DATZ-TM: Zero-Trust intrusion detection for VANETs with temporal graph neural networks

Vehicular Ad-Hoc Networks (VANETs) are vital for intelligent transport systems, yet they are extremely vulnerable to cyberattacks like Sybil, Blackhole, Denial-of-Service, Position Falsification, Message Suppression, and Replay. Existing intrusion detection systems use rigid metrics or constrained machine learning, resulting in unreliable detection across different threats. Critically, no previous work combines Zero-Trust security with dynamic traffic management in one vehicle framework. We introduce DATZ-TM (Dynamic Adaptive Zero-Trust Traffic Management), which consists of four integrated components: a calibrated behavioural Trust Scorer with temporal decay, a behavioural Anomaly Detector that couples a learned detection model with a Temporal Graph Neural Network (T-GNN), a Zero-Trust Enforcement layer, and a DQN + LLM traffic agent (Llama-3.3-70B). We evaluate the framework against ten baselines, spanning deep graph (GCN, GAT), temporal (LSTM, TCN), graph-anomaly (Graph-AE) and CNN-LSTM detectors, in two complementary settings: a controlled synthetic benchmark and real SUMO-collected traces of the Luxembourg road network (261 vehicles, 133,773 records), both under an observable-feature-only protocol with per-baseline threshold calibration for fairness and all results across three random seeds. On the synthetic benchmark DATZ-TM attains the best overall performance, leading every baseline on detection rate (86.0%), F1-score (88.5%), AUC (0.955), Matthews Correlation Coefficient (0.788), balanced accuracy (89.2%) and G-mean (89.1%) at a low 7.7% false-positive rate, surpassing the strongest baseline (Random Forest: F1 87.3%) on every aggregate metric. On the harder real-SUMO traces DATZ-TM remains the top detector on the detection-oriented metrics (F1 70.6% vs. Random Forest 69.1%; detection rate 60.7% vs. 56.9%; best balanced accuracy and G-mean) and is competitive on the precision-oriented metrics, while the deep graph baselines degrade sharply (GCN/GAT F1 51/42%). Beyond i.i.d. detection, the temporal model is necessary for coordinated temporal attacks (pulsing/on-off, regime-shift, low-and-slow), where it raises detection by +56.9 percentage points over a no-neural variant (e.g. pulsing attacks 21.6% $$\rightarrow$$ 100%) at a matched 5% false-positive rate. On real SUMO/TraCI simulation the DQN controller outperforms fixed-time and actuated control and is competitive with a max-pressure reference, and the LLM audit layer produces reliable, format-compliant records at $${\approx }0.4$$ s latency.