Multi-view graph contrastive representative learning for intrusion detection in EV charging station

Abstract

With the rapid proliferation of electric vehicles (EVs), the need to enhance EV charging infrastructure with integrated communication and software functionalities has become crucial. However, this integration also introduces new cybersecurity vulnerabilities, as sensitive data and operational control are increasingly exposed to potential attacks. Traditional intrusion detection systems often struggle with overfitting, low recall, and the scarcity of high-quality labeled data or fail to consider the correlation among different features, challenging the effectiveness of supervised learning approaches. To address these limitations, this paper proposes a novel Multi-View Graph Contrastive Representation Learning (MVGCRL) framework that leverages logs from Hardware Performance Counters (HPCs) collected from Electric Vehicle Supply Equipment (EVSE) and represents them as graph structure data. By constructing graph views for both hardware components and temporal windows, the framework utilizes a Graph Neural Network (GNN) model to capture correlations among various input features in a multi-view manner. This work designed a supervised intrusion detection system (IDS) for multi-class classification. Specifically, our method introduces hybrid graph augmentations through node feature masking and edge weight perturbation, and then employs a novel mask-attention Graph Transformer to capture complex feature correlations. Additionally, MVGCRL is extended to a self-supervised learning version by minimizing the distance between node embeddings and input features, followed by fine-tuning for improved classification. Experiments on real-world datasets demonstrate that our approach outperforms both traditional supervised methods and state-of-the-art self-supervised learning models, offering an effective solution for enhancing cybersecurity in EV charging infrastructures.