Privacy-preserving and Uncertainty-aware Federated Trajectory Prediction for Connected Autonomous Vehicles

Muzi Peng, Jiangwei Wang, Dongjin Song, Fei Miao, Lili Su

Abstract

Deep learning is the method of choice for trajectory prediction for autonomous vehicles. Unfortunately, its data-hungry nature implicitly requires the availability of sufficiently rich and high-quality centralized datasets, which easily leads to privacy leakage. Besides, uncertainty-awareness becomes increasingly important for safety-crucial cyber physical systems whose prediction module heavily relies on machine learning tools. In this paper, we relax the data collection requirement and enhance uncertainty-awareness by using Federated Learning on Connected Autonomous Vehicles with an uncertainty-aware global objective. We name our algorithm as FLTP. We further introduce ALFLTP which boosts FLTP via using active learning techniques in adaptatively selecting participating clients. We consider both negative log-likelihood (NLL) and aleatoric uncertainty (AU) as client selection metrics. Experiments on Argoverse dataset show that FLTP significantly outperforms the model trained on local data. In addition, ALFLTP-AU converges faster in training regression loss and performs better in terms of NLL, minADE and MR than FLTP in most rounds, and has more stable round-wise performance than ALFLTP-NLL.

Introduction

Accurate trajectory prediction—the ability to forecast the future positions of neighboring vehicles and pedestrians—is a cornerstone of autonomous vehicle (AV) safety. Modern deep learning models, such as transformers, have significantly advanced this field but remain inherently "data-hungry." Traditionally, developing these models required aggregating massive amounts of raw driving logs, including GPS, video, and LIDAR data, into a centralized server. This centralized approach poses severe privacy risks, as driving logs can inadvertently reveal sensitive information such as home addresses, daily routines, and vehicle identities.

Furthermore, in safety-critical driving environments, a model's point prediction is insufficient; it must also quantify its own uncertainty. Most existing trajectory prediction frameworks fail to integrate uncertainty-awareness into the training process itself. This paper introduces FLTP (Federated Learning based Trajectory Prediction) and ALFLTP (Active Learning-boosted FLTP) to address these dual challenges of privacy and uncertainty.

Methodology

FLTP: Federated Learning Framework

The FLTP framework enables vehicles (clients) to collaboratively train a global trajectory prediction model without sharing raw sensor data. Each client uses the Hierarchical Vector Transformer (HiVT) as its local model backbone. During training, clients perform local updates using the AdamW optimizer and transmit only the resulting model weights to a central server. The server aggregates these weights to update the global model, which is then redistributed to the clients. This decentralized approach ensures that raw data never leaves the vehicle, maintaining user privacy.

ALFLTP: Active Learning and Uncertainty-Awareness

Standard Federated Learning often selects participating clients at random, which can be inefficient. ALFLTP improves upon this by using Active Learning to adaptively select the most informative clients in each round. The selection is based on two primary uncertainty metrics:

  1. Negative Log-Likelihood (NLL): This metric identifies clients where the current model performs poorly, effectively focusing the global model on "hard" cases or edge scenarios.
  2. Aleatoric Uncertainty (AU): This measures the inherent noise in the sensor data. To capture this, the model outputs a Laplace mixture distribution, providing both a predicted coordinate and a "scale" parameter representing the model's confidence.

Experimental Setup

The researchers evaluated the proposed frameworks using the Argoverse Motion Forecasting v1.1 dataset. To simulate a realistic distributed network of Connected Autonomous Vehicles (CAVs), the data was partitioned among 100 clients based on city labels (Pittsburgh and Miami). This created a "non-IID" (non-independent and identically distributed) environment, reflecting the diverse traffic patterns different vehicles encounter in the real world. The task involved predicting 3 seconds of future movement based on 2 seconds of historical trajectory data.

Key Insights and Results

The "Median" Selection Strategy

A significant finding of this research is the effectiveness of the "median" strategy for Aleatoric Uncertainty. The authors discovered that selecting clients with median uncertainty levels yielded the best results. Clients with very low uncertainty provide "easy" data that the model has already mastered, while those with very high uncertainty often contribute noisy or unlearnable data. The median represents the most representative and informative data for improving the global model.

Performance and Convergence

  • Superiority over Local Training: FLTP significantly outperformed standalone models trained only on local data. Local models tended to overfit quickly and failed to generalize, whereas the federated approach benefited from the collective knowledge of the network.
  • Efficiency of ALFLTP-AU: The variant using Aleatoric Uncertainty (ALFLTP-AU) demonstrated faster convergence in regression loss compared to standard FLTP. It also exhibited greater stability across training rounds than the NLL-based selection method.
  • Competitive Accuracy: While centralized training still provides a performance ceiling, the FLTP models achieved results remarkably close to the state-of-the-art. At 250 rounds, ALFLTP-AU achieved a minADE (minimum Average Displacement Error) of 0.729, compared to 0.685 for the centralized HiVT model, while providing robust privacy guarantees.

Conclusion

This work demonstrates that Federated Learning is a viable and effective solution for trajectory prediction in autonomous driving. By integrating uncertainty-aware active learning, the ALFLTP framework not only preserves data privacy but also improves training efficiency and model robustness. Future research will focus on reducing the communication overhead associated with weight transmission and exploring the integration of epistemic uncertainty to further refine client selection strategies in intermittent connectivity environments.

Summary updated

Abstract

Deep learning is the method of choice for trajectory prediction for autonomous vehicles. Unfortunately, its data-hungry nature implicitly requires the availability of sufficiently rich and high-quality centralized datasets, which easily leads to privacy leakage. Besides, uncertainty-awareness becomes increasingly important for safety-crucial cyber physical systems whose prediction module heavily relies on machine learning tools. In this paper, we relax the data collection requirement and enhance uncertainty-awareness by using Federated Learning on Connected Autonomous Vehicles with an uncertainty-aware global objective. We name our algorithm as FLTP. We further introduce ALFLTP which boosts FLTP via using active learning techniques in adaptatively selecting participating clients. We consider both negative log-likelihood (NLL) and aleatoric uncertainty (AU) as client selection metrics. Experiments on Argoverse dataset show that FLTP significantly outperforms the model trained on local data. In addition, ALFLTP-AU converges faster in training regression loss and performs better in terms of NLL, minADE and MR than FLTP in most rounds, and has more stable round-wise performance than ALFLTP-NLL.
BibTeX
@inproceedings{peng2023privacy,
  author = {Muzi Peng and Jiangwei Wang and Dongjin Song and Fei Miao and Lili Su},
  title = {{Privacy-preserving and Uncertainty-aware Federated Trajectory Prediction for Connected Autonomous Vehicles}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  doi = {10.1109/iros55552.2023.10341638},
}
Figure