Federated Learning via Plurality Vote.

Area of Science:

• Machine Learning
• Data Privacy
• Distributed Systems

Background:

• Federated learning enables collaborative model training while preserving data privacy.
• Existing federated learning methods face challenges in optimizing communication, reliability, and deployment efficiency.
• Joint optimization of these factors remains an open research problem.

Purpose of the Study:

• To propose a novel federated learning scheme, FedVote, that addresses the joint optimization of communication overhead, learning reliability, and deployment efficiency.
• To introduce a method for efficient and robust federated model aggregation and deployment.

Main Methods:

• Clients transmit low-bit (binary or ternary) weights to the server, reducing communication overhead.
• Model parameters are aggregated using a weighted voting mechanism for enhanced resilience against Byzantine attacks.
• The resulting quantized model is resource-friendly for edge device deployment.

Main Results:

• FedVote significantly reduces communication overhead compared to traditional federated learning approaches.
• The weighted voting aggregation enhances model robustness against adversarial (Byzantine) attacks.
• Models trained with FedVote exhibit reduced quantization error and faster convergence rates.
• The quantized models are efficient for deployment on resource-constrained edge devices.

Conclusions:

• FedVote presents an effective solution for optimizing communication, reliability, and deployment efficiency in federated learning.
• The proposed scheme offers a practical approach for secure and efficient collaborative machine learning on edge devices.
• FedVote demonstrates superior performance in terms of convergence speed and quantization error reduction compared to existing methods.