Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Federated Learning with Heterogeneous Data Handling for Robust Vehicular Object Detection

Ahmad Khalil, Tizian Dege, Pegah Golchin, Rostyslav Olshevskyi, Antonio Fernandez Anta, Tobias Meuser

TL;DR

This work addresses non-IID data challenges in federated learning for vehicular object detection by introducing FedProx+LA, a hybrid that combines proximal updates with label-aware aggregation. Building on FedLA and FedProx, the method leverages per-label statistics and proximal regularization to mitigate heterogeneity in online, distributed training. Evaluations on NuScenes show FedProx+LA achieves up to 6% higher detection accuracy and around 30% faster convergence than baselines, with FedLA also delivering notable gains under label skew. The results demonstrate improved robustness and communication efficiency for autonomous-driving perception in heterogeneous vehicular networks, and the accompanying code is publicly available.

Abstract

In the pursuit of refining precise perception models for fully autonomous driving, continual online model training becomes essential. Federated Learning (FL) within vehicular networks offers an efficient mechanism for model training while preserving raw sensory data integrity. Yet, FL struggles with non-identically distributed data (e.g., quantity skew), leading to suboptimal convergence rates during model training. In previous work, we introduced FedLA, an innovative Label-Aware aggregation method addressing data heterogeneity in FL for generic scenarios. In this paper, we introduce FedProx+LA, a novel FL method building upon the state-of-the-art FedProx and FedLA to tackle data heterogeneity, which is specifically tailored for vehicular networks. We evaluate the efficacy of FedProx+LA in continuous online object detection model training. Through a comparative analysis against conventional and state-of-the-art methods, our findings reveal the superior convergence rate of FedProx+LA. Notably, if the label distribution is very heterogeneous, our FedProx+LA approach shows substantial improvements in detection performance compared to baseline methods, also outperforming our previous FedLA approach. Moreover, both FedLA and FedProx+LA increase convergence speed by 30% compared to baseline methods.

Federated Learning with Heterogeneous Data Handling for Robust Vehicular Object Detection

TL;DR

This work addresses non-IID data challenges in federated learning for vehicular object detection by introducing FedProx+LA, a hybrid that combines proximal updates with label-aware aggregation. Building on FedLA and FedProx, the method leverages per-label statistics and proximal regularization to mitigate heterogeneity in online, distributed training. Evaluations on NuScenes show FedProx+LA achieves up to 6% higher detection accuracy and around 30% faster convergence than baselines, with FedLA also delivering notable gains under label skew. The results demonstrate improved robustness and communication efficiency for autonomous-driving perception in heterogeneous vehicular networks, and the accompanying code is publicly available.

Abstract

In the pursuit of refining precise perception models for fully autonomous driving, continual online model training becomes essential. Federated Learning (FL) within vehicular networks offers an efficient mechanism for model training while preserving raw sensory data integrity. Yet, FL struggles with non-identically distributed data (e.g., quantity skew), leading to suboptimal convergence rates during model training. In previous work, we introduced FedLA, an innovative Label-Aware aggregation method addressing data heterogeneity in FL for generic scenarios. In this paper, we introduce FedProx+LA, a novel FL method building upon the state-of-the-art FedProx and FedLA to tackle data heterogeneity, which is specifically tailored for vehicular networks. We evaluate the efficacy of FedProx+LA in continuous online object detection model training. Through a comparative analysis against conventional and state-of-the-art methods, our findings reveal the superior convergence rate of FedProx+LA. Notably, if the label distribution is very heterogeneous, our FedProx+LA approach shows substantial improvements in detection performance compared to baseline methods, also outperforming our previous FedLA approach. Moreover, both FedLA and FedProx+LA increase convergence speed by 30% compared to baseline methods.
Paper Structure (15 sections, 5 figures, 1 table, 1 algorithm)

This paper contains 15 sections, 5 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Federated Label-Aware Averaging in Vehicular Networks
  4. FedLA
  5. FedProx+LA
  6. Evaluation
  7. Experimental Setups
  8. Data Set and Data Distribution
  9. Implementation Details
  10. Baselines
  11. Evaluation Metrics
  12. Results and Discussion
  13. Object Detection Performance
  14. Communication Efficiency and Convergence Rate
  15. Conclusion

Figures (5)

  • Figure 1: Online training of the general perception model using FL khalil2022situational. Each communication round involves (1) transmission of local updates to the edge server (e.g., ), where intermediate aggregation occurs. (2) The edge servers transfer the aggregated model to the central aggregation server for final aggregation. (3) and (4) The aggregated model is subsequently transferred back to individual clients (vehicles).
  • Figure 2: FedProx+LA method: At each global epoch (communication round), every client (vehicle) incorporates the proximal term $\mathcal{L} = \mathcal{L}_{ce} + \mu\mathcal{L}_{prox}$ into its local update and transmits it alongside the label distribution statistics. Upon receiving the model updates along with the label distribution statistics, the aggregation server considers them to generate the aggregated global model.
  • Figure 3: Different data distribution schemes are illustrated with Pedestrian (P) and Car (C) class labels. (a) depicts data distribution across clients. (b) showcases data distribution using the Dirichlet distribution. Both distributions ensure equal sample sizes across clients.
  • Figure 4: Comparison of ( $\geq$ 0.5) across various methods in the setting.
  • Figure 5: Comparison of convergence rates in the setting: FedAvg and FedAvgL fail to reach the target of 0.55 within global epochs $E_g = 50$. While FedProx+LA and FedLA achieve a considerable convergence rate.