Early detection of cyberattacks in renewable-integrated EV charging systems using transfer learning

Le résumé fourni par la source

The growing interconnectivity of renewable energy sources with Electric Vehicle (EV) charging stations has triggered the implementation of smart grids that are intelligent and interconnected. Although such developments improve energy efficiency, they also increase the cyberattack surface, making EV charging networks vulnerable to previously unseen threats. Traditional intrusion detection systems based on fixed feature patterns or static learning are usually not very responsive to changing and zero-day-oriented attacks, leading to reliability issues and slow identification of attacks. To address these constraints, this study presents a Progressive Domain-Refinement Learning Framework (PDRLF), a transfer-learning framework intended to enhance the accuracy of early detection in renewable-integrated EV charging systems. The proposed PDRLF uses large-scale pretraining on multisource datasets, such as CIC-IDS2017, UNSW-NB15, and CICEVSE2024, followed by fine-tuning with a hybrid CNN-BiLSTM feature extractor. The two-step adaptation facilitates strong modelling of spatiotemporal traffic behaviors and increases the system sensitivity to unidentified cyber behaviors. An analysis based on precision-recall curves, ROC curves, confusion matrices, latency measurements, and hyperparameter-sensitivity experiments demonstrated the superiority of the framework over the baseline classifiers. The proposed PDRLF demonstrated a high intrusion detection rate of 98.43% precision, 98.14% recall, 97,4% accuracy, 98.28% F1-score, and 0.98 ROC-AUC, with a low mean detection latency of 18.4 ms/sample. The proposed framework achieved 97.8% of accuracy under emulated unseen attack scenarios by indicating the adaptability towards the real-time smart grid cybersecurity. Since this evaluation relies on simulated and emulated attacks, the validation on external datasets and unseen real-world attacks will be considered in future work. The results obtained are to be interpreted as evidence of potential rather than definitive zero-day detection capacity.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.