Transformer-Based RF Jamming Detection in Vehicular Networks: A BERT Model Analysis

Vehicular networks are increasingly susceptible to RF jamming attacks that compromise communication reliability. This paper proposes four Transformer-based intrusion detection system (IDS) models – BERT, RoBERTa, DistilBERT, and ALBERT – and compares them against three baselines (KNN, RaFo, and LSTM) under vehicular mobility conditions of \(25\) m/s and \(15\) m/s. At \(25\) m/s, BERT and DistilBERT outperform all baselines, while RoBERTa and ALBERT show slightly lower performance. Under the more challenging \(15\) m/s scenario, all Transformer-based models exhibit superior generalization compared to the baselines, which experience significant degradation. The proposed models also maintain higher stability across key metrics—accuracy, precision, recall, and F1-score. Confusion matrix analysis confirms that no RF jamming traffic is misclassified as benign, even under degraded conditions, an essential property for timely threat detection in dynamic vehicular environments. Finally, a computational cost analysis indicates that BERT is better suited for deployment in resource-rich roadside units, while DistilBERT offers the most balanced trade-off for constrained environments, combining low inference time, reduced memory usage, and moderate resource demands. RoBERTa and ALBERT occupy intermediate positions in this spectrum.