Optimization of Moon Model-Contrastive Federated Learning

Changyu Chen, Weiheng Rao

2025

Abstract

Federated contrast learning has shown great potential in privacy-sensitive scenarios, enabling multiple parties to train models using their local data, rather than sharing privacy data. Model-contrastive Federated learning (MOON) effectively improves the accuracy of graphical contrast learning. However, after researching SimCLR which is a part of the origin of the MOON algorithm, it is found that the MOON federal learning algorithm does not consider the influence of the change of hyperparameter (temperature) on the accuracy of its model. This article will focus on the model comparison federated learning framework MOON, and propose an adaptive temperature control mechanism based on simulated annealing, aiming at the static set limit of its key hyperparameter, contrast loss temperature (τ). The temperature attenuation function is designed to achieve global-local optimization of dynamic balance - the initial high-temperature promotion model explores the global feature space and later low-temperature enhanced local fine-grained optimization. The experiments in this paper show that the dynamic temperature can slightly improve the accuracy of the MOON model. This work systematically quantifies the influence of temperature parameters on model contrast federation learning.

Paper Citation

in Harvard Style

Chen C. and Rao W. (2025). Optimization of Moon Model-Contrastive Federated Learning. In Proceedings of the 2nd International Conference on Data Science and Engineering - Volume 1: ICDSE; ISBN 978-989-758-765-8, SciTePress, pages 128-132. DOI: 10.5220/0013679800004670

in Bibtex Style

@conference{icdse25,
author={Changyu Chen and Weiheng Rao},
title={Optimization of Moon Model-Contrastive Federated Learning},
booktitle={Proceedings of the 2nd International Conference on Data Science and Engineering - Volume 1: ICDSE},
year={2025},
pages={128-132},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013679800004670},
isbn={978-989-758-765-8},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 2nd International Conference on Data Science and Engineering - Volume 1: ICDSE
TI - Optimization of Moon Model-Contrastive Federated Learning
SN - 978-989-758-765-8
AU - Chen C.
AU - Rao W.
PY - 2025
SP - 128
EP - 132
DO - 10.5220/0013679800004670
PB - SciTePress