Federated learning in edge computing settings in the medical domain

Abstract

In the current rapid pace of digitalization, the volume of personal medical data also increases, raising additional privacy and security concerns. A large amount of this sensitive data is stored on mobile devices and often unavailable for collection due to different regulations, such as the GDPR, as well as users’ reluctance to share. Federated Learning (FL) has emerged as a way to train a model in a decentralized manner without sharing raw data. However, in edge computing settings, FL faces major challenges of data heterogeneity and scarcity.This work explores the ability to train a model in an extreme FL setting, namely Single- Record Per-Client FL (SRPC FL), where each participant’s local dataset comprises one record. Such a scenario reflects eHealth applications like remote patient monitoring, fitness tracking, and personalized treatment support, where local data is highly individual and not representative of the global distribution. In our work, we empirically evaluate standard FL strategies for obtaining a global model, such as (FedAvg, FedProx), adaptive optimization methods (FedAdam, FedAdagrad), and clustering-based approaches (CFL, IFCA) in SRPC FL and compare them with models trained in centralized settings. Additionally, we propose clustering and weighting methods that tackle the challenges of this setup. Finally, we provide an extensive analysis of the robustness of the baseline and the proposed methods in SRPC FL conditions under different label skews in the global dataset.