Noise-Perturbed Federated Learning for Synthetic Wearable Health Monitoring: A Reproducible Multi-Seed Evaluation

Authors

• Billel Essaid, LSEA Laboratory, Electrical Engineering Department, University Dr. Yahia Fares of Medea, Medea 26000, AlgeriaFollow

Abstract

Federated learning (FL) keeps wearable records local, but shared updates can still disclose participant information. This study evaluates an auditable noise-perturbed FL pipeline for synthetic wearable risk monitoring and examines how client-side clipping, central Gaussian perturbation, participant-held-out validation, and two empirical privacy attacks behave across repeated cohorts. Twelve independently generated cohorts (24 participants; 5,760 half-hour windows each) were analyzed with matched-exposure centralized logistic regression, FedAvg, clipped FedAvg, and clipped-plus-noise FedAvg using 16/4/4 participant train/validation/test splits. Uncertainty was estimated with a two-stage hierarchical bootstrap that preserves each seed's validation-selected threshold. The clipped-plus-noise model achieved mean AUROC 0.903 (95% CI 0.867–0.913), AUPRC 0.535 (0.406–0.567), F1 0.484 (0.351–0.506), and recall 0.524 (0.349–0.574). Validation-fitted Platt scaling reduced mean Brier score from 0.0488 to 0.0445. A released-model membership attack was near chance (AUC 0.480, 0.392-0.563), whereas coordinator-visible clipped updates revealed chronic-condition status (AUC 0.999, 0.997–1.000); removing the explicit chronic feature reduced but did not eliminate leakage (AUC 0.974). The contribution is a reproducible evaluation framework that separates data locality, coordinator visibility, empirical attack evidence, and formal privacy guarantees. No differential-privacy guarantee or clinical effectiveness is claimed.

Recommended Citation

Essaid, Billel (2026) "Noise-Perturbed Federated Learning for Synthetic Wearable Health Monitoring: A Reproducible Multi-Seed Evaluation," Mesopotamian Journal of Artificial Intelligence in Healthcare: Vol. 4: Iss. 1, Article 5.