Performance Analysis of Aggregation Algorithms in Cross-Silo Federated Learning for Non-IID Data

Federated learning is a powerful machine learning paradigm that enables a large number of machine learning applications that must comply with strict and complex data privacy regulations. In this paper, we focus on one of the fundamental components of Federated Learning: federated aggregation algorithms. These algorithms play a pivotal role in consolidating insights and the model updates from various clients while preserving data privacy and security. We compare their efficacy across different types of dataset configurations, including balanced IID (Independent and identically distributed) data, unbalanced IID data, and non-IID data characterized by label distribution skew and feature distribution skew. For our specific use case, the experiments presented in this work show that FedProx is the overall best performing state-of-the-art algorithm for binary classification tasks of medical X-rays distributed in datasets of a small number of hospitals.