Federated Learning with Heterogeneous Data Handling for Robust Vehicular Object Detection

TL;DR

This paper proposes FedProx+LA, an advanced federated learning method tailored for vehicular networks, which improves convergence speed and detection accuracy in online autonomous driving object detection tasks, especially under heterogeneous data conditions.

Contribution

It introduces FedProx+LA, combining FedProx and FedLA, to effectively handle data heterogeneity in federated learning for vehicular object detection, outperforming previous methods.

Findings

FedProx+LA converges 30% faster than baseline methods.

It significantly improves detection performance with heterogeneous label distributions.

The method outperforms previous FedLA in convergence and accuracy.

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.