Auction-guided model diffusion for communication-efficient federated learning on non-IID data.

Area of Science:

• Artificial Intelligence
• Mobile Communication Systems
• Machine Learning

Background:

• 6G systems utilize AI-driven network functions, adopting Federated Learning (FL) to protect user data privacy.
• Non-independent and identically distributed (non-IID) datasets in FL can degrade global model performance due to gradient divergence, causing weight divergence issues.

Purpose of the Study:

• To propose a novel diffusion strategy, FedDif, to enhance machine learning (ML) model performance in 6G systems using non-IID data.
• To theoretically prove FedDif's ability to overcome weight divergence problems inherent in FL with non-IID data.

Main Methods:

• FedDif facilitates local models learning diverse data distributions via device-to-device communication before parameter aggregation.
• Auction theory is employed to develop a communication-efficient diffusion strategy, balancing learning performance and communication costs.
• Theoretical analysis demonstrates FedDif's capability to circumvent the weight divergence problem.

Main Results:

• FedDif significantly improves top-1 test accuracy by up to 20.07 percentage points.
• The proposed strategy reduces communication costs by as much as 45.27% compared to the standard FedAvg algorithm.
• Experimental validation confirms FedDif's effectiveness in handling non-IID data and optimizing resource allocation.

Conclusions:

• FedDif offers a robust solution for enhancing FL performance in 6G environments with non-IID data.
• The diffusion strategy effectively mitigates weight divergence, leading to superior global model accuracy.
• FedDif presents a communication-efficient approach, optimizing the trade-off between learning gains and transmission overhead.