ST-GAT: A Spatio-Temporal Graph Attention Network For Sybil Attack Detection in VANETs

Abstract

Sybil attacks are a major security threat in Vehicular Ad-Hoc Networks (VANETs). In this attack, a malicious vehicle creates multiple fake identities to mislead the network. Existing machine learning approaches analyze vehicles in isolation, failing to capture the spatial coordination characteristic of Sybil behavior. To address this, ST-GAT, a Spatio-Temporal Graph Attention Network, is proposed that models VANET snapshots as dynamic graphs for robust node-level classification. The proposed framework introduces adaptive kinematic edge-weighting, a physics-aware node feature set capturing neighbor-aggregate context, and a dual GATv2-GRU encoder trained with a composite Focal+Dice+WCE loss to handle severe class imbalance. Evaluated on the VeReMi dataset using a rigorous chronological split, ST-GAT achieves state-of-the-art performance, including 0.999 accuracy, 1.000 precision, and a 0.995 F1-score, effectively eliminating false positives. These results confirm that explicit graph-based neighborhood modeling provides highly superior, practically deployable Sybil detection for connected vehicles.