Publicly Auditable Zero-Trust Federated Learning for Privacy-Preserving Intrusion Detection in Implantable Medical Device Ecosystems

Implantable medical device (IMD) and Internet of Medical Things (IoMT) environments need intrusion detection systems that learn across distributed hospitals without centralizing sensitive data, while controlling admission, protecting shared model artifacts, filtering unreliable contributors, and supporting post-run auditability. However, many secure federated learning designs treat identity, privacy, robustness, and evidence verification as separate layers, leaving a gap between privacy-preserving execution and public accountability. This paper presents an implemented zero-trust hierarchical federated learning-based intrusion detection system (FL-IDS) framework for IMD/IoMT security analytics. Hospital clients train eXtreme Gradient Boosting (XGBoost) detectors; self-sovereign identity gates participation; contribution-level differential privacy (DP) perturbs exported booster leaf weights; country aggregators apply adaptive Krum-inspired selection; and the global server performs trust-weighted prediction-level fusion. The evidence layer binds artifacts using Module-Lattice-Based Digital Signature Algorithm signatures, canonical hashes, Merkle roots, decentralized publication, Ethereum Sepolia anchoring, and standalone auditor verification. The framework is evaluated on WUSTL-EHMS-2020, ECU-IoHT, and CICIoMT2024 under paired DP-disabled and DP-enabled modes. Under DP-enabled execution, CICIoMT2024 achieved an F1-score of 0.998789 and area under the receiver operating characteristic curve (AUROC) of 0.999814, ECU-IoHT achieved an AUROC of 0.999337, and WUSTL-EHMS-2020 remained DP-sensitive with an F1-score of 0.422880 and AUROC of 0.776685. All paired evidence runs passed standalone auditor verification, demonstrating that privacy-preserving learning and public accountability can be integrated within a single experimental FL-IDS pipeline.