Decentralized Federated Learning Over Noisy Labels: A Majority Voting Method

Guan Huang, Tao Shu

Keywords: decentralized federated learning, distributed learning, federated learning, majority voting, label-noise learning

TL;DR: We propose a majority voting method across different clients to deal with noisy labels in a decentralized federated learning.

Abstract: Contrary to centralized federated learning (CFL), decentralized federated learning (DFL) allows clients to cooperate in training their local models without relying on a central parameter server. As different clients have varying annotation skills and preferences, noisy labels are inevitable in decentralized data ownership. In centralized learning (CL) and CFL settings, learning from noisy labels has been extensively explored; however, such methods cannot be directly applied in DFL settings due to limited computational resources or privacy requirements. This paper introduces DFLMV \textit{(majority voting based decentralized federated learning)}, a general DFL framework for learning from noisy data without relying on any assumptions about local client noise models while maintaining data privacy for all clients. Specifically, (1) Clients first use traditional DFL to train their local models until they become stable. (2) Clients use each of their neighbors' models to make a prediction of every data point in their training datasets, then correct the labels based on majority voting. (3) Clients further fine-tune their models based on their updated training dataset. A theoretical analysis of DFLMV is also provided. Extensive experiments conducted on MNIST, Fashion-MNIST, CIFA-10, CIFAR-10N, CIFAR-100N, Clothing1M, and ANIMAL-10N validate the effectiveness of our proposed approach at various noise levels and different data settings in mitigating the adverse effects of noisy labels.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1177