FedSTS: A Stratified Client Selection Framework for Consistently Fast Federated Learning.

Area of Science:

• Machine Learning
• Distributed Systems
• Optimization

Background:

• Federated learning (FL) enables collaborative model training without sharing raw data.
• Client selection is crucial for efficient FL, with clustering being a common approach.
• Existing clustering methods struggle with high-dimensional gradients, leading to suboptimal and inconsistent client grouping.

Purpose of the Study:

• To propose FedSTS, a novel client selection scheme for horizontal federated learning.
• To accelerate FL convergence by reducing variance and improving client representativeness.
• To address the limitations of raw gradient-based clustering in FL.

Main Methods:

• FedSTS stratifies compressed model updates for effective client grouping.
• It reduces cross-client variance by reallocating sampling probabilities based on group heterogeneity.
• The scheme prioritizes client groups with lower similarity to enhance subset representativeness.

Main Results:

• Theoretical analysis demonstrates significant variance reduction and improved convergence guarantees.
• FedSTS achieves better grouping effectiveness compared to traditional methods.
• Experimental results validate the superior efficiency of FedSTS over alternative approaches.

Conclusions:

• FedSTS offers a robust solution for client selection in federated learning.
• The proposed method enhances training convergence speed and reliability.
• This approach effectively tackles the challenges of gradient-based clustering in FL.