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

Federated learning enables multiple autonomous vehicles (AVs) to collaboratively train machine learning models while preserving data privacy. However, performance degrades significantly under non-independent and identically distributed (non-IID) data conditions commonly encountered in real-world driving scenarios. Existing aggregation methods, particularly Federated Averaging (FedAvg), struggle to effectively handle client update divergence, leading to inefficient communication, unstable convergence, and increased privacy risks.

To address these challenges, we propose a Dynamic Variance-Aware Federated Tuning (DV-FedTune) framework for object detection in autonomous driving systems using YOLOv12. The proposed framework dynamically adjusts client contributions through a variance-aware aggregation strategy that jointly models update consistency, variance-based diversity, and loss-guided reliability using a round-adaptive weighting mechanism.

Comprehensive experiments were conducted on the KITTI object detection dataset under various non-IID federated learning configurations involving different numbers of clients, local training durations, and communication rounds. The results demonstrate that DV-FedTune consistently outperforms FedAvg, Exponential Moving Average in Federated Learning (EWHFed), and VINOEffiFedAV in terms of communication efficiency, computational cost, and model performance while maintaining stronger privacy preservation.

The proposed framework achieves stable aggregation behavior and effective parameter utilization as the federated network scales. These findings indicate that DV-FedTune provides an efficient and privacy-preserving federated learning solution for distributed object detection in autonomous vehicle environments operating under heterogeneous data distributions.