Differentially Private Client Selection and Resource Allocation in Federated Learning for Medical Applications Using...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Federated learning (FL) enables decentralized model training, crucial for healthcare data privacy.
• Medical applications face challenges with data heterogeneity and communication limits in FL.

Purpose of the Study:

• To propose DPS-GAT, a novel FL approach integrating graph attention networks (GATs) with differential privacy.
• To optimize client selection and resource allocation in FL for medical applications.
• To enhance model robustness, generalizability, and privacy preservation.

Main Methods:

• Integration of graph neural networks (GNNs) to model client relationships.
• Implementation of differentially private client selection and resource allocation.
• Experimental validation using the Regensburg pediatric appendicitis open dataset.

Main Results:

• DPS-GAT demonstrated superior accuracy, privacy, and resource efficiency compared to traditional FL methods.
• The approach maintained stable client selection across FL rounds and privacy budgets.
• Achieved a balance between strong privacy guarantees and high model performance.

Conclusions:

• DPS-GAT offers a promising direction for secure and efficient FL in medical applications.
• The method can improve patient care via enhanced predictive models and collaborative data use.
• Highlights the feasibility of robust privacy in FL without performance compromise.