This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 1

An Evolutionary Deep Learning-Based Anomaly

Detection Model for Securing Vehicles
Abdollah Kavousi-Fard , Senior Member, IEEE, Morteza Dabbaghjamanesh, Senior Member, IEEE,
Tao Jin , Senior Member, IEEE, Wencong Su, Senior Member, IEEE, and Mahmoud Roustaei

Abstract— This article proposes a deep learning based By penetrating into the car system, the hackers can not only
approach for cyber attack detection in the vehicles. The proposed disturb the car tasks but can also affect the entire smart city
method is constructed based on generative adversarial network due to the vehicle-2-vehicle and vehicle-2-grid plans available
(GAN) classification to assess the message frames transferring
between the electric control unit (ECU) and other hardware in these systems. In a vehicle, the electric control unit (ECU)
in the vehicle. To this end, two networks called generator (G) is in charge of controlling all hardware units through the
and discriminator (D) will run an adversarial game to fool information gathered from many sensors considered. After
each other. In such a process, the most optimal structure is analyzing these info, the ECU launch some comments to
found which distinguish between the model normal behavior control the vehicle in a protocol bed such as CAN, MOST,
and abnormalities. Due to the instabilities existing in the GAN
model, a new optimization method based on firefly algorithm is Ethernet, LIN or FLexray [7]–[9]. CAN protocol due to
proposed to create a class of generators in a feasible region, i.e. special features was the first most widely used protocol in
the discriminator D. A three-stage modification method is also the car industry since 1986 [10], [11]. This protocol uses
devised to increase the algorithm population diversity and reduce some identifiers (IDs) to escape from message traffic in the
the possibility of falling in local optima. The performance of the vehicle. Unfortunately, the CAN was designed at for simple
model is assessed on the experimental dataset recorded from the
OBD-II port of an undefined vehicle. inaccessible vehicles of those years and is not secure in the
modern vehicles which are equipped with many facilities for
Index Terms— Deep learning, generative adversarial networks, the owners. This has become a vulnerable point of the vehicles
controller-area-networks, firefly algorithm.
in recent years which a review on some of the works is
provided in the rest.
I. I NTRODUCTION In [12], a Tojan attacks the CAN bus protocol as a hidden

V EHICLES are considered as critical component of the

human modern life, due to the many benefits providing
for the every daily activities of the societies. Ranging from
cyber attack destroying its performance through a wireless
communication. The most vulnerable points of CAN bus are
discovered in [13] by launching different scenarios of attacking
the fossil fuel based vehicles to hybrid vehicles (both fossil in a vehicle. As a sequence, it can be deduced that these cyber
fuel and electric type) to all electric vehicles, these devices attacks can finally reach to the smart grid and run malicious
are inevitable players which their security can affect not activities. In [12], it is tried to find a defense mechanism
only the human security but also the car owner privacy, for stopping cyber hacking in the vehicle system based on
effectively [1]–[3]. With the growth of the smart city concept classifying. It also provides a very good review on some of
in recent years, many advanced vehicles with entertaining tools the well-known methods cyber recognition approaches in the
and facilities are appeared in the market which can be a very industry. In [14], all CAN messages are first investigated and
good target for hackers to run their malicious purposes [4]–[6]. compromised ones are filtered from approaching the ECU.
Three different architectures of central (with an ECU in the
Manuscript received March 19, 2020; revised June 15, 2020; accepted
July 30, 2020. The Associate Editor for this article was A. Jolfaei. (Cor- center), distributed (several ECUs) and hybrid (both central
responding author: Tao Jin.) ECU and some supplementary ECUs) are introduced for this
Abdollah Kavousi-Fard is with the Department of Electrical and Electronics filtering process. Message flooding cyber attack as a famous
Engineering, Shiraz University of Technology, Shiraz 71557-13876, Iran,
and also with the Department of Electrical Engineering, Fuzhou University, denial of service (DoS) attack in CAN is assessed in [15]. Here
Fuzhou 350116, China (e-mail: kavousi@sutech.ac.ir). two counters are deployed to distinguish between the normal
Morteza Dabbaghjamanesh is with the Department of Electrical and and abnormal message IDs. In [16], the frequency and message
Computer Engineering, The University of Texas at Dallas, Richardson,
TX 75080-3021 USA (e-mail: mortezadabba@utdallas.edu). ID are used as two features for detecting any abnormality in
Tao Jin is with the Department of Electrical Engineering, Fuzhou University, the messages. In the case that a cyber attack has happened,
Fuzhou 350116, China (e-mail: jintly@fzu.edu.cn). the system will inform the car owner and suggest turning off
Wencong Su is with the Department of Electrical and Computer Engineer-
ing, University of Michigan–Dearborn, Dearborn, MI 48128 USA (e-mail: the car for a few minutes. It is suggested in [17] to make use of
wencong@umich.edu). data encryption methods using one of the ECUs to keep the
Mahmoud Roustaei is with the Department of Electrical and Electronics transmitted data secret. Similarly, [18] suggests to measure
Engineering, Shiraz University of Technology, Shiraz 71557-13876, Iran
(e-mail: rousta.mahoud@gmail.com). the message entropy and compare it with the normal status
Digital Object Identifier 10.1109/TITS.2020.3015143 to check the health of the CAN bus message frames. This is
1524-9050 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Cornell University Library. Downloaded on August 27,2020 at 13:08:52 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

a good method which works based on the randomness flat in II. V EHICLE AS A C OMPLEX C YBER -P HYSICAL
the CAN traffic. In [19], it is suggested to install a firewall S YSTEM (CPS)
between the ECU and communicating modules to check the In-vehicle data communication has improved based on the
health of the message frames based on comparison and look-up progress of vehicles facilities and services. This has caused
table. that the domain of vehicles extends from legacy electronics
As it is inferred from the above literature survey, ECU to new software. In fact, the emerging features in the modern
performance and security is highly dependent on the CAN vehicles such as entertainments or driving assistance systems
bus security, without which the entire vehicle performance are based on interconnected software (cyber) and physi-
is affected. Unfortunately, the research in this area is still cal (electronics) systems rather than mechanical components.
at its infancy and requires much more efforts to get to an There are over 100 million lines of coding in a modern vehicle,
acceptable level. This research article focuses on this challenge which clearly shows the complex nature of these CPSs [19].
and proposes a new deep learning based model for assessing In such a CPS, over a hundreds of programmable ECUs exist
the health of message frames within the car. The suggested which communicate on the bed of CAN protocol. Such a
model makes use of the message frame frequency and ID distributed and assorted nature of vehicles makes them an
number to train generative adversarial network (GAN) when appealing target for the cyber hackers, which emphasizes on
offline. This model can be further used for online checking of the necessity of CPS-oriented defense. CAN in a vehicle
the CAN bus message traffic. GAN [20] belongs to the class of represents the nervous system in our body, without which
deep learning models and is composed of two networks, called the entire system collapse. Technically, CAN is designed
generator (G) and discriminator (D). When generator tries to to control over five hundred million chips. Unfortunately,
mislead the G by generating noisy samples, the discriminator the CAN protocol, designed around 30 years ago, does not
tries to improve its training process by comparing the real data provide the required security for the today modern vehicles.
with the fake data. Through such a process, a powerful clas- Ignoring the confidentiality and authentication mechanisms
sification model with very high appealing is achieved which is a natural disadvantage of this protocol which can attract
can be used as an anomaly detection model in the vehicle. hackers.
In order to get to most optimal structure, an evolutionary GAN This may happen in either a wired or wireless way. In the
based on firefly algorithm (FA) is proposed. This can help to wired cyber attack, OBD-II port can let the hackers penetrate
overcome the instabilities existing in the GAN by generating the car. This port, which is located beneath the steering wheel,
a set of generators {G} instead of a single generator. FA is is originally designed for onboard diagnosis but also provides
a meta-heuristic optimization algorithm which mimics the access to the CAN bus. From there, one can read or write some
mating behavior of firefly insects in the tropical regions [21]. coding on the ECU. In the wireless cyber attack, a smartphone
In addition, a three-stage modification method based on the or the car entertaining service can be the penetration point.
powerful math operators of other algorithms is deployed here In order to write a code on the ECU, one first need to know the
which can improve the search ability of FA. To summarize, message frame structure in the CAN bus protocol. Fig. 1 shows
the main paper contributions can be named as follows: the structure of a CAN message in a vehicle.
Each message has an ID which shows its priority, i.e.
• Recommending a new deep anomaly detection model a lower ID value shows a higher priority. Therefore, when
based on generative adversarial network for cyber-attack multiples of messages arrive to the ECU simultaneously,
detection in the vehicle ECU. The proposed method can message arbitration happens which means a message with a
extract out the main features and provide the most reliable lower priority is analyzed first. In Fig. 1, the message frame
results. consists of several parts: one principal bit as the start of the
• Introducing an evolutionary GAN based on FA for over- frame (SOF), 12 bits as arbitration field, six bit as control
coming the instability issues in the model. The proposed field, data field (in range of 0 to 64 bytes), CRC field with
evolutionary GAN improves the stability of the proposed 16-bits, ACK field with 2-bits, and end of frame (EOF) with
classification model. 7-bit. The message arbitration process is a key point for
• Proposing a three-stage modification solution for FA to hackers to cyber attack to the system. In other words, a hacker
avoid premature convergence and increase the population launches several messages with low ID but high frequency to
diversity. wind the arbitration and run his malicious purposes on the
The experimental dataset of an undefined vehicle is deployed vehicle. In order to stop this event, next section provides a
as the case study and assessing the proposed anomaly detection novel deep learning based cyber attack detection model. The
model performance. The rest of this article is organized as ECUs read messages measured by varied range of sensors and
follows: make relevant processing for varying intentions like pedestrian
This article is organized as follows: In section II, the cyber detection, path planning, auto-parking, collision avoidance,
security of vehicle as a complex cyber-physical system (CPS) etc.
is investigated. In section III, the proposed evolutionary deep
learning anomaly detection model based on GAN and modified III. A NOMALY D ETECTION M ODEL BASED ON D EEP
FA (MFA) is explained. The experimental simulation results G ENERATIVE A DVERSARIAL N ETWORKS
are discussed in Section IV. To finish, the main thoughts and In this section, a deep learning based anomaly detection
conclusions are given in Section V. model is proposed for filtering the CAN bus messages.

Authorized licensed use limited to: Cornell University Library. Downloaded on August 27,2020 at 13:08:52 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

KAVOUSI-FARD et al.: EVOLUTIONARY DEEP LEARNING-BASED ANOMALY DETECTION MODEL FOR SECURING VEHICLES 3

Fig. 1. Message frame structure in CAN bus protocol.

dataset PR (x). A signal is used as the error feedback which

returns to the generator and discriminator. In a recursive
process, the generator attempts to generate more similar fake
data to the real dataset x to fool the D network. Such a process
should continue until PG (N) becomes similar to PR (x). The
error feedback is the difference existing between PG (N) and
PR (x). Consider the parameter characteristics of D network
as θ (D), the loss function of the discriminator can be defined
as follows [20]:
 
V D, θ ( D) = −E x∼PR (x) [log D (x)]
 
− E N∼Pg (N) log (1 − D (G (N))) (1)
In the same way, the G network with the parameter character-
istics θ (G) would gain a loss function as below [20]:
   
V G, θ (G) = E N∼Pg (N) log (1 − D (G (N))) (2)
The above formulations show that each network has a spe-
cific objective function which needs to optimize, individually.
This means that the G network is trained by minimizing
Fig. 2. Arrangement of a GAN model. V (G) θ (G) , θ (D) with respect to θ (G) when considering fixed
values for θ (D) and
 the D network is trained when maximizing
Having the normal CAN traffic, the proposed model is trained V (D) D (G) , θ (D) with respect to θ (D) hen considering fixed
when off-line for further usage at on-line. values for the θ (G) . In order to unify these two objectives
under the same shelter, a new loss function can be formulated
A. Deep Learning Based on GAN as a min-max equation of this form [20]:
Technically, GAN is constructed of two neural networks: min max V (D, G) = E x∼PR (x) [log D (x)]
called generator (G) and discriminator (D). In an adversarial G D
 
game, these two networks try to improve their performance + E N∼Pg (N) log (1 − D (G (N))) (3)
based on the response of each other. The generator tries to
In (3), the first term is a maximization operator and the second
produce noise sample for misleading the discriminator. On the
term is a minimization operator. As the above formulation
other hand, the discriminator tries to improve the training
is solved, the generator G and discriminator D networks
process up to a level that can differ fake data from the
converge into the Nash equilibrium point.
real data. In Fig. 2, the structure of a GAN is depicted.
The discriminator will assign values in the range 0 to 1,
representing the probability that the input signal is fake. B. Evolutionary GAN Based on MFA
According to Fig. 1, the input noise vector N is used According to the above explanations, GAN is a powerful
by the generator G to produce a noisy vector G (N) with deep learning based model for classification and prediction
the probability distribution of PG (N). The signal G (N) is purposes. Such an adversarial gaming between the two net-
then fed into the discriminator to compare with the real works can benefit the model to yield a reliable and secure

Authorized licensed use limited to: Cornell University Library. Downloaded on August 27,2020 at 13:08:52 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

anomaly detection model. In fact, the case of fooling D net- searching mechanism. In other words, when the population
work by the G is exactly what the vehicle ECU may encounter diversity is enhanced, the algorithm has a higher chance to find
when cyber hacking. Nevertheless, the original GAN has some new fitting solutions in the population rather than trapping in
deficiencies such as instability of performance when facing the local optima. This will not stop but reduce the possibility
highly nonlinear dataset and mode collapse. In order to over- of premature convergence in the algorithm:
come these issues, a new optimization algorithm called MFA -Modification 1: This method is an accelerating movement
is introduced here which produces a set of generators {G}, for improving the convergence rate of the algorithm. To this
instead of a single generator in the discriminator domain. end, the average position of fireflies is first calculated and then
This would help to improve the chance of producing more the position of each firefly is updated based on its distance
qualified noisy data samples and thus a better matching may from the best solution:
be gained with the real data set x. Each solution X represents
X kI t er+1 = X kI t er + TF (X best − M Pop ) (7)
the parameter characteristics of a generator network, θ (G) .
The most fitting generator would be used for serving the -Modification 2: The second modification is a math formu-
discriminator and providing the most suitable classification. lation for improving the position of each firefly based on its
Similar to the other evolutionary algorithms, firefly algo- distance from the most suitable firefly X best . This would help
rithm starts with a random population. Being inspired from the less fitting solutions to upgrade their position sooner top attend
firefly brightening at the tropical regions, FA represents a mix the game.
of guided search and random search which can be very useful X kI t er+1 = X kI t er +  × (X best
I
− X kI t er ) (8)
for the optimization purposes. FA works based on three main
ideas: 1) fireflies do not show any gender and are assumed as wherein  obeys a Levy movement as follows:
unisex, 2) a firefly attractiveness is determined based on its λ(λ) sin(πλ/2) 1
≈ (s > 0) (9)
brightness seen by other insects and 3) if a firefly does not π s 1+λ
see any other type in its surrounding area, it can fly randomly -Modification 3: The third modification updates the value of
in the space. Compared to the particle swarm optimization, constant parameter α to provide a dynamic randomization in
FA is assumed as its advanced version. After calculating the the FA. It is well accepted in the society that a successful
objective function value for the firefly population, the best optimization algorithm should have a global search at the
solution is picked up and stored. The rest of the population beginning and then focuses on local search as iterations pass.
needs to improve. To this end, first the distance between any Therefore, a dynamic formulation is proposed for as follows:
two fireflies is calculated as follows [21]: 1 1
α I t er+1 = ( )( θ ×I ter ) α I t er ; θ ≥ 100 (10)
θ × I ter
  d
ri j =  X i − X j  = (X i,k − X j,k )2 (4) Fig. 3 shows the flowchart of the proposed MFA.
k=1 Using the above evolutionary deep learning model,
the required toll for a powerful classification is ready. In order
It is clear that as the firefly is further, its attractiveness
to have a criterion for measuring the performance quality of
reduces. The firefly attractiveness can then be calculated as
the anomaly detection model in the vehicles, some indices are
an exponential function as follows [21]:
defined here. The model can make four different decisions
β(r ) = β0 × exp(−γ r m ); m ≥ 1 (5) based on the nature of data (real or fake) and decision
(positive or negative). A decision is called positive when
where β0 shows the initial attractiveness of a firefly at the the model recognize a data as fake data (compromised by
precise locality of another type. Moreover, γ is a constant hacker). A decision is negative when the model recognize
value representing the absorption coefficient. The dimension a data as healthy. A decision is true, when the model has
of each vector is show by d. Then, the firefly population made a correction decision. Therefore, a false decision is a
is motivated to move toward a possible better position as wrong decision by the model. These four decisions appear
follows [21]: on a confusion matrix as shown in Fig. 4. True positive (hit
X j = X j + β0 × exp(−γ r m ) × (X i − X j ) + u j rate or HR), false positive (False alarm Rate or FR), false
1 negative (Miss Rate or MR) and true negative (Correct Reject
u j = α(r and − ) (6) rate or CR) are these four indices.
2
Considering C A and C N as the compromised data and real
In (6), the first term is the previous position of the insect, the data, the above four indices are formulated as follows:
second term is the insect attractiveness and the third term is
|Hi |
the random movement in the air. HR = ; Hi = {Y ∈ D |Y ∈ C A & Y ∈ C O } (11)
The conventional FA is a successful metaheuristic algorithm |C A |
in solving the optimization problems. But still it can get |FA |
FR = ; F A = {Y ∈ D |Y ∈ C N & Y ∈ C O } (12)
improved by the usage of some powerful math operator. |C N |
Therefore, a three-stage modification is proposed in the rest. |Mi |
MR = ; Mi = {Y ∈ D |Y ∈ C A & Y ∈ C I } (13)
The proposed three-phase modification method can avoid |C A |
premature convergence by increasing the diversity of the firefly |C R |
DR = ; C R = {Y ∈ D |Y ∈ C N & Y ∈ C I } (14)
population and thus increasing the chance of a more successful |C N |

Authorized licensed use limited to: Cornell University Library. Downloaded on August 27,2020 at 13:08:52 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

KAVOUSI-FARD et al.: EVOLUTIONARY DEEP LEARNING-BASED ANOMALY DETECTION MODEL FOR SECURING VEHICLES 5

TABLE I
S AMPLE C AN B US M ESSAGE T RAFFIC C HARACTERISTICS [25]

real data, C I represents the set of inliers and C O shows the

set of outliers.

IV. E XPERIMENTAL R ESULTS

In this section, the performance of the proposed deep
anomaly detection model would be assessed. To this end,
the experimental dataset of an undefined plate vehicle is
recorded from the OBD-II port for a 10-min normal driving
at different conditions. The CAN traffic is recorded in this
condition: the engine ignition is turned on, the vehicle stays
standstill for a few minutes, the gear is switched to “D” and
drive for around 8 minutes at a public area. The brake pedal
is pressed for several times. The gear is then changed to “R”
Fig. 3. Flowchart of the proposed MFA. and for a backward drive. Finally, the gear is changed to “P”
and the vehicle stays standstill for a few seconds and then
the car is turned off. The unique message ID with its relevant
frequency is used for training and then testing the proposed
model. For instance, the message ID of D21 has been repeated
by the frequency of 38.7804878. Still, it is observed that this
ID appears with other frequencies as well. Table I provides a
set of CAN message frames characteristics including ID and
frequency. The recorded data is divided into 3 groups, 50%
for training, 30% for validation and 20% for testing. For the
MFA, the population size is 20 and the termination criterion
is 100. As it is mentioned in the literature [23], [24], only
two features of the message ID and frequency suffice for a
reliable a secure classification, which will result in a smart
animally detection model. We consider the same procedure in
this work.
On the contrary to the other machine learning methods
Fig. 4. Confusion matrix to provide quality indices deep anomaly detection like neural network or support vector machine, GAN does
model.
not have a loss function to evaluate its performance quality.
This is required for the MFA since the quality of its members
where D represents the aggregated set of real data fake data, should be measured for improvisation process. In original
C A represents the set of fake data, C N represents the set of GAN, it is difficult to say when the generator produces a

Authorized licensed use limited to: Cornell University Library. Downloaded on August 27,2020 at 13:08:52 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

TABLE II
C ONFUSION M ATRIX VALUES FOR D IFFERENT A NOMALY D ETECTION
M ODEL , C OMPARED TO THE P ROPOSED E VOLUTIONARY GAN

Fig. 5. Convergence curve of the MFA, FA, PSO and GA when optimizing
the detection rate in GAN.

high quality sample. In order to overcome this issue, a fitness

function is deployed here based on the quality. The quality
is based on the generated sample capability in fooling the
discriminator. A generator with higher capability in misleading
the discriminator (showing a fake sample as a real sample) has
a higher quality. This can be shown as follows:
 
Fq = E N∼Pg (N) log (1 − D (G (N))) (15)
In order to assess the performance of the proposed MFA
method, the HR (as detection rate) is plotted versus the
iteration for MFA, FA, particle swarm optimization (PSO)
and genetic algorithm (GA). All algorithms start with similar Fig. 6. Hit, Miss, Correct Reject, and False Alarm area for message flooding
initial random population to have a fair comparison. The size for an specific ID of CAN in the vehicle.
of population is 20 for all algorithms. For PSO, the value of
the weighting factor and social parameters are assumed 2 and
0.8, respectively. For GA, the mutation and crossover operators values of MR% shows the low values of the false negative
are assumed as 0.9 and 0.1, respectively. Fig. 5 provides decisions made by the model. These results clearly advocate
the detection rate curve for the algorithms based on GAN. the importance of a high capability classification model, such
According to the figure, the MFA has not only first converged as the one proposed in this article, for learning the normal
among other algorithms (which shows its high convergence behavior of the CAN message frames. This for sure would
rate), but also it could result in better detection rate value be a great help for the ECU to avoid getting hacked. The
(showing the capability of escaping from the local optima). appropriate performance of the propose classification method
Therefore, MFA can be a reliable and stable algorithm to help as an anomaly detection model is validated here.
providing an appropriate anomaly detection model. It should In order to better perceive the performance of the anomaly
be noted that this is an offline process, and when the model detection model, some message frames with similar IDs but
is achieved, it will be used in the vehicle without any further different frequencies are simulated and send to the ECU.
change. This can result in a some healthy and compromised messages
In order to assess the cyber attack detection capability of the depending on the feasible operating frequency of a message.
proposed deep model, some fake message traffic are launched Due to the complex and wide frequency range, message frame
to the ECU and the detection capability of the model is with ID A7F is chosen here. The simulation results are shown
assessed. Table II provides the simulation results for this model in Fig. 6. It can be seen in Fig. 6 that the proposed evolutionary
based on the confusion matrix. Still, the performance of the deep anomaly detection model could detect the malicious
model is compared to some successful methods in the area message frames, successfully when rejecting fake ones. The
such one class support vector model (SVM), MOCSVM [26], very low values of miss rate and false negative decisions
MOCSVM based on BA and MOCSVM based on MBA. A set reveals the high reliability of this model. Still, we should keep
of different message IDs are considered for constructing the in the mind that these low values are happening due to the
false data, simulating the hacker cyber attacking. According message frames with frequencies so close to the real frequency
to these results, the proposed model shows higher HR% and of the real message. Since a hacker lunches messages from
CR% and lower MR% and FR%, compared to the other very high frequencies to win the arbitration, the proposed
models. The high value of HR% shows the high number model response is still accepted and useful for the practical
of true positive decisions made. On the other hand, the low cases.

Authorized licensed use limited to: Cornell University Library. Downloaded on August 27,2020 at 13:08:52 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

KAVOUSI-FARD et al.: EVOLUTIONARY DEEP LEARNING-BASED ANOMALY DETECTION MODEL FOR SECURING VEHICLES 7

Fig. 7. Performance of the anomaly detection model for different message IDs in a varied frequencies.

In the last part, the anomaly detection model when facing and RFID car keys or long-range wireless channels such as
message traffic of the same ID but different frequencies are broadcast channels and addressable channels. Be on that, this
shown in Fig. 7. Some of the message IDs which possess a article aimed to propose an evolutionary deep learning based
varied frequency behavior are deployed here. In this figure, anomaly detection model, so called MFA-GAM, to secure the
the feasible frequency domain, within which each message ECU against message flooding attack.
ID frequency is accepted, along with infeasible regions are In comparison with the other anomaly detection models
shown in the same frame. The small green balls show the available in the literature, the proposed model has some special
lower and upper bound of this region for each message ID features as follows:
frequency domain. This figure gives an estimation on the very -Advanced Classification: By developing an evolutionary
wide frequency range of the messages from 40 to 110 whereas deep learning model, a new powerful classification method
the proposed evolutionary GAN anomaly detection model still based on modified firefly algorithm and generative adversarial
shows very good performance. As it can be seen from this networks is devised. The proposed model can be used in a
figure, each message ID has a healthy frequency domain quite quite varied application, not limited to the automotive industry,
different from the other message ID. but power system, electric grids, smart grids, etc. The proposed
Such a heterogeneous and varied features necessitates the deep learning model is benefited from the high learning
use of a powerful deep learning based model for a secured capability of the GAN as well as the random search capability
and reliable performance of the vehicle. of the evolutionary method.
-Secured ECU: Since most of the vehicles internal com-
V. D ISCUSSION munications still rely on the CAN protocol, the proposed
The cyber security of a vehicle is a precious and very classification model could be used a as a powerful anomaly
technical area which needs urgent attention due to the fast detection model which can assess the message traffic in CAN.
growing nature of this industry. With the recent development This will help to secure the ECU and thus the overall car
of the smart city and the appearance of new communication operation. Recording the normal operation of a vehicle from
technologies between the vehicles, this is getting more and starting to driving and stopping will help us to have a clear
more demanding for the future of human life. The automo- picture of the normal message traffic which flows during a
tive industry is experiencing a complex hardware-software normal operation. Any changes in the traffic can activate an
development which is highly computerized to improve its alarm signal which will put the car in an aware model.
services to the customers. On this way, several detection and -Powerful Optimizer: The proposed MFA is a very fast
vision algorithms are introduced by the developers to help converging and high-search capability algorithm which may be
get into this goal. However, still there are some requirements used in many upcoming applications as a powerful optimizer.
for accurate and fast performance of these algorithms, with- This is necessary due to the very complex and nonlinear
out which the vehicle normal operation would be affected. nature of the big data facing in the smart city. For the
Researchers have shown in that there are still several gaps in GAN, by generation a random generators set, it could help
the cars which can let hackers penetrate into the system, either to overcome the GAN drawbacks such as instability and com-
wired methods such as OBD-II and USB or wireless methods plex training. Being equipped with a three-stage modification
such as Bluetooth, remote keyless, entry, tire pressure sensors method, MFA can be widely used in many other applications

Authorized licensed use limited to: Cornell University Library. Downloaded on August 27,2020 at 13:08:52 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

when offline training. Still it should be noted that training is [10] Y. Li, Y. Tu, Q. Fan, C. Dong, and W. Wang, “Influence of cyber-attacks
happening offline, and the trained anomaly detection model is on longitudinal safety of connected and automated vehicles,” Accident
Anal. Prevention, vol. 121, pp. 148–156, Dec. 2018.
then used in some master ECUs for online detection. [11] G. Loukas, “Cyber-physical attacks on implants and vehicles,” in Cyber-
Physical Attacks. London, U.K.: Elsevier, 2015, pp. 59–104.
[12] G. Loukas, E. Karapistoli, E. Panaousis, P. Sarigiannidis, A. Bezemskij,
VI. C ONCLUSION and T. Vuong, “A taxonomy and survey of cyber-physical intrusion
detection approaches for vehicles,” Ad Hoc Netw., vol. 84, pp. 124–147,
This article tried to investigate the cyber security of CAN in Mar. 2019.
the vehicle for the message flooding cyber attacking. To this [13] B. Mohandes, R. Al Hammadi, W. Sanusi, T. Mezher, and S. El Khatib,
“Advancing cyber–physical sustainability through integrated analysis of
end, a deep anomaly detection model based on MFA and smart power systems: A case study on electric vehicles,” Int. J. Crit.
adversarial networks was proposed. Other than getting benefits Infrastruct. Protection, vol. 23, pp. 33–48, Dec. 2018.
of the GAN for deep training and modeling, it is benefited [14] S. Schulze, M. Pukall, G. Saake, T. Hoppe, and J. Dittmann, “On
the need of data management in automotive systems,” in Proc. BTW,
from the high search ability of MFA for finding the most vol. 144, 2009, pp. 217–266.
suitable generator in the GAN. The experimental results on [15] C. Ling and D. Feng, “An algorithm for detection of malicious messages
the recorded CAN message traffic reveals the high capability on can buses,” in Proc. Nat. Conf. Inf. Technol. Comput. Sci. Detroit,
MI, USA: Atlantis Press, 2012, pp. 1–4.
of the model for detecting true cyber attacks with high HR% [16] T. Hoppe, S. Kiltz, and J. Dittmann, “Security threats to automotive CAN
and CR%. Compared to other well-known classification algo- networks–practical examples and selected short-term countermeasures,”
rithms, this is secure model for detecting abnormal behaviors Reliab. Eng. Syst. Saf., vol. 96, no. 1, pp. 11–25, 2011.
[17] H. Oguma, A. Yoshioka, M. Nishikawa, R. Shigetomi, A. Otsuka, and
in the vehicle. It was seen that even though that there is a H. Imai, “New attestation based security architecture for in-vehicle com-
wide range of acceptable frequencies for each message ID, still munication,” in Proc. IEEE Global Telecommun. Conf. (GLOBECOM),
the proposed model has appropriate performance for covering Nov./Dec. 2008, pp. 1–6.
[18] M.-J. Kang and J.-W. Kang, “Intrusion detection system using deep
the entire domain. The vey low FP and FN, which generally neural network for in-vehicle network security,” PLoS ONE, vol. 11,
happens at the very neighboring areas of the normal frequency no. 6, Jun. 2016, Art. no. e0155781.
of the vehicle, advocate the high dependability of the proposed [19] L. Pan, X. Zheng, H. X. Chen, T. Luan, H. Bootwala, and L. Batten,
“Cyber security attacks to modern vehicular systems,” J. Inf. Secur.
cyber attack detection model. From the optimization point of Appl., vol. 36, pp. 90–100, Oct. 2017.
view, the high search capability of MFA over original FA, [20] H. M. D. Kabir, A. Khosravi, S. Nahavandi, and A. Kavousi-Fard,
PSO and GA for helping in the training process of GAN was “Partial adversarial training for neural network-based uncertainty quan-
tification,” IEEE Trans. Emerg. Topics Comput. Intell., early access,
demonstrated. The simulation results show that the proposed Sep. 5, 2019, doi: 10.1109/TETCI.2019.2936546.
anomaly detection model is so capable for stopping various [21] A. K. Fard and T. Niknam, “Optimal stochastic capacitor placement
cyber attacks in CAN bus. problem from the reliability and cost views using firefly algorithm,”
IET Sci., Meas. Technol., vol. 8, no. 5, pp. 260–269, 2014.
[22] R. Charette, “This car runs on code,” IEEE Spectr., vol. 46, no. 3, p. 3,
Feb. 2009.
R EFERENCES [23] M. Al-Saud, A. M. Eltamaly, M. A. Mohamed, and A. Kavousi-Fard,
“An intelligent data-driven model to secure intravehicle communications
[1] E. Mousavinejad, F. Yang, Q.-L. Han, X. Ge, and L. Vlacic, “Distributed based on machine learning,” IEEE Trans. Ind. Electron., vol. 67, no. 6,
cyber attacks detection and recovery mechanism for vehicle platoon- pp. 5112–5119, Jun. 2020.
ing,” IEEE Trans. Intell. Transp. Syst., early access, Aug. 21, 2019, [24] K. F. Abdollah, W. Su, and T. Jin, “A machine learning based
doi: 10.1109/TITS.2019.2934481. cyber attack detection model for wireless sensor networks in micro-
[2] S. Dey and M. Khanra, “Cybersecurity of plug-in electric vehicles: grids,” IEEE Trans. Ind. Informat., early access, Jan. 7, 2020,
Cyber attack detection during charging,” IEEE Trans. Ind. Electron., doi: 10.1109/TII.2020.2964704.
early access, Jan. 15, 2020, doi: 10.1109/TIE.2020.2965497. [25] A. Kavousi-Fard, T. Jin, W. Su, and N. Parsa, “An effective
[3] T. Chen, B. Zhang, H. Pourbabak, A. Kavousi-Fard, and W. Su, “Optimal anomaly detection model for securing communications in electric
routing and charging of an electric vehicle fleet for high-efficiency vehicles,” IEEE Trans. Ind. Appl., early access, Jun. 25, 2020,
dynamic transit systems,” IEEE Trans. Smart Grid, vol. 9, no. 4, doi: 10.1109/TIA.2020.3005062.
pp. 3563–3572, Jul. 2018. [26] O. Avatefipour et al., “An intelligent secured framework for cyberattack
[4] A. Kavousi-Fard, A. Abbasi, T. Niknam, and H. Taherpoor, “Multi- detection in electric vehicles’ CAN bus using machine learning,” IEEE
objective probabilistic reconfiguration considering uncertainty and multi- Access, vol. 7, pp. 127580–127592, 2019.
level load model,” IET Sci., Meas. Technol., vol. 9, no. 1, pp. 44–55,
Jan. 2015.
[5] S. Parkinson, P. Ward, K. Wilson, and J. Miller, “Cyber threats Abdollah Kavousi-Fard (Senior Member, IEEE)
facing autonomous and connected vehicles: Future challenges,” IEEE received the B.Sc. degree from the Shiraz University
Transactions on Intelligent Transportation Systems, vol. 18, no. 11, of Technology, Shiraz, Iran, in 2009, the M.Sc.
pp. 2898–2915, Nov. 2017. degree from Shiraz University, Shiraz, in 2011, and
[6] F. van Wyk, Y. Wang, A. Khojandi, and N. Masoud, “Real-time sensor the Ph.D. degree from the Shiraz University of
anomaly detection and identification in automated vehicles,” IEEE Trans. Technology in 2016, all in electrical engineering.
Intell. Transp. Syst., vol. 21, no. 3, pp. 1264–1276, Mar. 2020. He was a Post-Doctoral Research Assistant with the
[7] M. Ghanavati, A. Chakravarthy, and P. P. Menon, “Analysis of auto- University of Michigan, MI, USA, from 2016 to
motive cyber-attacks on highways using partial differential equation 2018. He was a Researcher with the University
models,” IEEE Trans. Control Netw. Syst., vol. 5, no. 4, pp. 1775–1786, of Denver, Denver, CO, USA, from 2015 to 2016,
Dec. 2018. conducting research on microgrids. He is currently
[8] A. Monot, N. Navet, B. Bavoux, and F. Simonot-Lion, “Multisource an Assistant Professor with the Shiraz University of Technology. He has
software on multicore automotive ECUs—Combining runnable sequenc- published more than 100 research papers in prestigious international journals
ing with task scheduling,” IEEE Trans. Ind. Electron., vol. 59, no. 10, and peer-reviewed conference proceedings. His current research interests
pp. 3934–3942, Oct. 2012. include operation, management and cyber security analysis of smart grids,
[9] F. Ahmad, F. Kurugollu, A. Adnane, R. Hussain, and F. Hussain, microgrids, smart city, electric vehicles and protection of power systems,
“MARINE: Man-in-the-Middle attack resistant trust model in connected reliability, artificial intelligence, and machine learning. He is an Associate
vehicles,” IEEE Internet Things J., vol. 7, no. 4, pp. 3310–3322, Editor of ISTE ISI Journal and the IEEE T RANSACTIONS ON I NDUSTRY
Apr. 2020. A PPLICATIONS.

Authorized licensed use limited to: Cornell University Library. Downloaded on August 27,2020 at 13:08:52 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

KAVOUSI-FARD et al.: EVOLUTIONARY DEEP LEARNING-BASED ANOMALY DETECTION MODEL FOR SECURING VEHICLES 9

Morteza Dabbaghjamanesh (Senior Member, Wencong Su (Senior Member, IEEE) received

IEEE) received the M.Sc. degree in electrical engi- the B.S. degree (Hons.) from Clarkson Univer-
neering from Northern Illinois University, DeKalb, sity, Potsdam, NY, USA, in 2008, the M.S. degree
IL, USA, in 2014, and the Ph.D. degree in elec- from Virginia Tech, Blacksburg, VA, USA, in 2009,
trical and computer engineering form Louisiana and the Ph.D. degree from North Carolina State
State University, Baton Rouge, LA, USA, in 2019. University, Raleigh, NC, USA, in 2013. He is
In 2015, he joined the Renewable Energy and currently an Associate Professor with the Depart-
Smart Grid Laboratory, Louisiana State University. ment of Electrical and Computer Engineering, Uni-
In 2019, he joined the Design and Optimization of versity of Michigan–Dearborn, USA. His current
Energy Systems (DOES) Laboratory, The University research interests include power systems, electrified
of Texas at Dallas, Richardson, TX, USA. He is transportation systems, and cyber-physical systems.
currently a Research Associate with The University of Texas at Dallas. He has published more than 100 research papers in prestigious international
His current research interests include power system operation, reliability, journals and peer-reviewed conference proceedings. He is a fellow of IET.
and resiliency, renewable energy sources, cyber security analysis, machine He is an Editor of the IEEE T RANSACTIONS ON S MART G RID and an
learning, smart grids, and microgrids. Associate Editor of IEEE A CCESS and the IEEE D ATA P ORT. He was a
recipient of the 2015 IEEE Power and Energy Society (PES) Technical
Committee Prize Paper Award and the 2013 IEEE Industrial Electronics
Society (IES) Student Best Paper Award. He is a registered Professional
Engineer (P.E.) in the State of Michigan, USA.
Tao Jin (Senior Member, IEEE) received the
B.S. and M.S. degrees from Yanshan University
in 1998 and 2001, respectively, and the Ph.D. degree
in electrical engineering from Shanghai Jiao Tong
University in 2005. From 2005 to 2007, he worked
as a Post-Doctoral Researcher with Shanghai Jiao
Tong University. He was in charge of a research
group in the biggest dry-type transformer company
in Asia, Sunten Electric Company Ltd., to develop
new transformer technology with distribution grid.
From 2008 to 2009, he held research scientist posi-
tion at Virginia Tech, Blacksburg, VA, USA, where he was involved in the Mahmoud Roustaei was born in Shiraz, Iran,
design and test of PMU technology and GPS/Internet-based power system in April 1993. He received the M.Sc. degree in
frequency monitoring networks. In 2010, he joined the Imperial College electrical engineering from the Sharif University of
London, U.K., as an European Union Marie Curie Research Fellow, where he Technology, Tehran, Iran, in 2017. He is currently
was focused on electrical technologies related to smart grid. He is currently a pursuing the Ph.D. degree in the power systems with
Professor with the College of Electrical Engineering and Automation, Fuzhou the Shiraz University of Technology, Shiraz.
University, China. He has published about 150 articles. He is a member of He is cooperating with Software Energy Com-
the IEEE Power and Energy Society and the IEEE Industrial Electronics pany LLC, Detroit, MI, USA. His research inter-
Society and a Special Committee Member of the Chinese Society of Electrical ests include smart energy hub, smart city, smart
Engineering, and the China Electrotechnical Society. He serves as an Associate grid and microgrid, distributed optimization, power
Editor for MPCE, PCMP, Measurement and Testing Technology China, and system cyber security, power system resiliency, and
other journals. advanced machine learning.

Authorized licensed use limited to: Cornell University Library. Downloaded on August 27,2020 at 13:08:52 UTC from IEEE Xplore. Restrictions apply.