FLClustering: A Fair and Robust Federated Learning Approach in Heterogeneous Environments
Lists
Only you see this
Abstract
Federated learning in heterogeneous environments faces the fairness problem of global models performing differently across participants due to differences in participant sample resources and participation. The existing fairness algorithms which mainly adjust the aggregate weight are easy to give malicious participants more attack space, resulting in reduced robustness of the model. The mainstream attack defense schemes are mostly based on the assumption of participants’ data with independence and same distribution, which easily filters out legal but uncommon updates, leading to more unfairness. Aiming at this contradiction, this article proposes a fair and robust federated learning approach called FLClustering. Based on similar model parameters, a cluster of participants is generated to mitigate the heterogeneity among participants in the same cluster. Then a Byzantine attack defense mechanism is introduced in the cluster, which can resist label flipping attacks, model poisoning attacks and other Byzantine attacks while avoiding the error of discarding legitimate participants. Especially, both a dynamic adjustment strategy and a fair participant selection mechanism are proposed to further reduce the performance difference between participants. Experiments on Modified National Institute of Standards and Technology database (MNIST) and Fashion-MNIST datasets show that the proposed approach can obtain a fair and robust global model. Compared with the existing fairness algorithms, the accuracy under attack is improved by more than 3% and the standard deviation is reduced by more than 4%.
BibTeX
@article{lin2025flclustering,
author = {Li Lin and Wei Shen and Chi-Ho Song and Shuai Ma},
title = {{FLClustering: A Fair and Robust Federated Learning Approach in Heterogeneous Environments}},
journal = {IEEE Transactions on Computational Social Systems},
year = {2025},
doi = {10.1109/tcss.2025.3602647},
}Loading related content...