Empirical Analysis of FedAvg, FedProx and SCAFFOLD for Heterogeneous Data Distributions*

With the rise of big data and AI advancements, Federated Learning (FL) has gained significant research interest. FL operates on decentralized, often heterogeneous (non-IID) data, posing challenges in distributed databases. While various FL algorithms address non-IID issues, few comprehensively evaluate their impact across different non-IID distributions or propose effective improvements. This paper empirically analyzes FedAvg, FedProx, and SCAFFOLD using two datasets and two non-IID data partitioning schemes-Dirichlet-based and classbased. Our study highlights practical bottlenecks and theoretical insights, concluding with future research directions to enhance FL algorithm performance.