Federated learning with transient central initialization and decentralized edge aggregation

doi:10.1016/j.engappai.2025.113527

Federated learning with transient central initialization and decentralized edge aggregation

2025-12-01

2

PRE

AI

Abena Achiaa Atwereboannah

DOI：10.1016/j.engappai.2025.113527

Original

Original request for help

Save

Abstract

En 中文

Federated Learning (FL) enables privacy preservation and collaborative model training across several distributed datasets and clients while heavily relying on a centralized server to aggregate updates from the clients. However, the over-reliance on centralized aggregation paradigm faces challenges such as high communication overhead, single point of failure, and reduced model performance caused by the inclusion of low-quality clients. To address these issues, we propose a Decentralized Aggregation Federated Learning (DAFL) framework, which largely shifts aggregation tasks to multiple edge servers, dynamically organizing clients into hierarchical clusters during training. Representative clients, selected based on the alignment of their model updates with the global updates, optimize communication efficiency and model accuracy. Empirical evaluations across computer vision tasks with convolutional neural networks, and natural language processing applications (AG News text classification with long short-term memory neural network) demonstrate that DAFL saves over 60% of communication rounds on average while reaching high target accuracies on the datasets used.

Keywords:

Federated learning

Convolutional neural network

Long short-term memory neural network

Computer vision

Natural language processing

Federated learning with transient central initialization and decentralized edge aggregation

Abstract

Journal

Organization