A HYBRID INTRUSION DETECTION SYSTEM FOR ELECTRIC VEHICLE CHARGING INFRASTRUCTURE
1
A Hybrid Intrusion Detection System for Electric Vehicle Charging Infrastructure
arXiv:2606.23236v1 [cs.CR] 22 Jun 2026
Charukeshi Joglekar , Member, IEEE, Chijioke Eze , Graduate Student Member, IEEE, Danni Xiang , and Antonello Monti , Senior Member, IEEE
Abstract—The integration of Electric Vehicle Charging Stations (EVCSs) into the smart grid necessitates sophisticated digital infrastructure for their management and coordination, which expands the attack surface and makes both the power grid and EVCSs vulnerable to cyberattacks. This research addresses critical gaps in existing EVCS Intrusion Detection Systems (IDS) by proposing a hybrid IDS that integrates attack detection on both the cyber and physical layer of the EVCS ecosystem. The proposed hybrid IDS utilizes a dual-layer integration method, which combines network-based IDS (NIDS) and host-based IDS (HIDS). This approach facilitates for comprehensive monitoring of both network traffic through the NIDS and host-level activities via the HIDS, effectively addressing the unique challenges posed by the interconnected nature of EVCS ecosystems. Utilizing the recent CICEVSE2024 dataset, the IDS presented in this work performs multiclass classification across various attack types, including False Data Injection Attacks (FDIAs), reconnaissance, denial of service, backdoor, and cryptojacking attacks. Experimental results demonstrate that our approach achieves excellent detection accuracy, with the NIDS component reaching 99.99% accuracy for network-based attacks and the HIDS component achieving 83.47% accuracy on FDIA, cryptojacking, backdoor, all DoS, all Recon except Slowloris Scan attacks. This duallayer detection significantly outperforms single-source detection approaches previously presented in literature. Index Terms—Electric Vehicle Charging Stations, Cybersecurity, Intrusion detection system, Machine learning, Network security, Host security, Power consumption
I. I NTRODUCTION The global transition toward sustainable transportation has accelerated the deployment of Electric Vehicle Charging Stations (EVCSs). At the end of 2024, there were approximately 5.2 million public EVCSs worldwide, with the number continuing to grow rapidly [1]. As the integration of Electric Vehicle Charging Stations (EVCSs) into the grid increases, it require sophisticated digital infrastructures as well as communication This work was supported by the project End-to-end Cybersecurity to NEMO meta-OS (CyberNEMO), funded by the European Union’s HORIZON Innovation Actions under Grant Agreement No.: 101168182. (Charukeshi Joglekar and Chijioke Eze contributed equally to this work.) (Corresponding author: Charukeshi Joglekar.) Charukeshi Joglekar and Antonello Monti are with the Fraunhofer Institute for Applied Information Technology, 52068 Aachen, Germany, and also with the Institute for Automation of Complex Power Systems, RWTH Aachen University, 52074 Aachen, Germany (e-mail: [email protected], [email protected]). Chijioke Eze and Danni Xiang are with the Institute for Automation of Complex Power Systems, RWTH Aachen University, 52074 Aachen, Germany (e-mail:[email protected]; [email protected]).
between different actors to ensure coordinated charging from both the grid as well as consumer perspective. As shown in Fig. 1, the EVCS ecosystem is a complex system composed of a number of components managed by various actors. The components include: (1) Electric Vehicles (EVs), (2) EVCSs, which deliver power to EVs, (3) the Charging Station Management System (CSMS), which provides functionalities for managing the operation of all connected EVCSs; (4) applications/web interfaces, which allow users to locate, reserve, and pay for charging services; and (5) the power grid, which supplies electricity to the entire system.
Fig. 1: EV Charging ecosystems The two primary actors in this ecosystem are: Charging Point Operators (CPOs), who own and operate the charging stations; and EV drivers, who use the charging services. These components and actors interact through various communication protocols. The communication between the CSMS, which is managed by CPO and the EVCS uses the Open Charge Point Protocol (OCPP), whereas the communication between EVCSs and EVs uses the ISO 15118 protocol. The ISO 15118 protocol supports V2G communication and relevant functionalities. Other actors such as Electromobility Roaming Service Providers, Distribution System Operators(DSOs) also communicate with the aforementioned actors for the control and coordination of EV charging to ensure grid reliability while fulfilling EV consumer requirements. The EV charging ecosystem is thus a cyber-physical system where digital communications control physical power flows. The interconnected nature of the system creates multiple potential attack surfaces that malicious actors can exploit to disrupt EVCS as well as grid operation [2] (See Section II-A). As Sharma et al. [3] point out, sophisticated intrusion detection mechanisms are required that facilitate attack detection at multiple layers of the cyber-physical ecosystem. Intrusion Detection Systems (IDSs) proposed in literature for EVCSs have generally been single-source detection approaches i.e. either Network-based IDS (NIDS) or Host-based
This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.
A HYBRID INTRUSION DETECTION SYSTEM FOR ELECTRIC VEHICLE CHARGING INFRASTRUCTURE
IDS (HIDS). Each of these have notable limitations regarding detection comprehensiveness and dataset relevance. Machine Learning (ML) classification algorithms, primarily supervised learning, have been widely used in prior studies [4]–[6], where labeled data is required for training and prediction. These algorithms are validated for efficient and accurate attack detection and are commonly used in IDS development. Prior studies on NIDSs [4], [5] have typically relied on attack datasets containing both benign and malicious traffic. However, these datasets were derived from non-EVCS environments, limiting their applicability to EVCS-specific network environments. Conversely, existing HIDSs [6], [7] have used datasets specific to EVCSs, but these contain only benign data, necessitating the use of simulated attack data that may not fully reflect real-world attack patterns. Recent advancements in deep learning, as seen in the work of Li et al. [8] with multi-view graph contrastive learning and Benfarhat et al. [9] with temporal convolutional networks, have shown promising results in EVCS intrusion detection. However, these approaches typically focus on either networklevel or specific temporal patterns, not the comprehensive view that a hybrid approach can provide. To address these limitations, this paper proposes a Hybrid IDS for EVCSs based on the recently released CICEVSE2024 dataset [10], which encompasses network traffic data, power consumption data, and host activities of EVCSs under both benign and attack scenarios. The proposed Hybrid IDS comprises a NIDS deployed at the CSMS in the cyber layer and HIDSs placed at individual EVCSs. The NIDS monitors overall network traffic, detecting network-based attacks that may involve multiple EVCSs, while HIDSs at each EVCS provide localized detection by monitoring log events and power consumption, primarily focused on identifying hostbased attacks. The main contributions of this paper are: (1) A dual-layered Hybrid IDS architecture that aligns with the cyber-physical structure of the EVCS ecosystem, providing comprehensive security monitoring; (2) Implementation and evaluation of a machine learning-based multiclass classification for detecting and categorizing network-based and host-based attacks, including backdoor, cryptojacking, FDIAs, and different variants of reconnaissance, and DoS attacks etc.; (3) Development and validation of the proposed IDS using the CICEVSE2024 dataset, which contains real-world attack data specific to EVCSs rather than simulated attacks or general network or IoT related attacks, and (4) Empirical evaluation demonstrating that the data preprocessing and feature engineering pipeline proposed in this work results in improved attack detection performance: higher detection accuracy and better attack classification compared to the baseline approaches. The remainder of this paper is organized as follows: Section II discusses EVCS threat landscape, and presents related work on IDS for EVCSs. Section III presents the conceptual architecture of our proposed Hybrid IDS and describes the CICEVSE2024 dataset. Section IV details the implementation methodology and experimentation setup. Section V presents the evaluation results and comparative analysis. Finally, Section VI concludes the paper and suggests future research
2
directions. II. BACKGROUND AND R ELATED W ORK A. Threat Landscape for EVCS The EVCS ecosystem faces numerous security threats. Benfarhat et al. [9] identify up to 16 different attack types against EVCS systems, including both traditional network attacks and emerging threats specific to charging infrastructure. Considering the complex nature of the EVCS ecosystem, the Intrusion Kill Chain [11] introduced by Hutchins et al. provides a structured framework to describe the stages attackers go through to execute attacks on such a system. These are reconnaissance, weaponization, delivery, exploitation, installation, command and control, and actions on objectives [11]. The first step in an attack is reconnaissance, which can be used to identify vulnerabilities in the communication between EVs, EVCSs, and the CSMS. In the context of EVCSs, reconnaissance attacks can be used by attackers to identify vulnerabilities in the communication between EVs, EVCSs, and the CSMS. These vulnerabilities can then be exploited to launch more sophisticated attacks on these components. In the weaponisation stage, an attacker can execute hostbased attacks, such as backdoor attacks and cryptojacking, which directly target the EVCS hardware. Backdoor attacks can provide unauthorized access to attackers, while cryptojacking uses the computing resources of EVCSs for cryptocurrency mining without authorization. Particularly, backdoor attacks can be escalated further to orchestrate disruptive attacks that have a adverse consequences for the individual EVS as well as interconnected devices in the EVCS ecosystems. Therefore, the detection of host-based attacks at the level of individual EVCS’ can reduce the likelihood of subsequent attacks. In the context of attacks on the communication interfaces among EVCS ecosystem components, the work of Zhdanova et al. [12] identifies two main classes of network attacks that pose threats to V2G communication : (i) Vulnerabilities in the ISO 15118 protocol, which can be exploited to conduct Denial of Service (DoS) attacks by flooding communication channels and jamming attacks by transmitting jamming signals between the EVCS and EV. (ii) Attackers can exploit the vulnerabilities of the Open Charge Point Protocol (OCPP) execute False Data Injection Attacks (FDIAs) by manipulating voltage, current, or power data, or by launching Man-in-the-Middle (MitM) attacks to intercept and communication between the CPO and EVCS. Sharma et al. [3] categorize EVCS threats into three key domains: grid-side attacks targeting the power grid infrastructure, communication-side attacks exploiting vulnerabilities in protocols like OCPP, and user-side attacks that target charging station access points and user interfaces. Their research highlights that positioning of EVCS at the intersection of energy and transportation infrastructure makes it particularly vulnerable to sophisticated attacks that could impact both domains simultaneously. These attacks can have severe consequences, including power grid instability [2], financial losses for service providers [13], and compromised availability of charging services for
A HYBRID INTRUSION DETECTION SYSTEM FOR ELECTRIC VEHICLE CHARGING INFRASTRUCTURE
EV users [14]. These impacts of cyberattacks on EVCSs emphasizes the need for robust security measures that protect both cyber and physical layers of the ecosystem. Further, Warraich and Morsi [15] specifically examine the threat landscape for fast charging stations, noting that the high power rates involved create additional risks, as compromised charging sessions could lead to more immediate and severe impacts on grid stability, particularly in microgrid environments where V2G operations are implemented. Therefore, intrusion detection systems for the EVCS ecosystem must be able to classify network-based and host-based attacks to mitigate their impact on individual EVCS as on the power grid. B. Existing IDS Approaches Existing research on IDSs for EVCSs can be broadly categorized into two approaches which are discussed in the subsections II-B1 and II-B2. 1) Network-based IDS (NIDS) for EVCS: The network layer of the EVCS is one of the key points that attackers use to orchestrate attacks against the system. This has attracted the attention of researchers within the EVCS community. As a result, many detection approaches have been proposed. For example, ElKashlan et al. [4] developed a machine learningbased NIDS for IoT EV charging stations. In their work, they propose a random forest classifier trained using features derived from the IoT23 dataset, which contains benign and malicious network traffic. Their approach achieved good accuracy but its reliance on a general IoT dataset [16] dataset rather than EVCS-specific data, implies it may not guarantee good performance in a typical EVCS environment . Similarly, Basnet and Ali [5] implemented a deep learning-based IDS for electric vehicle charging stations. They used the CICIDS2017 dataset, containing benign and various attack traffic patterns, and demonstrated that deep learning models could effectively detect network-based attacks. However, the dataset was not specific to EVCS environments. More recently, Li et al. [8] proposed a multi-view graph contrastive representative learning approach for intrusion detection in EV charging stations. Their method extracts multiple views of network traffic data and leverages graph contrastive learning to better capture the relationships between different network flows. This approach showed improved performance in identifying subtle attack patterns that might not be detected by traditional methods. However, it also leaves open questions in terms of real-world deployment and generalization ability of the method. Additionally, Benfarhat et al. [9] proposed an advanced Temporal Convolutional Network (TCN) framework specifically for intrusion detection in EVCSs. Their proposed MRG-ID-SA-TCN model integrates multi-receptive fields, gating mechanisms, iterative dilation, and self-attention to effectively capture temporal dependencies in network traffic. This approach achieved state-of-the-art results on the CICEVSE2024 dataset, with particularly strong performance in detecting denial-of-service and reconnaissance attacks. However, despite its good performance on multi-class attack detection for EVCS, again, the method leaves critical
3
gaps in terms of real-world deployment considerations, including cross-dataset generalization, explainability, and continuous learning for evolving threats.In Jiang et al. [17], the authors present a knowledge distillation enhanced semi-supervised Federated Learning (FL) framework for intrusion detection in EV charging networks. Although their model outperforms baseline methods from semi-supervised as well as federated learning, their proposed approach was only evaluated using the network traffic data from the CICEVSE2024 dataset; hostbased attacks are not taken into consideration in their work. Unlike these group of works, the present paper is not only validated on realistic EVCS datatset to meet the unique nature of network based intrusions in EVCS environment but also includes detection of host based intrusions at the same time. 2) Host-based IDS (HIDS) for EVCS: Another important entry point for attacks against EVCS is the host devices installed within the charging infrastructure. An example is backdoor attack, which can be exploited to launch attacks that affect the charging infrastructure or things connected to steal data or cause damage to the EVCS. Many methods have been proposed in the literature to address the host intrusion attacks in EVCS. For instance, Cumplido et al. [6] implement a centralized HIDS to classify EV charging sessions as normal or anomalous. Their system combines regressionbased charging behavior prediction with ensemble anomaly detection, incorporating classification-based and novelty-based detection methods. They used real-world EV charging session data from ACN [18] and ElaadNL [19] datasets, which contain only benign charging data to validate their proposed method. Thus, raising the question as to whether applicability of their method in disntinguishing different HIDS attacks. In Girdhar et al. [7], the authors propose a Hidden Markov Modelsbased anomaly correlation approach for detecting suspicious activities in EVCSs. Their system used system logs and monitoring data to identify temporal patterns indicative of attacks. Similar to other HIDS approaches, they alos worked with only benign data and simulated attack scenarios. In Warraich and Morsi [15], the authors propose propose an early detection system for cyber-physical attacks on fast charging stations. Their approach is particularly notable for its focus on the physical aspects of charging infrastructure and consideration of Vehicle-to-Grid operations in microgrid environments. They demonstrated that monitoring the power consumption patterns and electrical characteristics of EVCS could provide early warning signs of attacks, even before they manifest themselves as obvious network anomalies. Sharma et al. [3] proposed an AI-augmented architecture that incorporates both host-level monitoring and grid-level anomaly detection. Their system uses a hierarchical approach to correlate events across multiple layers of the charging infrastructure, enabling more accurate detection of sophisticated attacks that might affect both the cyber and physical layers simultaneously. However, the implementation of the architecture proposed by these authors is a centralized AIdriven anomaly detection framework, rather than a true duallayer hybrid IDS. In other words, their framework lacks separate NIDS and HIDS components. It does not have dedicated network traffic analysis components that monitor flow-
A HYBRID INTRUSION DETECTION SYSTEM FOR ELECTRIC VEHICLE CHARGING INFRASTRUCTURE
level and packet-level features, nor independent host-based detectors that analyze system calls and hardware performance counters. Their detection approach relies primarily on power consumption patterns and charging session behavior using Gaussian-based thresholding. It can be argued that such a monolithic design cannot detect network-level reconnaissance attacks, protocol-layer exploits, or early stage attack indicators, which function reactively only after attacks impact operations rather than proactively identifying threats at both the network and host levels. These existing approaches highlight a critical gap: existing NIDS approaches for EVCS use generalized network datasets, which do not reflect the realities in EVCS environments. Although existing HIDS approaches use EVCS-specific data, they rely on simulated attacks rather than actual attack data. Additionally, most of the existing approaches focus exclusively on either network or host-based detection, missing the potential benefits of a combined approach. Additionally, recent works such as Benfarhat et al. [9] and Li et al. [8] use EVCS-specific datasets with real attack data, however, they focus mainly on network-level detection using advanced deep learning techniques. Similarly, Sharma et al. [3] and Warraich and Morsi [15] consider both cyber and physical aspects, but again, they do not fully implement a dual-layer detection architecture that can enable fine-grained network- and hostbased monitoring. To overcome these gaps, we propose a dual-layered architecture that aligns with the cyber-physical structure of the EVCS ecosystem. Specifically, the network-based IDS is deployed at the CSMS, while multiple host-based IDSs are deployed at individual EVCSs to detect atomic attacks at both layers. Our proposed Hybrid IDS addresses three main research gaps in existing IDS for EVCSs: (1) the lack of complementary monitoring across cyber and physical layers, and (2) limitations in data specificity and classification granularity, and (3) efficient data preprocessing and feature engineering pipeline that ensure optimal performance of the attack detection models. The recently released CICEVSE2024 dataset [10] addresses data limitations by providing a comprehensive dataset specific to EVCS environments that includes both benign and actual attack data collected at the network and host layers. Hence, the hybrid IDS proposed in this paper uses this dataset in training and validation of detection models for attacks at both network and host layers of the EVCS environment to ensure it adequately reflect real-world realities. III. P ROPOSED H YBRID IDS A PPROACH Building upon the analysis of existing approaches and their limitations, we hereby present our proposed solution—a duallayered hybrid intrusion detection system designed specifically for the unique requirements of intrusion detection in EVCS environments. A. Conceptual Architecture As illustrated in Fig. 2, the Hybrid IDS architecture consists of two main components: the NIDS and the HIDS components. The NIDS component captures network traffic from the CSMS
4
while the HIDS component collects host event logs and power consumption data from individual EVCSs. The collected data are preprocessed to remove missing values. In the next step, the preprocessed data are analyzed using machine learning classifiers, which identify each sample as benign or a specific attack type. Upon detection of an attack, an alert is generated and transmitted to the network administrator, who can then respond accordingly.
Fig. 2: The Conceptual Hybrid IDS Architecture This coordinated approach enables comprehensive detection at both the network and host levels, effectively addressing threats towards the cyber-physical system. This is particularly important, as attacks such as denial of service attacks at the network layer can result in abnormal activities at the host level, for example, interruption of charging and ultimately power flow to the EV. Hence, by ensuring simultaneous attack detection at both levels, the proposed framework helps to reduce the chance of attacks going undetected. B. Dataset Description The CICEVSE2024 dataset [10] is composed of power consumption data, network traffic, and host event logs of EVCSs in both benign and attack scenarios. In the lab setup used to generate the data, there were two Electric Vehicle Supply Equipments(EVSEs), serving as the physical interface within EVCS: EVSE-A, a real EV charger, and EVSE-B, an EVSE emulated on a Raspberry Pi. The dataset was collected in different scenarios: benign, host-based attacks, and networkbased attacks, each in both idle and charging states generating a total of six scenarios. A total of 16 attacks were executed, of which cryptojacking and backdoor are host-based attacks, and the remaining 14 are network-based attacks, including 7 various types of reconnaissance attacks and 7 different kinds of DoS attacks. The three data sources (network traffic, host events, and power consumption) have different data sizes, features, and were collected for different attack scenarios. 1) Network Traffic Data: The CICEVSE2024 dataset contains network traffic data from both EVSE-A and EVSE-B. The data was captured using a network topology where the Electric Vehicle Communication Controller (EVCC) communicates bidirectionally with EVSE-B through a switch using the ISO 15118 protocol, while EVSE-A and EVSE-B use the OCPP protocol to communicate with remote and local CSMS, respectively. Two types of network traffic data were collected: flowlevel data and packet-level data. Flow-level data extracted
A HYBRID INTRUSION DETECTION SYSTEM FOR ELECTRIC VEHICLE CHARGING INFRASTRUCTURE
using the NFStream Python library, provides features such as flow duration and the number of packets within each flow; whereas Packet-level data offers more detailed information about individual packets, including protocol types, packet sizes, and header information. 2) Host Events Data: In CICEVSE2024, host events data were collected only from EVSE-B. The event logs of Hardware Performance Counters (HPC) and kernel were collected under six different scenarios (benign, network attack, hostbased attack each in idle and charging state). C. Attack Detection Framework The goal of the proposed Hybrid IDS is to identify and distinguish between various network- and host-based attacks. For network-based attacks, the system is trained to identify (detect) normal traffic, Reconnaissance (Recon) attacks, Denial of Service (DoS) attacks, and False Data Injection Attacks (FDIAs) etc. In the case of host-based attacks, the system is trained to detect normal events, backdoor attacks, and cryptojacking, etc. Formally, given a set of input features X = {x1 , x2 , ..., xn } extracted from network traffic, host events, or power consumption data, the classification task performed by the proposed hybrid IDS can be defined as finding a function f that maps these features to a set of attack classes C = {c1 , c2 , ..., cm }: f :X→C
(1)
where c1 typically represents benign traffic/events and c2 through cm represent different attack types. The objective of the Hybrid IDS is to minimize the classification error: min f
N X
I(f (xi ) ̸= yi )
(2)
i=1
where yi is the true label of sample i, N is the total number of samples, and I is the indicator function. We employ the following ML algorithms in the classification process: Random Forest (RF), Light Gradient Boosting Machine (LGBM), XGBoost (XGB), Decision Tree (DT), and Support Vector Machine (SVM). The reason for choosing these ML algorithms is due to their good performance as reported in prior intrusion detection researches and their ability to handle the types of feature present in the chosen dataset. IV. E XPERIMENTS In this section, we discuss the procedure used to validate the performance of the proposed hybrid IDS. We first discuss the data processing and feature engineering approach we employed, after which we present the training and validation of the ML classifiers. A. Data Processing and Feature Engineering The data processing and feature engineering workflow for each data source follows a systematic process comprising three main phases: (1) data reorganization and integration, (2) preprocessing and feature extraction, and (3) feature selection and transformation. Each data type required specific handling techniques to address its unique characteristics and challenges.
5
1) Network Traffic Data Processing: For network traffic data, we focused on data relevant to EVSE-A due to its completeness and representativeness. Two complementary approaches were implemented: Flow-based Analysis: The original flow-level datasets were integrated into a unified dataset with standardized labels. The original CICEVSE2024 dataset contains 86 features extracted using NFStream, out of which we used 67 after eliminating non-numeric features, columns with missing data or zero variance. Some of the features used include temporal attributes (e.g., flow duration, inter-arrival time), volumetric features (e.g., packet count, byte count), and protocol-specific metrics. Address information (e.g., IP, MAC, and OUI) was excluded from the dataset to prevent potential data leakage. This decision is made based on the dynamic nature of real-world network environments, which contrasts with the relatively static and limited number of devices in the simulation. Quantitative analysis revealed a significant class imbalance, with benign traffic constituting 0.015% of all flows. To address this, we applied Synthetic Minority Over-sampling Technique (SMOTE) with a sampling strategy optimized to achieve a more balanced representation while avoiding overfitting. However, it either decreased the performance or had no effect on choosing the best classifier. Packet-based Analysis: We processed raw packet captures through a multi-stage pipeline: (1) packet extraction and parsing, (2) feature derivation, and (3) dimensionality optimization. The packets were organized into a structured format, with 38 distinct features extracted per packet. After performing feature engineering, 28 features were left, including TCP flags, and payload characteristics. For FDIA simulation, we developed a pattern-based manipulation technique based on observed attack signatures in similar IoT environments, selectively altering packet payloads and headers to represent falsified data transmissions while maintaining protocol validity. 2) Host Events Data Processing: In the case of host events data, we constructed a multi-dimensional feature space from heterogeneous log sources. The preprocessing pipeline combined Hardware Performance Counters (HPC) data, capturing microarchitectural events (e.g., cache misses, branch prediction outcomes) with kernel-level system events (e.g., system calls, I/O operations). 3) Power Consumption Data Processing: Power consumption data required specialized processing to extract meaningful patterns from time-series measurements owing to the low training and evaluation performances initially. We developed a feature extraction framework consisting of data cleaning: (1) Data Cleaning: removal of duplicate rows and handling of missing values through cubic spline interpolation, (2) Feature Engineering: derivation of new features using a rolling window technique (3) Scaling and Class Imbalance Management: application of scaling techniques like MinMaxScaler and StandardScaler, along with handling class imbalance through methods such as class weight and SMOTE, which ultimately demonstrated limited success in improving performance, and (4) Final Preprocessing: selection of StandardScaler as the optimal scaling method based on its effectiveness in addressing low standard deviation issues within key features.
A HYBRID INTRUSION DETECTION SYSTEM FOR ELECTRIC VEHICLE CHARGING INFRASTRUCTURE
B. ML classifier training and validation For each data source, we implemented and compared multiple machine learning classifiers: Random Forest (RF), Light Gradient Boosting Machine (LGBM), XGBoost (XGB), Decision Tree (DT), and Support Vector Machine (SVM). The network traffic datasets were split into training (80%), validation (10%), and testing (10%) sets due to their large size (547,834 flow-level and 1,309,252 packet-level records), whereas the host events (12,499 records) and power consumption (115,298 records) datasets were split into training (60%), validation (20%), and testing (20%) sets. Hyperparameter optimization was performed using Bayesian optimization with tree-structured Parzen estimators to find the optimal configuration for each classifier. The primary evaluation metrics included accuracy, precision, recall, F1-score, and G-mean. For a multiclass classification problem with k classes, these metrics are defined as follows:
6
v u k uY k G-mean = t Recall
i
(7)
i=1
where k is the number of classes and Recalli is the recall for class i. This metric is particularly useful for imbalanced datasets as it is sensitive to performance on minority classes. The testing results demonstrated only 5 misclassifications out of 54,784 samples, all of which involved misclassifying one attack type as another, while maintaining perfect distinction between normal and attack traffic. As shown in Fig. 3, the confusion matrix for the packet-level NIDS shows that the system effectively distinguishes between different attack types, with some minor confusion between similar attack variants.
Pk Accuracy = Pk
i=1 T Pi
i=1 (T Pi + F Pi )
T Pi T Pi + F Pi T Pi Recalli = T Pi + F Ni Precisioni × Recalli F1-scorei = 2 × Precisioni + Recalli
Precisioni =
(3) (4) (5) (6)
where T Pi , F Pi , and F Ni represent true positives, false positives, and false negatives for class i, respectively. For multiclass problems, we report the macro-average of these metrics across all classes owing to the class imbalance, and thus the need to ensure that larger class with good performance is not going to incorrectly bring the overall metric too high.
Fig. 3: Confusion Matrix for Network-based Detection using Packet-level Data
B. Host-based Detection Results V. R ESULTS After conducting validation experiments based on the data processing pipelines and model training methodology described in Section IV, we hereby present the results of the experiments conducted and analyze the performance of our hybrid approach in comparison with the methods proposed in related works. A. Network-based Detection Results The performance of different classifiers for network-based detection varies depending on whether flow-level or packetlevel data is used. During validation, we found that flow-level network traffic data had better performance and lower prediction time in general. Based on testing with flow-level network traffic data, the XGBoost classifier achieved the highest performance for the NIDS, with an accuracy of 99.99%, precision of 99.98%, recall of 99.96%, F1-score of 99.97% and G-mean of 99.98% with the prediction time 0.706 µs per sample. The G-mean metric, which provides a balanced measure of performance across all classes, is calculated as:
For host events data, the XGBoost classifier achieved the best performance, with an accuracy of 96.60%, precision of 85.25%, recall of 85.58%, and F1-score of 85.31% and Gmean of 92.41%. The host events model showed particularly strong performance in detecting benign activities, cryptojacking, backdoor attacks, and certain DoS attacks like PSHACK Flood, SYN Flood, and SynonymousIP Flood, all with F1scores above 0.96%. For power consumption data, the LGBM classifier achieved an accuracy of 70.34%, with precision of 73.78%, recall of 71.40%, and F1-score of 70.57%. While this performance is lower than the host events model, it excelled at detecting specific attack types, particularly FDIAs (F1-score of 0.96), cryptojacking (F1-score of 0.94), and backdoor attacks (F1score of 0.92). C. Performance of proposed Hybrid IDS When combining the results from network-based and hostbased detection, our Hybrid IDS demonstrated comprehensive coverage of the EVCS threat landscape. The overall HIDS reached an accuracy of 83.47%, precision of 82.18%, recall
A HYBRID INTRUSION DETECTION SYSTEM FOR ELECTRIC VEHICLE CHARGING INFRASTRUCTURE
7
TABLE I: Hybrid IDS Component Performance by Attack Type Detection type Benign Classification FDIA Detection Cryptojacking Detection Backdoor Detection
Precision (%) 99.5 99.0 98.0 93.0
Recall (%) 96.0 93.0 96.0 100
F1-score (%) 97.5 96.0 97.0 96.0
TABLE II: Performance Comparison with Existing Approaches Approach
Accuracy (%)
ElKashlan et al. [4] Basnet et al. [5] NADM in [20] FL-based ADS [21] Li et al. [8] Jiang et al. [17] Benfarhat et al. [9] Proposed NIDS
99.20 99.95 98.00 96.97 97.82 92.48 93.90 99.99
Cumplido et al. [6] Girdhar et al. [7] Warraich & Morsi [15] Sharma et al. [3] Proposed HIDS
95.30 96.80 91.20 89.70 83.47
Precision (%) Recall (%) F1-score (%) Network-based Detection Systems 98.50 99.00 98.70 100 99.80 99.80 99.00 98.00 98.00 97.40 98.40 97.60 98.00 93.13 92.20 94.10 93.30 93.90 99.98 99.96 99.97 Host-based Detection Systems 99.984 92.50 90.80 91.60 90.30 89.10 89.70 82.18 82.00 81.47
of 82.00%, and F1-score of 81.47%, with a G-mean of 87.48%. The network-based component successfully detected reconnaissance, DoS, and network-based attacks, while the host-based component successfully detected host-based attacks such as backdoor (F1-score of 0.96) and cryptojacking (F1score of 0.97), as well as FDIAs (F1-score of 0.96). Table I summarizes the performance of the Hybrid IDS components by attack type, highlighting the precision, recall, and F1-score for each detection category. As indicated, the system demonstrated high performance in distinguishing between benign and malicious activities, further solidifying its effectiveness in enhancing the security of the EVCS ecosystem. D. Comparative Analysis The performance comparison with existing approaches is summarized in Table II. While direct comparison is challenging due to differences in datasets and evaluation methods, our approach shows competitive or superior performance across different attack categories. Compared to existing approaches that focus solely on either network or host-based detection, our Hybrid IDS offers several advantages: (1) More comprehensive coverage of attack vectors by monitoring both network traffic and host activities; (2) Higher detection accuracy for specific attack types by utilizing data specific to EVCS environments; (3) Ability to detect a wider range of attacks, including those that might evade single-source detection; and (4) Better context awareness by correlating information from multiple sources. A notable recent contribution in this field is the advanced Temporal Convolutional Network framework proposed by Benfarhat et al. [9]. Their MRG-ID-SA-TCN model achieves impressive performance on the same CICEVSE2024 dataset, with an overall accuracy of 93.9%, precision of 94.1%, recall of 93.3%, and F1-score of 93.9% for the 17-class problem (including benign and 16 attack types). Their approach excels at detecting high-priority attacks like ICMP flood, SYN flood,
Dataset
Attack Types
IoT23 CICIDS2017 EVCS-specific EVCS-specific CICEVSE2024 CICEVSE2024 CICEVSE2024 CICEVSE2024
Network-based Network-based Network-based Network-based Network-based Network-based Network-based Network-based
ACN, ElaadNL Simulated Simulated Power data CICEVSE2024
Host-based Host-based Cyber-physical Host-based Host-based
and Backdoor attacks with perfect detection in most cases. However, it shows some limitations with less common attacks such as Vulnerability Scan and Service Version Detection. In contrast, our hybrid approach demonstrates better performance for these specialized attack types. Our comprehensive evaluation results demonstrate the effectiveness of our hybrid IDS approach in detecting a wide variety of attacks against EVCS infrastructure. Comparative analysis shows that our system consistently outperforms existing approaches. Specifically, our proposed approach achieves superior performance across all metrics, particularly for specialized attacks that target the unique characteristics of EVCS environments. VI. C ONCLUSION AND F UTURE W ORK In this paper, we proposed a Hybrid IDS for EVCSs that combines network-based and host-based detection to provide comprehensive security monitoring across both cyber and physical layers. Leveraging the CICEVSE2024 dataset, which contains network traffic, host events, and power consumption data from actual EVCS environments, our system demonstrated high detection accuracy for various attack types. The evaluation results show that the NIDS component of the proposed Hybrid-IDS achieves 99.99% accuracy for networkbased attacks, while the HIDS component of the proposed Hybrid-IDS achieves 83.47% accuracy for host-based attacks. The HIDS demonstrated exceptional performance for specific attack types, particularly FDIAs (F1-score of 0.96), backdoor (F1-score of 0.96), and cryptojacking (F1-score of 0.97). These results demonstrate the effectiveness of our approach in detecting and classifying a wide range of attacks targeting EVCSs. Using the dual-layer integration, this approach faciliates intrusion detection at the network-level and the host-level. While recent works such as Benfarhat et al. [9] have achieved good performance (93.9% accuracy) using advanced deep learning approaches like temporal convolutional networks, our hybrid approach provides more comprehensive coverage by monitoring both network traffic and host activities.
A HYBRID INTRUSION DETECTION SYSTEM FOR ELECTRIC VEHICLE CHARGING INFRASTRUCTURE
Similarly, the multi-view graph contrastive learning approach proposed by Li et al. [8] shows promise for network-level detection but lacks the host-level insights that our hybrid approach provides. The early detection framework for cyber-physical attacks presented by Warraich and Morsi [15] and the AI-augmented architecture proposed by Sharma et al. [3] offer valuable perspectives on microgrid-specific and architectural considerations, respectively, which could be integrated into future iterations of our system. There are several directions for future work. First, real-time detection capabilities could be enhanced by optimizing feature extraction and classification processes. Second, the correlation between different data sources could be further explored to improve the detection accuracy and reduce false positives. Third, automated response mechanisms could be integrated to provide immediate mitigation of detected attacks. Finally, the framework could be extended to incorporate data from additional sources, such as EV charging behavior data, to enhance the detection of sophisticated attacks targeting the entire EVCS ecosystem. R EFERENCES [1] International Energy Agency, “Global stock of public charging points by speed, 2018-2024,” https://www.iea.org/data-and-statistics/charts/ global-stock-of-public-charging-points-by-speed-2018-2024, 2025. [2] S. Acharya, H. A. U. Khan, R. Karri, and Y. Dvorkin, “MaDEVIoT: Cyberattacks on EV charging can disrupt power grid operation,” in 2024 IEEE Power & Energy Society Innovative Smart Grid Technologies Conference (ISGT), 2024, pp. 1–5. [3] A. Sharma, S. Rani, and M. Shabaz, “Artificial intelligenceaugmented smart grid architecture for cyber intrusion detection and mitigation in electric vehicle charging infrastructure,” Scientific Reports, vol. 15, p. 21653, 2025. [Online]. Available: https: //doi.org/10.1038/s41598-025-04984-4 [4] M. ElKashlan, M. S. Elsayed, A. D. Jurcut, and M. Azer, “A machine learning-based intrusion detection system for iot electric vehicle charging stations (evcss),” Electronics, vol. 12, no. 4, 2023. [Online]. Available: https://www.mdpi.com/2079-9292/12/4/1044 [5] M. Basnet and M. Hasan Ali, “Deep learning-based intrusion detection system for electric vehicle charging station,” in 2020 2nd International Conference on Smart Power & Internet Energy Systems (SPIES), 2020, pp. 408–413. [6] D. Kern, C. Krauß, and M. Hollick, “Detection of anomalies in electric vehicle charging sessions,” in Proceedings of the 39th Annual Computer Security Applications Conference, ser. ACSAC ’23. New York, NY, USA: Association for Computing Machinery, 2023, p. 298–309. [Online]. Available: https://doi.org/10.1145/3627106.3627127 [7] M. Girdhar, J. Hong, H. Lee, and T.-J. Song, “Hidden markov modelsbased anomaly correlations for the cyber-physical security of ev charging stations,” IEEE Transactions on Smart Grid, vol. 13, no. 5, pp. 3903– 3914, 2022. [8] Y. Li, G. Chen, and Z. Dong, “Multi-view graph contrastive representative learning for intrusion detection in ev charging station,” Applied Energy, vol. 385, p. 125439, 2025. [9] I. Benfarhat, V. T. Goh, C. L. Siow, I. E. Lee, M. Sheraz, E. E. Ngu, and T. C. Chuah, “Advanced temporal convolutional network framework for intrusion detection in electric vehicle charging stations,” IEEE Open Journal of Vehicular Technology, 2025. [10] E. D. Buedi, A. A. Ghorbani, S. Dadkhah, and R. L. Ferreira, “Enhancing ev charging station security using a multi-dimensional dataset: Cicevse2024,” in Data and Applications Security and Privacy XXXVIII, A. L. Ferrara and R. Krishnan, Eds. Cham: Springer Nature Switzerland, 2024, pp. 171–190. [11] E. M. Hutchins, M. J. Cloppert, R. M. Amin et al., “Intelligence-driven computer network defense informed by analysis of adversary campaigns and intrusion kill chains,” Leading Issues in Information Warfare & Security Research, vol. 1, no. 1, p. 80, 2011.
8
[12] M. Zhdanova, J. Urbansky, A. Hagemeier, D. Zelle, I. Herrmann, and D. Höffner, “Local power grids at risk – an experimental and simulation-based analysis of attacks on vehicle-to-grid communication,” in Proceedings of the 38th Annual Computer Security Applications Conference, ser. ACSAC ’22. New York, NY, USA: Association for Computing Machinery, 2022, p. 42–55. [Online]. Available: https://doi.org/10.1145/3564625.3568136 [13] E. Gumrukcu, A. Arsalan, G. Muriithi, C. Joglekar, A. Aboulebdeh, M. Alparslan Zehir, B. Papari, and A. Monti, “Impact of cyber-attacks on ev charging coordination: The case of single point of failure,” in 2022 4th Global Power, Energy and Communication Conference (GPECOM), 2022, pp. 506–511. [14] E. Gümrükcü, C. Joglekar, G. Muriithi, A. Arsalan, A. Aboulebdeh, B. Papari, A. Zehir, F. Ponci, and A. Monti, “Dynamic capacity sharing for cyber–physical resilience of EV charging,” Energies, vol. 17, no. 24, p. 6277, 2024. [15] Z. Warraich and W. Morsi, “Early detection of cyber–physical attacks on fast charging stations using machine learning considering vehicle-to-grid operation in microgrids,” Sustainable Energy, Grids and Networks, vol. 34, p. 101027, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2352467723000358 [16] S. Garcia, A. Parmisano, and M. J. Erquiaga, “IoT-23: A labeled dataset with malicious and benign IoT network traffic (Version 1.0.0),” https: //doi.org/10.5281/zenodo.4743746, 2020, zenodo. [Data set]. [17] L. Jiang, Q. li, X. Che, and X. Chen, “A knowledge distillation enhanced semi-supervised federated learning framework for intrusion detection in ev charging networks,” IEEE Internet of Things Journal, vol. 12, no. 16, pp. 34 360–34 373, 2025. [18] Z. J. Lee, T. Li, and S. H. Low, “ACN-Data: Analysis and Applications of an Open EV Charging Dataset,” in Proceedings of the Tenth International Conference on Future Energy Systems, ser. e-Energy ’19, Jun. 2019. [19] ElaadNL, “ElaadNL Open EV Charging Transactions,” https://platform. elaad.io/download-data/, 2019, accessed: 2025-07-16. [Data set]. [20] E. D. Buedi, “Enhancing ev charging station security: A multi-stage approach,” 2024, master’s thesis. [21] S. Purohit and M. Govindarasu, “Fl-evcs: Federated learning based anomaly detection for ev charging ecosystem,” in 2024 33rd International Conference on Computer Communications and Networks (ICCCN), 2024, pp. 1–9.