Tsukada Lab. Computer Network Group
← Publications Int'l conference

Overcoming Environmental Challenges in CAVs Through MEC-Based Federated Learning

Zekun Wang, Jin Nakazato, Muhammad Asad, Ehsan Javanmardi, Manabu Tsukada

14th International Conference on Ubiquitous and Future Networks (ICUFN 2023), Paris, France, pp. 1–6, Jul 4, 2023

Summary in three lines

  • Variable weather and operating conditions can affect models used by connected automated vehicles.
  • An edge-based federated learning framework caches and updates models for different conditions.
  • Edge nodes aggregate local learning and the proposed algorithms categorize and manage specialized models.

Figures from the paper

Figure 1: Federated learning across areas with different conditions
Figure 1. Federated learning across areas with different conditionsSource: Figure 1 from Wang et al., IEEE ICUFN 2023 (publisher version, © 2023 IEEE)The example shows edge nodes serving regions with different weather and road scenes.
Figure 5: General and dedicated models by condition
Figure 5. General and dedicated models by conditionSource: Figure 5 from Wang et al., IEEE ICUFN 2023 (publisher version, © 2023 IEEE)The results compare the pretrained general model with dedicated models across condition classes.

Abstract

Connected autonomous vehicles (CAVs), through vehicle-to-everything communication and computing resources, enable the vital exchange of information. Although deep learning is crucial in this landscape, it requires extensive and intricate datasets covering all potential scenarios. Furthermore, this situation poses a hazard, as the likelihood of accidents associated with imbalanced datasets increases, particularly in scenarios where processing analysis is compromised due to fluctuating weather conditions. We propose a Federated Learning (FL) framework undergirded by Multi-Access Edge Computing (MEC) to counter these challenges. This local device-focused framework enhances task-specific models' caching and continual updating across various conditions. In a more specific sense, edge nodes (ENs) operate as MEC, each caching multiple dedicated models and serving as the aggregator as part of the FL process. Additionally, we have engineered two innovative algorithms that categorize various states into multiple classes, thereby ensuring the efficient utilization of computing resources in ENs. Simulation results substantiate the effectiveness of our approach, showing that the proposed dedicated model consistently outperforms a general model designed for all situations.

Related research and software

How to cite

Z. Wang, J. Nakazato, M. Asad, E. Javanmardi, and M. Tsukada, “Overcoming Environmental Challenges in CAVs Through MEC-Based Federated Learning,” in 14th International Conference on Ubiquitous and Future Networks (ICUFN 2023), Paris, France, 2023, doi: 10.1109/ICUFN57995.2023.10200688.

BibTeX

@inproceedings{Wang2023,
  title         = {Overcoming Environmental Challenges in CAVs Through MEC-Based Federated Learning},
  author        = {Zekun Wang and Jin Nakazato and Muhammad Asad and Ehsan Javanmardi and Manabu Tsukada},
  url           = {https://www.researchgate.net/profile/Jin-Nakazato/publication/371685830_Overcoming_Environmental_Challenges_in_CAVs_through_MEC-based_Federated_Learning/links/64a420ea8de7ed28ba7465c7/Overcoming-Environmental-Challenges-in-CAVs-through-MEC-based-Federated-Learning.pdf},
  doi           = {10.1109/ICUFN57995.2023.10200688},
  year          = {2023},
  date          = {2023-07-04},
  urldate       = {2023-07-04},
  booktitle     = {14th International Conference on Ubiquitous and Future Networks (ICUFN 2023)},
  pages         = {1--6},
  address       = {Paris, France},
  abstract      = {Connected autonomous vehicles (CAVs), through vehicle-to-everything communication and computing resources, enable the vital exchange of information. Although deep learning is crucial in this landscape, it requires extensive and intricate datasets covering all potential scenarios. Furthermore, this situation poses a hazard, as the likelihood of accidents associated with imbalanced datasets increases, particularly in scenarios where processing analysis is compromised due to fluctuating weather conditions. We propose a Federated Learning (FL) framework undergirded by Multi-Access Edge Computing (MEC) to counter these challenges. This local device-focused framework enhances task-specific models' caching and continual updating across various conditions. In a more specific sense, edge nodes (ENs) operate as MEC, each caching multiple dedicated models and serving as the aggregator as part of the FL process. Additionally, we have engineered two innovative algorithms that categorize various states into multiple classes, thereby ensuring the efficient utilization of computing resources in ENs. Simulation results substantiate the effectiveness of our approach, showing that the proposed dedicated model consistently outperforms a general model designed for all situations.},
  keywords      = {ITS},
  pubstate      = {published},
  tppubtype     = {inproceedings}
}