A Generalized Look at Federated Learning: Survey and Perspectives

TL;DR

This paper provides a comprehensive survey of federated learning, discussing its principles, benefits, challenges, and future directions in distributed machine learning without sharing local data.

Contribution

It offers a systematic summary of existing research, challenges, and potential applications of federated learning, highlighting areas for future exploration.

Findings

Federated learning enhances data privacy and scalability.

Statistical and system heterogeneity pose significant challenges.

Communication bottlenecks and security issues remain open problems.

Abstract

Federated learning (FL) refers to a distributed machine learning framework involving learning from several decentralized edge clients without sharing local dataset. This distributed strategy prevents data leakage and enables on-device training as it updates the global model based on the local model updates. Despite offering several advantages, including data privacy and scalability, FL poses challenges such as statistical and system heterogeneity of data in federated networks, communication bottlenecks, privacy and security issues. This survey contains a systematic summarization of previous work, studies, and experiments on FL and presents a list of possibilities for FL across a range of applications and use cases. Other than that, various challenges of implementing FL and promising directions revolving around the corresponding challenges are provided.