Multi-Global Models for Edge Computing Environment

This study introduces a novel three-tier Federated Learning (FL) architecture to
tackle the challenge of non-IID data in Edge Computing. Traditional FL models,
typically limited by a singular global model, face difficulties with data diversity,
impacting learning efficacy. This research proposes an advanced framework that
utilizes multiple global models within a clustered setup, significantly improving
learning outcomes by aligning with the varied data characteristics found across
edge devices.
Employing the MNIST dataset, known for benchmarking FL algorithms,
this approach was tested to address label skewness—a prevalent form of non-
IID data. The findings highlight a substantial enhancement in learning accu-
racy, showcasing the framework’s ability to efficiently navigate distributed data
environments, a notable advancement over conventional FL methodologies.
Transitioning to a multi-global model architecture, the research offers a fit-
ting solution to the complex data scenarios characteristic of Edge Computing.
This strategic shift not only addresses the limitations imposed by non-IID data
but also bolsters the scalability and resilience of FL systems. It marks a sig-
nificant advancement in deploying decentralized machine learning technologies,
facilitating broader FL application across varied sectors. This framework’s suc-
cess in managing data heterogeneity effectively paves the way for enhanced FL
deployments, demonstrating its potential as a robust solution for real-world
challenges in decentralized learning environments.