Exploring the Impact of Optimization Algorithms in Federated Learning Under Non-IID Contexts

Abstract

As machine learning becomes increasingly widespread across different sectors, it encounters key challenges such as data privacy, uneven data distribution, and limited data access. Federated learning (FL) emerges as a decentralized and privacy-preserving solution that enables model training across distributed devices without sharing raw data. This study presents a comparative analysis of two optimization algorithms, FedAvg and FedProx, within the FL framework. The focus is placed on evaluating their behavior under non-IID settings, simulated through different usage contexts: IoT (1 class/client), mobile (2 classes/client), and desktop (5 classes/client). Using the MNIST dataset and the LeNet model, the evaluation investigates metrics such as accuracy and convergence over 200 communication rounds. The findings aim to highlight how each algorithm handles data heterogeneity and to guide future FL deployments in real-world scenarios.