Dynamic variance-aware federated tuning for efficient autonomous vehicle perception under non-IID settings.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Federated learning (FL) enables collaborative model training for autonomous vehicles (AVs) while protecting data privacy.
• Non-independent and identically distributed (non-IID) data in real-world driving degrades FL performance, challenging methods like Federated Averaging (FedAvg) due to client update divergence.

Purpose of the Study:

• To introduce a novel federated learning framework, Dynamic Variance-Aware Federated Tuning (DV-FedTune), for robust object detection in AVs.
• To address the limitations of existing FL aggregation methods in handling non-IID data and client divergence.

Main Methods:

• DV-FedTune employs a variance-aware aggregation strategy for dynamic client contribution adjustment.
• The framework integrates update consistency, diversity, and loss-guided reliability with a round-adaptive weighting mechanism.
• Object detection is performed using the YOLOv12 model within the DV-FedTune framework.

Main Results:

• Experiments on the KITTI dataset under various non-IID settings show DV-FedTune outperforms FedAvg, EWHFed, and VINOEffiFedAV.
• DV-FedTune demonstrates superior communication efficiency, reduced computational cost, and enhanced model performance.
• The framework maintains stronger privacy preservation compared to baseline methods.

Conclusions:

• DV-FedTune offers stable aggregation and effective parameter utilization in scaled federated networks.
• The proposed framework provides an efficient and privacy-preserving solution for distributed object detection in AVs under heterogeneous data.
• DV-FedTune is a promising approach for real-world autonomous driving applications with non-IID data.