Joint Client Selection and CPU Frequency Control in Wireless Federated Learning Networks with Power Constraints

Area of Science:

• Machine Learning
• Distributed Systems
• Wireless Networks

Background:

• Federated learning (FL) enables privacy-preserving distributed training but faces challenges from heterogeneous resources causing stragglers.
• Non-independent and identically distributed (non-IID) data and resource constraints further slow down FL training.
• Existing methods often optimize client selection or resource allocation separately, not jointly.

Purpose of the Study:

• To develop a joint optimization strategy for client selection and computational power (CPU frequency) in wireless federated learning.
• To minimize a cost function that balances learning latency and non-IID data characteristics.
• To address resource constraints and improve the efficiency of federated learning.

Main Methods:

• Defined a cost function incorporating learning latency and non-IID data properties.
• Formulated a joint client selection and CPU frequency control problem to minimize the time-averaged cost under power constraints.
• Applied Lyapunov optimization theory to convert the long-term problem into sequential short-term problems.
• Developed an algorithm for optimal client selection and CPU frequency allocation for clients and the edge server.

Main Results:

• The proposed algorithm achieves optimal client selection and CPU frequency control.
• Theoretical analysis provides performance guarantees for the algorithm.
• Simulation results demonstrate superior test accuracy compared to existing algorithms.
• The approach effectively maintains low power consumption.

Conclusions:

• The joint optimization of client selection and CPU frequency is crucial for efficient federated learning in wireless networks.
• The proposed algorithm effectively mitigates challenges posed by stragglers and non-IID data.
• This method offers a promising solution for improving federated learning performance and accuracy while managing resources.