CARLA-V2X-MBD: A Perception-Aided V2X Misbehavior Detection Dataset (v1.0.0)

CARLA-V2X-MBD is, to our knowledge, the first open V2X misbehavior detection dataset generated in a 3D driving simulator (CARLA 0.9.16) using the simulator’s native ETSI CAM (Cooperative Awareness Message) stack. Unlike network-simulator-based datasets (VeReMi, VeReMi Extension/NextGen, MisbehaviorX), every reception event carries the receiver’s own pose and a perception ground-truth summary of its neighborhood, enabling perception-aided and cooperative plausibility checks. Contents: 72 scenarios (urban Town10 + highway Town04; 20–60 vehicles; 5–20% attackers; 2 attack mixes; 2 seeds; 60 s each), 35,336,104 labeled CAM reception events. Five attack types in VeReMi taxonomy naming: ConstPos, ConstSpeedOffset, Replay (DataReplay), Sybil (ghost vehicles), DoS (transmission-level flooding). Leakage-free configuration-level train/val/test splits and a VeReMi field mapping are included in manifest.json. Files: six data archives grouped by town × vehicle count (.tar.zst; per-run cam_log.jsonl.gz, cam_log.parquet, metadata.json), a 24 MB sample run containing all five attack types, the full generation/feature/training pipeline (code archive, MIT licensed), a datasheet (DATASHEET.md), and SHA256SUMS for integrity verification. Reproduction: the code archive regenerates or extends the dataset from a single campaign config (requires CARLA 0.9.16) and includes RandomForest/XGBoost and sequence-transformer baselines with a local-vs-cooperative perception cross-check ablation.