Related Experiment Videos
Adaptive Compressed Sensing Differential Privacy Federated Learning Based on Orbital Spatiotemporal Characteristics
Weibang Li1, Ling Li2, Lidong Zhu2
1College of Computer Science and Artificial Intelligence, Southwest Minzu University, Chengdu 610041, China.
Sensors (Basel, Switzerland)
|March 28, 2026
Summary
This study introduces a novel framework for federated learning in 6G Space-Air-Ground Integrated Networks (SAGINs), enhancing efficiency and privacy. The approach optimizes communication and protects data by adapting to satellite movement and resource variations.
Area of Science:
- Network Engineering
- Artificial Intelligence
- Cybersecurity
Background:
- 6G communication technology necessitates Space-Air-Ground Integrated Networks (SAGINs) for global intelligent computing.
- Federated learning in SAGINs faces challenges from satellite dynamics, heterogeneous resources, and data privacy vulnerabilities.
- Existing solutions often lack adaptability to the spatiotemporal characteristics of SAGINs.
Purpose of the Study:
- To propose an adaptive compressed sensing differential privacy federated learning framework for 6G SAGINs.
- To address the challenges of high dynamics, resource heterogeneity, and privacy vulnerabilities in SAGIN federated learning.
- To optimize communication efficiency and ensure robust privacy protection by integrating orbital spatiotemporal characteristics.
Main Methods:
- Designed orbital periodicity-driven time-varying sparse sensing matrices for dynamic compression strategies.
- Developed an orbital predictability-based privacy budget temporal allocation mechanism with compressed domain differential privacy injection.
- Constructed an energy-communication-privacy ternary collaborative mechanism using model predictive control and reinforcement learning for dynamic routing and aggregation.
Main Results:
- Achieved 3-12% improvement in model accuracy and 30-50% enhancement in communication efficiency compared to existing methods.
- Maintained differential privacy protection with dynamic privacy budget ε∈[0.1,10.0] and compression ratio ρ∈[0.2,0.8].
- Reduced reconstruction error by up to 19.4% using orbital-driven time-varying sensing matrices compared to fixed-matrix baselines.
Conclusions:
- The proposed framework effectively integrates orbital spatiotemporal characteristics with federated learning for 6G SAGINs.
- The adaptive and dynamic approach significantly improves model accuracy, communication efficiency, and privacy protection.
- The synergistic effectiveness validates the framework for future intelligent collaborative computing in dynamic network environments.
