Active Client Selection for Clustered Federated Learning.

Area of Science:

• Machine Learning
• Distributed Systems
• Artificial Intelligence

Background:

• Federated learning (FL) is a distributed machine learning framework with privacy and communication constraints.
• Clustered FL (CFL) addresses data heterogeneity by learning customized models for client groups.
• Existing CFL methods suffer from slow convergence and limited performance due to ineffective client selection, especially with non-IID data.

Purpose of the Study:

• To introduce a novel active client selection strategy for clustered federated learning (ACFL).
• To enhance the efficiency and accuracy of CFL by selecting the most informative clients for model updates.

Main Methods:

• Proposed ACFL method utilizes active learning (AL) metrics (e.g., uncertainty sampling, QBC, loss) for client selection within each cluster.
• In each round, clusters identify and aggregate updates from a small subset of informative clients.
• Empirical evaluation on MNIST, CIFAR-10, and LEAF datasets under class-imbalanced conditions.

Main Results:

• ACFL significantly accelerates the learning process compared to baseline FL and CFL methods.
• The method requires less client participation while achieving improved model accuracy.
• ACFL demonstrates a notable reduction in communication overhead.

Conclusions:

• Active client selection is an effective strategy for improving clustered federated learning.
• ACFL offers a promising solution for efficient and accurate distributed machine learning in heterogeneous environments.
• The proposed method enhances CFL performance by intelligently selecting clients based on informativeness.