SeAFed: Severity-Aware Client Ranking for Federated Learning in Healthcare

Abstract

Federated Learning (FL) is increasingly important in Internet of Things (IoT) for healthcare, as it facilitates decentralized training of machine learning models while ensuring privacy of sensitive data. However, FL faces difficulties with pathological non-Independent and Identically Distributed (non-IID) setting, where differences in data from various sources can reduce the effectiveness of the models. This work proposes SeAFed, a client ranking-based Federated Learning (FL) system designed for pathological non-IID environment. Current FL systems are often challenged with non-IID data, leading to inefficient and biased model training. SeAFed addresses these limitations by employing a server that coordinates with the clients using a ranking algorithm that prioritizes clients based on disease severity. SeAFed deploys client ranking using an Analytic Hierarchy Process (AHP) on the server by utilizing several criteria such as disease severity level, local accuracy, training time, and data size. This ranking demonstrates the contribution that each client provides to global accuracy. Using this approach, any number of clients can be selected after the ranking for training. We evaluate the system's performance in terms of accuracy, process time, CPU usage, and memory consumption. Our approach achieves 22.5% less process time, 29.8% less CPU usage and 19.5% less memory consumption. We observe 70% test accuracy on the non-IID data for all the clients without applying SeAFed as well as for 3 clients after applying SeAFed. We find that SeAFed achieves comparable accuracy while blackucing the resource consumption. Therefore, our approach improves the efficiency of FL in healthcare by effectively utilizing critical data to improve model performance.