Edge-efficient deep learning-based IDS for VANETs using structured pruning and knowledge-preserving weight transfer

VANETs support real-time safety communication but are vulnerable to cyber attacks due to their dynamic nature. Deep learning-based IDS detect threats effectively but are unsuitable for RSUs due to size and latency. Balancing detection accuracy with deployment efficiency remains a significant challenge. This paper addresses this gap by proposing a DL-based IDS framework optimized for edge-enabled VANET deployment. The proposed method utilizes a hybrid CNN-BiLSTM framework to capture spatiotemporal patterns of communication patterns and structured pruning to simplify the model while preserving key features. To examine the performance-efficiency trade-off, three model variations are created: (i) a baseline model,(ii) a pruned model, and (iii) a compact model created from weight transfer from the pruned model. The VeReMi Extension dataset is used to assess the proposed framework in coarse-grained and fine-grained classification scenarios. The compact model achieves good performance with high efficiency. In the coarse-grained scenario, the model attains an average 98.08% accuracy, 98% precision, 96% recall, and 97% F1-score with an inference latency of 60.07 ms, a model size of 2.96 MB and fine-grained classification achieves 95.37% accuracy with an inference time of 56.99 ms while maintaining the same model size.