Augmenting the FedProx Algorithm by Minimizing Convergence

TL;DR

This paper introduces G Federated Proximity, an improved version of FedProx, which enhances convergence speed and efficiency for IoT applications by leveraging normalization and FTL techniques, achieving about 90% better throughput.

Contribution

The paper proposes G Federated Proximity, a novel modification to FedProx, to improve convergence and efficiency in federated learning for IoT environments.

Findings

Approximately 90% better convergence throughput.

Enhanced model accuracy on real-time devices.

Improved efficiency in heterogeneous networks.

Abstract

The Internet of Things has experienced significant growth and has become an integral part of various industries. This expansion has given rise to the Industrial IoT initiative where industries are utilizing IoT technology to enhance communication and connectivity through innovative solutions such as data analytics and cloud computing. However this widespread adoption of IoT is demanding of algorithms that provide better efficiency for the same training environment without speed being a factor. In this paper we present a novel approach called G Federated Proximity. Building upon the existing FedProx technique our implementation introduces slight modifications to enhance its efficiency and effectiveness. By leveraging FTL our proposed system aims to improve the accuracy of model obtained after the training dataset with the help of normalization techniques such that it performs better on real time devices and heterogeneous networks Our results indicate a significant increase in the throughput of approximately 90% better convergence compared to existing model performance.