across identity, device posture, workload attestation, update screening, and explanation-access policy [1], [2]. Rather than presenting institution-specific results, we provide a reproducible evaluation protocol that jointly measures utility and calibration, privacy loss (ε, δ), security control coverage, explanation stability, and robustness against adversarial updates and inference attacks [9], [10], [11], enabling benchmarking on multi-site healthcare tasks without centralizing PHI.