Synthetic Data Augmentation for Imbalanced EV Charging Network Security: An Empirical Study of Generative Models and Feature Transformation
Year: 2026
Type: conference-paper
Abstract: Class imbalance and limited data availability are major challenges for intrusion detection systems in Electric Vehicle (EV) environments, particularly in cyber-physical datasets exhibiting distinct operational regimes. This work investigates synthetic data augmentation strategies for the EV power consumption subset of the CICEVSE2024 dataset, which exhibits a significant class imbalance ratio of approximately 7:1 (attack:benign), using Conditional Tabular Generative Adversarial Networks (CTGAN), Tabular Variational Autoencoders (TVAE), and Gaussian Copula models from the Synthetic Data Vault (SDV) framework. Additionally, the impact of feature transformation is examined by comparing raw and log1p-transformed numerical features. Experimental results show that raw-feature generative pipelines achieve higher statistical fidelity by better preserving quantized operational patterns, whereas log1p-transformed pipelines provide improved classification utility and competitive macro F1-scores compared to traditional imbalance handling approaches. The results highlight a trade-off between statistical realism and predictive utility, suggesting that the suitability of synthetic augmentation strategies depends on the intended cyber-physical application.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225612232