Privacy-Preserving Federated Learning for UAV-Enabled Networks: Learning-Based Joint Scheduling and Resource Management

Helin Yang, Jun Zhao, Zehui Xiong, Kwok-Yan Lam, Sumei Sun, Liang Xiao

TL;DR

This paper introduces a privacy-preserving Asynchronous Federated Learning (AFL) framework specifically designed for networks where Unmanned Aerial Vehicles (UAVs) act as flying base stations. By combining AFL with a Deep Reinforcement Learning algorithm (A3C), the authors successfully coordinate device selection, UAV positioning, and resource allocation to achieve faster convergence and higher accuracy than traditional synchronous methods.

What Problem Does This Paper Address?

Deploying AI in the field often requires training models on data generated by mobile devices (like smartphones or IoT sensors). However, two major hurdles exist:

  1. Privacy: Users are rightfully hesitant to send raw, sensitive data to a central server.
  2. Resource Constraints: UAVs, while flexible as "flying base stations," have limited battery life and computing power.

Standard Federated Learning (FL) solves the privacy issue by keeping data on-device and only sharing model updates. However, standard FL is usually synchronous—the server waits for every single device to finish its local training before updating the global model. In a dynamic UAV network, "straggler" devices with poor connections or slow processors create a massive bottleneck. This paper tackles the challenge of making FL efficient and reliable in these highly variable, resource-constrained aerial environments.

The architecture of a Federated Learning-based UAV-enabled wireless network, showing the interaction between ground devices and aerial servers.

Methods

The authors propose a two-pronged solution:

  • Asynchronous Federated Learning (AFL): Unlike synchronous FL, the UAV servers in this framework update the global model as soon as individual devices finish their local tasks. This prevents the "slowest hiker" from holding up the entire group.
  • A3C-based Optimization: To manage the complexity of the network, they use the Asynchronous Advantage Actor-Critic (A3C) algorithm. This Deep Reinforcement Learning (DRL) approach handles the "joint" optimization of three critical variables:
  • Device Selection: Choosing only high-quality devices (those with good data and strong hardware) to participate.
  • UAV Placement: Dynamically moving UAVs to the best horizontal coordinates to maximize signal strength for the chosen devices.
  • Resource Management: Allocating subchannels and transmit power levels.

The use of A3C is particularly clever here because it can handle continuous action spaces—like the specific X and Y coordinates of a flying UAV—which traditional Q-learning struggles with.

Datasets

The framework was evaluated using:

  • MNIST Dataset: The classic benchmark of handwritten digits (60,000 training images, 10,000 testing images).
  • CNN Model: A Convolutional Neural Network was used as the local model trained on the devices.
  • Simulated UAV Environment: A square 400m x 400m area with four UAVs and up to 300 ground devices, incorporating realistic air-to-ground path loss models (LoS and NLoS).

Key Insights and Ideas

  • Filtering "Low-Quality" Participants: One of the most important findings is that more data isn't always better. Devices with "low-quality" parameters (due to noisy data or poor local training) can actually degrade the global model. The A3C agent learns to ignore these devices, improving overall accuracy.
  • Asynchronous is Essential for Mobility: In a network where both the base stations (UAVs) and the users are moving, maintaining a synchronous "heartbeat" is nearly impossible. Asynchrony allows the system to remain resilient to temporary connection drops.
  • The Power of Joint Optimization: Rather than solving for UAV placement and then solving for resource allocation, the A3C agent treats them as a single, unified problem. This allows the system to find "win-win" configurations that a piecemeal approach would miss.

Key Results

  • Accuracy: The proposed A3C-AFL reached 90% accuracy in approximately 25 global rounds.
  • Speed: Compared to Synchronous Federated Learning (SFL), the AFL approach significantly reduced "wall-clock time" because it didn't have to wait for stragglers.
  • Scalability: As the number of devices increased from 50 to 300, the A3C-AFL algorithm maintained high accuracy (~94%), while simpler benchmarks (like AFL without device selection) saw accuracy drop or stall around 78%.
  • Convergence: The A3C agent's training converged steadily after about 500 episodes, proving that the DRL approach can effectively learn the complexities of the UAV environment.

Limitations and Future Work

  • Dataset Complexity: While MNIST is a standard starting point, it is relatively simple. Future work should test the framework on more complex, non-IID (Independent and Identically Distributed) data typical of real-world edge environments.
  • UAV Altitude: The current model focuses on 2D horizontal placement at a fixed height. Incorporating 3D trajectory optimization (changing altitude) could further improve signal quality but would add significant complexity to the A3C action space.
  • Energy Efficiency: While the paper mentions resource constraints, a more explicit "energy-per-bit" or battery-life-extension objective in the reward function could make the framework even more practical for long-term UAV deployments.

Summary updated

Abstract

Unmanned aerial vehicles (UAVs) are capable of serving as flying base stations (BSs) for supporting data collection, machine learning (ML) model training, and wireless communications. However, due to the privacy concerns of devices and limited computation or communication resource of UAVs, it is impractical to send raw data of devices to UAV servers for model training. Moreover, due to the dynamic channel condition and heterogeneous computing capacity of devices in UAV-enabled networks, the reliability and efficiency of data sharing require to be further improved. In this paper, we develop an asynchronous federated learning (AFL) framework for multi-UAV-enabled networks, which can provide asynchronous distributed computing by enabling model training locally without transmitting raw sensitive data to UAV servers. The device selection strategy is also introduced into the AFL framework to keep the low-quality devices from affecting the learning efficiency and accuracy. Moreover, we propose an asynchronous advantage actor-critic (A3C) based joint device selection, UAVs placement, and resource management algorithm to enhance the federated convergence speed and accuracy. Simulation results demonstrate that our proposed framework and algorithm achieve higher learning accuracy and faster federated execution time compared to other existing solutions.
BibTeX
@article{yang2021privacy,
  author = {Helin Yang and Jun Zhao and Zehui Xiong and Kwok-Yan Lam and Sumei Sun and Liang Xiao},
  title = {{Privacy-Preserving Federated Learning for UAV-Enabled Networks: Learning-Based Joint Scheduling and Resource Management}},
  journal = {IEEE Journal on Selected Areas in Communications},
  year = {2021},
  url = {https://openreview.net/forum?id=dYduCOK2mq},
}

Loading related content...

Figure