Sun, Zhanbo (Southwest Jiaotong University), Du, Yueyuan (Southwest Jiaotong University), Wang, Feilong (Southwest Jiaotong University), Zhao, Rong (Dongguan University of Technology)

Exploiting Adjacent Vehicle Interactions to Mislead Trajectory Predictions

Scheduled for presentation during the Invited Session "S23b-Trustworthy AI for Traffic Sensing and Control" (TH-EA-T23), Thursday, November 20, 2025, 14:50−15:30, Coolangata 2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Graph Neural Networks (GNNs) have proven effective in vehicle trajectory prediction by capturing multi-agent interactions. While recent studies have revealed that GNN models could be compromised by exploiting these interactions, their robustness under adversarial perturbations in the context of trajectory prediction remains under-explored. This study assesses the vulnerability of GNN-based trajectory prediction models via a spatiotemporal adversarial attack framework. We introduce a gradient-guided, time-dependent method to identify the most influential adjacent vehicle, followed by a spatiotemporal projected gradient descent approach to generate effective and practical perturbations that are injected into the identified vehicle to compromise the ego vehicle. Experiments conducted on a real-world dataset show that the proposed attack framework significantly degrades prediction accuracy, increasing the prediction deviation by 51.92%. It is found that the framework produces larger disruptions than baseline attack strategies. Interestingly, while an advanced prediction model like Graph Attention Network (GAT) can improve the prediction accuracy over the baseline GNN, our findings reveal that it also exhibits increased vulnerability. These findings highlight the critical trade-off between model accuracy and robustness, with implications for improving the safety of intelligent driving systems enhanced by adjacent vehicle interactions.