SecureVAN: Deep Learning-Driven Anomaly and Intrusion Detection in Vehicular Ad-hoc Networks

Vehicular Ad-hoc Networks (VANETs) support safety-critical applications such as collision avoidance and cooperative traffic management, yet their open wireless medium and highly dynamic topology expose them to spoofing, Sybil, black-hole, denial-of-service, and position-falsification attacks that conventional signature-based intrusion detection systems struggle to catch in real time. This paper proposes SecureVAN, a deep learning-driven framework that fuses a Convolutional Neural Network (CNN) for spatial feature extraction with a Long Short-Term Memory (LSTM) network for temporal sequence modelling to detect anomalous and malicious behaviour from Basic Safety Messages (BSMs) and CAN-bus traffic. The hybrid model is trained and evaluated on the VeReMi dataset augmented with SUMO/NS-3 simulated attack traces covering five major misbehaviour classes. Experimental results show that SecureVAN achieves 97.8% detection accuracy and a 97.3% F1-score, outperforming SVM, Random Forest, plain CNN, and plain LSTM baselines, while reducing the false-positive rate to 1.9% and sustaining sub-50-millisecond inference latency suitable for on-board deployment. These findings indicate that combining spatial and temporal deep representations offers a practical and scalable route to real-time intrusion and anomaly detection in next-generation connected-vehicle networks.