Dynamic Corrected Split Federated Learning With Homomorphic Encryption for U-Shaped Medical Image Networks

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Science

Background:

• U-shaped networks are widely used in medical image analysis tasks like segmentation and restoration.
• Centralized learning in these networks raises significant data privacy concerns.
• Existing federated learning (FL) and split learning (SL) methods struggle to balance computational cost, privacy, and parallel training.

Purpose of the Study:

• To propose a novel hybrid learning paradigm, Dynamic Corrected Split Federated Learning (DC-SFL), for U-shaped medical image networks.
• To address privacy concerns associated with centralized learning in medical imaging.
• To improve the balance between computational cost, model privacy, and parallel training in distributed learning.

Main Methods:

• Proposed a three-part network split across different parties to preserve input, model parameters, label, and output privacy.
• Introduced a Dynamic Weight Correction Strategy (DWCS) to stabilize training and mitigate model drift from data heterogeneity.
• Incorporated additively homomorphic encryption into client-side model aggregation for enhanced privacy and protection against collusion.

Main Results:

• The proposed DC-SFL method was evaluated on various medical image tasks.
• Experimental results demonstrated the effectiveness of the DC-SFL approach.
• Achieved competitive performance compared to state-of-the-art distributed learning methods.

Conclusions:

• DC-SFL offers a robust solution for privacy-preserving U-shaped medical image network training.
• The hybrid approach effectively balances privacy, computational efficiency, and training stability.
• The method provides a trustworthy distributed learning paradigm for sensitive medical data.