Cybersecurity Challenges and Intrusion Detection

Systems for Electric Vehicles

A Comprehensive Report based on Recent Research

Generated for User

Based on Research by Singh et al. and Al Isawi et al.

Abstract—As the automotive industry transitions towards B. The Core Problem: Connectivity vs. Security
Electric Vehicles (EVs), the integration of complex electronic
To facilitate communication between ECUs, the auto-
systems and connectivity features has introduced significant
cybersecurity vulnerabilities. This report synthesizes findings motive industry standardized the Controller Area Network
from two pivotal research papers: ”Intrusion Detection System (CAN) bus protocol. While robust and efficient for real-
for Smart Vehicles” by Singh and Agarwal, and ”Electric time control, the CAN bus was designed in an era before
Vehicles CAN Bus Cyber Attacks Detection Using Adaptive vehicle connectivity was a primary concern. Consequently,
Neuro Fuzzy Inference System” by Al Isawi et al.
communication over CAN is unencrypted and unauthen-
The report provides an in-depth analysis of the Controller
Area Network (CAN) bus protocol, highlighting its inherent ticated, making it prone to malicious attacks [1].
lack of encryption and authentication which exposes EVs to As EVs become increasingly integrated into the Internet
Denial of Service (DoS), Impersonation, and Fuzzy attacks. of Things (IoT) ecosystem—connecting with charging
It examines the ”CIA” (Confidentiality, Integrity, Availability) infrastructure, smartphones, and traffic networks—the
triad in the context of vehicular security and explores various interconnected nature of components exposes them to
attack vectors including physical and remote access points.
Furthermore, this document details the methodology and
potential cyber threats [1]. The vulnerabilities are not
performance of modern Intrusion Detection Systems (IDS). limited to the vehicle itself but extend to the charging
It compares standard Machine Learning algorithms (Decision infrastructure, where protocols like the Open Charge Point
Trees, Random Forest) against advanced Deep Learning and Protocol (OCPP) also face security challenges [2].
Hybrid approaches. Special attention is given to the Adaptive
Neuro-Fuzzy Inference System (ANFIS) combined with Sub- C. Objectives of this Report
tractive Clustering (SC) and Fuzzy C-Means (FCM), which This report aims to consolidate the findings of two
has demonstrated superior accuracy (99.68%) and real-time
processing capabilities. The report also discusses data balanc- distinct studies to provide a comprehensive overview of:
ing techniques such as the Synthetic Minority Oversampling 1) The architectural vulnerabilities of the CAN bus
Technique (SMOTE) and their critical role in mitigating class system.
imbalance in cybersecurity datasets. 2) The specific types of cyber-attacks threatening EV
Index Terms—Electric Vehicles, CAN Bus, Intrusion Detec- integrity, confidentiality, and availability.
tion System, Cybersecurity, Machine Learning, ANFIS.
3) The design and implementation of Intrusion Detec-
tion Systems (IDS).
I. Introduction 4) A comparative analysis of Machine Learning (ML)
A. The Rise of Electric Vehicles and Smart Systems and Adaptive Neuro-Fuzzy (ANFIS) techniques for
detecting these attacks.
Electric Vehicle Systems (EVS) are rapidly gaining
popularity and adoption due to their environmental II. The Controller Area Network (CAN) Architecture
benefits and technological advancements. Modern EVs A. Historical Context and Standardization
are no longer just mechanical modes of transport; they The Controller Area Network (CAN) was developed
are sophisticated cyber-physical systems equipped with by Robert Bosch GmbH in 1986. It was designed as a
advanced features such as Adaptive Cruise Control (ACC), serial bus message-broadcast system based on the Carrier
automatic parking, emergency braking, lane-keep assist, Sense Multiple Access with Collision Detection (CSMA-
and blind-spot monitors [1]. CD) protocol [1]. In 1993, the International Organization
These critical functions are controlled by Electronic for Standardization (ISO) adopted it as a standard (ISO
Control Units (ECUs). An ECU is an embedded device 11898) [1].
that provides control signals to different components of Today, CAN is the most prominently used intra-vehicle
an electric vehicle. A modern smart vehicle may contain communication network. It is preferred over other proto-
dozens of these ECUs, which not only control individual cols (such as Ethernet, FlexRay, or LIN) for many critical
components but must also communicate with each other control tasks because it eliminates the need for complex
to share necessary information (e.g., engine speed, battery and expensive dual-port infrastructures and provides high
status) to ensure the smooth functioning of the vehicle [1]. reliability [2].
B. Technical Specifications 2) Integrity: Protecting data from being altered by
The CAN bus is a high-speed, two-wire, half-duplex unauthorized parties. Attacks that inject false data
communication system. It functions as a central net- into the CAN bus violate system integrity.
working system where ECUs are embedded via the CAN 3) Availability: Ensuring that vehicle services and data
interface [2]. are available when needed. Denial of Service (DoS)
attacks specifically target this by flooding the bus
1) Message Format (11-bit Mode): Data on the CAN
[2].
bus is transmitted in frames. The standard 11-bit message
format is critical for understanding both the operation of
the vehicle and the nature of cyber-attacks. The key fields B. Attack Surfaces and Entry Points
are defined as follows [2]:
Attackers can infiltrate the EV network through two
• Start of Frame (SOF): A single bit (1 bit) that signals primary mediums [1]:
to all ECUs that a message transmission has started.
1) Physical Access: Attackers can gain direct access
CAN ID (11 bits): This field serves two purposes:
to the vehicle’s internal network via physical ports. The
•

1) It identifies the content of the message (e.g., most common entry point is the OBD-II (On-Board
engine RPM). Diagnostics) port, which provides direct access to the CAN
2) It determines the priority of the message. In bus. USB ports in the infotainment system can also be
CAN arbitration, a lower CAN ID value corre- exploited.
sponds to a higher priority. This feature is often 2) Remote Access: Modern EVs are connected vehicles.
exploited in Denial of Service (DoS) attacks. Attackers can exploit wireless interfaces such as:
• Data Length Code (DLC): A 4-bit field that deter-
Wi-Fi and Bluetooth connections.
mines the byte length of the data payload. The values
•
Telematics units.
range between 0 and 8 bytes.
•
Companion smartphone applications.
Data Field: This contains the actual payload (0 to 8
•
•
Keyless entry systems.
bytes) intended for the receiving ECUs.
•

• CRC Field (16 bits): The Cyclic Redundancy Check Once an attacker gains entry (e.g., via a compromised
is used to detect transmission errors. infotainment unit), they can pivot to the critical CAN
• ACK Field (2 bits): The Acknowledge field confirms network to launch active attacks.
that the intended ECUs have received the message.
• End of Frame (EOF): A 7-bit field marking the C. Classification of Cyber Attacks
conclusion of the CAN message.
1) Denial of Service (DoS) Attack: In a DoS attack, the
C. Inherent Vulnerabilities attacker floods the CAN bus with high-frequency, high-
Despite its widespread use, the CAN bus lacks intrinsic priority messages. By exploiting the arbitration mecha-
security measures. Specifically, it has no support for: nism—where lower IDs have higher priority—an attacker
can inject messages with a CAN ID of 0x000 (the highest
• Authentication: There is no mechanism to verify the priority). This prevents legitimate ECUs from transmit-
identity of the sender. Any node can masquerade as ting their data, causing traffic congestion and effectively
another. disabling vehicle functions [2]. This attack targets the
• Encryption: Data is broadcast in plaintext, allowing Availability of the system.
any node on the bus to sniff sensitive information.
2) Impersonation (Masquerade) Attack: An imperson-
• Message Verification: There is no built-in protection
ation attack involves an attacker stopping a targeted ECU
against replay attacks or message modification [2].
from transmitting and then ”impersonating” it by sending
These limitations mean that if an attacker compromises messages with the targeted ECU’s ID but with malicious
a single ECU, they can broadcast malicious messages that data payloads.
are treated as legitimate by the rest of the network.
• Target: Confidentiality and Integrity.
III. Cyber Threat Landscape for EVs • Mechanism: The receiving nodes believe the messages
originate from a trusted source and act on false
A. The CIA Triad in EV Security information (e.g., indicating the battery is full when
The security goals for EVs are often described using the it is empty) [2].
Confidentiality, Integrity, and Availability (CIA) model [2]: 3) Fuzzy Attack: In a fuzzy attack, the adversary
1) Confidentiality: Ensuring that sensitive data (e.g., injects messages with random CAN IDs and random data
user location, charging credentials) is accessible only values. This serves to observe the vehicle’s response to
to authorized entities. Attacks like eavesdropping or unexpected input and can cause unpredictable behavior
impersonation target this pillar. or faults in the ECUs [1].
 4) Active Attack Variants: Paper 1 further classifies 3) Data Labeling and Balancing (SMOTE): One of
active attacks into three categories [1]: the major challenges in IDS development is Class Im-
• Fabrication Attack: Sending fabricated messages with balance. In a real-world scenario, ”Normal” traffic vastly
a forged ID to trick the receiver. outnumbers ”Attack” traffic. If left unaddressed, models
• Suspension Attack: Preventing a legitimate ECU from will overlearn the majority class.
sending messages. To solve this, Paper 2 utilized the Synthetic Minority
• Masquerade Attack: A complex combination where Oversampling Technique (SMOTE). SMOTE generates
the attacker suspends the legitimate node and sends synthetic samples for the minority classes (DoS and
fabricated messages in its place. Impersonation) by interpolating between existing samples
in the feature space.
D. Charging Infrastructure Vulnerabilities • Parameter: Paper 2 used k = 30 nearest neighbors
The security of EVs extends to Electric Vehicle Charg- for SMOTE.
ing Stations (EVCS). The Open Charge Point Protocol • Result: The dataset was balanced such that all classes
(OCPP) governs communication between the charger and (Normal, DoS, Impersonation) had an equal number
the central server. Research indicates that vulnerabilities of samples (14,036 samples each in the training set)
in OCPP can lead to Man-in-the-Middle (MitM) attacks. [2].
Furthermore, web interfaces for chargers (e.g., Schneider
Electric EV-Link) and smartphone apps have been found V. Machine Learning Approaches for IDS
to contain flaws allowing SQL injection or credential
bypassing [2]. Paper 1 focused on comparing standard Machine Learn-
ing (ML) algorithms for detecting intrusions.
IV. Intrusion Detection Systems (IDS): Design and
Methodology
A. Algorithms Evaluated
To mitigate the risks outlined above, an Intrusion
The authors of Paper 1 implemented and compared the
Detection System (IDS) is essential. An IDS monitors
following algorithms:
network traffic to detect abnormal or suspicious activities.
• Decision Tree (DT): A tree-like model of decisions. It
A. General IDS Architecture splits data into branches to form a hierarchy of rules.
A typical IDS architecture for smart vehicles involves • Random Forest (RF): An ensemble learning method
several stages, as outlined in Paper 1 [1]: that constructs a multitude of decision trees at
training time and outputs the class that is the mode
of the classes of the individual trees.
• Extra Tree (XT): Extremely Randomized Trees, sim-
ilar to Random Forest but with more randomness in
[Figure Representation: IDS Pipeline] splitting.
• XGBoost (XGB): Extreme Gradient Boosting, a scal-
Data Collection → Preprocessing → Feature able implementation of gradient tree boosting.
Extraction →
Model Training → Anomaly Detection → Alert B. Performance Analysis (Paper 1)
Generation
The models were evaluated based on Accuracy, Preci-
sion, Recall, and F1 Score.

Fig. 1. The end-to-end pipeline of an Intrusion Detection System TABLE I

[1]. Comparison of ML Algorithms from Paper 1 [1]

1) Data Collection: Data is gathered from various Algorithm Acc Prec Recall F1
Decision Tree 0.99 0.99 0.99 0.99
sources, including sensors, network taps, and device logs. Random Forest 0.99 0.99 0.99 0.99
This data includes packet details such as CAN IDs, DLC, Extra Tree 0.98 0.98 0.98 0.98
and payloads. XGBoost 0.98 0.98 0.98 0.98
2) Preprocessing and Feature Extraction: Raw data
must be cleaned (handling missing values) and normal- Findings: Both Decision Tree and Random Forest ex-
ized. Feature extraction is crucial to select relevant data celled, achieving near-perfect scores of 0.99 across all
points (e.g., timestamp, ID) to avoid unnecessary jitter metrics. This indicates that for the specific dataset used,
during model training. Techniques like forward/backward standard tree-based ensemble methods are highly effective
selection or ANOVA are used to refine the feature set [1]. at distinguishing benign traffic from malicious traffic.
 VI. Adaptive Neuro-Fuzzy Inference System (ANFIS) B. Multiclass Classification Results
Paper 2 proposed a more advanced, hybrid approach: The models were tested on their ability to classify data
the Adaptive Neuro-Fuzzy Inference System (ANFIS). into three categories: DoS, Impersonation, and Normal.
This technique combines the learning capabilities of Ar-
tificial Neural Networks (ANN) with the reasoning capa-
TABLE II
bilities of Fuzzy Logic. Multiclass Classification Performance Comparison [2]
A. ANFIS Architecture
Algorithm Acc Prec Recall F1
The ANFIS structure consists of five distinct layers [2]: CNN [2] 82.63% 0.480 0.515 0.497
1) Layer 1 (Fuzzification): Inputs (x and y) are fuzzified CNN-LSTM [2] 86.45% 0.484 0.436 0.459
by determining membership function values (e.g., A GRU [2] 99.50% 0.979 0.985 0.982
and B). LSTM [2] 99.55% 0.977 0.992 0.984
ANFIS-SC 99.65% 0.978 0.999 0.988
2) Layer 2 (Product): This layer multiplies the input
ANFIS-FCM 99.68% 0.979 0.999 0.989
signals to determine the firing strength (weights) of
each rule.
3) Layer 3 (Normalization): The firing strengths are 1) Analysis of Results:
normalized (calculating the ratio of the i-th rule’s fir- • Accuracy Dominance: The ANFIS-FCM model
ing strength to the sum of all rules’ firing strengths). achieved the highest accuracy of 99.68%, surpassing
4) Layer 4 (De-fuzzification): This layer calculates the all Deep Learning models.
contribution of each rule to the final output. • Reliability: With an F1 score of 0.989, ANFIS-FCM
5) Layer 5 (Output): The single node in this layer sums demonstrated superior balance between precision and
all incoming signals to compute the final crisp output recall.
(Z). • Deep Learning Flaws: The CNN and CNN-LSTM
B. Clustering Techniques models performed poorly (accuracies below 90%),
likely due to the specific nature of CAN bus data
To optimize the ANFIS structure and generate the
which may not benefit from the spatial feature ex-
initial Fuzzy Inference System (FIS), Paper 2 integrated
traction strengths of standard CNNs without careful
two clustering algorithms:
tuning.
1) Subtractive Clustering (SC): SC is a fast, one-pass
algorithm for estimating the number of clusters in a
dataset. C. Computational Efficiency
• Parameters: Cluster range of influence set to 0.20 For an IDS to be effective in a vehicle, it must operate
(binary) and 0.25 (multiclass). in real-time. Paper 2 analyzed the computational time for
• Function: It determines the cluster centers which are training and testing:
then used to define the fuzzy rules. • Testing Time (ANFIS-SC): 0.150 seconds.
2) Fuzzy C-Means (FCM): FCM allows data points • Testing Time (ANFIS-FCM): 0.210 seconds.
to belong to multiple clusters with varying degrees of • Testing Time (LSTM): 0.467 seconds.
membership.
Conclusion: The ANFIS models are significantly faster
• Parameters: Number of clusters set to 10; Maximum
during the testing (detection) phase compared to recur-
iterations set to 100. rent neural networks like LSTM and GRU. This speed
• Function: It refines the cluster centers iteratively to
advantage, combined with higher accuracy, makes ANFIS
minimize the objective function. a more suitable candidate for deployment on resource-
VII. Deep Learning vs. ANFIS: A Comparative Study constrained ECUs [2].
Paper 2 provided a rigorous comparison between their
proposed ANFIS models (ANFIS-SC and ANFIS-FCM) VIII. Critical Analysis and Discussion
and state-of-the-art Deep Learning models found in the A. The Importance of Data Balancing
literature.
A key insight from Paper 2 is the necessity of SMOTE.
A. Deep Learning Baselines Real-world attack data is sparse. Without oversampling,
The authors compared against: classifiers tend to be biased toward the ”Normal” class.
• Convolutional Neural Network (CNN) The high performance metrics observed in the ANFIS
• CNN-LSTM: A hybrid of CNN and Long Short-Term study are partly attributable to the effective use of
Memory networks. SMOTE to create a balanced training environment, en-
• GRU: Gated Recurrent Units. suring the model learns the distinct features of DoS and
• LSTM: Long Short-Term Memory networks. Impersonation attacks equally well.
B. Model Selection: Tree-based vs. Neuro-Fuzzy [2] O. A. Al Isawi, K. A. Al Jaafari, and A. S. Al-Sumaiti, “Electric
Vehicles CAN Bus Cyber Attacks Detection Using Adaptive
Paper 1 showed that Random Forest is highly effective Neuro Fuzzy Inference System,” IEEE Access, vol. 13, pp.
(99% accuracy). Paper 2 showed that ANFIS is also highly 90862–90874, 2025.
effective (99.68% accuracy).
• Simplicity vs. Adaptability: Random Forests are gen-
erally easier to implement and interpret. However,
ANFIS offers the benefit of fuzzy logic, which can
handle uncertainty and imprecise data better than
crisp decision boundaries.
• Hybrid Power: The success of ANFIS lies in its ability
to map non-linear relationships using neural network
training while maintaining the interpretability of
fuzzy rules.

C. Limitations
While the results are promising, both papers rely on
specific datasets (e.g., OTIDS). The real-world variability
of CAN traffic (driven by different driving styles, road
conditions, and vehicle models) may introduce noise that
was not fully captured in the experimental setups. Further-
more, while detection is solved, the automated mitigation
(stopping the car safely) remains a complex engineering
challenge.

IX. Conclusion and Future Directions

The electrification of the automotive industry has
brought cybersecurity to the forefront of vehicle design.
The Controller Area Network (CAN), while efficient, is
fundamentally insecure, lacking encryption and authenti-
cation. This report has highlighted that:
1) Vulnerabilities are Systemic: From the lack of CAN
security to vulnerabilities in charging infrastructure
(OCPP), the threat surface is vast.
2) Attacks are Diverse: Vehicles face DoS attacks tar-
geting availability, and Impersonation attacks tar-
geting confidentiality and integrity.
3) AI is the Defense: Intrusion Detection Systems (IDS)
powered by Artificial Intelligence are the most viable
defense.
4) ANFIS is Superior: The Adaptive Neuro-Fuzzy In-
ference System (specifically ANFIS-FCM) demon-
strated the best performance, achieving 99.68% ac-
curacy with low computational cost, outperforming
traditional Deep Learning models.
Future Work: Future research should focus on opti-
mizing these models using metaheuristic algorithms (e.g.,
Particle Swarm Optimization) as suggested by Al Isawi et
al., and validating these systems on hardware-in-the-loop
(HIL) simulations to ensure they can operate under the
strict latency constraints of a moving vehicle.

References
[1] N. Singh and R. Agarwal, “Intrusion Detection System for
Smart Vehicles,” Department of Information Technology, Delhi
Technological University, Delhi, India.