Dispersed Federated Learning: Vision, Taxonomy, and Future Directions

TL;DR

This paper introduces dispersed federated learning (DFL), a decentralized approach to address privacy, robustness, and communication challenges in IoT-based machine learning, with a comprehensive taxonomy and future research outlook.

Contribution

The paper proposes a novel decentralized DFL framework, providing a taxonomy, theoretical analysis, and a matching theory-based solution for IoT applications.

Findings

DFL enhances privacy by decentralizing data processing.

The taxonomy categorizes various DFL schemes effectively.

A matching theory-based framework optimizes IoT DFL deployment.

Abstract

The ongoing deployment of the Internet of Things (IoT)-based smart applications is spurring the adoption of machine learning as a key technology enabler. To overcome the privacy and overhead challenges of centralized machine learning, there has been a significant recent interest in the concept of federated learning. Federated learning offers on-device, privacy-preserving machine learning without the need to transfer end-devices data to a third party location. However, federated learning still has privacy concerns due to sensitive information inferring capability of the aggregation server using end-devices local learning models. Furthermore, the federated learning process might fail due to a failure in the aggregation server (e.g., due to a malicious attack or physical defect). Other than privacy and robustness issues, federated learning over IoT networks requires a significant amount of communication resources for training. To cope with these issues, we propose a novel concept of dispersed federated learning (DFL) that is based on the true decentralization. We opine that DFL will serve as a practical implementation of federated learning for various IoT-based smart applications such as smart industries and intelligent transportation systems. First, the fundamentals of the DFL are presented. Second, a taxonomy is devised with a qualitative analysis of various DFL schemes. Third, a DFL framework for IoT networks is proposed with a matching theory-based solution. Finally, an outlook on future research directions is presented.