Privacy-Preserving Federated Learning for Space-Air-Ground Integrated Networks: A Bi-Level Reinforcement Learning and...

Area of Science:

• Computer Science
• Network Engineering
• Artificial Intelligence

Background:

• Space-Air-Ground Integrated Networks (SAGINs) are crucial for future intelligent communication.
• Federated Learning (FL) faces challenges in SAGINs due to high latency, dynamic topology, and privacy concerns.

Purpose of the Study:

• To propose a Privacy-Preserving Federated Learning framework for SAGINs (PPFL-SAGIN).
• To address data heterogeneity, device dynamics, and privacy leakage in SAGINs.

Main Methods:

• Integration of differential privacy, adaptive transfer learning, and bi-level reinforcement learning.
• Adaptive knowledge-sharing mechanism for device heterogeneity and data distribution divergence.
• Bi-level reinforcement learning for optimized device selection and model convergence.
• Dynamic privacy budget allocation and robust aggregation algorithms.

Main Results:

• PPFL-SAGIN demonstrates superior performance over baseline methods (FedAvg, FedAsync, FedAsyncISL).
• Significant improvements in model accuracy, convergence speed, and privacy protection strength were observed.
• Effectiveness in privacy preservation, device selection, and global aggregation within SAGINs.

Conclusions:

• PPFL-SAGIN effectively tackles key challenges in applying FL to SAGINs.
• The proposed framework offers a robust solution for intelligent communication in integrated network environments.