International Journal of Computer Science and Technology

International Journal of Computer Science and Technology | Vol. 3, No. 7, July 2022 | pp. 49–56

Research Article

Title: A Decentralized Federated Learning Architecture with Non-IID Data Mitigation for Distributed Edge Computing

Names of Authors: Chao Zhang¹, Xianfeng Wang², and Jun Lin³

Authors’ Affiliations:
¹School of Computer Science, Wuhan University, Wuhan, China
²Department of Electronic Information, Huazhong University of Science and Technology, Wuhan, China
³State Key Laboratory of Software Development Environment, Beihang University, Beijing, China

Abstract: Abstract: Centralized federated learning architectures rely heavily on a single orchestrating server to aggregate local gradients, introducing structural single points of failure and significant wide-area network bottlenecks. Furthermore, data collected across heterogeneous edge devices is often non-independently and identically distributed (non-IID), which compromises global model accuracy and decelerates algorithm convergence. This paper introduces an optimized, completely decentralized peer-to-peer federated learning framework designed for edge computing environments without central server dependencies. Individual edge nodes train deep neural network parameters locally on native datasets and exchange model weights exclusively with single-hop topological neighbors using an asynchronous gossip consensus protocol. To counter client-side data divergence, a localized dynamic regularization factor is integrated into the native objective function, punishing severe weight drift away from neighboring consensus baselines. Communication overhead is minimized through an adaptive gradient quantization routine that compresses transmission payload sizes dynamically based on instantaneous wireless link states. Matrix simulations using skewed MNIST and CIFAR-10 data distributions demonstrate that the proposed decentralized protocol achieves a high classification accuracy of 92.4%, matching centralized baseline performance while lowering total wide-area network data transmission volume by 46.2%. The framework provides stable, private machine learning capabilities for high-mobility mobile ad-hoc environments.

Keywords: Federated learning, Edge computing, Gossip protocol, Non-IID data, Gradient compression, Decentralized optimization

Manuscript Timeline: Received: April 14, 2022; Revised: May 20, 2022; Accepted: June 15, 2022; Published: July 1, 2022

Citation: Zhang, C., Wang, X., & Lin, J. (2022). A decentralized federated learning architecture with non-IID data mitigation for distributed edge computing. International Journal of Computer Science and Technology, 3(7), 49–56. DOI: 10.46882/2022/IJCST/000031

Subscribe to read the full article: https://internationalscholarsjournals.org/subscribe-to-read