Communication-efficient federated learning.

Area of Science:

• Computer Science
• Machine Learning
• Communication Engineering

Background:

• Federated learning (FL) allows collaborative model training without data sharing, crucial for privacy-sensitive applications.
• Communication delays from iterative parameter exchange in FL significantly hinder model convergence, especially in resource-limited networks.
• Existing FL methods face bottlenecks due to large parameter transmissions over wireless networks.

Purpose of the Study:

• To develop a communication-efficient federated learning framework to reduce convergence time and training loss.
• To address the major bottleneck of communication delay in federated learning systems.
• To improve the overall efficiency and performance of federated learning.

Main Methods:

• Proposed a probabilistic device selection scheme to prioritize devices that enhance convergence speed and reduce training loss.
• Introduced a quantization method to decrease the volume of exchanged model parameters.
• Developed an efficient wireless resource allocation scheme to optimize communication.

Main Results:

• The proposed FL framework demonstrated significant improvements in identification accuracy and convergence time.
• Achieved up to 3.6% improvement in identification accuracy compared to standard FL.
• Achieved up to 87% improvement in convergence time compared to standard FL.

Conclusions:

• The developed communication-efficient FL framework effectively mitigates communication delays.
• The framework jointly optimizes FL convergence time and training loss, outperforming standard FL.
• Probabilistic device selection, parameter quantization, and resource allocation are key to efficient FL.