An explainable data-local federated learning framework for intrusion detection in connected healthcare systems

Introduction Connected healthcare networks link medical sensors, gateways, electronic health record systems and cloud services, which enlarges their exposure to cyberattacks. Centralized intrusion detection systems (IDSs) require traffic from all sites to be pooled in one repository, which conflicts with data-residency constraints and creates a single high-value target, and their alerts are often not interpretable. Methods This paper proposes an explainable, privacy-aware federated learning (FL) framework for intrusion detection in which privacy is limited to data locality: clients exchange trained models and summary statistics instead of raw records, but model updates are not encrypted, securely aggregated or differentially private. The framework combines (i) federated training, (ii) client-side SHAP-guided feature ranking reconciled at the server into a common global feature set, and (iii) SHAP and LIME explanations of the aggregated model that are evaluated quantitatively. The framework was evaluated on the WUSTL-EHMS-2020 healthcare dataset, which combines network-flow and patient biometric features, partitioned into five simulated clients under IID and non-IID settings, using a leakage-controlled protocol in which host identifiers and record indices were excluded and all preprocessing was fitted on training data only. Results Under non-IID partitioning, the federated gradient-boosting model reached an accuracy of 0.935 ± 0.009, a Matthews correlation coefficient (MCC) of 0.673 ± 0.045 and a ROC-AUC of 0.895 ± 0.020, compared with an MCC of 0.700 ± 0.013 for centralized training and 0.540 ± 0.036 for local-only training. All data-alteration attacks were detected, whereas most spoofing attacks were missed. SHAP-guided feature selection and reliability-weighted aggregation did not improve detection and are therefore presented as design options; LIME explanations had low local fidelity ( R² = 0.332), and the usefulness of the explanations for security analysts was not evaluated. Discussion All results come from a single laboratory dataset partitioned into simulated clients; they demonstrate experimental feasibility only and do not establish generalizability to other healthcare networks or readiness for clinical deployment.