Data-Driven Intrusion Detection in Vehicles: Integrating Unscented

Kalman Filter (UKF) with Machine Learning

Shuhao Bian

, Milad Farsi

, Nasser L. Azad

and Chris Hobbs

Systems Design Engineering Dep., University of Waterloo, 200 University Ave W, Waterloo, Canada

{s6bian, mfarsi, nlashgarianazad}@uwaterloo.ca, cwlh@farmhall.ca

Keywords:

UKF, Machine Learning (ML), Cyber–Physical System (CPS), Advanced Driver Assistance Systems (ADAS).

Abstract:

In the realm of Cyber–Physical System (CPS), accurately identifying attacks without detailed knowledge of

the system’s parameters remains a major challenge. When it comes to Advanced Driver Assistance Systems

(ADAS), identifying the parameters of vehicle dynamics could be impractical or prohibitively costly. To tackle

this challenge, we propose a novel framework for attack detection in vehicles that effectively addresses the

uncertainty in their dynamics. Our method integrates the widely used Unscented Kalman Filter (UKF), a well-

known technique for nonlinear state estimation in dynamic systems, with machine learning algorithms. This

combination eliminates the requirement for precise vehicle modeling in the detection process, enhancing the

system’s adaptability and accuracy. To validate the efﬁcacy and practicality of our proposed framework, we

conducted extensive comparative simulations by introducing Denial of Service (DoS) attacks on the vehicle

systems’ sensors and actuators.

1 INTRODUCTION

Cyber–Physical System (CPS) refers to the integra-

tion of computation with physical processes. There-

fore, cyber attacks on these systems can cause se-

vere consequences. The reliable operation of Ad-

vanced Driver Assistance Systems (ADAS) depends

on the accurate functioning of various sensors and

power management systems. If these elements are

targeted by malicious attackers, passengers, pedestri-

ans and drivers could be exposed to signiﬁcant safety

risks, potentially endangering their lives. This situa-

tion underscores the critical necessity for attack detec-

tion methods within autonomous driving systems. By

identifying potential threats as they occur, the system

can initiate appropriate protective actions to safeguard

passengers, pedestrians and drivers.

Denial of Service (DoS) attack is one of the most

well-known cyber attacks, and it has become more

prevalent since 2004 (Mirkovic and Reiher, 2004).

These attacks purposefully ﬂood networks with too

much trafﬁc, overwhelming systems and compromis-

ing service availability. DoS attacks can cause sig-

https://orcid.org/0009-0000-7518-0972

https://orcid.org/0000-0002-5909-4377

https://orcid.org/0000-0003-1412-7961

https://orcid.org/0009-0007-6175-5449

niﬁcant operational disruptions in Autonomous Ve-

hicles (AVs) and Connected Autonomous Vehicles

(CAVs), which rely heavily on continuous and secure

communication channels for services such as naviga-

tion, real-time trafﬁc updates, and vehicle-to-vehicle

(V2V) communication. The absence of connectivity

not only affects the safety elements essential to AV

operations but also degrades the system’s ability to

make intelligent decisions on the road. Multiple re-

cent studies can be found in the literature, providing

insight into the different classes of attacks and defense

mechanisms developed, such as (Naqvi et al., 2022;

Marcillo et al., 2022; Al-Jarrah et al., 2019; Banaf-

shehvaragh and Rahmani, 2023; Farsi. et al., 2023).

Concerning vehicle security, there exist various

applications of machine learning techniques in the lit-

erature for detecting anomalies in different networks.

One group of approaches considers the possibility of

attacks on communications between vehicles them-

selves and between vehicles and roadside infrastruc-

ture. (Canh and HoangVan, 2023) seeks to build and

evaluate a particular attack detection system that em-

ploys four speciﬁc discriminating features. A col-

lected dataset is then utilized to train and evaluate sev-

eral machine learning and statistical models, allowing

for a comparative examination of their efﬁcacy. The

suggested strategy focuses on early detection, allow-

ing for timely and effective countermeasures.

714

Bian, S., Farsi, M., Azad, N. and Hobbs, C.

Data-Driven Intrusion Detection in Vehicles: Integrating Unscented Kalman Filter (UKF) with Machine Learning.

DOI: 10.5220/0013063900003822

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 21st International Conference on Informatics in Control, Automation and Robotics (ICINCO 2024) - Volume 1, pages 714-723

ISBN: 978-989-758-717-7; ISSN: 2184-2809

Another class of intrusion detection has focused

on in-vehicle communications used for exchang-

ing data between different control units of vehicles.

(Berger et al., 2018) evaluates various machine learn-

ing methods, including deep learning, for in-vehicle

intrusion detection systems. In a more recent tech-

nique, (Aldhyani and Alkahtani, 2022) implemented

deep learning approaches like Convolutional Neural

Network (CNN)s and CNN-Long Short-Term Mem-

ory (LSTM) hybrid models to detect attacks such as

spooﬁng, ﬂooding, and replay attacks on the Con-

troller Area Network (CAN) bus. Other similar tech-

niques can be found in (Pawar et al., 2022). A com-

prehensive survey of the techniques presented, in the

literature, can be found in (Rajapaksha et al., 2023).

The majority of these approaches focus on net-

work trafﬁc and data package analysis to discover re-

curring patterns in normal operation during the train-

ing stage, and then use them to detect anomalies dur-

ing the exploitation stage. Therefore, the dynamic as-

pects of different subsystems in the vehicles receive

less attention. On the other hand, dynamic models

are the foundation of vehicle design, allowing engi-

neers to predict and optimize the performance of nu-

merous vehicle systems under a variety of operating

situations. Exploiting such models for intrusion de-

tection allows us to identify anomalies by continu-

ously comparing real-time data against the predicted

normal behavior. Through this proactive strategy, in-

trusions can be detected early and promptly addressed

to reduce potential damage. (Ju et al., 2022) provides

a review of such techniques from a control perspec-

tive. In line with these approaches, in this study, we

employ a rather system dynamics model-based tech-

nique as a defense mechanism. Moreover, since the

parameters of the vehicle, such as weight, tire con-

ditions, and engine characteristics, can change over

time or under different conditions, we have found ma-

chine learning techniques particularly advantageous

as they offer high adaptability.

The Unscented Kalman Filter (UKF) has been em-

ployed successfully in various ﬁelds of applications,

such as power and automotive systems. In power sys-

tems, the UKF contributes to stability and operational

integrity by accurately evaluating the state of electri-

cal grids and identifying potential interruptions (Du

et al., 2022; Rashed et al., 2022). The UKF has played

an important role in improving cyber threat detection

in-vehicle systems (Zhang et al., 2021), particularly

AVs and CAVs. The UKF improves the system’s ro-

bustness against DoS attacks by enabling real-time,

precise anomaly detection, preventing possible threats

from causing harm. In (

Zivkovi

c and Sari

c, 2018),

the authors employed the UKF to predict and update

state variables from a previously known state to de-

tect False Data Injection (FDI) attacks. They com-

pared the results to those obtained using a common

weighted least squares-based state estimation tech-

nique. They observed that the state variables under

attack signiﬁcantly deviated between them, which can

be used to detect the attack.

In this paper, we describe a novel approach for de-

veloping an attack detection system tailored to vehi-

cles. Our proposed approach integrates the UKF with

a learning-based module to obtain a resilient adaptive

framework. This feature eliminates the requirement

for detailed vehicle modeling in the attack detection

process, simplifying implementation while retaining

accuracy. The framework’s effectiveness is strength-

ened by the use of a Cumulative Sum (CUSUM) algo-

rithm with a sliding window for responsive anomaly

detection, and by incorporating the learned dynam-

ics to predict and compare real-time data against ex-

pected behavior. By leveraging the UKF’s capabilities

in handling non-linear dynamics, our proposed algo-

rithm signiﬁcantly improves the robustness and accu-

racy of intrusion detection in cyber-physical systems,

particularly in ADAS. We conducted extensive sim-

ulations using CARLA, a simulation platform com-

monly used in autonomous driving research, to eval-

uate the efﬁcacy and feasibility of our system. This

made it possible for us thoroughly to test our struc-

ture in practical settings and make sure it can operate

dependably in actual situations.

The rest of the paper is organized in the follow-

ing order: In Section 2, we formulate the problem

by assuming a DoS attack on the actuator. Section 3

presents the attack detection framework, introducing

the main algorithm. In the following Section 4, we

discuss the detailed simulation results.

2 PROBLEM FORMULATION

The system model is described by the following equa-

tions

k+1

= f (x

= h(x

(1)

where x

∈ R

, represents the state space of the vehi-

cle, and u

∈ R

, denotes the control input. y

∈ R

the measurement vector. Moreover, f : R

×R

→ R

represents the state transition function that deﬁnes the

next state x

k+1

based on the current state x

and the

control input u

. The function h : R

→ R

denotes

the measurement function that maps the current state

to the measurement vector y

. k gives the time step.

Data-Driven Intrusion Detection in Vehicles: Integrating Unscented Kalman Filter (UKF) with Machine Learning

715

Next, we will model the DoS attack on various

elements of the control loop.

2.1 Attack Model

Considering that the attacks are implemented in the

cyber layer, we construct the cyber attack in a discrete

space, as shown below.

2.1.1 Actuator Attack

To model the effect of the attack, the dynamics in (1)

are modiﬁed as

k+1

(

f (x

), k ∈ α,

f (x

), k /∈ α,

(2)

where we denote the set of time steps during which

the attack is active using α. Moreover, we assume

that DoS can arbitrarily affect a subset of components

of u

. This can be further described by

[i] =

(

0, i ∈ Γ

[i], otherwise,

(3)

where Γ

denotes the set of indices corresponding to

the actuators that are affected by the attack, and i

ranges over the set Γ

⊆ {1,2,..., p}. Accordingly, a

particular attack strategy can be represented by choos-

ing nonempty sets α and Γ

, which specify the time

of the attack and the indices of targeted actuators, re-

spectively.

2.1.2 Sensor Attack

In this section, we deﬁne the sensor attack in a similar

fashion. We modify the measurements relation given

in equation (1) to reﬂect the DoS attack on some spe-

ciﬁc set of sensors. The equation

(

, k ∈ β,

h(x

), k /∈ β,

(4)

switches the measurements to y

that may be manipu-

lated by the anomaly, at different time steps speciﬁed

by set β. Moreover, based on a particular attack pat-

tern, a list of targeted sensors is given by Γ

, with the

attacked components being set to zero. This is shown

below

[i] =

(

0, i ∈ Γ

[i], otherwise,

(5)

where Γ

⊆ {1,2,...,m} denotes the set of indices

corresponding to the components of y

that are

blocked by the attacker. Here, m refers to the dimen-

sion of the measurement vector.

2.2 Vehicle Model

In order to demonstrate the performance of the devel-

oped approach, we use the Kalman ﬁlter as a base-

line approach for comparison in Section 4. It should

be noted that we aim to obtain a sample-based tech-

nique, meaning that we treat the dynamics as a black

box without using the analytical model. However,

since the dynamic model are essential for running the

Kalman Filter, we concisely present the equations of

the vehicle in what follows.

According to (Takahama and Akasaka, 2018), a

vehicle dynamics model can be obtained as below.

The longitudinal dynamics of the vehicle is given by

vehicle

˙v

= M

vehicle

− r

travel

, (6)

where M

vehicle

is the mass of the vehicle, a

is the

traction force converted to acceleration and r

travel

the travel resistance. The model of the r

travel

can be

described as

travel

= r

air

) + r

roll

) + r

accel

( ˙v

) + r

grad

(θ),

(7)

where r

air

ρC

is the air drag, ρ is the air den-

sity, C

is the drag coefﬁcient, A is the frontal area of

the vehicle, and v

is the vehicle speed. The rolling

resistance is given by r

roll

= C

vehicle

, where C

and g denote the rolling resistance coefﬁcient and the

acceleration due to gravity, respectively.

Then, the acceleration resistance is r

accel

vehicle

˙v

, where ˙v

is the acceleration of the vehicle.

Finally, r

grad

= M

vehicle

gsin(θ), where θ is the slope

angle of the road.

3 ATTACK DETECTION

FRAMEWORK

In this section, we present the required components to

construct the attack detection framework.

3.1 Unscented Kalman Filter

In the UKF, it is not necessary to know the detailed

model. The UKF uses sigma points to sample the

input and obtain the corresponding output. It is a

method similar to the Monte Carlo approach but re-

quires only a small number of sigma points. Next, we

will provide a detailed introduction to the UKF.

3.1.1 Unscented Transformation (UT)

As described by (Wan and Van Der Merwe, 2000),

the UT is a technique developed for the generation of

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

716

Observer/Filter

Environment

Detector

Predict

Select Sigma

Points

Update

CUSUM

Extract Sliding

Window

Standard Error

Figure 1: The ﬂow chart of the algorithm.

sigma points that are capable of undergoing nonlinear

transformations expressed as f (x

). This approach

is particularly valuable when dealing with a multitude

of random vectors, each residing in an n-dimensional

space (x

∈ R

), characterized by a mean ¯x

and a co-

variance matrix P

Manual Control

collect data

Figure 2: CARLA Environment.

3.1.2 Select Sigma Points

The sigma points χ

are chosen such that they capture

the true mean and covariance of the random variable

, enabling an accurate propagation through the non-

linear function f (x

). This selection ensures that

the mean and covariance of the sigma points match

¯x

and P

, respectively, providing an effective mech-

anism for estimating the statistical properties of x

k+1

The sigma points are selected according to











= ¯x

+ (

(n + λ)P

)

i = 1, ...,n,

= ¯x

− (

(n + λ)P

)

i = n + 1, ...,2n,

(m)

n+λ

(c)

n+λ

+ (1 − φ

+ β),

(m)

= W

(c)

2(n+λ)

i = 1, ...,2n.

(8)

According to (Wang et al., 2023), λ is a key pa-

rameter calculated using λ = φ

(n + κ) − n, φ deter-

mines the dispersion degree of σ points, and κ is typi-

cally 3−n. β is for integrating prior knowledge about

x’s distribution.

3.1.3 Predict

The predicted state and covariance are

¯x

k+1|k

∑

i=0

(m)

k+1|k

∑

i=0

(c)



k+1|k

− ¯x

k+1|k



k+1|k

− ¯x

k+1|k



¯y

k+1|k

∑

i=0

(m)

k+1|k

k+1

∑

i=0

(c)



k+1|k

− ¯y

k+1|k



k+1|k

− ¯y

k+1|k



k+1

∑

i=0

(c)



k+1|k

− ¯x

k+1|k



k+1|k

− ¯y

k+1|k



(9)

where χ

k+1|k

= f (χ

), Y

k+1|k

= h(χ

), for i =

0,...,2n.

3.1.4 Update

The update step can be deﬁned as

k+1

= P

k+1



k+1



−1

k+1

= y

k+1

− ¯y

k+1|k

k+1

= ¯x

k+1|k

+ K

k+1

= P

k+1|k

− K

k+1

(10)

where K is the Kalman gain, r is the residual. These

values are updated when sensors provide new mea-

surements.

Data-Driven Intrusion Detection in Vehicles: Integrating Unscented Kalman Filter (UKF) with Machine Learning

717

Data: Sensor signal, Input signal

Result: Attack State: f

(0 for no attack, 1

for attack detected)

Initialisation: ;

Initialize the cumulative sum, s

← 0 ;

Initialize the attack detection ﬂag, f

← 0;

Initialize the threshold value, threshold ;

Deﬁne the window size N ;

Initialize an empty list Q, to store the last N

residuals ;

Detection Loop: ;

while each new sensor signal input y do

Perform the selecting sigma points step;

Perform the prediction step of UKF to

estimate the next state ;

Perform the update step of UKF with

signal y to obtain residual r ;

Append r to Q ;

Update the UKF model // using

equation (11) ;

if Size of Q > N then

Remove the oldest residual from Q ;

end

Update cumulative sum using residuals

within Q: s

←

∑

Q // using

equation (12) ;

Update standard error using residuals

within Q: s

← std(Q) // using

equation (13) ;

if abs(s

) > threshold

and

> threshold

then

← 1 // attack detected

immediately upon detection ;

else

← 0;

Continue monitoring ;

end

Algorithm 1: CUSUM Algorithm with Sliding Window for

Attack Detection.

3.2 Machine Learning (ML)

Obtaining the parameters of a vehicle might often be

infeasible or economically prohibitive. In the update

phase of the UKF, we employ ML techniques to pre-

dict the vehicle’s subsequent state.

For this purpose, we selected the multi-layer per-

ceptron (MLP) network. This neural network receives

control commands (u) — encompassing the throttle-

brake and steering angle — as well as the vehicle’s

current state, which includes velocity, angular veloc-

ity, and acceleration, as its input. The network is de-

signed to output the vehicle’s acceleration for the next

time step. Designed to predict the vehicle’s accelera-

tion at the next time step, it was necessary to account

for the delay between the input commands and the re-

sultant state changes. To address this, according to

(Xu et al., 2019), we selected a model output that re-

ﬂects the vehicle’s acceleration 50 milliseconds after

the input, ensuring the training data adequately cap-

tures the dynamics of the system.

The MLP is structured with three distinct layers:

an input layer comprising 5 units, a hidden layer con-

taining 20 units, and a single-unit output layer. The

activation function utilized within the hidden layer is

the Rectiﬁed Linear Unit (ReLU).

With this predicted acceleration, we can determine

the vehicle’s upcoming velocity using Newton’s sec-

ond law, thereby integrating ML predictions seam-

lessly into the UKF’s update mechanism for enhanced

estimation accuracy.

3.3 Model Update Methods

In order to make the detector adaptive, we deployed

an algorithm that can make the detector update auto-

matically. In the proposed detector, we updated the

model of the vehicle when it got new sensor signals.

To make sure the model is not updated when the at-

tacker implements the DoS attack on the car model

and is updated when the car is running normally, we

update the learning rate according to

l = 1 −

1 + e

−S

rate

∥r∥

. (11)

Here, l denotes the learning rate, r represents the

residual of the car, S

rate

is the scale of the residual,

and ∥ · ∥

denotes the L

norm of a vector.

3.4 Attack Detection Method

As a widely used attack detector in many literature

(Liu et al., 2019), CUSUM is selected as the detec-

tor. In (Van Eykeren et al., 2012), they also calculated

a moving window average for the residual. In prac-

tice, there is often a gap between theoretical models

and engineering applications. Therefore, we intro-

duce W

and W

to ﬁne-tune the target parameters.

Additionally, we used two tests to detect whether the

target is under attack.

Test 1 is shown by

∑

i=k−N+1

, (12)

where r is the residual updated by the UKF, W

∈ R

N is the length of the sliding window.

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

718

Test 2 is described as

∑

i=k−N+1

− ¯r)W

− ¯r)

, (13)

where ¯r is the average of the sliding window for the

residuals, W

∈ R

m×m

is the weight of the residual.

Then s

and s

will be compared with two thresh-

olds, t

and t

, respectively. Then the alarm is trig-

gered according to

A =

(

, s

> t

and s

> t

, s

≤ t

or s

≤ t

(14)

As illustrated in Figure 1, the trained model de-

rived from historical data is subsequently integrated

into the UKF. The mechanism of the detector is de-

tailed in Algorithm 1.

4 SIMULATION RESULTS

To demonstrate the efﬁcacy of the proposed approach

under DoS attacks, we chose CARLA (Dosovitskiy

et al., 2017) as our simulator, which can simulate real-

world dynamics and generate sensor signals in real-

time. This allows for accurate and timely responses

in the simulation environment. Moreover, to highlight

the superiority of the proposed method, we compare

it with the traditional Kalman ﬁlter.

All neural network training was conducted in

Python on the Ubuntu operating system. The hard-

ware conﬁguration used included an AMD Ryzen 9

processor with 16 cores, clocked at 3.40 GHz, and

64GB of RAM. The Kalman ﬁlter and UKF were im-

plemented using the FilterPy library (Labbe, 2024),

while the neural network was developed with PyTorch

(Paszke et al., 2019).

In the following subsection, we detail the data

preprocessing methods, the implementation of attack

scenarios, the simulation parameters and the resulting

simulation outcomes.

4.1 Data Preprocessing

Both manual driving data and autonomous driving

data generate command signals (throttle, brake, and

steering angle), which can be used as training data.

To obtain representative data, we used manual con-

trol to generate data in the CARLA environment, as

shown in Figure 2.

In order to achieve better training results, we need

to ensure that the data are valid by removing all out-

liers.

The brake and throttle inputs are combined and

normalized into a uniﬁed control signal ranging from

0 to 1 according to

u =

T − B + 1

, (15)

where u is the uniﬁed control signal, T ∈ [0, 1] is the

throttle input, B ∈ [0,1] is the brake input, T

max

the maximum throttle value and B

min

is the minimum

brake value. When u = 0, it represents full braking,

and when u = 1, it represents full throttle.

4.2 Attack Implementation

We implemented the attack at 20 seconds in the sim-

ulation. To validate the algorithm, we chose a DoS

attack signal in the form of a Pulse Width Modulation

(PWM) structure. When the signal is 1, it indicates

that the monitor is blocked.

4.3 Parameters

To have a comparison we chose a typical Kalman ﬁl-

ter. The parameters of equation (7) is shown as ta-

ble 1.

Table 1: Parameters of Kalman ﬁlter.

Name Value

air

68.9 N

roll

271.6 N

grad

0 N

If there is no slope, r

grad

= 0. Therefore, the to-

tal travel resistance r

travel

is given by r

travel

= r

air

roll

+ r

grad

= 340.5N. In order to make the detec-

tor more accurate, we have W



1 0.01 0







1 0 0

0 0.01 0

0 0 0





and we focus on the acceler-

ation residual to detect the attack.

The learning parameters are detailed in Table 2.

The optimizer used is Adam (Kingma and Ba, 2014).

Table 2: Parameters of Machine Learning.

Name Value

Learning rate 0.001

Train size/Total 0.8

Epochs 1000

MLP batch size 64

MLP epochs 1000

To improve the clarity of the results, we simpliﬁed

the model, the value of t

is based on t

, maintaining

a ﬁxed proportional relationship as shown by

= γt

, (16)

Data-Driven Intrusion Detection in Vehicles: Integrating Unscented Kalman Filter (UKF) with Machine Learning

719

Figure 3: The results of the attack detector are as follows: for the Kalman ﬁlter, a1 represents the residual and CUSUM of

residuals in the sliding window, a2 denotes the standard error of residuals in the sliding window, and a3 indicates the attack

state; for the proposed method, b1 signiﬁes the residual and CUSUM of residuals in the sliding window, b2 refers to the

standard error of residuals in the sliding window, and b3 denotes the attack state of the UKF.

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

720

where γ is a tuning parameter. In this simulation, γ is

set to 0.04.

4.4 Results

In this study, the performance of the Kalman ﬁlter

and UKF for detecting attacks in autonomous vehi-

cles was evaluated. The experiments were conducted

under two scenarios: actuator attack and sensor at-

tack.

4.4.1 Actuator Attack

To validate the detector, we use four metrics to mea-

sure the performance: false positive alarm rate, true

positive alarm rate, false negative alarm rate and true

negative alarm rate.

As shown in Figure 4, despite the threshold

changes from 10 to 25, the true positive alarm rate

and true negative alarm rate of the proposed method

are signiﬁcantly higher than those of the Kalman ﬁlter

when using the same threshold.

Figure 4: The metrics of different detectors are deﬁned as

follows: the true positive rate (TP), the true negative rate

(TN), the false positive rate (FP), and the false negative rate

(FN).

To achieve a more precise quantiﬁcation of the

research data, we utilized the F1-score (Bishop and

Nasrabadi, 2006). As shown in Figure 5, the F1-score

of the Kalman ﬁlter shows a decreasing trend, while

the F1-score of the UKF ﬁrst increases and then de-

creases. Based on this pattern, we selected the ap-

propriate thresholds for the Kalman ﬁlter and the pro-

posed method. The thresholds are 10 for the Kalman

ﬁlter and 13.33 for the proposed method.

Table 3 presents the optimal thresholds and cor-

responding values for the Kalman ﬁlter and the pro-

posed method. In this table, all metrics of the pro-

posed method signiﬁcantly outperform those of the

Kalman ﬁlter.

Figure 5: F1 Score of Kalman Filter and the Proposed

Method.

Table 3: Optimal Thresholds and Confusion Matrix Values

for Kalman Filter and the Proposed Method.

Metric Kalman ﬁlter Ours

Optimal Threshold 10.0 13.33

F1-Score 0.4784 0.8826

True Positives (TP) 42.88% 93.58%

False Positives (FP) 36.41% 18.48%

True Negatives (TN) 63.59% 81.52%

False Negatives (FN) 57.12% 6.42%

4.4.2 Sensor Attack

Figure 6: The metrics of different detectors are deﬁned as

follows: the true positive rate (TP), the true negative rate

(TN), the false positive rate (FP), and the false negative rate

(FN).

We obtained similar results for Attack Detection (AD)

on the sensor, as shown in Figure 6. Both the true

positive and true negative rates of the proposed UKF

Data-Driven Intrusion Detection in Vehicles: Integrating Unscented Kalman Filter (UKF) with Machine Learning

721

method are signiﬁcantly higher than those of the

Kalman ﬁlter when choosing the best F1-Score.

5 CONCLUSION

In conclusion, we proposed a novel framework for at-

tack detection in vehicles with unknown systems. We

exploited a learning-based model to predict and com-

pare observations against expected behavior. There-

fore, compared to the Kalman ﬁlter, the proposed ap-

proach based on the UKF is capable of detecting DoS

attacks from the sensor and actuator without prior

knowledge of the system parameters. Accordingly,

by exploiting UKF’s capabilities in handling nonlin-

earity, our proposed algorithm demonstrated a signif-

icant advantage over the traditional Kalman ﬁlter for

detecting DoS attacks on sensors and actuators. In

detail, through extensive simulations of our proposed

algorithm, we observed that our method outperforms

the Kalman ﬁlter by demonstrating substantial results

in both true positive alarm rate and true negative alarm

rate. Enhancing the ﬁltering design for vehicle in-

cursion detection can be the primary focus of future

research. By precisely simulating intricate dynam-

ics, investigating robust particle ﬁlters may improve

anomaly identiﬁcation even further and strengthen ve-

hicle security. Moreover, the presented framework

can be extended to other types of attacks such as FDI.

ACKNOWLEDGEMENT

We gratefully acknowledge the ﬁnancial support pro-

vided by BlackBerry and the Natural Sciences and

Engineering Research Council of Canada (NSERC).

Furthermore, we acknowledge BlackBerry QNX’s as-

sistance and cooperation with this study. Their assis-

tance has been crucial to the accomplishment of this

task.

REFERENCES

Al-Jarrah, O. Y., Maple, C., Dianati, M., Oxtoby, D., and

Mouzakitis, A. (2019). Intrusion detection systems

for intra-vehicle networks: A review. Ieee Access,

7:21266–21289.

Aldhyani, T. H. and Alkahtani, H. (2022). Attacks to au-

tomatous vehicles: A deep learning algorithm for cy-

bersecurity. Sensors, 22(1):360.

Banafshehvaragh, S. T. and Rahmani, A. M. (2023). Intru-

sion, anomaly, and attack detection in smart vehicles.

Microprocessors and Microsystems, 96:104726.

Berger, I., Rieke, R., Kolomeets, M., Chechulin, A., and

Kotenko, I. (2018). Comparative study of machine

learning methods for in-vehicle intrusion detection. In

International Workshop on Security and Privacy Re-

quirements Engineering, pages 85–101. Springer.

Bishop, C. M. and Nasrabadi, N. M. (2006). Pattern recog-

nition and machine learning, volume 4. Springer.

Canh, T. N. and HoangVan, X. (2023). Machine learning-

based malicious vehicle detection for security threats

and attacks in vehicle ad-hoc network (vanet) commu-

nications. In 2023 RIVF International Conference on

Computing and Communication Technologies (RIVF),

pages 206–211. IEEE.

Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., and

Koltun, V. (2017). Carla: An open urban driving sim-

ulator.

Du, D., Zhu, M., Li, X., Fei, M., Bu, S., Wu, L., and Li, K.

(2022). A review on cybersecurity analysis, attack de-

tection, and attack defense methods in cyber-physical

power systems. Journal of Modern Power Systems and

Clean Energy, 11(3):727–743.

Farsi., M., Bian., S., Azad., N. L., Shi., X., and Walenstein.,

A. (2023). An efﬁcient resilient mpc scheme via con-

straint tightening against cyberattacks: Application to

vehicle cruise control. In Proceedings of the 20th In-

ternational Conference on Informatics in Control, Au-

tomation and Robotics - Volume 1: ICINCO, pages

674–682. INSTICC, SciTePress.

Ju, Z., Zhang, H., Li, X., Chen, X., Han, J., and Yang, M.

(2022). A survey on attack detection and resilience

for connected and automated vehicles: From vehicle

dynamics and control perspective. IEEE Transactions

on Intelligent Vehicles, 7(4):815–837.

Kingma, D. P. and Ba, J. (2014). Adam: A

method for stochastic optimization. arXiv preprint

arXiv:1412.6980.

Labbe, R. (2024). ﬁlterpy: Kalman ﬁlters and other optimal

and non-optimal estimation ﬁlters in python. https:

//github.com/rlabbe/ﬁlterpy. Accessed: 20 June 2024.

Liu, Q., Mo, Y., Mo, X., Lv, C., Mihankhah, E., and Wang,

D. (2019). Secure pose estimation for autonomous

vehicles under cyber attacks. In 2019 IEEE Intelligent

Vehicles Symposium (IV), pages 1583–1588. IEEE.

Marcillo, P., Tamayo-Urgil

es, D., Valdivieso Caraguay,

A. L., and Hern

andez-

Alvarez, M. (2022). Security

in v2i communications: a systematic literature review.

Sensors, 22(23):9123.

Mirkovic, J. and Reiher, P. (2004). A taxonomy of ddos at-

tack and ddos defense mechanisms. ACM SIGCOMM

Computer Communication Review, 34(2):39–53.

Naqvi, I., Chaudhary, A., and Kumar, A. (2022). A sys-

tematic review of the intrusion detection techniques in

vanets. TEM Journal, 11(2):900.

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,

Chanan, G., Killeen, T., Lin, Z., Gimelshein, N.,

Antiga, L., et al. (2019). Pytorch: An imperative style,

high-performance deep learning library. Advances in

neural information processing systems, 32.

Pawar, Y. S., Honnavalli, P., and Eswaran, S. (2022). Cy-

ber attack detection on self-driving cars using machine

ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics

722

learning techniques. In 2022 IEEE 3rd Global Confer-

ence for Advancement in Technology (GCAT), pages

1–5. IEEE.

Rajapaksha, S., Kalutarage, H., Al-Kadri, M. O., Petrovski,

A., Madzudzo, G., and Cheah, M. (2023). Ai-based

intrusion detection systems for in-vehicle networks: A

survey. ACM Computing Surveys, 55(11):1–40.

Rashed, M., Kamruzzaman, J., Gondal, I., and Islam, S.

(2022). False data detection in a clustered smart

grid using unscented kalman ﬁlter. IEEE Access,

10:78548–78556.

Takahama, T. and Akasaka, D. (2018). Model predictive

control approach to design practical adaptive cruise

control for trafﬁc jam. International Journal of Au-

tomotive Engineering, 9(3):99–104.

Van Eykeren, L., Chu, Q., and Mulder, J. (2012). Sen-

sor fault detection and isolation using adaptive ex-

tended kalman ﬁlter. IFAC Proceedings Volumes,

45(20):1155–1160.

Wan, E. A. and Van Der Merwe, R. (2000). The unscented

kalman ﬁlter for nonlinear estimation. In Proceed-

ings of the IEEE 2000 adaptive systems for signal

processing, communications, and control symposium

(Cat. No. 00EX373), pages 153–158. Ieee.

Wang, C., Chen, L., Li, L., Yan, Y.-H., Sun, J., Yu, C.,

Deng, X., Liang, C.-P., Zhang, X.-L., and Peng, W.-

M. (2023). Combining unscented kalman ﬁlter and

wavelet neural network for anti-slug. Petroleum Sci-

ence, 20(6):3752–3765.

Xu, J., Luo, Q., Xu, K., Xiao, X., Yu, S., Hu, J., Miao, J.,

and Wang, J. (2019). An automated learning-based

procedure for large-scale vehicle dynamics modeling

on baidu apollo platform. In 2019 IEEE/RSJ Interna-

tional Conference on Intelligent Robots and Systems

(IROS), pages 5049–5056. IEEE.

Zhang, D., Lv, C., Yang, T., and Hang, P. (2021). Cyber-

attack detection for autonomous driving using vehi-

cle dynamic state estimation. Automotive Innovation,

4:262–273.

Zivkovi

c, N. and Sari

c, A. T. (2018). Detection of false

data injection attacks using unscented kalman ﬁlter.

Journal of Modern Power Systems and Clean Energy,

6(5):847–859.

Data-Driven Intrusion Detection in Vehicles: Integrating Unscented Kalman Filter (UKF) with Machine Learning

723