Abstract
As electric vehicles (EVs) and Intelligent Transportation Systems (ITS) become more connected through IoTbased vehicular components and V2X communication, ensuring robust cybersecurity against adversarial attacks is critical for the safety and reliability of transportation networks. Vehicle networks are vital to smart cities, but they face increasing security threats, especially adversarial attacks. This work demonstrates a hybrid deep learning model, Inception + Skip Connection + BiGRU + Attention, for detecting such threats in IoT-based VANET datasets. The novel architecture presents a compact and efficient approach to detect misbehavior by employing an integrated approach that consists of multi-scale feature extraction, temporal sequence learning, and attention-based refinement. Global Average Pooling (GAP) enhances efficiency while preserving key features. Novelty of the proposed model lies in detecting diverse patterns from misbehavior dataset and overcoming the single model approach. Experiments on binary classification across five attack types, the model achieves F1-Scores up to 99.93% (Type-4) and maintains strong performance on others like Type-1 (97.25%) and Type-8 (98.74%). In the multi-class scenario, the model achieves 91.21% accuracy and a weighted F1 score of 90.3%. However, the recall for the type-16 attack is substantially lower, likely due to the subtle nature of this adversarial attack.