RaftFed: A Lightweight Federated Learning Framework for Vehicular Crowd Intelligence

Changan Yang, Yaxing Chen, Yao Zhang, Helei Cui, Zhiwen Yu, Bin Guo, Zheng Yan, Zijiang YangarXiv 2023 · arXiv · OpenReview

Abstract

Vehicular crowd intelligence (VCI) is an emerging research field. Facilitated by state-of-the-art vehicular ad-hoc networks and artificial intelligence, various VCI applications come to place, e.g., collaborative sensing, positioning, and mapping. The collaborative property of VCI applications generally requires data to be shared among participants, thus forming network-wide intelligence. How to fulfill this process without compromising data privacy remains a challenging issue. Although federated learning (FL) is a promising tool to solve the problem, adapting conventional FL frameworks to VCI is nontrivial. First, the centralized model aggregation is unreliable in VCI because of the existence of stragglers with unfavorable channel conditions. Second, existing FL schemes are vulnerable to Non-IID data, which is intensified by the data heterogeneity in VCI. This paper proposes a novel federated learning framework called RaftFed to facilitate privacy-preserving VCI. The experimental results show that RaftFed performs better than baselines regarding communication overhead, model accuracy, and model convergence.

Introduction

In the rapidly evolving landscape of connected vehicles, Vehicular Crowd Intelligence (VCI) enables cars to share critical "intelligence"—such as real-time traffic patterns or hazard maps—without compromising the privacy of raw sensor data. Federated Learning (FL) has emerged as the standard approach for this, allowing models to be trained locally on vehicles and aggregated into a global intelligence.

However, traditional FL architectures face significant hurdles in vehicular environments. First, reliance on a central cloud or edge server introduces high latency and reliability issues in high-speed vehicular ad-hoc networks (VANETs). Second, the data collected by vehicles is often Non-Independent and Identically Distributed (Non-IID), as different vehicles encounter vastly different environments. This data heterogeneity makes it difficult for a single global model to converge effectively.

System Architecture

The VCI ecosystem is typically conceptualized across three layers: the Real World Infrastructure, the Network Layer (VANET), and the Intelligence Layer. In traditional frameworks, the Intelligence Layer is often hosted on remote cloud servers, creating a communication bottleneck.

The multi-layered architecture of Vehicular Crowd Intelligence, illustrating the interaction between physical infrastructure, network communication, and the intelligence layer.

RaftFed addresses these challenges by decentralizing the Intelligence Layer, shifting the model aggregation process from the cloud directly onto the vehicles themselves.

The RaftFed Methodology

RaftFed introduces a two-phase approach to ensure resilient and accurate learning in dynamic environments.

1. Clustering Election Phase

To eliminate the need for a fixed central server, RaftFed utilizes a consensus-based leadership protocol inspired by the Raft algorithm. Vehicles elect a "leader" node based on network stability. This leader then performs dynamic clustering, grouping nearby vehicles using a "Data Distribution Overlap Rate" metric. By grouping vehicles with similar data distributions, the framework mitigates the negative impacts of Non-IID data.

2. Model Training and Aggregation

Once clusters are established, training proceeds through a hierarchical aggregation process. As illustrated in the workflow below, local vehicles (e.g., Vehicle 1 and Vehicle 2) train models on their private sensing data and upload the parameters to the elected Central Node (Vehicle 3).

The RaftFed decentralized federated learning workflow, where an elected vehicle acts as the central node for model aggregation and distribution.

This decentralized approach uses two-layer aggregation: models are first synchronized within a cluster (Intra-cluster) and then across different clusters (Inter-cluster). To further accelerate the process, RaftFed employs a "warm-start" initial model trained on non-sensitive public data.

Experimental Evaluation

The researchers evaluated RaftFed using modified MNIST and CIFAR-10 datasets to simulate the messy, non-uniform data found in real-world vehicular networks.

  • Superior Accuracy: On the MNIST dataset, RaftFed achieved 98.30% accuracy within 100 rounds, outperforming the FedCluster baseline (93.86%).
  • Faster Convergence: On the more complex CIFAR-10 task using a ResNet18 architecture, RaftFed reached 75.83% accuracy, roughly 9% higher than standard FedVANET after 500 rounds.
  • Efficiency: Theoretical analysis and experimental results confirm that RaftFed significantly reduces communication overhead by optimizing synchronization and removing the need for constant long-distance cloud communication.

Conclusion and Future Work

RaftFed represents a significant step toward practical, privacy-preserving vehicular intelligence. By decentralizing aggregation and utilizing dynamic clustering, it overcomes the latency and data heterogeneity issues inherent in traditional FL. Future research will focus on enhancing the framework's security through Trusted Execution Environments (TEEs) and improving robustness against adversarial poisoning attacks in high-mobility scenarios.

Summary updated

Abstract

Vehicular crowd intelligence (VCI) is an emerging research field. Facilitated by state-of-the-art vehicular ad-hoc networks and artificial intelligence, various VCI applications come to place, e.g., collaborative sensing, positioning, and mapping. The collaborative property of VCI applications generally requires data to be shared among participants, thus forming network-wide intelligence. How to fulfill this process without compromising data privacy remains a challenging issue. Although federated learning (FL) is a promising tool to solve the problem, adapting conventional FL frameworks to VCI is nontrivial. First, the centralized model aggregation is unreliable in VCI because of the existence of stragglers with unfavorable channel conditions. Second, existing FL schemes are vulnerable to Non-IID data, which is intensified by the data heterogeneity in VCI. This paper proposes a novel federated learning framework called RaftFed to facilitate privacy-preserving VCI. The experimental results show that RaftFed performs better than baselines regarding communication overhead, model accuracy, and model convergence.
BibTeX
@misc{yang2023raftfed,
  author = {Changan Yang and Yaxing Chen and Yao Zhang and Helei Cui and Zhiwen Yu and Bin Guo and Zheng Yan and Zijiang Yang},
  title = {{RaftFed: A Lightweight Federated Learning Framework for Vehicular Crowd Intelligence}},
  year = {2023},
  eprint = {2310.07268},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.07268},
}
Figure