Robust Federated Learning With Noisy Labeled Data Through Loss Function Correction

Federated learning (FL) is a communication-efficient machine learning paradigm to leverage distributed data at the network edge. Nevertheless, FL usually fails to train a high-quality model from the networks, where the edge nodes collect noisy labeled data. To tackle this challenge, this paper focus...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on network science and engineering Vol. 10; no. 3; pp. 1 - 11
Main Authors: Chen, Li, Ang, Fan, Chen, Yunfei
Format: Journal Article
Language:English
Published: Piscataway IEEE 01.05.2023
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Subjects:
Convergence
Data models
Datasets
Distributed networks
Federated learning
label noise
Loss measurement
Machine learning
Nodes
Noise measurement
non-convex optimization
parallel and distributed algorithms
robust design
Robustness
Servers
Training
ISSN:2327-4697, 2334-329X
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Federated learning (FL) is a communication-efficient machine learning paradigm to leverage distributed data at the network edge. Nevertheless, FL usually fails to train a high-quality model from the networks, where the edge nodes collect noisy labeled data. To tackle this challenge, this paper focuses on developing an innovative robust FL. We consider two kinds of networks with different data distribution. Firstly, we design a reweighted FL under a full-data network, where all edge nodes are equipped with both numerous noisy labeled dataset and small clean dataset. The key idea is that edge devices learn to assign the local weights of loss functions in noisy labeled dataset, and cooperate with central server to update global weights. Secondly, we consider a part-data network where some edge nodes exclude clean dataset, and can not compute the weights locally. The broadcasting of the global weights is added to help those edge nodes without clean dataset to reweight their noisy loss functions. Both designs have a convergence rate of <inline-formula><tex-math notation="LaTeX">\mathcal {O}(1/T^{2})</tex-math></inline-formula>. Simulation results illustrate that the both proposed training processes improve the prediction accuracy due to the proper weights assignments of noisy loss function.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:2327-4697
2334-329X
DOI:10.1109/TNSE.2022.3227287