Energy-Efficient Hierarchical Federated Learning in UAV Networks with Partial AI Model Upload Under Non-Convex Loss.

Area of Science:

• Mobile Computing
• Artificial Intelligence
• Network Communications

Background:

• Hierarchical Federated Learning (HFL) faces challenges in mobile scenarios due to high energy consumption, communication costs, and convergence issues.
• Existing HFL methods assume convex loss and full parameter uploads, which are impractical for real-world non-convex tasks and edge constraints.

Purpose of the Study:

• To propose an energy-efficient optimization scheme for HFL in mobile/UAV networks under non-convex loss.
• To address limitations of current HFL by integrating dynamic partial-dimension model uploads and minimizing total energy consumption.

Main Methods:

• Developed a dynamically adjustable partial-dimension model upload mechanism by screening key update dimensions.
• Constructed a total energy minimization model incorporating communication/computation energy and an attendance rate constraint.
• Utilized Lyapunov optimization to transform the long-term problem into single-round solvable subproblems.

Main Results:

• The proposed scheme significantly reduces energy consumption compared to the original HFL algorithm.
• High test accuracy is maintained, demonstrating the effectiveness of the energy-saving approach.
• Mobility positively impacts system performance, as verified by simulation results.

Conclusions:

• The novel HFL scheme effectively balances energy efficiency and learning performance in mobile environments.
• Partial-dimension model uploads are crucial for reducing overhead in resource-constrained mobile/UAV networks.
• The integration of mobility considerations enhances the practicality and performance of federated learning systems.