| 摘 要: 为了解决联邦学习中数据异构导致模型准确率下降的问题,提出了一种基于Logistic最优化鲁棒性的聚类联邦学习(Logistic-based More Robust Clustered Federated Learning,LMRCFL)方法,将具有相似数据分布的客户端分组到相同的集群中,不需要访问其私有数据,可为每个客户端集群训练模型;在目标函数中引入正则项更新本地损失函数,缓解Non-IID(非独立同分布)数据带来的客户端偏移问题,通过减小模型差异提升模型准确率。在CIFAR-10、fashion-MNIST、SHVN数据集上与其他联邦学习算法进行了对比,实验结果表明,LMRCFL算法在Non-IID数据分布下的准确率提高了8.13百分点~33.20百分点且具有鲁棒性。 |
| 关键词: 联邦学习;数据异构;聚类;非独立同分布;正则化 |
|
中图分类号: TP301
文献标识码: A
|
|
| Logistic-based More Robust Clustered Federated Learning |
|
SHI Yuqian, WU Zhaoxia
|
(School of Statistics and Data Science, Xinjiang University of Finance and Economics, Urumqi 830012, China)
1454945606@qq.com; wuzhaoxia828@163.com
|
| Abstract: In order to address the issue of decreased model accuracy in federated learning due to heterogeneous data, this paper proposes a method called Logistic-based More Robust Clustered Federated Learning (LMRCFL). This algorithm groups clients with similar data distributions into the same clusters without accessing their private data, enabling model training for each client cluster. A regularization term in the objective function is introduced to update the local loss function, so as to alleviate the problem of client drift caused by Non-IID (non-identically distributed) data and enhance model accuracy by reducing model discrepancies. Comparative experiments with other federated learning algorithms on CIFAR-10, Fashion-MNIST, and SHVN datasets demonstrate that the LMRCFL algorithm improves accuracy by 8.13 percentage points to 33.20 percentage points under Non-IID data distributions while maintaining robustness. |
| Keywords: federated learning; data heterogeneity; clustering; Non-IID; regularization |
