Securing Federated Learning in Health IoT with Edge-Assisted Homomorphic Encryption

The rapid expansion of wearable Internet of Things (IoT) devices in healthcare enables continuous monitoring of physiological signals but introduces critical privacy challenges, especially when training machine learning (ML) models on sensitive, distributed data. Federated learning (FL) addresses this by allowing collaborative model training without sharing raw patient data. However, most existing FL systems either transmit model updates in plaintext exposing them to inference attacks or adopt encryption schemes that are computationally prohibitive for resource-constrained devices. In this paper, we present a modular, privacy-preserving FL framework for health IoT that integrates Cheon-Kim-Kim-Song (CKKS) based Homomorphic Encryption (HE) at trusted edge nodes (e.g., home routers). This design offloads encryption from wearable devices, enabling secure model aggregation and update distribution while maintaining low energy and latency overheads. We evaluate our system on real-world datasets, measuring model utility (accuracy, F1), encryption latency, and end-to-end delay. Results demonstrate that our scheme preserves privacy with minimal performance degradation, supporting practical and secure deployment in realtime health monitoring scenarios.