Federated Learning in the Sky: Joint Power Allocation and Scheduling with UAV Swarms

Tengchan Zeng, Omid Semiari, Mohammad Mozaffari, Mingzhe Chen, Walid Saad, Mehdi Bennis

Abstract

Unmanned aerial vehicle (UAV) swarms must exploit machine learning (ML) in order to execute various tasks ranging from coordinated trajectory planning to cooperative target recognition. However, due to the lack of continuous connections between the UAV swarm and ground base stations (BSs), using centralized ML will be challenging, particularly when dealing with a large volume of data. In this paper, a novel framework is proposed to implement distributed federated learning (FL) algorithms within a UAV swarm that consists of a leading UAV and several following UAVs. Each following UAV trains a local FL model based on its collected data and then sends this trained local model to the leading UAV who will aggregate the received models, generate a global FL model, and transmit it to followers over the intra-swarm network. To identify how wireless factors, like fading, transmission delay, and UAV antenna angle deviations resulting from wind and mechanical vibrations, impact the performance of FL, a rigorous convergence analysis for FL is performed. Then, a joint power allocation and scheduling design is proposed to optimize the convergence rate of FL while taking into account the energy consumption during convergence and the delay requirement imposed by the swarm's control system. Simulation results validate the effectiveness of the FL convergence analysis and show that the joint design strategy can reduce the number of communication rounds needed for convergence by as much as 35% compared with the baseline design.

Introduction

UAV swarms are increasingly tasked with complex missions requiring autonomous decision-making, such as navigation and target recognition. Traditionally, these tasks rely on centralized machine learning (ML), where data is offloaded to a ground base station. However, this approach is often infeasible in aerial environments due to intermittent connectivity and the massive volume of data generated in flight. Federated Learning (FL) offers a promising alternative by allowing UAVs to train models locally and share only model updates. Yet, implementing FL in the sky introduces unique challenges: unstable wireless links caused by mobility and antenna misalignment, severe battery constraints, and strict low-latency requirements for swarm control stability.

System Model

The proposed framework utilizes a leader-follower architecture. In this setup, "follower" UAVs perform local training on their respective datasets, while a designated "leader" UAV aggregates these updates to form a global model. This global model is then broadcast back to the followers for the next iteration of training.

Leader-follower architecture for federated learning in a UAV swarm.

A critical aspect of this model is the inclusion of UAV-specific physical factors. The authors account for antenna angle deviations caused by wind and mechanical vibrations, which are modeled as Gaussian random variables that impact signal gain. Furthermore, a flying power model is integrated to account for the energy required to maintain flight at various velocities, creating a direct link between mechanical flight parameters and communication performance.

Convergence Analysis and Optimization

The core of the paper is a rigorous convergence analysis that links FL performance to wireless factors. The authors derive a theoretical bound (Theorem 1) for the number of convergence rounds, showing it depends on the probability of successful uplink and downlink transmissions within specific time windows.

To maximize learning efficiency, a joint optimization problem is formulated to minimize the number of communication rounds. The optimization variables include transmission power, the ratio of time allocated to uplink versus downlink (scheduling), and flight velocity. This non-convex problem is addressed using Sample Average Approximation (SAA) and dual methods from convex optimization, allowing the swarm to balance the trade-offs between flight endurance, communication reliability, and learning speed.

Experimental Results

The framework was validated using synthetic data and a linear regression FL task. The simulations utilized standard UAV parameters, such as a four-rotor configuration and Rician fading channel models, with a swarm size of five followers.

Comparison of the proposed joint design against baseline scheduling and power allocation schemes.

Key findings include:

  • Theoretical Accuracy: Simulation results closely matched the derived theoretical convergence rounds, with differences typically under 5%.
  • Performance Gains: The proposed joint design reduced the number of communication rounds by up to 35% compared to baseline schemes that optimize only power or scheduling independently.
  • Bandwidth Sensitivity: While the proposed design is superior across the board, the performance gap narrows at higher bandwidths as the delay budget becomes easier to satisfy regardless of the strategy.

Discussion and Conclusion

This research highlights the significant impact of physical disturbances—such as wind-induced vibration—on algorithmic performance in aerial networks. By treating flight velocity as a learning variable, the authors demonstrate a holistic approach to swarm management.

However, limitations remain. The current study focused on a small swarm and strongly convex loss functions. Scaling to larger swarms may introduce interference bottlenecks, and deep learning tasks with non-convex loss functions will require further analysis. Additionally, moving from pre-computed optimization to fully autonomous, on-the-fly parameter adjustment remains an important area for future work. In conclusion, this joint design provides a robust foundation for deploying efficient federated learning in resource-constrained UAV swarms.

Summary updated

Abstract

Unmanned aerial vehicle (UAV) swarms must exploit machine learning (ML) in order to execute various tasks ranging from coordinated trajectory planning to cooperative target recognition. However, due to the lack of continuous connections between the UAV swarm and ground base stations (BSs), using centralized ML will be challenging, particularly when dealing with a large volume of data. In this paper, a novel framework is proposed to implement distributed federated learning (FL) algorithms within a UAV swarm that consists of a leading UAV and several following UAVs. Each following UAV trains a local FL model based on its collected data and then sends this trained local model to the leading UAV who will aggregate the received models, generate a global FL model, and transmit it to followers over the intra-swarm network. To identify how wireless factors, like fading, transmission delay, and UAV antenna angle deviations resulting from wind and mechanical vibrations, impact the performance of FL, a rigorous convergence analysis for FL is performed. Then, a joint power allocation and scheduling design is proposed to optimize the convergence rate of FL while taking into account the energy consumption during convergence and the delay requirement imposed by the swarm's control system. Simulation results validate the effectiveness of the FL convergence analysis and show that the joint design strategy can reduce the number of communication rounds needed for convergence by as much as 35% compared with the baseline design.
BibTeX
@inproceedings{zeng2020federated,
  author = {Tengchan Zeng and Omid Semiari and Mohammad Mozaffari and Mingzhe Chen and Walid Saad and Mehdi Bennis},
  title = {{Federated Learning in the Sky: Joint Power Allocation and Scheduling with UAV Swarms}},
  booktitle = {ICC},
  year = {2020},
  url = {https://openreview.net/forum?id=cNSFjDlkjLN},
}
Figure