StoCFL: A stochastically clustered federated learning framework for Non-IID data with dynamic client participation

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Federated learning (FL) systems face performance degradation due to decentralized and Non-Independently Identically Distributed (Non-IID) data.
• Existing clustered federated learning (CFL) methods lack cross-cluster communication, leading to inefficiency and poor model performance.
• Current CFL approaches are sensitive to accurate, pre-defined client clustering, which is often impractical.

Purpose of the Study:

• To propose StoCFL, a novel clustered federated learning framework designed to mitigate the impact of Non-IID data.
• To develop a flexible CFL framework accommodating varying client participation and new client additions.
• To enhance data efficiency and model performance in decentralized learning environments.

Main Methods:

• StoCFL implements a flexible clustered federated learning framework with an information-sharing mechanism across clusters.
• The framework supports arbitrary client participation rates and accommodates newly joined clients in dynamic FL systems.
• Experiments were conducted using four Non-IID settings and a real-world dataset to evaluate performance.

Main Results:

• StoCFL demonstrates promising client clustering results, even when the number of clusters is not predefined.
• Models trained using StoCFL significantly outperform baseline approaches across various Non-IID scenarios.
• The framework achieves substantial improvements in model performance and data efficiency.

Conclusions:

• StoCFL effectively addresses generic Non-IID issues in federated learning through its novel clustered approach.
• The proposed framework offers flexibility and robustness for real-world, dynamic federated learning systems.
• StoCFL represents a significant advancement in improving the performance and efficiency of clustered federated learning.