End-to-End Privacy-Aware Federated Learning for Wearable Health Devices via Encrypted Aggregation in Programmable Networks

The widespread use of wearable Internet of Things (IoT) devices has transformed modern healthcare through the real-time monitoring of physiological signals. However, real- time responsiveness and data privacy are big challenges. Federated Learning (FL) keeps direct data exposure to a minimum but is susceptible to inference attacks on model updates and heavy communication overhead. In-network computing (INC) solutions currently offer greater efficiency but without cryptographic security, whereas homomorphic encryption (HE) offers high privacy but at the expense of latency and scalability. To bridge this gap, we present Edge-Assisted Homomorphic Federated Learning (EAH-FL), a framework that unifies Cheon–Kim–Kim–Song (CKKS) HE with in-network aggregation. Lightweight clients outsource encryption and decryption to trusted edge devices, whereas programmable switches carry out aggregation in the encrypted domain. Massive-scale simulations over realistic healthcare data sets demonstrate that EAH-FL preserves near-plaintext model accuracy (F1-score > 0.93), delivers packet delivery ratios > 0.95, and converges well for various client scales. The encryption expense is mostly incurred by the edge layer rather than resource-constrained wearables. Through the use of encryption, in- network acceleration, and smart routing, EAH-FL provides the first practical solution that achieves strong privacy, low latency, and scalability for real-time federated learning in healthcare in a single solution. These results validate its viability as a deployable and secure building block for next-generation digital health monitoring.