VeReMi-Graph v1.0: A Temporal Attributed Graph Dataset for Vehicular Misbehavior Detection
Description
VeReMi-Graph v1.0 is a standardized temporal attributed graph dataset derived from the VeReMi vehicular misbehavior dataset for research on graph-based vehicular misbehavior detection and related machine-learning tasks. The resource transforms the original message- and vehicle-level VeReMi records into a sequence of directed temporal attributed graphs using fixed one-second observation windows. Vehicles are represented as graph nodes, while received vehicular communication messages are represented as directed message-level edges between transmitting and receiving vehicles. The complete release contains 225 simulation archives, 22,500 temporal graph snapshots, 2,016,487 node instances, and 39,136,291 message-level edge instances. A frozen learning schema provides 26 node features, 23 edge features, and 13 preceding-snapshot graph-context features. Binary and multiclass edge-level attack labels are supplied for supervised learning. 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.
Files
Steps to reproduce
VeReMi-Graph v1.0 is distributed in canonical Parquet and pre-materialized PyTorch Geometric (PyG) representations. Users may reproduce the released graph objects using the accompanying schemas, manifests, and demonstration package. First, extract the metadata/reproducibility package and consult README.md and the ordered schema files. The canonical node, edge, and graph-context tables are organized by original VeReMi simulation archive. Graph snapshots correspond to fixed one-second observation windows. Use the supplied archive-disjoint split manifests to preserve the predefined training, validation, and test partitions. The frozen learning representation consists of 26 node features, 23 edge features, and 13 lagged graph-context features in the exact order specified by the distributed schema files. Missing historical information is represented using the accompanying node, edge, and graph-context masks. Pre-materialized PyTorch Geometric objects are supplied separately for the training, validation, and test partitions and can be loaded directly with PyTorch/PyTorch Geometric. SHA-256 manifests are provided to verify package integrity. The demonstration notebook/script included in the reproducibility package provides an example of loading and validating the released data structures.