Hybrid VAE-Transformer-based multi-task learning for cyberattack detection and fault classification in electric vehicle charging stations.
- 1School of Electrical Engineering, Vellore Institute of Technology, Chennai, Tamil Nadu, India.
Abstract
Electric vehicle (EV) charging stations have become a complex cyber-physical system, where the sensor network is susceptible to both hardware failures and cyberattacks. Both produce similar data patterns, making them difficult to distinguish. The majority of the approaches have addressed only cyberattacks or faults and rely on recurrent architectures, which are hard to capture long-term temporal patterns. In this study, a multi-task learning framework for cyberattack detection and fault classification is proposed based on four hybrid deep learning models, namely Variational Autoencoder-Long Short-Term Memory (VAE-LSTM), VAE-Gated Recurrent Unit (VAE-GRU), VAE-Generative Adversarial Network (VAE-GAN), and VAE-Transformer. The framework is tested on a publicly available EV Internet of Things (IoT) predictive-maintenance dataset with six threat scenarios like Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), Replay attacks, Man-in-the-Middle (MitM) attacks, Message Manipulation attacks, and a Flatline sensor fault. The results show that the VAE-Transformer achieved an Area Under the Curve-Receiver Operating Characteristic (AUC-ROC) of 99.76%, a binary F1-score of 99.03%, an attack-type Macro-F1 of 97.29%, an FPR@95% of 0.41%, and a Matthews correlation coefficient (MCC) of 98.20%. Furthermore, the low values of orthogonality and Mutual Information (MI) loss demonstrate that the proposed hybrid deep learning algorithm successfully separates cyberattack and physical fault features, thereby improving overall detection performance.
Related
- Normal-Data-Only Fault Detection and Faulty-Cell Localization in Heterogeneous Electric-Vehicle Battery Packs Using a Hotspot-Suppressed and Alarm-Consistent Dual-Branch Attention Transformer
- Transformer-based multimodal representation learning for blind acoustic classification
- SkinFusionNet: MobileNetV₂ and vision transformer hybrid deep learning model with automated skin cancer classification
- Comparative and explainable deep learning for cherry (Prunus avium L.) cultivar classification using convolutional neural networks (CNN) and transformer architectures
- A federated learning framework for anomaly detection in electric vehicle charging networks
BibTeX
@article{vinitha2026hybrid,
title = {{Hybrid VAE-Transformer-based multi-task learning for cyberattack detection and fault classification in electric vehicle charging stations.}},
author = {R Vinitha and R Senthil Kumar},
journal = {Frontiers in artificial intelligence},
year = {2026},
month = oct,
doi = {10.3389/frai.2026.1934106},
url = {https://europepmc.org/article/MED/42909096},
note = {CC BY},
}
RIS
TY - JOUR
TI - Hybrid VAE-Transformer-based multi-task learning for cyberattack detection and fault classification in electric vehicle charging stations.
AU - R Vinitha
AU - R Senthil Kumar
PY - 2026
DA - 2026/10/07
JO - Frontiers in artificial intelligence
DO - 10.3389/frai.2026.1934106
UR - https://europepmc.org/article/MED/42909096
AB - Electric vehicle (EV) charging stations have become a complex cyber-physical system, where the sensor network is susceptible to both hardware failures and cyberattacks. Both produce similar data patterns, making them difficult to distinguish. The majority of the approaches have addressed only cyberattacks or faults and rely on recurrent architectures, which are hard to capture long-term temporal patterns. In this study, a multi-task learning framework for cyberattack detection and fault classification is proposed based on four hybrid deep learning models, namely Variational Autoencoder-Long Short-Term Memory (VAE-LSTM), VAE-Gated Recurrent Unit (VAE-GRU), VAE-Generative Adversarial Network (VAE-GAN), and VAE-Transformer. The framework is tested on a publicly available EV Internet of Things (IoT) predictive-maintenance dataset with six threat scenarios like Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), Replay attacks, Man-in-the-Middle (MitM) attacks, Message Manipulation attacks, and a Flatline sensor fault. The results show that the VAE-Transformer achieved an Area Under the Curve-Receiver Operating Characteristic (AUC-ROC) of 99.76%, a binary F1-score of 99.03%, an attack-type Macro-F1 of 97.29%, an FPR@95% of 0.41%, and a Matthews correlation coefficient (MCC) of 98.20%. Furthermore, the low values of orthogonality and Mutual Information (MI) loss demonstrate that the proposed hybrid deep learning algorithm successfully separates cyberattack and physical fault features, thereby improving overall detection performance.
SN - 2624-8212
ER -