VeReMi-Graph v1.0: A Temporal Attributed Graph Dataset for Vehicular Misbehavior Detection

To support leakage-aware temporal learning, topology-dependent node attributes and graph-context variables used in the primary learning representation are derived from the preceding graph snapshot. Historical availability is explicitly represented through node-, edge-, and graph-context masks rather than silently treating unavailable historical values as observations. The dataset is distributed in two complementary representations: canonical Parquet tables for general data analysis and pre-materialized PyTorch Geometric (PyG) graph objects for direct use in graph-learning pipelines.

Fixed archive-disjoint benchmark partitions are provided, comprising 157 training archives, 34 validation archives, and 34 test archives, corresponding to 15,700, 3,400, and 3,400 graph snapshots, respectively. The release additionally includes ordered feature schemas, a complete feature dictionary, archive and snapshot manifests, benchmark split definitions, provenance metadata, validation reports, SHA-256 checksum manifests, and a demonstration/reproducibility package.

VeReMi-Graph does not introduce a new vehicular simulation or new attack scenarios. It is a derived graph-oriented representation of VeReMi intended to make the underlying vehicular communication data directly reusable for temporal graph analysis, graph machine learning, and reproducible misbehavior-detection research. Users should cite both this dataset and the original VeReMi dataset/publication when using VeReMi-Graph.