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.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to improve the robustness and convergence speed of the object detection model training in federated learning (FL) in vehicle networks under the heterogeneous data environment. Specifically: 1. **Heterogeneous data problem**: In federated learning, when the data distribution of clients is inconsistent (i.e., non - independent and identically distributed, non - IID), it will lead to a slower convergence speed of model training and a decline in performance. Especially in vehicle networks, the data collected by different vehicles may have significant differences, such as sensor data in different locations, times or environmental conditions. 2. **Limitations of existing methods**: Traditional federated learning methods such as FedAvg perform poorly when dealing with non - IID data, and although the existing improved methods have been improved in some aspects, they still have deficiencies in vehicle perception scenarios. To solve these problems, the author proposes a new method named **FedProx + LA**, which combines the advantages of previously proposed FedLA and FedProx. FedProx + LA copes with data heterogeneity by introducing label - aware aggregation and proximal term, thereby achieving better performance and faster convergence speed in vehicle object detection tasks. ### Main contributions: - **Propose FedProx + LA**: A new method that combines the advantages of FedProx and FedLA, specifically targeting the heterogeneous data problem in vehicle networks. - **Experimental verification**: Through comparative experiments with traditional methods (such as FedAvg, FedProx, etc.), it is proved that FedProx + LA has superior performance under non - IID data distribution. - **Performance improvement**: The experimental results on the NuScenes dataset show that FedProx + LA improves the detection performance (mAP) by 6% compared with the baseline method and increases the convergence speed by 30%. ### Summary: The core objective of this paper is to solve the challenges brought by data heterogeneity in vehicle networks through improving the federated learning algorithm, thereby achieving more efficient and accurate object detection model training.

Federated Learning with Heterogeneous Data Handling for Robust Vehicular Object Detection

A Distributed Aggregation Approach for Vehicular Federated Learning

FedVisual: Heterogeneity-Aware Model Aggregation for Federated Learning in Visual-Based Vehicular Crowdsensing

FedPylot: Navigating Federated Learning for Real-Time Object Detection in Internet of Vehicles

Federated Deep Learning Meets Autonomous Vehicle Perception: Design and Verification

On Addressing Heterogeneity in Federated Learning for Autonomous Vehicles Connected to a Drone Orchestrator

LFGurad: A defense against label flipping attack in federated learning for vehicular network

FedRAV: Hierarchically Federated Region-Learning for Traffic Object Classification of Autonomous Vehicles

Clustered Vehicular Federated Learning: Process and Optimization

Mobility, Communication and Computation Aware Federated Learning for Internet of Vehicles

Mobility-Aware Federated Self-supervised Learning in Vehicular Network

Navigating Data Heterogeneity in Federated Learning A Semi-Supervised Federated Object Detection

Secure Hierarchical Federated Learning in Vehicular Networks Using Dynamic Client Selection and Anomaly Detection

Federated learning system on autonomous vehicles for lane segmentation

Fast-Convergent and Communication-Alleviated Heterogeneous Hierarchical Federated Learning in Autonomous Driving

Distributed Dynamic Map Fusion via Federated Learning for Intelligent Networked Vehicles

Hierarchical Federated Learning: Architecture, Challenges, and Its Implementation in Vehicular Networks

Federated Learning for Misbehaviour Detection with Variational Autoencoders and Gaussian Mixture Models

Communication and Computation Efficient Federated Learning for Internet of Vehicles with a Constrained Latency

Adaptive and Parallel Split Federated Learning in Vehicular Edge Computing

Enhancing Federated Learning in Heterogeneous Internet of Vehicles: A Collaborative Training Approach