Research Article
Improving Model-Contrastive Federated Learning under Non-IID Data Constraints
@INPROCEEDINGS{10.4108/eai.22-5-2026.2365159, author={Hao Pi}, title={Improving Model-Contrastive Federated Learning under Non-IID Data Constraints}, proceedings={Proceedings of the 4th International Conference on Image, Algorithms, and Artificial Intelligence, ICIAAI 2026, 22-24 May 2026, Singapore, Singapore}, publisher={EAI}, proceedings_a={ICIAAI}, year={2026}, month={8}, keywords={Federated learning; Model-contrastive learning; Non-IID data; MOON}, doi={10.4108/eai.22-5-2026.2365159} }- Hao Pi
Year: 2026
Improving Model-Contrastive Federated Learning under Non-IID Data Constraints
ICIAAI
EAI
DOI: 10.4108/eai.22-5-2026.2365159
Abstract
To address the performance degradation and convergence instability issues caused by the non-IID nature of client data in federated learning, this paper proposes an enhanced model-contrastive federated learning framework and optimizes the existing methods at the client-side local training stage by introducing the dynamic scheduling strategy for the contrastive loss weight, temperature parameter annealing mechanism and multi-negative sample comparison strategy, which improves classification accuracy and achieves a better trade-off between supervised learning and contrast constraints in different training stages. For MINST, the proposed method achieves a classification accuracy of 95.18%, outperforming FedAvg and MOON by 0.19% and 0.11%, respectively. For Fashion-MNIST, the classification accuracy reached 81.87%, outperforming FedAvg and MOON by 1.42% and 1.19%, respectively. The experimental results demonstrate the effectiveness and feasibility of the proposed improved strategy under non-IID settings.
