EfficientNet-based universum-inspired supervised contrastive learning and transfer learning for in-vehicle intrusion detection systems

In-vehicle networks are highly vulnerable to cyberattacks because of their lack of authentication and encryption mechanisms in the controller area network (CAN) protocol. Although machine learning (ML)-based intrusion detection systems (IDS) have been promising in the aspect of overcoming the aforementioned vulnerability, their effectiveness is often hindered by class imbalance, data scarcity, and poor generalization across different vehicle models. To address these challenges, we propose E–UniCon, a lightweight IDS that integrates EfficientNet-Lite0 with Universum-Inspired Supervised Contrastive Learning to enhance feature discrimination and improve robustness against unseen attacks. The model leverages contrastive learning with universum samples to create structured, generalizable decision boundaries while maintaining suitable computational efficiency for real-time deployment on embedded automotive devices. Furthermore, we evaluate E–UniCon using the CAN-ML dataset under base and transfer learning settings. Experiments show that E–UniCon achieves overall F1 scores of 0.9939 and 0.9938 in the base-model setup and transfer learning, respectively, when fine-tuned with only 10 % of the labeled data from another vehicle. Furthermore, the model achieves a low false negative rate (1.0 %) even under domain shift conditions via transfer learning, and processes 64 CAN messages in only 7.5 ms, confirming its suitability for real-time in-vehicle IDS applications.