APB-FLDPA: Adaptive Personalized Blockchain-Federated Learning With Differential Privacy and Attention for...

Area of Science:

• Medical Artificial Intelligence (AI)
• Federated Learning (FL)
• Data Privacy and Security

Background:

• Developing robust medical AI necessitates multi-institutional collaboration, but regulations like HIPAA and GDPR impede centralized patient data sharing.
• Existing federated learning (FL) methods suffer significant performance degradation (15-30%) in clinical settings due to data heterogeneity, security risks, and privacy concerns.

Purpose of the Study:

• To introduce APB-FLDPA, a novel privacy-preserving federated learning framework designed for secure multi-hospital disease prediction.
• To address the limitations of current FL methods in real-world clinical applications, focusing on data heterogeneity, security, and privacy.

Main Methods:

• APB-FLDPA integrates adaptive Byzantine-resilient aggregation with dynamic client trust scoring.
• Incorporates self-attention for automated clinical feature importance, selective differential privacy, and cluster-aware personalization.
• Utilizes a lightweight blockchain module for enhanced model integrity and security.

Main Results:

• APB-FLDPA achieved 90.8% accuracy for diabetes and 83.8% accuracy for thyroid disease prediction across five institutions.
• Demonstrated minimal performance loss (<0.2%) compared to centralized learning, with statistical tests confirming significant improvements.
• Selective differential privacy enhanced accuracy by 5.6% over conventional methods.

Conclusions:

• APB-FLDPA offers a scalable, high-performance, and privacy-compliant solution for federated medical AI.
• The framework effectively overcomes data heterogeneity and privacy constraints in multi-institutional disease prediction.
• This approach facilitates collaborative medical AI development while adhering to strict data protection regulations.