FedAGL: A Communication-Efficient Federated Vehicular Network

Su Liu, Yushuai Li, Peiyuan Guan, Tianyi Li, Jiong Yu, Amir Taherkordi, Christian S. Jensen

Published in: IEEE Transactions on Intelligent Vehicles

Abstract

Federated Learning (FL) is a promising approach for training AI models in vehicular networks while preserving data privacy. However, high communication overhead and transmission latency remain significant barriers in these fast-moving environments. This paper presents FedAGL, a framework that optimizes communication efficiency through Adaptive Parameter Control (APC) and a hybrid bandwidth allocation strategy. By selectively freezing stable parameters and employing an LSTM-based online allocation model, FedAGL achieves substantial reductions in latency and data overhead without compromising global model accuracy.

Introduction

Modern autonomous and connected vehicles generate massive volumes of sensor data. While this data is essential for training advanced AI models, privacy concerns and the sheer volume of data make centralized training impractical. Federated Learning (FL) addresses privacy by keeping data on-device and only sharing model updates. However, in a vehicular context, FL faces two primary challenges: the massive communication overhead of modern AI models and the "straggler problem," where network latency is dictated by the slowest vehicle.

System architecture of the Federated Vehicular Network (FVN) showing the interaction between vehicles, Road Side Units (RSU), and the Edge Server (ES).

The FedAGL Framework

FedAGL introduces a three-pronged approach to mitigate communication bottlenecks and optimize resource utilization in Federated Vehicular Networks (FVN).

1. Adaptive Parameter Control (APC)

Rather than transmitting the entire model in every round, FedAGL identifies "stable" parameters—those that exhibit minimal change during training. Using an Exponential Moving Average (EMA) to track "efficient disturbance," the system identifies parameters that have matured. These parameters are frozen and excluded from transmission for a set duration, significantly reducing the data footprint of each update.

2. Bandwidth Allocation Strategy

Because APC results in varying update sizes across different vehicles, uniform bandwidth allocation is inefficient. FedAGL addresses this through a two-stage strategy:

  • Offline Optimization: A Genetic Algorithm (GA) is used to solve a complex min-max optimization problem, finding the ideal bandwidth split to minimize total network latency.
  • Online Execution: Since GA is too computationally intensive for real-time vehicular use, the authors trained a Long Short-Term Memory (LSTM) network using the GA's results. This allows the system to perform near-optimal bandwidth allocation in milliseconds based on current network conditions.
The FedAGL workflow: integrating local training with Adaptive Parameter Control (APC), bandwidth allocation strategies, and global model aggregation.

Experimental Evaluation

The framework was evaluated using three standard computer vision benchmarks: MNIST (hand-written digits), CIFAR-10 (general objects), and SVHN (Street View House Numbers). To simulate realistic conditions, the data was distributed across vehicles using a Dirichlet distribution to create a "Non-IID" (heterogeneous) environment.

Key Results

FedAGL demonstrated superior performance compared to standard FedAvg and other state-of-the-art parameter-reduction techniques:

  • Overhead Reduction: Achieved up to a 10.44% reduction in total communication overhead.
  • Latency Improvement: Reduced transmission latency by up to 22.62%.
  • Faster Convergence: FedAGL reached target accuracy levels significantly faster. For example, on the CIFAR-10 dataset, FedAGL required only 35 rounds to reach 80% accuracy, compared to 121 rounds for standard FedAvg.
  • Optimization Rate: The overall communication optimization rate was found to be 3.47 times better than traditional methods on average.

Conclusion and Future Work

FedAGL proves that intelligent parameter management and predictive resource allocation can make Federated Learning viable for high-stakes vehicular environments. While the current model focuses on bandwidth, future research will look to incorporate constraints on vehicle memory and CPU power, as well as addressing the challenges of high-speed mobility, such as packet loss and signal fluctuations. Additionally, the integration of Vehicle-to-Vehicle (V2V) communication could further offload the burden from centralized Road Side Units.

Summary updated

Abstract

With the development of the technologies deployed on vehicles, there is a significant increase in the amount of data, which comes from various applications, such as battery management, VR, autopilot, etc. However, privacy is a critical obstacle to utilizing such information since many vehicle-based applications involve locations, conversations, driving behaviors, etc. Federated Learning (FL) is a promising technology perfect for filling the gap, as it keeps users' data in their devices, which gives rise to Federated Vehicular Networks (FVN). As a distributed paradigm, communication efficiency is an essential issue for researchers, since it impacts every aspect of FVN. This article presents a communication-efficient federated vehicular network called FedAGL integrating A daptive parameter control, G enetic Algorithm (GA), and L ong Short-Term Memory (LSTM). This is a framework designed to accelerate the training process of FVN by minimizing communication overhead and latency. In this article, we introduce a parameter-control algorithm to alleviate transmission burdens, thereby reducing communication overhead. This approach does not involve freezing the parameters of the entire layer but identifies certain parameters with minor updates. Further, we employ the GA to allocate communication bandwidth for each vehicle, considering the varying sizes of vehicles' active parameters. Thus, the maximum transmission time in each FL round decreases. In addition, due to the complexity of the genetic algorithm, the LSTM model is developed to accelerate the bandwidth allocation with quick response and sufficient precision, which uses data from the genetic algorithm as training labels. Abundant experiments indicate that the proposed framework achieves up to 10.44% and 22.62% reduction in communication overhead and latency, respectively, outperforming the benchmarks.
BibTeX
@article{liu2024fedagl,
  author = {Su Liu and Yushuai Li and Peiyuan Guan and Tianyi Li and Jiong Yu and Amir Taherkordi and Christian S. Jensen},
  title = {{FedAGL: A Communication-Efficient Federated Vehicular Network}},
  journal = {IEEE Trans. Intell. Veh.},
  year = {2024},
  doi = {10.1109/tiv.2024.3349655},
}
Figure