Federated Learning in Vehicular Networks

Ahmet M. Elbir, Burak Soner, Sinem Coleri, Deniz Gündüz, Mehdi Bennis

Abstract

Machine learning (ML) has recently been adopted in vehicular networks for applications such as autonomous driving, road safety prediction and vehicular object detection, due to its model-free characteristic, allowing adaptive fast response. However, most of these ML applications employ centralized learning (CL), which brings significant overhead for data transmission between the parameter server and vehicular edge devices. Federated learning (FL) framework has been recently introduced as an efficient tool with the goal of reducing transmission overhead while achieving privacy through the transmission of model updates instead of the whole dataset. In this paper, we investigate the usage of FL over CL in vehicular network applications to develop intelligent transportation systems. We provide a comprehensive analysis on the feasibility of FL for the ML based vehicular applications, as well as investigating object detection by utilizing image-based datasets as a case study. Then, we identify the major challenges from both learning perspective, i.e., data labeling and model training, and from the communications point of view, i.e., data rate, reliability, transmission overhead, privacy and resource management. Finally, we highlight related future research directions for FL in vehicular networks.

Introduction

Modern autonomous vehicles generate a staggering amount of sensor data—roughly one gigabyte per second from cameras, RADAR, and LIDAR. Traditionally, training Machine Learning (ML) models for autonomous driving relies on Centralized Learning (CL), where all this raw data is transmitted to a central cloud server. In a dynamic vehicular network, CL is highly problematic. Transmitting massive datasets incurs prohibitive bandwidth costs, introduces unacceptable latency, and raises severe privacy concerns. Federated Learning (FL) offers a solution by bringing the training process to the edge (the vehicles themselves). Instead of sharing raw data, vehicles train models locally and only share the calculated gradients (model updates) with the server.

Learning Frameworks for Intelligent Transportation Systems

To address the challenges of mobility and resource constraints, the authors evaluate three distinct architectural approaches for vehicular networks:

  • Centralized Learning (CL): The standard baseline where all raw data is sent to a central server for training.
  • Federated Learning (FL): Vehicles compute model updates locally using their own data and transmit only the parameter updates to a central server for aggregation.
  • Hybrid Federated Centralized Learning (HFCL): A novel middle-ground approach introduced to handle device heterogeneity. Vehicles with sufficient computational resources perform FL locally. Meanwhile, resource-constrained vehicles act as passive clients, sending their raw data to a Parameter Server (PS) which computes the model updates on their behalf.
Comparison of Centralized, Federated, and Hybrid Federated Centralized Learning architectures.

Experimental Setup and Case Studies

The performance of these frameworks was tested using Convolutional Neural Networks (CNNs) and U-net architectures across two primary datasets:

  1. Lyft Level 5 AV Dataset: Used for 3D object detection. It contains LIDAR and camera data preprocessed into top-view images featuring 9 object classes.
  2. Custom LIDAR/mm-Wave Dataset: Used for Vehicle-to-Infrastructure (V2I) millimeter-wave beam selection, mapping LIDAR point cloud measurements to optimal beam directions.

Performance Evaluation and Key Results

The study demonstrates that FL effectively shifts the system bottleneck from the wireless communication channel to the edge device's local hardware. Quantitative results highlight significant gains in efficiency:

  • 3D Object Detection: CL required transmitting ~4.5 billion data symbols. FL and HFCL reduced this transmission overhead by approximately 28 times and 3 times, respectively. HFCL provided a performance balance between the high accuracy of CL and the low overhead of FL.
  • mm-Wave Beam Selection: FL achieved an 88.6% beam selection accuracy, highly competitive with the 89.1% accuracy of CL, while requiring 300 times lower transmission overhead.
Accuracy comparison between CL, FL, and HFCL over communication rounds.

Discussion and Future Directions

While FL shows great promise, several hurdles remain for its deployment in vehicular environments. Higher-level tasks like intent estimation and driving decisions are prime candidates for FL due to their need for rapid local adaptation. However, lower-level tasks still often require the heavy lifting of CL due to labeling requirements.

Future research must address:

  • Data Labeling: Exploring Reinforcement Learning (RL) to bypass the need for manual human annotation on-the-fly.
  • Non-IID Data: Using Transfer Learning (TL) to help models adapt to diverse geographic data distributions.
  • Model Compression: Implementing sparsification and quantization to further reduce the size of model parameter updates.
  • Resource Management: Developing multi-hop FL frameworks to handle the high mobility and frequent handovers inherent in vehicular topologies.

Summary updated

Abstract

Machine learning (ML) has recently been adopted in vehicular networks for applications such as autonomous driving, road safety prediction and vehicular object detection, due to its model-free characteristic, allowing adaptive fast response. However, most of these ML applications employ centralized learning (CL), which brings significant overhead for data trans-mission between the parameter server and vehicular edge devices. Federated learning (FL) framework has been recently introduced as an efficient tool with the goal of reducing transmission overhead while achieving privacy through the transmission of model updates instead of the whole dataset. In this paper, we investigate the usage of FL over CL in vehicular network applications to develop intelligent transportation systems. We provide a comprehensive analysis on the feasibility of FL for the ML based vehicular applications, as well as investigating object detection by utilizing image-based datasets as a case study. Then, we identify the major challenges from both learning perspective, i.e., data labeling and model training, and from the communications point of view, i.e., data rate, reliability, transmission overhead, privacy and resource management. Finally, we highlight related future research directions for FL in vehicular networks.
BibTeX
@inproceedings{elbir2022federated,
  author = {Ahmet M. Elbir and Burak Soner and Sinem Coleri and Deniz Gündüz and Mehdi Bennis},
  title = {{Federated Learning in Vehicular Networks}},
  booktitle = {MeditCom},
  year = {2022},
  url = {https://openreview.net/forum?id=JzTUjtbSMjy},
}
Figure