Personalized, robust federated learning with fed+

Publication Type

Book Chapter

Publication Date

7-2022

Abstract

Fed+ is a unified family of methods designed to better accommodate the real-world characteristics found in federated learning training, such as the lack of IID data across parties and the need for robustness to outliers. Fed+ does not require all parties to reach a consensus, allowing each party to train local, personalized models through a form of regularization while benefiting from the federation to improve accuracy and performance. The methods included in the Fed+ family are shown to be provably convergent. Experiments indicate that Fed+ outperform other methods when data is not IID, and the robust versions of Fed+ outperform other methods in the presence of outliers.

Discipline

Numerical Analysis and Scientific Computing

Research Areas

Intelligent Systems and Optimization

Publication

Federated learning: A comprehensive overview of methods and applications

Editor

LUDWIG, Heiko; BARACALDO, Nathalie

First Page

99

Last Page

123

ISBN

9783030968960

Identifier

10.1007/978-3-030-96896-0_5

Publisher

Springer International Publishing

City or Country

Cham

Additional URL

https://doi.org/10.1007/978-3-030-96896-0_5

This document is currently not available here.

Share

COinS