Evaluation Framework for Centralized and Decentralized Aggregation Algorithm in Federated Systems

TL;DR

This paper compares centralized and decentralized federated learning architectures, demonstrating that decentralized methods like AFL and CFL outperform centralized HFL in accuracy and efficiency on benchmark datasets.

Contribution

It provides a comprehensive evaluation of decentralized federated learning architectures, highlighting their advantages over centralized approaches in performance and privacy.

Findings

Decentralized methods outperform centralized HFL in accuracy metrics.

AFL and CFL show better privacy preservation and scalability.

Decentralized aggregation improves collaborative model training.

Abstract

In recent years, the landscape of federated learning has witnessed significant advancements, particularly in decentralized methodologies. This research paper presents a comprehensive comparison of Centralized Hierarchical Federated Learning (HFL) with Decentralized Aggregated Federated Learning (AFL) and Decentralized Continual Federated Learning (CFL) architectures. While HFL, in its centralized approach, faces challenges such as communication bottlenecks and privacy concerns due to centralized data aggregation, AFL and CFL provide promising alternatives by distributing computation and aggregation processes across devices. Through evaluation of Fashion MNIST and MNIST datasets, this study demonstrates the advantages of decentralized methodologies, showcasing how AFL and CFL outperform HFL in precision, recall, F1 score, and balanced accuracy. The analysis highlights the importance of decentralized aggregation mechanisms in AFL and CFL, which effectively enables collaborative model training across distributed devices. This comparative study contributes valuable insights into the evolving landscape of federated learning, guiding researchers and practitioners towards decentralized methodologies for enhanced performance in collaborative model training scenarios.