A new federated learning-based wireless communication and client scheduling solution for combating COVID-19.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Federated learning (FL) offers privacy-preserving training for medical image analysis.
• High communication costs and statistical heterogeneity in FL degrade model performance.

Purpose of the Study:

• To propose FedUC, an algorithm to control client updates in federated learning.
• To mitigate statistical heterogeneity and reduce communication overhead in medical image analysis.

Main Methods:

• FedUC employs client scheduling based on weight divergence, update increment, and loss.
• Image augmentation balances local data, while gradient compression reduces communication costs.
• Server-side dynamic weighting of model parameters during aggregation.

Main Results:

• FedUC demonstrates superior training performance compared to existing FL methods.
• Improved model accuracy and significant reduction in wireless communication costs were observed.
• Effective mitigation of non-independently and identically distributed data challenges.

Conclusions:

• FedUC effectively addresses statistical heterogeneity and communication costs in federated learning for medical imaging.
• The proposed algorithm enhances model accuracy and efficiency in privacy-preserving distributed learning scenarios.