GlobeCom High-Quality Trajectory Generation for Autonomous Driving: A Lightweight Federated Learning-Based Diffusion Model

Abstract

Vehicle trajectory data plays a pivotal role in simulation testing for autonomous driving. Hence, there exist well-established trajectory generation methods employing deep generative models to generate trajectories mapping the distribution of the original dataset, thereby augmenting existing trajectory datasets. However, these methods typically rely on large datasets gathered by governmental or organizational entities for central training, which may pose data privacy, security, and accessibility issues. Therefore, it is challenging to generate high-quality traffic trajectory data while preserving privacy which involves a delicate balance between these two objectives. To deal with this challenge, we introduce Federated Learning into the diffusion model and propose a Federated Learning-based diffusion model (FedDifftraj) to generate traffic trajectory data. Unlike existing central training methods, FedDifftraj aggregates model parameters uploaded by different vehicles and then updates a global model. Additionally, there is a substantial communication overhead incurred during the training of the federated diffusion model. Therefore, we quantize the local diffusion model before uploading it to the parameter server. Through extensive simulations on real-world datasets, FedDifftraj can generate high-quality traffic trajectory data that is consistent with the results of the central training while preserving privacy and reducing communication overhead by 93.74% when utilizing 8-bit quantization.

Source: IEEE