IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO.

5, MARCH 1, 2021 3157

AIT: An AI-Enabled Trust Management System for

Vehicular Networks Using Blockchain Technology
Chenyue Zhang, Wenjia Li , Senior Member, IEEE, Yuansheng Luo ,
and Yupeng Hu , Senior Member, IEEE

Abstract—Currently, connected vehicles have gradually cabs would be on the road by the end of 2020 [2]. To enhance
stepped into our daily lives, and they generally rely on vehic- the safety and efficiency of these connected and autonomous
ular networks to generate and exchange traffic-related messages vehicles, it has become essential that they could sense and
to improve the overall travel safety and efficiency. However, due
to the open nature of vehicular networks, these traffic-related collect various travel related information and exchange them
messages could be erroneous, which may be caused by var- via vehicular networks, which is enabled by the increasing
ious reasons, ranging from an onboard device (OBD) sensor onboard sensing, computation, and communication capabili-
malfunctioning and reporting incorrect reading to the message ties. In vehicular networks, different types of nodes, including
being tampered by a malicious vehicle. To address these rapidly vehicles and road-side units (RSUs), could communicate and
increasing security challenges, we have proposed an AI-enabled
trust management system (AIT) in this article, which is an AI- share information with each other through vehicle-to-vehicle
enabled trust management system for vehicular networks using (V2V) and vehicle-to-infrastructure (V2I) communications.
the blockchain technique. In the AIT system, each vehicle first The information that they share generally contains alerts and
senses, generates, and exchanges messages with other vehicles. updates regarding road accidents, traffic conditions (such as
These messages then get validated by the neighboring vehicles. As congestion, construction, and hazardous weather conditions),
vehicles receive and validate messages from other nearby vehicles,
they will establish and manage the trust of those nearby vehi- and other related transportation events. All these traffic alerts
cles, which is enabled by utilizing the deep learning algorithm. and updates could make vehicles aware of various traffic con-
Once a vehicle identifies untrustworthy vehicles, it reports them ditions in a timely manner, thus improving the transportation
to the nearby roadside unit (RSU), and the RSU will validate the safety and efficiency.
authenticity of the report as well as the identity of the vehicle by Despite that vehicular networks may help disseminate these
using the emerging blockchain technique. The security creden-
tials of untrustworthy vehicles will then be revoked by the RSU. types of important information, caution should still be taken
We have conducted an extensive experimental study to evaluate when interpreting and utilizing them, as you generally have
the AIT system. Simulation results clearly indicate that AIT per- little idea whom you are communicating with in a neighbor-
forms better than existing approaches and can manage the trust ing vehicle because of its high mobility. The situation further
of vehicles and detect malicious ones in an accurate and efficient deteriorates when there are malicious vehicles that have been
manner.
compromised and controlled by adversaries, who may delib-
Index Terms—Blockchain, deep learning, security, trust man- erately share fake traffic alerts and updates to confuse other
agement, vehicular networks. vehicles. An example of the fake traffic update shared by a
malicious vehicle is depicted in Fig. 1.
From Fig. 1(a), it shows that a malicious vehicle shares a
I. I NTRODUCTION fake traffic update stating that there is no traffic congestion,
N THE past few years, we have witnessed an explosive while there are some road hazards, such as traffic conges-
I growth to the number of connected and/or autonomous
vehicles on the road. News articles revealed that there were
tion in reality. As shown in Fig. 1(b), the fake traffic update
that is disseminated via vehicular networks makes the road
already one million connected vehicles made by general even more congested, which clearly demonstrates that the fake
motors (GMs) on the road in 2015 [1], and Tesla’s CEO Elon traffic updates and alerts could greatly jeopardize the safety
Musk estimated in 2019 that over a million of self-driving taxi and efficiency of the transportation system. As a result, it is
critical to properly evaluate the trustworthiness of both traffic-
Manuscript received May 11, 2020; revised September 1, 2020 and related messages and vehicles, which share them in vehicular
November 3, 2020; accepted November 30, 2020. Date of publica- networks.
tion December 14, 2020; date of current version February 19, 2021.
(Corresponding author: Wenjia Li.) In recent years, various research efforts have been made
Chenyue Zhang and Wenjia Li are with the Department of Computer to develop effective trust management systems for wireless
Science, New York Institute of Technology, New York, NY 10023 USA networks [3]–[5], including vehicular networks [6]–[8], sen-
(e-mail: [email protected]).
Yuansheng Luo is with the School of Computer and Communication sor networks [9]–[11], and Internet of Things [12]–[14]. While
Engineering, Changsha University of Science and Technology, Changsha they primarily focused on evaluating the trust of nodes (such as
410114, China. vehicles, sensors, etc.), there have been some research works
Yupeng Hu is with the College of Computer Science and Electronic
Engineering, Hunan University, Changsha 410082, China. in which the trust of data that the nodes generate is also
Digital Object Identifier 10.1109/JIOT.2020.3044296 evaluated [7], [15], [16]. In the specific context of vehicular
2327-4662 
c 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: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.
3158 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 5, MARCH 1, 2021

Fig. 1. Example of fake traffic update shared by a malicious vehicle and its outcome in vehicular networks. (a) Malicious vehicle sends a fake traffic update.
(b) Outcome of the fake traffic update: severe traffic congestion.

networks, all the vehicles normally travel constantly at a rel- design of our AIT system. Section V provides an in-depth dis-
atively high speed, resulting in a rapidly changing network cussion on the experiments that we have conducted to validate
topology. Consequently, it is challenging for a vehicle to eval- the AIT system. Finally, Section VI concludes this article.
uate the trust of all other vehicles that it interacts with in
a timely fashion. Moreover, due to its highly mobile and II. R ELATED W ORK
dynamic nature, vehicular networks usually generate a large
Due to the widespread and increasing security threats that
amount of data, such as traffic alerts and updates reported by
recently arise in vehicular networks, many research efforts
both vehicles and RSUs. However, the reliability of these data
have been made to detect and mitigate them. Researchers
is weakened by both imperfect reporting devices and mali-
have mainly been focusing on developing two categories of
cious vehicles that have been compromised by adversaries,
approaches in response to those security threats: 1) trust
which further undermines the trust management system that
management and 2) malicious node detection.
largely depends on the data to properly function. Hence, it is
essential yet challenging to design effective trust management
systems for vehicular networks. A. Trust Management System
To address this urgent need, we have proposed an AI- In recent years, trust and reputation management has
enabled trust management system for vehicular networks, attracted great attention from the research community.
namely, AIT, in this article. In contrast to the existing trust Guo et al. [17] proposed a trust management model that
management systems, such as [3] and [8]–[10], the AIT system is aware of the context and able to authenticate the message
does not rely on a fixed formula to evaluate the trust ratings received by evaluating the trust value of the sender. In this
of individual messages or calculate the overall trust of vehi- approach, the trust is decided by both the related information
cles. Instead, the emerging deep learning algorithm is used to that is available and the current evaluation strategy. In addition,
determine the trust of vehicles in an automatic manner. a reinforcement learning model was developed so that vehi-
The main contributions of this research work are as follows. cles could dynamically set the evaluation strategy in different
1) We proposed a novel trust management system that is driving scenarios.
based on deep learning to evaluate the trust of nodes El Sayed et al. [18] studied a hierarchical trust management
(including both vehicles and RSUs) as well as data (such system for vehicular networks, which performs trust computa-
as messages) in an automatic and dynamic manner. tion for each vehicle. More specifically, the proposed system
2) We applied the emerging blockchain technology to the takes the steps of trust evaluation, trust propagation, and trust
trust management system so that both the identity of aggregation so that the trustworthiness of vehicles could be
both vehicles and RSUs and the authenticity of mes- accurately derived.
sages sent in the vehicular networks could be validated, In [19], an anti-attack trust management scheme named
thereby remarkably enhancing the security of vehicular AATMS was proposed by Zhang et al. In AATMS, the local
networks. trust will first be computed, which is used as the criterion to
3) To validate and evaluate the performance of the AIT determine a set of trustworthy seed vehicles. The trustworthy
system, we have conducted extensive simulations, and seed vehicles are then used together with the local trust link
the experimental results clearly show that the proposed to evaluate the global trust of all vehicles.
trust management system could evaluate the trust in an Li and Song proposed ART in [7], and its primary focus
accurate and efficient manner. was to ensure that the trust could be evaluated in an efficient
The remainder of this article is organized as follows. and accurate manner in the presence of adversaries. Thus,
Section II summarizes the existing research findings that the proposed attack-resistant trust management scheme has
are related to this work. The research problem is defined the functionality of both malicious node detection and trust
in Section III, which contains details regarding both system management. Moreover, the trustworthiness of both data and
model and adversary model. Section IV looks at the detailed vehicles is evaluated by the ART scheme.

Authorized licensed use limited to: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: AIT: AI-ENABLED TRUST MANAGEMENT SYSTEM FOR VEHICULAR NETWORKS USING BLOCKCHAIN TECHNOLOGY 3159

B. Malicious Node Detection

The other category of research works that could help better
secure vehicular networks is malicious node detection, which
aims at detecting and possibly mitigating malicious nodes.
Gu et al. proposed a method to detect malicious nodes in
fog computing-based VANETs [24], in which the fog server
identifies suspicious vehicles via trust evaluation based on the
correlation of data and network topology.
Fig. 2. System model for the AIT in vehicular networks. In [25], a machine learning-based misbehavior detection
system was proposed for vehicular networks, which aimed at
detecting and coping with insider attacks.
Chen et al. [26] studied a secure message dissemination
Ahmed et al. [20] proposed a trust management model for scheme for vehicular networks in the presence of mali-
connected vehicles, which is resilient to the well-known man- cious vehicles. More specifically, the proposed scheme could
in-the-middle (MiTM) attack. In this model, dishonest vehicles identify true message(s) when the messages received from
which perform MiTM attacks could effectively be detected and multiple nearby vehicles are conflicting, which was achieved
their security credentials would then be revoked. by incorporating the underlying network topology information.
Yang et al. [8] proposed to use the blockchain to help bet- In [27], a security framework was proposed for VANETs,
ter evaluate the trustworthiness of vehicles. In the proposed in which the SVM algorithm was used to automatically dis-
scheme, vehicles first perform message validation using the tinguish malicious vehicles from benign ones. Moreover, the
Bayesian inference model. As the next step, the vehicle will proposed approach was aware of the context in which the mis-
generate a rating for each vehicle that originates the mes- behaviors occurred. By this means, the detection accuracy of
sage. Then, the RSUs perform trust evaluation on the vehicles. malicious vehicles was further improved.
Based on these ratings, they will add the trust offsets as a block Sedjelmaci et al. [28] proposed a misbehavior detection and
to the blockchain, which is maintained by all the RSUs. prevention scheme for vehicular networks, and the misbehav-
In [21], a hierarchical blockchain structure is studied for ior prediction is based on game theory so that the occurrence
managing reputation in vehicular networks. The first layer of of malicious vehicles could be prevented.
the blockchain hierarchy is the vehicles that are located within
the same district, and they record reputation values for their III. P ROBLEM D EFINITION
own district. Through voting, vehicles can become a miner in
the vehicular network. Once a block is generated by the miner, In this section, we describe the research problem that is
it will be sent over to the RSU that is in the same district. being investigated in this article in more detail. More specif-
Then, the RSU sends this block to every vehicle in the same ically, we are defining both the network model and the
district. RSU just sends blocks but they are not responsible for adversary model in a detailed fashion.
storing the blocks and managing the blockchain. The blocks
are stored in each vehicle instead. The second layer covers a A. System Model
district in the first layer along with its neighboring areas. As depicted in Fig. 2, the trust management system in vehic-
Although the blockchain technology is used in both [8] ular networks is generally composed of the following two
and [21] to manage trust for vehicular networks, the proposed types of nodes: 1) RSUs and 2) vehicles.
AIT system has some fundamental uniqueness and differences 1) RSU: RSUs generally possess large amounts of process-
when compared to them. First, as described in [8], all the RSU ing, storage, and communication capacities, and they are
maintain one single blockchain for all vehicles. As a result, stationary. They can keep track of all current vehicle
it is costly in terms of both time and computational power to trust levels, based on messages sent by vehicles about
locate and mine a block for a specific vehicle because all the other vehicles and the current traffic conditions. They
blocks for all vehicles are in the same blockchain. As for the have a long communication range that covers a large
hierarchical blockchain structure proposed in [21], each vehi- section of roads, which allows them to receive mes-
cle maintains its own blockchain, which makes the blockchain sages from other RSUs and vehicles from a far distance.
more vulnerable for data tampering and loss. Unlike these two After they receive messages, the messages are analyzed
previous research efforts, we have designed and implemented by them. After evaluating the information received, the
the distributed blockchain based on the tree structure that is RSUs will adjust all vehicles’ trust values so that they
maintained by each RSU, and each branch of the tree cor- can be sent to other RSUs. Because different RSUs could
responds to a vehicle within the RSU’s direct communication get different messages from different vehicles about sus-
range. Moreover, we use the Merkle value to prevent data tam- picious activity, their evaluation of vehicle trust may be
pering and loss. In this way, the overall security and efficiency different. The trust values from all RSUs regarding the
of the blockchain in the AIT system is further enhanced. same vehicle will be averaged to determine if a vehicle
In addition, there have also been some other research efforts is malicious. Once found to be malicious, the vehicle’s
that aim at trust and reputation management for vehicular and public key will be revoked, blocking them from sending
other types of wireless networks [13], [22], [23]. new messages in the network. In the meanwhile, if they

Authorized licensed use limited to: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.
3160 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 5, MARCH 1, 2021

Vehicles could communicate with other vehicles by

sending messages, and they pass information to each
other to provide current road condition updates. If a haz-
ard is spotted, a vehicle can pass the location and type of
the hazard to other vehicles or RSUs. Each time when
a vehicle gets a communication request from another
vehicle, a trust evaluation is done on the other vehicle
to make sure it is not malicious prior to the communica-
tion. If the other vehicle is detected to be malicious, the
communication request will be denied, and a message is
sent to the nearest RSU to report this malicious vehicle.

B. Adversary Model
In this research, we assume that both vehicles and RSUs
could be compromised by adversaries, and they will then
exhibit various malicious behaviors, such as intentionally shar-
ing a fake traffic alert or falsely reporting another vehicle to
be malicious, etc. More specifically, the following three types
of malicious attacks are considered in this article.
1) Simple Attack (SA): The primary goal of malicious
attackers when performing SA is to interfere with the
messaging service in vehicular networks by different
means, such as sending out an excessive number of
messages, so that other benign nodes cannot send out
any message successfully during this time. In addition,
the compromised nodes may choose to drop or alter the
incoming messages so that the current traffic status will
be distorted when being shared with other nodes.
2) Bad Mouth Attack (BMA): In BMA, malicious attackers
deliberately share fake trust ratings for other nodes, such
as claiming that a benign vehicle is malicious and vice
versa. By this means, the security credentials of benign
vehicles could potentially be revoked, and the vehicular
networks may be dominated by malicious vehicles.
3) Zigzag Attack (ZA): The advanced attackers may attempt
to avoid being detected by conducting the malicious
behaviors in an intermittent manner. For example, an
attacker may choose to spoof the incoming messages
for some time, then stop for a while before switching to
conducting the BMA. Given that each attack is occur-
ring at a lower frequency, we would envision that the
ZA (also known as the on-and-off attack) is the most
challenging attack to be caught by the trust management
system.

IV. D ETAILED D ESIGN OF THE AIT S YSTEM FOR

V EHICULAR N ETWORKS
Fig. 3. Overall workflow of the AIT system.
In this section, we discuss the proposed AIT system in
detail. The overall workflow of the AIT system is shown in
want to regain the public key, they will send a request Fig. 3.
to RSU and the RSU will evaluate its trustworthiness to As depicted in Fig. 3, there are five important steps in the
decide whether to give the vehicle a new public key. AIT system, namely: 1) traffic data collection and message
2) Vehicle: Vehicles have limited processing, storage, and generation; 2) local trust level (LTL) calculation and sharing
communication capacities when compared to RSUs, and with local RSU; 3) global trust level (GTL) calculation by
they are moving constantly in the road network. Each local RSU; 4) trust validation and archiving by blockchain;
vehicle is represented as a node in the vehicular network. and 5) GTL voting and dissemination by all the RSUs.

Authorized licensed use limited to: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: AIT: AI-ENABLED TRUST MANAGEMENT SYSTEM FOR VEHICULAR NETWORKS USING BLOCKCHAIN TECHNOLOGY 3161

TABLE I
I NPUT OF THE F EEDFORWARD N EURAL N ETWORK A LGORITHM

Fig. 4. Structure of the feedforward neural network.

A. Traffic Data Collection and Message Generation

As the first step of the trust management process, each
vehicle in the vehicular networks collects the nearby traf-
fic updates and then generates messages that best describe
the traffic updates. Given that there may be faulty or even
malicious vehicles that could share misleading messages, it
is important to take different factors into consideration, such
as the distance between vehicles that reports the same traffic
update, the number of vehicles in the range where the traffic
incident is visible, the location of the traffic incident, the loca-
tion of the nearest RSU, the communication range that RSU
can cover, the direction of the traffic, the location of each vehi-
cle that reports the incident, and so on. All these data may be
integrated into the message so that the recipient could better
identify the context in which the traffic incident is reported.

B. Local Trust Levels Calculation and Sharing With Local

RSU TABLE II
O UTPUT OF THE F EEDFORWARD N EURAL N ETWORK A LGORITHM
Once a vehicle receives messages from a nearby vehicle
regarding some traffic updates, it will calculate the LTL of the
message sender by applying the feedforward neural network
algorithm [29], which is depicted in Fig. 4.
As shown in Fig. 4, the input layer is composed of 18 input
nodes, such as the location of the reporting vehicle, the loca-
tion of the traffic incident, the distance that the traffic incident
is visible, the current trust rating of the reporting vehicle, and
so on. The complete list of the eighteen input data is shown
in Table I.
There are two hidden layers with totally 14 hidden neu-
rons. The output layer produces the following three outputs: It is important to learn which weights and biases that min-
1) updated trust rating of the reporting vehicle; 2) prediction of imize a certain cost function. In the following functions [31],
next message time; and 3) level of need for message relay as the activation of the last neuron is represented as a, with a
the output, which are shown in Table II. Note that these three superscript L, which represents the current layer. Then, the
outputs of the feedforward neural network algorithm will be activation of the previous layer is a(L−1) . Cost can be cal-
stored in the body of a block in the blockchain. culated from finding the differences to each neuron from the

Authorized licensed use limited to: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.
3162 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 5, MARCH 1, 2021

output it is expected to give minus the current output value. which is (∂C/∂ωL )
Then, add up the square of those differences between each
L −1
n
component onto all the layers in a neural network. Equation (1) ∂C0 ∂zL ∂aL ∂C0
= . (5)
describes how the cost in a neural network is computed. Note ∂ωL ∂ωL ∂zL ∂aL
j=0
that yj stands for the jth output in the output layer. Suppose
there are totally L hidden layers, then nL − 1 represents the This chain rule function describes how sensitive the cost
neuron in the last hidden layer right before the output layer is to a specific weight. It implies the sensitivity of the cost
regarding small changes of weights, even slight changes can
L −1
n 2 cause a huge impact. The goal is to figure out how sensitive
C0 = aLj − yj . (1) the cost function is too small changes in our weight ωL . After
j=0
calculating how sensitive the cost function is to the activations,
We then calculate the cost for all layers and then average processes are repeated for all the weights and biases feeding
them, so that the total cost of the neural network is calculated. into the layer. The second term in the chain rule is also called
The derivative of the full cost function is equal to the aver- a sigmoid function, which is a critical component of artificial
age of all training examples, which is shown in the following intelligence. With the help of sigmoid function, most of the
equation: functions can be predicted and estimated.
The sigmoid function can be both amplified and shifted.
1  ∂Ck
n−1 By changing the gains of the hidden layers, functions can
∂C
= . (2) be amplified while through changing the constant input or
∂ωL n ∂ωL
k=0 bias, functions can be shifted. With the help of multiple
As the next step, the square of the difference we calcu- Sigmoid functions, by amplifying and shifting, we can then
lated previously is summed up, and then we average the add them together and deliver a good approximate function
result, which gives us the total cost of the network. Then, we that is estimated and used in the neural network. In addition,
can adjust the weight and differences of the current neurons the backpropagation algorithm is effective in calculating the
accordingly. It is important to look for gradients in this cost derivative on weights.
function because it implies how to change all the weights and By applying the feedforward neural network algorithm, the
biases of all the connections between neurons. In the gradient trust management system can successfully identify which vehi-
vector, each element is showing how sensitive the cost func- cles are malicious and learn the potential correlation among
tion is to each weight and bias. It explains how to get weight those malicious vehicles. As a result, the LTL of the reporting
and biases for a gradient descent. vehicle will be calculated and shared with the RSU after this
Backpropagation is an algorithm to compute the gradient process.
that is used to minimize the cost and biases. Each component
of the cost function shows how sensitive the cost function is C. Global Trust Levels Calculation by Local RSU
to each weight and biases. ω represents the weight Each vehicle’s behaviors may vary over time, especially if
  the malicious vehicle is adopting the ZA as its attack pattern,
aL = σ ω0L−1 a0L−1 + ω1L−1 a1L−1 + · · · + ωn−1
L−1 L−1
an−1 + b . in which the malicious vehicle can show different malicious
(3) behaviors at different times with some intervals in between. In
this case, an instantaneous local trust view could be inaccurate
For the neuron a, which lies in layer L, is defined as aL . N and misleading if the malicious vehicle carefully spreads its
stands for the total number of nodes in a layer. The weighted activities over many neighboring vehicles at different times.
sum is calculated from each individual weights times the neu- Thus, as the next step, the GTLs of all vehicles will be
(L−1) (L−1) L
ron value that is ω0 a0 . a is calculated from a certain calculated by RSUs. The trust of each message coming into the
weighted sum of all the activations in the previous layer, plus RSU is also calculated with the same deep learning algorithm
b, which is a bias. They are plugged into a sigmoid function, as the vehicles use for their local trust evaluation. The result of
which is represented as σ . the GTL calculation is an adjustment to the initial LTL for each
From the equation, there are three elements that can be individual vehicle. This process produces a more accurate and
changed so as to change the value of aL , and they are weights, comprehensive view of each vehicle’s trust level, which can
biases, and the value of the previous node. The following (4) help better protect vehicular networks against those malicious
contains all of these three elements of the weighted sum and is behaviors. Suppose the RSU r1 receives an LTL update of
called zL , which represents the total weighted sum regarding vehicle vi from the reporting vehicle vj , the evaluation process
the weights of the activation of the previous layer plus a bias of GTL is shown in Algorithm 1. Note that the change in trust
of the current layer rating (Ti ) corresponds to the output (1) of the feedforward
neural network.
zL = ωL a(L−1) + bL . (4)

It is important to figure out how sensitive the changes of D. Trust Validation and Archiving by Blockchain
weight will influence the cost function. That implies the key Once the GTLs for all the vehicles are calculated by the
is to figure out what is the derivative of C in regarding to ωL , local RSU, the identity of each vehicle is first validated by

Authorized licensed use limited to: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: AIT: AI-ENABLED TRUST MANAGEMENT SYSTEM FOR VEHICULAR NETWORKS USING BLOCKCHAIN TECHNOLOGY 3163

Algorithm 1: GTL Update

Input: VehicleID of vi , Ti
Output: Updated trust rating of vi (Ti )
if VehicleID is valid andTi = 0 then
Ti ← Ti + Ti ;
else
stop the update process;
end
if Ti ≥ 0.5 then
trust validation for vi succeeds;
else
trust validation for vi fails; Fig. 5. How RSU manages blockchain for the neighboring vehicles.
end
validated and becomes successful. Otherwise, the validation
fails, and the blockchain has to be resynchronized by back-
the blockchain technology, and the GTL will be encoded and tracking previous transactions until the hashes match, which
added as a new block to the blockchain. This provides for is achieved by both vi and r.
an unbreakable chain of GTLs. Without blockchain, fake trust As shown in Fig. 5, each RSU uses a tree structure to store
levels can be distributed by malicious RSUs. the blockchains for all the neighboring vehicles, in which each
More specifically, RSU keeps track of all the transaction branch of the tree corresponds to one vehicle’s blockchain. As
history from all the vehicles that are located within the direct a result, the computational complexity to locate each vehicle’s
communication range of the RSU by using the Merkle root blockchain will be O(1). Therefore, the overall computational
hash value. In addition, the blockchain maintained in each complexity to use blockchain to validate the hash values will
RSU is built by following the tree structure. The root node r be O(n).
of this tree represents the RSU, and its immediate child nodes
vi (i = 0, 1, 2, . . . , n) represent the vehicles that are located E. Global Trust-Level Voting and Dissemination by All the
within that RSU’s direct communication range. Therefore, the RSUs
number of immediate child nodes for a specific RSU should In Section IV-D, each RSU calculates GTLs for all the vehi-
be equal to the number of vehicles that are in its direct com- cles in its transmission range using information collected from
munication range. Each vi corresponds to a blockchain for the all other vehicles in that range. In general, the composition
corresponding vehicle. Fig. 5 demonstrates how the blockchain of vehicles is different for each RSU, because of transmis-
technology is applied to maintain all the trust ratings in the sion range limitation. Therefore, GTL for the same vehicle
AIT system. could be different when being calculated by different RSUs.
Unlike a recent research effort in this domain [21], in which To address this issue, all RSUs will share the GTLs they have
vehicles maintain blockchain and store all the blocks, which calculated for vehicles that are in their communication range
is unsafe and inefficient, in our approach each RSU manages at a predefined time interval. The average of all the trust levels
the blockchain for all the vehicles in its direct communication for each vehicle is then calculated and used as its GTL. This
range, and it only provides information to vehicles as needed. new GTL will also be encoded and added to the blockchain
Vehicles only keeps the Merkle root, while the RSU keeps all as a new block. Different evaluation time durations could be
the blocks. By this means, the resource intensive tasks, such used in practice because short-term trust may be okay for less
as creation and mining of blocks, are only performed on the sensitive messages, such as temperature, but a longer term of
RSUs instead of vehicles, which could improve the scalability trust is required for more critical messages, such as traffic con-
of vehicular networks. gestion or accidents. The new GTLs are distributed as links
Moreover, the Merkle root, which is a hash value gener- to the blockchain so that the GTL cannot be spoofed by a
ated by the hash function, serves as a stub between vehicles malicious RSU anymore.
and RSU to protect the blocks from malicious attacks from When the trust level of a vehicle drops below 0.5, it is
the RSU itself and other malicious vehicles that attempt to considered untrustworthy, and any message that it sends out
change the blocks. After a vehicle vi generates a new mes- will be ignored by other vehicles and RSUs for forwarding.
sage, it will be sent to the RSU r to generate a block and then However, even if a vehicle is untrustworthy, its messages are
be attached to the end of the blockchain that corresponds to vi still evaluated for trust level update purpose. If its trust level
in the tree structure. Every time with a newly added block, the rises above 0.5 later, it becomes trustworthy again. The same
merkle value that corresponds to each vehicle in this RSU’s principle applies to RSUs: they are also dynamically evaluated
tree structure will be updated. The hash will be the new stub as of their trust levels. If the trust level of a RSU falls below 0.5,
the transaction continues. As a result, both vi and r contain the then it will be viewed as untrustworthy, and its messages will
up to date Merkle value. In the future, before every transac- be ignored. It could only become trustworthy later again when
tion, vi will send a request to r asking for its current Merkle its trust level rises above 0.5, which could be simply decided
root value. If the hash values match, then the transaction is by a single vehicle or RSU.

Authorized licensed use limited to: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.
3164 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 5, MARCH 1, 2021

TABLE III
S IMULATION PARAMETERS FOR E XPERIMENTAL S TUDY

it more difficult to locate the malicious vehicles that may only

be in the simulation area for a short period of time during
the simulation. To address this issue, we adjusted the SUMO
data to keep the vehicles from leaving the simulation area: if
a vehicle is about to leave the simulation area, its direction of
Fig. 6. Map generated by SUMO, which represents part of the New York
city road network. travel is reversed, and it will travel back to where it started.
The network simulations were set up by placing 50, 100, and
200 vehicles in the simulation area during the simulation runs.
Unlike vehicles, if an RSU is compromised by the adversary The raw SUMO simulation uses 900 vehicles to generate a
and considered as malicious after trust evaluation by a vehi- training set of 20 000 inputs/outputs. Then, a simulation was
cle or another RSU, the untrustworthy RSU will be reported run using SUMO and all the data generated was outputted to
to other RSUs. The other RSUs have the authority to make a .xml file containing each individual vehicle’s information,
the final decision on whether this RSU is malicious or not such as real-time location and speed. The data used are split
based on their prior observations and knowledge on the com- to 60% training, 20% generalization, and 20% validation.
promised RSU. Once the compromised RSU is confirmed as We use NS-2 as the network simulator [31], and the
untrustworthy by other RSUs, its security credentials will be simulation parameters are listed in Table III.
revoked and it will no longer be able to participate in vehic- To evaluate the performance of the proposed AIT system,
ular networks. All the RSUs have to decide whether or not we use the following three metrics, namely, precision (P),
to allow an RSU rejoin the network after its trust level rises recall (R), and communication overhead (CO). Precision and
above 0.5 again. By this means, we could ensure that RSUs recall are widely used performance metrics to evaluate the
always maintain higher security requirements when compared accuracy in machine learning and classification [32]. More
to vehicles. specifically, in this research, precision and recall are defined
as follows:
V. E XPERIMENTAL S TUDY Number of truly malicious vehicles detected
P= (6)
In this section, we provide a detailed discussion on how we Total number of malicious vehicles reported
set up the simulation for the experimental study, and also the Number of truly malicious vehicles detected
R= . (7)
experimental results based on the network simulation. Total number of truly malicious vehicles
An example of network simulation that we built is shown
A. Real-Time Virtual Map Simulation in Fig. 7. We have compared the performance of the proposed
In this article, we use simulator of urban mobility AIT system with the following two baseline approaches: 1) the
(SUMO) [30] to download and generate maps for our vehicu- well-known weighted voting approach, which was widely
lar network simulation using its Web Wizard utility. SUMO is adopted in various existing trust management schemes for
a widely used road traffic simulation tool, which is designed wireless networks [15], [33], [34] and 2) the ART scheme [7].
to handle large road networks. SUMO uses the downloaded
road information to generate vehicles that are running on the B. Experimental Results
downloaded map with different directions, speeds, and ini- There are three series of experiments, and each of them
tial locations. The data set generated by SUMO provides a aims at evaluating the proposed AIT system from different
good simulation that mimics what is happening in the real perspectives. The first series of experiments aims at evaluating
world. Fig. 6 shows a map that is downloaded and generated the performance of the AIT system under different adversary
by SUMO, which represents the part of the road networks in models. As discussed in Section III-B, we have considered
New York city. The size of the map is 602 by 622 m. the following three types of malicious attacks in this work,
Using this map data generated from SUMO, a real-time namely, SA, BMA, and ZA. The experimental results for the
vehicular network simulation can be performed. The data first series of experiments are shown in Figs. 8–10.
generated from SUMO last for 900 s, which includes the As we can find from Figs. 8–10, the proposed AIT gen-
information for the location, ID, and velocity of vehicles. In erally outperforms both two baseline approaches. The high
the raw simulation data from SUMO, vehicles could enter and precision and recall values for the AIT system are achieved
leave the simulation area during the simulation, which makes owing to the application of the deep learning algorithm and

Authorized licensed use limited to: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: AIT: AI-ENABLED TRUST MANAGEMENT SYSTEM FOR VEHICULAR NETWORKS USING BLOCKCHAIN TECHNOLOGY 3165

Fig. 7. Example of the network simulation for our experimental study.

Fig. 8. Performance comparison of AIT versus baseline approaches with the SA pattern. (a) Precision. (b) Recall.

Fig. 9. Performance comparison of AIT versus baseline approaches with the BMA pattern. (a) Precision. (b) Recall.

the blockchain technology, because the deep learning algo- resulting in a more accurate trust evaluation and detection of
rithm works well especially when there is a large volume of malicious vehicles.
data, which is the case in the vehicular networks. In addition, Moreover, as clearly shown in Fig. 8, we find that the gap
the application of the blockchain technology will ensure the between AIT and the baseline approaches is not that significant
validity and authenticity of both vehicles and messages, thus when the adversary is adopting the SA pattern, which indicates

Authorized licensed use limited to: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.
3166 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 5, MARCH 1, 2021

Fig. 10. Performance comparison of AIT versus baseline approaches with the ZA pattern. (a) Precision. (b) Recall.

Fig. 11. Performance comparison of AIT versus baseline approaches with different numbers of vehicles. (a) Precision. (b) Recall. (c) CO.

Fig. 12. Performance comparison of AIT versus baseline approaches with different numbers of malicious vehicles. (a) Precision. (b) Recall. (c) CO.

that the SA, as its name suggests, may not be very difficult to when there are more benign vehicles in the network. Finally, as
be detected. In contrast, the difference between AIT and the depicted in Fig. 11(c), the CO brought by AIT is slight lower
two baseline approaches is more significant with the BMA and than both weighted voting and ART, and even for 200 node
ZA patterns, which is demonstrated in Figs. 9 and 10. This case, the CO for AIT is under 8%, which suggests that the
is the case because AIT benefits from the application of both proposed AIT system does not generate too much additional
deep learning and blockchain, which makes it more resistant network traffic.
to these two types of more sophisticated attacks. Fig. 12 shows the effect of malicious nodes on AIT and the
The second series of experiments aim at evaluating the two baseline approaches. As Fig. 12(a) illustrates, the precision
performance of the proposed AIT system in different exper- for AIT is generally better compared to those for the two base-
imental settings, such as the different numbers of nodes, line approaches. When there are 10% of malicious nodes in
different numbers of malicious nodes, and different travel- the vehicular network, the ART scheme works slightly bet-
ing speeds. The experimental results for the second series of ter than AIT, but the difference is minimal. Fig. 12(b) shows
experiments are depicted in Figs. 11–13. that the recall for AIT is always higher than the two baseline
From Fig. 11(a), we can find that the AIT system generally approaches. Based on the comparison in terms of precision
works better when compared with the weighted voting method and recall, we also notice a performance drop when the per-
and ART scheme in terms of precision. As for the recall, centage of malicious nodes increases. This is rational because
Fig. 11(b) shows that AIT always outperforms both weighted with a higher percentage of malicious nodes in the networks,
voting and the ART scheme when the number of nodes differs. it is generally more difficult to receive the true traffic mes-
Moreover, the precision and recall are both higher when there sage from a trustworthy neighbor, which makes it harder to
are more nodes in the network. This is the case because it is accurately evaluate the trust of vehicles and also successfully
more likely to receive true traffic updates from other vehicles identify all the malicious vehicles.

Authorized licensed use limited to: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: AIT: AI-ENABLED TRUST MANAGEMENT SYSTEM FOR VEHICULAR NETWORKS USING BLOCKCHAIN TECHNOLOGY 3167

Fig. 13. Performance comparison of AIT versus baseline approaches with different traveling speeds of vehicles. (a) Precision. (b) Recall. (c) CO.

TABLE IV
P ERFORMANCE C OMPARISON A MONG FF, RNN, AND CNN We compare the performance of feedforward neural network
(FF) with two other well-known deep learning models, namely,
recurrent neural network (RNN) and convolutional neural
network (CNN). The result is shown in Table IV.
From Table IV, we can find that all the three deep learning
models perform similarly in terms of the accuracy, and the dif-
ference is minimal. In contrast, when it comes to the training
time cost, FF clearly outperforms the other two deep learning
Fig. 13 demonstrates the performance of AIT and the two models. This is true because FF has the simplest connection
baseline approaches with different node traveling speeds. As structure among nodes when compared to RNN: connections
the map generated by SUMO represents part of the road between the nodes in FF do not form a cycle, whereas con-
networks in New York city, we are fully aware that in reality nections between nodes form a directed graph in RNN. As
vehicles may not travel as fast as 20 m/s (which is equivalent to for CNN, the use of convolutional function generally incurs
72 km/h or approximately 45 m/h), due to the traffic condition additional time overhead when compared to the threshold func-
as well as the designated city speed limit. However, we would tion that FF generally uses for the hidden layers. Therefore,
still like to observe the performance of the three approaches we conclude that FF is a suitable deep learning model for the
when vehicles are traveling at that speed, so that we could get a AIT system because of its high accuracy and low training time
more holistic view of how well they perform, especially when cost.
the traveling speed is relatively high. As shown in Fig. 13(a),
AIT always outperforms both two baseline approaches in terms
of precision, which clearly indicates that it works well when VI. D ISCUSSION
vehicles are traveling at different speeds. Fig. 13(b) shows that A. Privacy Concern in Vehicular Networks
in terms of recall, AIT works better when vehicles are trav- One of the assumptions that we have made in this research
eling at a lower speed, and ART may work slightly better in is that each vehicle’s identity is unique and it could be used
terms of the recall when the traveling speed is higher. But it to distinguish one vehicle from others, which could pos-
is worth noting that the recall value for AIT and ART is sim- sibly cause privacy leakage. In general, privacy has been
ilar when the traveling speed is 10 and 20 m/s. As shown viewed as one of the major concerns when securing vehicular
in Fig. 13(c), the baseline, ART, and ATI systems all use networks [35].
communications bandwidth in order to evaluate trust on the To address the privacy concern in vehicular networks, one
network. The communication charts all show the percentage possible solution is to simply use the public key of each
of total communication bandwidth that is used by each system. vehicle as its identifier [8]. Alternatively, a more sophis-
If malicious cars are found quicker using local trust, the extra ticated solution could be obtaining a pseudonym from the
communications needed for future evaluations is eliminated. well-designed pseudonym scheme and using it to represent
To summarize, we can see from Figs. 11–13 that the the vehicle [36]. However, it is worth noting that both solu-
proposed AIT system generally outperforms the weighted vot- tions will incur a substantial amount of computational and CO,
ing and ART schemes under different circumstances, which which may also be a valid concern for vehicular networks due
also indicates that the AIT system can accurately evaluate to the resource constraint nature. As a result, an open research
the trust of vehicles and detect malicious ones with small problem that should be further explored is how to effectively
overhead. evaluate trust in vehicular networks while maintaining the
In the third series of experiments, our goal is to study the privacy of vehicles.
effect of different deep learning models and check to see which
of them may be most suitable for the AIT system.
It is well understood that nodes in vehicular networks are VII. C ONCLUSION
constantly moving, and the network topology is rapidly chang- In this article, we proposed AIT, an AIT for vehicular
ing due to the node mobility. Consequently, it is critical for networks using block technology. In the proposed AIT system,
the deep learning model to achieve high accuracy while taking both vehicles and RSUs will participate in the trust man-
less time to train the model. agement process, and both local trust evaluation and global

Authorized licensed use limited to: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.
3168 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 5, MARCH 1, 2021

trust evaluation help determine the trust of vehicles in a more [20] F. Ahmad, F. Kurugollu, A. Adnane, R. Hussain, and F. Hussain,
accurate fashion. Moreover, the application of blockchain tech- “MARINE: Man-in-the-middle attack resistant trust model in con-
nected vehicles,” IEEE Internet Things J., vol. 7, no. 4, pp. 3310–3322,
nology can ensure the validity and authenticity of information, Apr. 2020.
such as the identity of nodes and the calculated trust values. [21] W. Dong, Y. Li, R. Hou, X. Lv, H. Li, and B. Sun, “A blockchain-based
To validate and evaluate the performance of the AIT system, hierarchical reputation management scheme in vehicular network,” in
Proc. IEEE Global Commun. Conf. (GLOBECOM), 2019, pp. 1–6.
extensive experiments have been conducted through network [22] X. Liu, H. Huang, F. Xiao, and Z. Ma, “A blockchain-based trust
simulation. The experimental results clearly show that the AIT management with conditional privacy-preserving announcement scheme
system can evaluate the trust of nodes and detect malicious for VANETs,” IEEE Internet Things J., vol. 7, no. 5, pp. 4101–4112,
May 2020.
ones in an accurate and efficient manner. [23] H. Xia, S.-S. Zhang, Y. Li, Z.-K. Pan, X. Peng, and X.-Z. Cheng, “An
attack-resistant trust inference model for securing routing in vehicular ad
R EFERENCES hoc networks,” IEEE Trans. Veh. Technol., vol. 68, no. 7, pp. 7108–7120,
Jul. 2019.
[1] H. Clancy. (Oct. 2015). GM Already Has 1 Million Connected Cars [24] K. Gu, X. Dong, and W. Jia, “Malicious node detection scheme based
on the Road. [Online]. Available: https://fortune.com/2015/10/02/gm-1- on correlation of data and network topology in fog computing-based
million-connected-cars/ vanets,” IEEE Trans. Cloud Comput., early access, Apr. 3, 2019,
[2] F. Lambert. (Apr. 2019). Tesla Unveils Robotaxi Plan for Self- doi: 10.1109/TCC.2020.2985050.
Driving Ride-Sharing Network Next Year. [Online]. Available: [25] S. Gyawali and Y. Qian, “Misbehavior detection using machine learn-
https://electrek.co/2019/04/22/tesla-robotaxi-network-self-driving-fleet- ing in vehicular communication networks,” in Proc. IEEE Int. Conf.
ride-sharing-cars/ Commun. (ICC), 2019, pp. 1–6.
[3] W. Li, A. Joshi, and T. Finin, “Coping with node misbehaviors in [26] J. Chen, G. Mao, C. Li, and D. Zhang, “A topological approach to
ad hoc networks: A multi-dimensional trust management approach,” in secure message dissemination in vehicular networks,” IEEE Trans. Intell.
Proc. 11th IEEE Int. Conf. Mobile Data Manag. (IEEE MDM), 2010, Transp. Syst., vol. 21, no. 1, pp. 135–148, Jan. 2020.
pp. 85–94. [27] W. Li, A. Joshi, and T. Finin, “SVM-case: An SVM-based context aware
[4] W. Li, J. Parker, and A. Joshi, “Security through collaboration and trust security framework for vehicular ad-hoc networks,” in Proc. IEEE 82nd
in manets,” Mobile Netw. Appl., vol. 17, no. 3, pp. 342–352, 2012. Veh. Technol. Conf. (VTC-Fall), 2015, pp. 1–5.
[5] Z. Wei, H. Tang, F. R. Yu, M. Wang, and P. Mason, “Security [28] H. Sedjelmaci, S. M. Senouci, and T. Bouali, “Predict and prevent
enhancements for mobile ad hoc networks with trust management from misbehaving intruders in heterogeneous vehicular networks,” Veh.
using uncertain reasoning,” IEEE Trans. Veh. Technol., vol. 63, no. 9, Commun., vol. 10, pp. 74–83, Oct. 2017.
pp. 4647–4658, Nov. 2014. [29] D. J. Montana and L. Davis, “Training feedforward neural networks
[6] K. Rostamzadeh, H. Nicanfar, N. Torabi, S. Gopalakrishnan, and using genetic algorithms,” in Proc. IJCAI, vol. 89, 1989, pp. 762–767.
V. C. Leung, “A context-aware trust-based information dissemination [30] P. A. Lopez et al., “Microscopic traffic simulation using SUMO,” in
framework for vehicular networks,” IEEE Internet Things J., vol. 2, Proc. 21st Int. Conf. Intell. Transp. Syst. (ITSC), 2018, pp. 2575–2582.
no. 2, pp. 121–132, Apr. 2015. [31] NS-2 Network Simulator. Accessed: Dec. 27, 2020. [Online]. Available:
[7] W. Li and H. Song, “ART: An attack-resistant trust management scheme http://nsnam.sourceforge.net/wiki/index.php/Main_Page
for securing vehicular ad hoc networks,” IEEE Trans. Intell. Transp. [32] J. Davis and M. Goadrich, “The relationship between precision-
Syst., vol. 17, no. 4, pp. 960–969, Apr. 2016. recall and roc curves,” in Proc. 23rd Int. Conf. Mach. Learn., 2006,
[8] Z. Yang, K. Yang, L. Lei, K. Zheng, and V. C. M. Leung, “Blockchain- pp. 233–240.
based decentralized trust management in vehicular networks,” IEEE [33] S. Buchegger and J.-Y. Le Boudec, “Performance analysis of the con-
Internet Things J., vol. 6, no. 2, pp. 1495–1505, Apr. 2019. fidant protocol,” in Proc. 3rd ACM Int. Symp. Mobile Ad Hoc Netw.
[9] D. He, C. Chen, S. Chan, J. Bu, and A. V. Vasilakos, “Retrust: Comput., 2002, pp. 226–236.
Attack-resistant and lightweight trust management for medical sen- [34] I.-R. Chen, F. Bao, M. Chang, and J.-H. Cho, “Dynamic trust manage-
sor networks,” IEEE Trans. Inf. Technol. Biomed., vol. 16, no. 4, ment for delay tolerant networks and its application to secure routing,”
pp. 623–632, Jul. 2012. IEEE Trans. Parallel Distrib. Syst., vol. 25, no. 5, pp. 1200–1210,
[10] F. Bao, R. Chen, M. Chang, and J.-H. Cho, “Hierarchical trust man- May 2014.
agement for wireless sensor networks and its applications to trust-based [35] B. Parno and A. Perrig, “Challenges in securing vehicular networks,” in
routing and intrusion detection,” IEEE Trans. Netw. Service Manag., Proc. Workshop Hot Topics Netw. (HotNets-IV), 2005, pp. 1–6.
vol. 9, no. 2, pp. 169–183, Jun. 2012. [36] J. Petit, F. Schaub, M. Feiri, and F. Kargl, “Pseudonym schemes in
[11] Y. Ren, V. I. Zadorozhny, V. A. Oleshchuk, and F. Y. Li, “A novel vehicular networks: A survey,” IEEE Commun. Surveys Tuts., vol. 17,
approach to trust management in unattended wireless sensor networks,” no. 1, pp. 228–255, 1st Quart., 2015.
IEEE Trans. Mobile Comput., vol. 13, no. 7, pp. 1409–1423, Jul. 2014.
[12] I.-R. Chen, F. Bao, and J. Guo, “Trust-based service management
for social Internet of Things systems,” IEEE Trans. Depend. Secure
Comput., vol. 13, no. 6, pp. 684–696, Nov./Dec. 2016.
[13] W. Li, H. Song, and F. Zeng, “Policy-based secure and trustworthy sens-
ing for Internet of Things in smart cities,” IEEE Internet Things J., vol. 5,
no. 2, pp. 716–723, Apr. 2018.
[14] M. A. Azad, S. Bag, F. Hao, and A. Shalaginov, “Decentralized self-
enforcing trust management system for social Internet of Things,” IEEE
Internet Things J., vol. 7, no. 4, pp. 2690–2703, Apr. 2020.
[15] M. Raya, P. Papadimitratos, V. D. Gligor, and J.-P. Hubaux, “On data-
centric trust establishment in ephemeral ad hoc networks,” in Proc. IEEE
INFOCOM 27th Conf. Comput. Commun., 2008, pp. 1238–1246.
[16] X. Yao, X. Zhang, H. Ning, and P. Li, “Using trust model to ensure reli- Chenyue Zhang received the B.S. degree in
able data acquisition in VANETs,” Ad Hoc Netw., vol. 55, pp. 107–118, mechanical engineering from Case Western Reserve
Feb. 2017. University, Cleveland, OH, USA, in May 2016, and
[17] J. Guo et al., “TROVE: A context awareness trust model for VANETs the master’s degree in computer science from New
using reinforcement learning,” IEEE Internet Things J., vol. 7, no. 7, York Institute of Technology, New York, NY, USA,
pp. 6647–6662, Jul. 2020. in December 2020.
[18] H. El Sayed, S. Zeadally, and D. Puthal, “Design and evaluation of a She worked on developing prototypes for indus-
novel hierarchical trust assessment approach for vehicular networks,” trial machines and manufacturing robotics, which
Veh. Commun., vol. 24, Aug. 2020, Art. no. 100227. were built and developed based upon biological
[19] J. Zhang, K. Zheng, D. Zhang, and B. Yan, “AATMS: An anti-attack trust features and principles. Her research interests are
management scheme in vanet,” IEEE Access, vol. 8, pp. 21077–21090, in cyber security, blockchain, vehicular networks,
2020. machine vision, and machine learning.

Authorized licensed use limited to: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: AIT: AI-ENABLED TRUST MANAGEMENT SYSTEM FOR VEHICULAR NETWORKS USING BLOCKCHAIN TECHNOLOGY 3169

Wenjia Li (Senior Member, IEEE) received Yuansheng Luo received the B.S. and M.S. degrees
the Ph.D. degree in computer science from from Hunan University, Changsha, China, in 2002
the University of Maryland Baltimore County, and 2005, respectively, and the Ph.D. degree from
Baltimore, MD, USA, in 2011. Xi’an Jiaotong University, Xi’an, China, in 2010.
He was a tenure-track Assistant Professor He is currently a Lecturer with the School
of Computer Science with Georgia Southern of Computer and Communication Engineering,
University, Statesboro, GA, USA, from 2011 Changsha University of Science and Technology,
to 2014. In 2014, he joined the Department Changsha. His current interests include fog/edge
of Computer Science, New York Institute of computing, service computing, and wireless
Technology, New York, NY, USA, as a tenure-track networks.
Assistant Professor, where he has been a Tenured
Associate Professor since September 2020. His research has been supported
by the National Institute of Health and the U.S. Department of Transportation
Yupeng Hu (Senior Member, IEEE) received the
Region two University Transportation Research Center. He has authored or
M.S. and Ph.D. degrees in computer science from
coauthored over 80 peer-reviewed publications in various journals and confer-
Hunan University, Changsha, China, in 2005, and
ence proceedings. His current research interests include cyber security, mobile
2008, respectively.
computing, and wireless networking, particularly security, trust, and policy
He is currently a Professor with the College
issues for wireless networks, cyber–physical systems, Internet of Things, and
of Computer Science and Electronic Engineering,
intelligent transportation systems.
Hunan University, where he is the Dean of the
Dr. Li was a recipient of the 2019 IEEE Region 1 Technological Innovation
Department of Cyberspace Security. He was with the
(Academic) Award. He has served as the Organizing Committee Member for
Department of Computer Science and Engineering,
many international conferences, such as ACM WiSec, IEEE DySPAN, IEEE
University of Texas at Arlington, Arlington, TX,
MDM, IEEE IPCCC, and IEEE Sarnoff, and he also served as a Reviewer
USA, as a Visiting Scholar from 2015 to 2016. He
for many prestigious journals, such as the IEEE T RANSACTIONS ON
was also with IBM China Development Laboratory, Shanghai, China, as an
W IRELESS C OMMUNICATIONS, the IEEE T RANSACTIONS ON PARALLEL
Academic Visitor in 2012. He has published more than 60 journal articles,
AND D ISTRIBUTED S YSTEMS , the IEEE T RANSACTIONS ON I NFORMATION
book chapters, and refereed conference papers. His research interests include
F ORENSICS AND S ECURITY, the IEEE T RANSACTIONS ON D EPENDABLE
big data and storage systems security, erasure coding, AI security, and network
AND S ECURE C OMPUTING , and the IEEE I NTERNET OF T HINGS J OURNAL .
and system security.
Prof. Hu is a Senior Member of ACM.

Authorized licensed use limited to: Carleton University. Downloaded on June 03,2021 at 21:03:50 UTC from IEEE Xplore. Restrictions apply.