Performance Analysis of a Hybrid Federated-Centralized Learning Framework

Centralized Learning (CL) in Machine Learning (ML) raises privacy concerns due to data aggregation on a central server. Federated Learning (FL) addresses this by allowing clients to share model updates instead of raw data. However, the varying capabilities of edge devices hinder FL's effectiveness in real-world scenarios. This work proposes a Hybrid Federated Centralized Learning (HFCL) model to overcome these limitations. HFCL optimizes task distribution based on individual client capabilities, enabling collaboration regardless of resource limitations. We implement CL, FL, and HFCL using MNIST and CIFAR-10 datasets with Convolutional Neural Networks (CNNs). We analyze the impact of varying active/passive client ratios within HFCL and different learning rates on accuracy. Results demonstrate that HFCL significantly improves learning performance compared to pure FL or CL. This highlights its potential to address limitations in edge computing environments with disparate computational resources. Future research directions include further exploration of HFCL within edge-based FL.