Class-aware oversampling for robust intrusion detection in VANETs using deep learning

Vehicular Ad Hoc Networks (VANETs) Intrusion Detection is essential in ensuring road usage and safe communication. Intrusion Detection Systems based on Artificial Intelligence in VANETs have demonstrated astounding advancements, reporting their growing abilities in enhancing network security and reliability. However, class imbalance remains a critical problem in developing accurate IDS, often leading to biased learning and poor detection of minority classes. To address this problem, this research introduces a visualization and Silhouette score guided oversampling framework. Unlike conventional blind oversampling methods, the proposed approach examines class topology to choose appropriate oversampling techniques, making the augmentation process both targeted and structurally informed. We derived six experimental scenarios from CICIDS2017, fine and coarse-grained intrusion classifications. For each scenario, we employed a visualization and silhouette-driven oversampling framework, followed by classification assessment using a Feedforward Neural Network (FNN). Class-wise precision, recall, F1-score, and log loss were assessed before and after oversampling. Consequently, the model achieved a macro-average accuracy of 99.75% and nearly $$\sim $$ 100% weighted-average accuracy, demonstrating robust minority-class recall through cluster-structure guided adaptive oversampling under extreme imbalance. This approach integrates quantitative cluster structure analysis to support adaptive oversampling decisions, improving robustness under extreme minority-class scarcity.