Trust-Related Attacks and Their Detection: A Trust Management Model for the Social IoT

Abstract

The integration of social networking concepts into the Internet of Things (IoT) has led to the so called Social Internet of Things paradigm, according to which the objects are capable of establishing social relationships in an autonomous way with respect to their owners. Within this scenario“, things" interact opportunistically with their peers to seek needed services. However, attacks and malfunctions in the IoT can outweigh any of its benefits if not handled adequately. In this paper, we focus on the possible types of trust attacks that can affect the IoT and propose a trust management model able to overcome all the analyzed attacks. Simulations show how the proposed model can effectively isolate almost any malicious nodes in the network at the expense of an increase in the number of transactions needed for the model to converge.

Full text

1
Trust-related Attacks and their Detection: a Trust
Management Model for the Social IoT
C. Marche, M. Nitti
DIEE, University of Cagliari, Italy
National Telecommunication Inter University Consortium - Research Unit of Cagliari - Italy
{claudio.marche, michele.nitti}@unica.it
Abstract—The integration of social networking concepts into
the Internet of Things (IoT) has led to the so called Social
Internet of Things paradigm, according to which the objects
are capable of establishing social relationships in an autonomous
way with respect to their owners. Within this scenario, “things”
interact opportunistically with their peers to seek needed services.
However, attacks and malfunctions in the IoT can outweigh any
of its benefits if not handled adequately. In this paper, we focus
on the possible types of trust attacks that can affect the IoT
and propose a trust management model able to overcome all the
analyzed attacks. Simulations show how the proposed model can
effectively isolate almost any malicious nodes in the network at
the expense of an increase in the number of transactions needed
for the model to converge.
Index Terms—Internet of Things, Social Internet of Things,
Trustworthiness Management, Machine Learning
I. INTRODUCTION
The Internet of Things (IoT) has become a reality with
billions of devices able to send key information about the
physical world and implementing simple actions, which leads
to the paradigm of the anytime and anyplace connectivity for
anything [1]. The massive amount of data flowing through the
IoT has pushed forward the development of new applications
in several domains, such as the management of industrial
production plants, the logistics and transport supply chain, the
e-health, the smart building, just to cite a few.
Such future IoT applications are likely developed making
use of a service oriented architecture where each device can
play the role of a service provider or a service requester, or
both. IoT is moving towards a model where things look for
other things to provide composite services for the benefit of
human beings (object-object interaction). With such an inter-
action model, it is essential to understand how the information
provided by each object can be processed automatically by any
other peer in the system. This cannot clearly disregard the level
of trustworthiness of the object providing information and
services, which should take into account the profile and history
of it. Although we experience and rely on trust during our
interactions in everyday life, trust can have many definitions
so that it is challenging to define it accurately. The literature
on trust is also quite confusing, since it manifests itself in
fairly different forms. In this paper, we adopt the following
definition for trust:
Trust is the subjective probability by which an individual,
the trustor, expects that another individual, the trustee, per-
forms a given action on which its welfare depends [2].
In the IoT scenario, the requester has the role of the trustor
and has to trust that the provider, which is then the trustee, will
provide the required service. However, misbehaving devices
may perform several types of attacks for their own gain
towards other IoT nodes: they can provide false services or
false recommendations, they can act alone or create a group
of colluding devices to monopoly a class of services. If not
handled adequately, attacks and malfunctions would outweigh
any of the benefits of the IoT [3] [4]. For example, in February
2020, Simon Weckert transported 99 smartphones in a hand-
cart and was able to generate virtual traffic jam in Google
Maps1. In this scenario, trustworthiness management models
have to solve the important issue to identify and understand
which, among the nodes in the network, are trustworthy and
can then lead to successful collaborations.
Several works have been proposed to address the problem
of trust management in the IoT; however, all these works are
usually tested considering only a subset of the possible attack
patterns. Indeed, attack patterns are highly heterogeneous
so that malicious nodes try to exploit the weak points of
trustworthiness algorithms so as to operate unnoticed.
An approach, which is recently gaining increasing popu-
larity and has the potential to properly address this issue, is
based on the exploitation of social networking notions into
the IoT, as formalized by the Social IoT (SIoT) concept [5].
According to this vision, things create relationships among
them as humans do: this approach introduces the vision of
social relationships among different devices, so that they
are more willing to collaborate with friends w.r.t. strangers.
This is expected to make the exchange of information and
services among different devices easier and to perform the
identification of malicious nodes by creating a society-based
view about the trust level of each member of the community.
Our paper works in this direction with the goal to recognize
the trustworthiness attacks and thus provides the following
contributions:
•First, we analyze all the possible types of trust attacks
described in the literature that can affect the IoT and
briefly review the resiliency of existing models against
the identified trust related attacks.
•Second, we propose a decentralized trust management
model, based on a Machine Learning algorithm, which
makes use of novel parameters, namely the goodness, the
1http://www.simonweckert.com/googlemapshacks.html

2
usefulness and the perseverance score. Thanks to these
scores, the model trains and adapts itself and it is able to
identify and react to all possible malicious attacks.
•Third, by using a dataset of real IoT objects, we conduct
extensive experiments to show how our model reacts
to each type of attack. Furthermore, we compare our
algorithm with two well-acknowledged state-of-the-art
models: the experiment results show that even if our algo-
rithm shows the slightly worse performance when under
attack by simple mechanisms, it is able to outperform the
other two models when considering a network with a mix
of different types of attack.
The rest of the paper is organized as follows: Section
II presents the scenario of social IoT, a brief survey on
trust management models and the possible types of attacks.
In Section III, we define the problem, introduce the used
notations and illustrate the proposed trust management model.
Section IV presents the system performance against all the
type of attacks analyzed in Section II while Section V draws
final remarks.
II. BACKGROU ND
A. The Social Internet of Things
The SIoT represents the convergence of the technologies
belonging to two domains: IoT and Social Networking. The
result is the creation of social networks in which things are
nodes that establish social links as humans do [5]. According
to the SIoT model, every node is an object that is capable of es-
tablishing social relationships with other things autonomously,
according to rules set by the owner. This concept is fast
gaining ground thanks to the key benefits deriving from the
potentials of the social networks within the IoT domain, such
as: simplification in the navigability of a dynamic network of
billions of objects [5]; efficiency in the dynamic discovery,
selection and composition of services (and of information
segments) provided by distributed objects and networks [6];
robustness in the management of the trustworthiness of objects
when providing information and services [7].
When it comes to the IoT paradigm, the idea is to exploit
social awareness as a means to turn communicating objects
into autonomous decision-making entities. The new social
dimension shall, somehow, be able to mimic interactions
among users and to motivate a drift from an egoistic behavior
to altruism or reciprocity. The main principle is to enable
objects to autonomously establish social links with each other
(by adhering to rules set by their owners) so that “friend”
objects exchange data in a trustworthy manner. According to
this model, a set of forms of socialization among objects is
foreseen. The parental object relationship (POR) is defined
among similar objects, built in the same period by the same
manufacturer (the role of the family is played by the pro-
duction batch). Moreover, objects can establish a co-location
object relationship (CLOR) and co-work object relationship
(CWOR), like humans do when they share personal (e.g.,
cohabitation) or public (e.g., work) experiences. A further type
of relationship is defined for objects owned by the same user
(mobile phones, game consoles, etc.) that is named ownership
object relationship (OOR). The last relationship is established
when objects come into contact, sporadically or continuously,
for reasons purely related to relations among their owners
(e.g., devices/sensors belonging to friends); it is named social
object relationship (SOR). These relationships are created and
updated on the basis of the objects’ features (such as type,
computational power, mobility capabilities, brand, etc.) and
activities (frequency in meeting the other objects, mainly).
However, to fully exploit the benefits of a SIoT network,
a trustworthiness management model, able to defend against
malicious attacks, is needed, which we investigate in this
paper.
B. Trustworthiness Models
This subsection provides a brief overview regarding the
background of trustworthiness management in the IoT. In the
last years, many researchers have tackled this problem, so that
the literature is now quite rich. In this Section, we want to
show the most appreciated models in the literature and do not
intend to cover all the published papers. We classified them
into three categories based on the metric used to compute the
trust value: metrics obtained from social aspects, metrics based
on the Quality of Service (QoS) and mixed approaches, i.e.
papers considering both social and QoS aspects.
Among the works considering social aspects, in [8] the
authors propose an adaptive decentralized trust mechanism
based on social trust. Through a weighted sum, the authors
combine factors that concern the cooperativeness and the
social communities and demonstrate the effectiveness of the
model making use of two real-world social IoT scenarios.
Another trust model concerning social trust is presented in
[9]. Authors propose a machine learning-based approach to
formalize the trust evaluation as a classification problem. The
feature vector in a social network is constructed according to
social factors like the reputation and the centrality. Another
social approach is used in [10]. Throughout a few SIoT trust
metrics as centrality, community interest and cooperativeness,
the authors illustrate a trust management scheme to facilitate
an automatic trustworthy decision making based on the be-
havior of smart objects. Two social scenarios are described
in [11] and [12]. In the first work, the authors take into
account metrics such as social similarity and the importance
of the service. The resulting trust management algorithm is
developed using social relationships to compute the trust level
of the nodes in a SIoT network. In the second one, the authors
propose a centralized trust-based protocol for mobile objects.
To guarantee the trust accuracy between the devices the system
makes use of friendships and social contacts.
Concepts of QoS are used for example in [13]: authors
present a remote attestation mechanism for the sensing layer
node in the IoT. A real-time trust measurement is realized
through a combination of QoS factors, such as transmission
delay, historical data and feedback originated from other
objects. Firstly, a node verifies the identities of the other nodes
and only then measures whether the computing environment
is trustworthy. In [14], the authors compute the trust scores
based on the exchange of feedback, which are provided taking

3
into account QoS factors such as the monetary cost of the
resources, the computation capabilities and the communication
failures. Two other QoS approaches based on centralized
architectures are described in [15] and [16]. In the first work,
the authors propose a policy-based secure scheme for IoT, in
which the trustworthiness of data and devices are evaluated
according to the reporting history and the context in which
the data are collected. In the second one, the centralized
architecture is used for information sharing among health IoT
devices. The proposed trust protocol considers the loss of
probability of health data and the reliability of the IoT devices.
Another approach concerning QoS factors is presented in
[17]. Authors introduce an approach to evaluate the trust of
services combining several QoS attributes (such as availability
and response time) and user’s ratings. The model focuses on
satisfying the users’ choices on web services and it is evaluated
considering the influence of malicious rating.
The last group of papers makes use of both QoS and social
trust metrics to compose the trust value. Among them, in
[18], the authors propose a decentralized trust mechanism
in a social scenario. In that model, each node computes
the trustworthiness of the service providers on the basis of
its own experience and on the opinion of its friends. The
authors analyze the QoS factors, as computation capabilities,
and social factors, such as centrality and credibility. QoS and
social metrics are both considered also by Chen et al. in [19].
They adopt a distributed scheme where each node maintains its
own trust assessments. The QoS factors (i.e., quality reputation
and energy status) are related to the social relationships and
recommendations from the other nodes. Two other mixed
approaches are described in [20] and [21]. In [20], authors
propose a trust evaluation model incorporating heterogeneous
information from direct observation, personal experiences and
global reputation. The subjective algorithm makes use of social
factors, e.g. cooperativeness and community-interest, and of
QoS factors, aggregated with a weighted sum mechanism
and a machine learning to change the weights according to
the particular context. In [21], the authors illustrate an IoT
protocol that uses trust for the evaluation of nodes to make
optimal routing decisions. It computes the trust of nodes by
examining QoS factors, such as the number of exchanged
packets, and the recommendations from the neighbors. A
recent model is described in [22]. Authors propose guidelines
for the design of a decentralized trust management model,
which can be used for assisting humans and devices in the
decision making process.
All the analyzed models are designed and tested to isolate
nodes that implement a subset of the possible types of attacks.
However, the heterogeneity of IoT scenarios call for models
with no weak points, while existing works show a common
limitation: they are not able to properly identify all the type of
malicious attacks. The next subsection shows all the possible
malicious behaviours that can be implemented in a network.
C. Trustworthiness Attacks
Two different behaviors can be considered in a network
[23]: one is always benevolent and cooperative, while the other
one is a strategic behavior corresponding to an opportunistic
participant who cheats whenever it is advantageous for it to
do so. The goal of a node performing maliciously is usually to
provide low quality or false services in order to save its own
resources; at the same time, it aims to maintain a high value
of trust toward the rest of the network so that other nodes
will be agreeable to provide their services when requested.
This strategy, even if successful for a single node at first
sight, involves a huge risk for the network because trusting
the information from malicious devices could lead to serious
compromises within the network and this has a direct impact
on the applications that can be delivered to users [24]. A
trust model has to identify this behaviour to discourage nodes
from implementing it; however, such malicious nodes can
perform several types of trust-related attacks, which represent
the different solutions they adopt to avoid being detected. We
classify trustworthiness attacks based on two dimensions: the
first dimension is related to the target of the attack, i.e. if
the malicious node aims to confuse the network by providing
false services, false recommendations or both. The second
dimension is connected to the size of the attack, i.e. if the
trustworthiness attack is carried on by a single node or by a
group. In the following, we briefly describe the different types
of attacks known in the literature.
The largest group of attacks is composed of single nodes
that indiscriminately provide both bad services and recommen-
dations. In this group, trustworthiness attacks differ based on
the mechanism they adopt in order not to be recognized:
Malicious with Everyone (ME): a malicious node acts
maliciously with everyone. This is the most basic attack:
a node always provides bad services and recommendations,
regardless of the requester [18].
Discrimination Attack (DA): a malicious node modifies
its behavior based on the service requester. This means that a
node can discriminate non-friends nodes or nodes with weak
social ties. As a result, some devices can consider the node as
benevolent while others can label it as malevolent [25].
On-Off Attack (OOA): a node periodically changes its be-
havior, by alternatively being benevolent (ON) and malevolent
(OFF). During the ON state, the node builds up its trust, which
is then used to attack the network [26].
Whitewashing Attack (WA): a node with a bad reputation
leaves the network and then registers again with a different
identity. When the node re-join the network its reputation is
reset to a default value [27].
Self-Promoting Attack (SPA): a malicious node provides
good recommendations for itself in order to be selected as a
service provider. After it is selected as a provider, it provides
only bad services [28].
The other types of attacks concentrate on a single target,
i.e. malicious nodes only provide bad services or bad recom-
mendations.
Bad Mouthing Attack (BMA): this attack is addressed
to ruin the reputation of other nodes; a malicious node only
provides false recommendations to decrease the chance of
benevolent nodes being selected as providers. Usually, this
attack is part of a collusive behavior where a group of nodes

4
TABLE I
CLA SSI FIC ATION O F DI FFER EN T TYP ES OF T RUS TW ORTH IN ESS ATTAC KS .
Size
Single Group
[OSA]
Service [ME] [DA] [OOA]
[WA] [SPA] [SA] [BSA]
Target
Recommendation [BMA]
TABLE II
RES ILI EN CY OF E XIS TI NG MO DE LS AG AIN ST I DEN TI FIED T RUS T REL ATED
ATTACK S.
Ref DA OOA BSA WA BMA SA SPA
[8] X-X X X -X
[9] - - - - - - -
[10] - X- - - - -
[11] - - X-X-X
[12] X-X-X-X
[13] - - - X-X X
[14] - - - - - -
[15] - X X -X- -
[16] - - X-X- -
[17] - - - - X- -
[18] X X - - - - X
[19] X X - - X-X
[20] - - X-X-X
[21] - - - - - X-
[22] - - X-X- -
works together to ruin the reputation of a good node but it can
also be carried on by a single node [29].
Ballot Stuffing Attack (BSA): this is a type of collusive
attack, where a malicious node provides good recommenda-
tions toward another malicious node to boost its reputation
and increase its chances to be selected as the provider [30].
Sybil Attack (SA): a malicious node uses multiple identities
to provide different types of recommendations on the same
service. These multiple identities are usually fake and they
are all responsible for the attack process [31].
Opportunistic Service Attack (OSA): a malicious node
provides good services only when it senses that its trust
reputation is dropping. In this way, the node tries to maintain
an acceptable level of trust in order to still be selected as a
service provider [32].
To sum it up, Table I shows a classification of trust-related
attacks based on the two dimensions identified, while Table
II compares the analyzed models with the attacks they are
able to identify. To the best of our knowledge, all available
trustworthiness models are able to isolate only a subset of
the presented attacks, i.e. they are designed to recognize and
isolate some specific attacks, but none of them is able to
defend efficiently against all the attacks. Table II does not
show the ME attack, which is used as a reference attack by
all the models, and the OSA attack since a node performing
it can not be completely isolated but it is only possible to
reduce the number of times a node acts maliciously due to the
reliability needed to build up the trust.
These attacks span from simple ones, which have a constant
behaviour over time, such as ME, to more complex ones which
are able to change their behaviour over time: among them,
for example, there is the On-Off Attack, the Discriminatory
Attack or the Opportunistic Service Attack, which have all
been tested in our paper. In particular, the OSA is considered
the most complex attack in the literature since it knows exactly
how the trust model implemented in a system works, so it is
able to accurately predict how its trust value will change based
on its behaviour and then behaves accordingly.
III. PROP OS ED SOLUTION
A. Notation and Problem Definition
The focus of this paper is to propose a trust manage-
ment model able to identify all the trust attacks analyzed in
Section II-C and isolate the nodes performing them. In our
modeling, the set of nodes in the Social IoT is represented
by N={n1, ..., ni, ...nI}with cardinality I, where niis
the generic node. The resulting social network, created by
the devices’ relationships, can be described by an undirected
graph G={N ,E}, where E ⊆ {N × N } is the set of
edges, each representing a social relation between a couple
of nodes. The friends of the generic node niare represented
in our model by Ni={nj∈ N :ni, nj∈ E}, that is the set
of nodes that share a relation with it; moreover, we define
Hij ={nh∈ N :nh∈ Ni∩ Nj}as the set of common
friends between niand nj.
Every node in our network can provide one or more
services, so that Sjis the set of service that can be provided
by nj. The reference scenario is then represented by a node
nirequesting a particular service Sh: a Service Discovery
component in the network is able to return to nia list of
potential providers Ph={nj∈ N :Sh∈ Sj}. At this point,
the requester has to select one of the providers in Phbased on
their level of trust. The trust level is usually computed based on
the previous interactions among the nodes. Indeed, after every
transaction l, the requester niassigns feedback to the selected
provider njto evaluate the service: we can then define the
set of feedback Fij =nf1
ij , ..., f l
ij , ...f Lij
ij o, where lindexes
from the latest transactions (l= 1) to the oldest one (l=Lij ),
so that Lij represents the total number of transactions between
the two nodes. Each feedback can be expressed using values in
the continuous range [0,1], where 1 is used when the requester
is fully satisfied by the service and 0 otherwise.
Figure 1 provides a simple example of a generic graph N=
{n1, ..., n9}, with each node capable of providing one or more
services, as highlighted in the grey clouds; n1is the node
that is requesting the service S7, as highlighted in the white
cloud; Ph={n5, n6}is the set of nodes that can provide
the requested service. In this figure, we also highlight the set
N1={n2, n3, n4}of nodes that are friends of n1(in light blue
color). Within note that the set H15 ={n2, n4}and the set
H16 ={n4}of nodes represent the common friends between
n1and n5and between n1and n6, respectively. For each of the
provider in Ph, the requester n1computes the trustworthiness
levels, T15 and T16, and then chooses the provider with the
highest value, which is n5in our example.
The goal of any trustworthiness management model is to
compute and list the trust level of all the providers. This step is

5
S2, S5
S1
S7?
S5
S3, S4
S1, S2
S1, S7, S8
S2
S6, S7
Trust levels:
1. T15
2. T16
Fig. 1. Trust Management Model.
fundamental to help the requester to identify the most reliable
node to whom require the service and to avoid any malicious
node. In our model, we envision that each node nicomputes
the trustworthiness level Tij of all the possible providers nj
on its own, so that different nodes can make different choices
when selecting a provider based on their past experiences.
B. Trust Management Model
According to the presented scenario, we propose a de-
centralized scenario, where each node calculates and stores
information about the other nodes, so to have its own opinion
about the network: in this way, malicious attacks that change
their behaviour based on the requester, such as DA, are
easily identified. Whenever a node nihas to evaluate the
trustworthiness of another node nj, it computes the trust value
as follows:
Tij =αLij Cj+βLij Rij +γLij Oij +δLij Sij (1)
All these addends are in the range [0,1] and the weights are
selected based on the total number of transactions Lij between
node niand nj. Moreover, the weights’ sum, namely αLij +
βLij +γLij +δLij , is always equal to 1, in order to normalize
the trust value in the interval [0,1], and their relative value
can be changed to give more impact to a particular parameter.
A generic node nievaluates the trustworthiness Tij based
on four parameters: the Computation Capabilities Cjof the
service provider, the Relationship Factor Rij between the two
nodes, the External Opinions Oij provided by ni’s friends and
the Dynamic Knowledge Sij acquired by the requester. The
Dynamic Knowledge represents the core of our system, which
has to learn how to identify malicious nodes. This ability is
tied to its experience, i.e. to the past transactions of the node.
Accordingly, the proposed trustworthiness model is divided
into two phases: a training phase and a steady-state phase. In
the training phase, the contribution of the Dynamic Knowledge
is limited, because the requester is trying to learn the behavior
of the provider: since the requester has to understand the
behavior of each node it interacts with, the four weights are
dependent by both the requester and the provider; we omit
this dependency to avoid too much confusion in the presented
equations. In particular, the value of δLij grows with the total
number of transactions Lij between the requester niand the
provider nj, as follows:
δLij =((Lij −1)/Ltr for Lij ≤Ltr
1for Lij > Ltr (2)
where Ltr represents the number of transactions needed to
train the Dynamic Knowledge. The residual weight, i.e. 1−
δLij , is then shared among the other weights.
1) Training Phase: The goal of this phase is to let the
Dynamic Knowledge factor collects enough experience. Until
this happens, the trust value of the potential providers is
calculated based on the elements described below.
The Computation Capabilities Cjis a static characteristic
of an object which does not vary over time. This factor ac-
counts for the heterogeneity of the IoT where some devices are
more powerful than others so their ability to act maliciously
is higher and they can lead to more uncertain transactions.
To take into account this possibility, the model assigns lower
values to objects with great computational capabilities w.r.t.
devices with only sensing and actuation capabilities.
The Relationship Factor Rij is a unique characteristic
of the SIoT and it is related to the relationships that ties
node niand nj. Using [18] as a starting point, we set the
greatest value for the OOR relationship and decreasing values
for the other relations. If two nodes are tied by two or more
relationships, e.g. they have created both an OOR and a SOR,
we consider the strongest relation which then they have with
the highest value. If two nodes have no direct relation, the
model computes the sequence of social links between them and
consider the weakest link in the path, i.e. the minimum value
of all the relationship factor. To account for the uncertainty of
the intermediates nodes, this value is further divided for the
number of hops that separate node niand node nj.
The External Opinion Oij evaluates the recommendations
provided to niby the friends in common with nj, namely the
nodes in Hij and is expressed as:
Oij =
|Hij |
X
h=1
Tih ·Thj ,|Hij |
X
h=1
Tih (3)
where Thj represents the opinion, i.e. the trust value, that
each of the common friends nhhas for node nj. These
values are weighted with the trust values that node nihas
already computed towards its friends, so that the opinion
of trustworthy nodes is considered more than the one from
low trustworthy nodes. Indeed, recommendations represent an
effective strategy, adopted by many trust algorithms, to easily
obtain information regarding other nodes. This is especially
true when a node’s direct experience is still scarce. However,
they are also exploited by many trustworthiness attacks, such
as BMA and BSA, to confuse the network: using the external
opinion only in the training phase, our model is resilient to all
these types of attacks.
Moreover, at the end of each transaction, niassigns a
feedback not only to the provider but also to the friends in

6
Hij , which have contributed to the computation of the external
opinion. According to Eq. 4, if a node provided a positive
opinion, it receives the same feedback as the provider, i.e. a
positive feedback if the transaction was satisfactory, fl
ij ≥0.5,
and a negative one otherwise, fl
ij <0.5. Instead, if nhgave
a negative opinion, then it receives a negative feedback if the
transaction was satisfactory and a positive one otherwise.
fl
ih =(fl
ij if Thj ≥0.5
1−fl
ij if Thj <0.5(4)
Moreover, to further reduce the possibility of attacks on the
recommendations, in our algorithm, a node uses them only in
the training phase to accumulate experience and then it only
relies on its Dynamic Knowledge.
2) Steady-State phase: After the training phase, only the
Dynamic Knowledge is used to evaluate the possible providers.
According to the presented scenario, certain types of malicious
nodes, e.g. OOA and OSA, continuously change their be-
haviour. In order to address this issue, the Dynamic Knowledge
must be able to continuously learn and adapt to the myriad
of possible malicious behaviours. To compute its value, we
make use of an incremental Support Vector Machine (iSVM),
so that a node can constantly extends its knowledge after a
new transaction: in particular, a SVM is a supervised learning
model that analyzes a set of data, in our case the first Ltr
transactions, to provide some sort of classification. SVM
algorithms have been applied to solve a variety of applications
[33]. With respect to other machine learning algorithms, the
risk of over-fitting is less, it is relatively memory efficient
and is effective when there is a margin of separation between
classes. The accuracy of this classification is tied to the number
of historical data obtained [34]: in our case, the output of
the SVM represents the probability that a service provider is
benevolent or not, i.e. its trust value. More details regarding the
validation process of the iSVM and a comparison with other
incremental machine learning algorithms will be presented in
Section IV-B.
After every transaction, the Dynamic Knowledge is updated,
so that it is able to learn from its past experience and can
provide a more accurate evaluation. Since we make use of
an incremental SVM, with each new transaction the model’s
knowledge is extended and updated, without the need to train
the SVM from scratch. This way, each node can implement
a dynamic Machine Learning algorithm even with limited
resources and active learning is much faster w.r.t. a traditional
approach. In order to train the SVM, past transactions are
expressed in terms of scores, which have the goal to highlight
different aspects of the interaction among nodes. Three scores
are used as inputs for the Dynamic Knowledge, which are able
to evaluate the entire history of the nodes as well as their recent
behavior. In this way, the attacks with a dynamic behaviour
over time, such as OOA and OSA, can be recognized. The
first score is the Goodness Score: this score enables the SVM
to evaluate nodes on a long-term period and measures how
benevolent the node has been during all its transactions. The
score is evaluated as the fraction of all the “good” transactions,
i.e. all the transactions evaluated in a positive way by the
requester:
Gij =|fl
ij ∈ F :fl
ij > T H |
Lij
(5)
where T H is the threshold a requester set to consider ser-
vices as “good”. High values of this score mean that the service
requester is overall satisfied by the services obtained from the
provider. This factor is also useful to identify benevolent nodes
which provide services with low accuracy that a requester
would like to avoid and that are then labeled with a low value
of the Goodness Score.
However, the Goodness Score is not able to react to sudden
changes in the behavior of a node, as it happens for dynamic
attacks such as OOA and OSA. To overcome these attacks,
we make use of two other scores, which evaluate the behavior
of the service provider considering a small temporal window,
which makes use of the last Lstransactions.
The Usefulness Score is used to evaluate only the recent
behavior of a node, as follows:
Uij =
Ls
X
l=1
wl·fl
ij (6)
where, in order to give more relevance to the latest trans-
action w.r.t. the oldest one, the weights wlof each feedback
follows a geometric distribution with parameter ρ
wl=ρ(1 −ρ)l−1+ (ξres/Ls)(7)
to maintain the score in the range [0,1], we introduce the
term ξres which account for all the residual weight of the
distribution due to the transactions older than Ls.ξres is then
computed as:
ξres =
Lij
X
r=Ls+1
ρ(1 −ρ)r−1(8)
The Perseverance Score evaluates the constancy of a node
in providing good services and it is computed as:
(PLij
ij = 0.5if Lij = 1
PLij
ij =PLij −1
ij +Vij if Lij >1(9)
where Vij is a parameter that reward/punish a node based
on its constancy in providing good/bad services, as described
by:
Vij =(vij ufor fLij
ij ≥T H
−vij dfor fLij
ij < T H (10)
uand drepresent the basic increase/decrease of the score;
however, consecutive good or bad transactions can further
reward/penalize a node, which is then encouraged to stay
benevolent, according to the value of vij: this value is cal-
culated as the number of consecutive transactions evaluated
positively/negatively by the requester. As the other scores, also
the Perseverance Score is limited in the interval [0; 1]; in the
event the score obtained from Eq. 9 is out of these bounds,
its value is set to the nearest bound.

7
IV. EXP ER IM EN TAL EVALUATIO N
A. Simulation Setup
In order to test our trustworthiness model, we need a large
dataset of a SIoT scenario. To this, we make use of the
dataset made available by [35]; it consists of a network of
16216 devices owned by 4000 users and by the municipality
of Santander (Spain), which create their own relations over
11 days. Moreover, the authors share a set of real services
and applications offered and requested by the nodes, which
are useful to emulate interactions among nodes. We decide to
consider only a connected sub-network of around 800 nodes
to increase the probability of two nodes interacting with each
other. Furthermore, a model of interaction among the nodes
is also needed to understand which devices are more likely to
interact; trust models are usually tested considering random
interactions among nodes without taking into account the
behavior of objects that generate queries of services when
interacting with the other peers. To this, we have adopted the
query generation model presented in [36], so that at the start
of each transaction, the simulator can choose the requester and
select all the possible providers.
Two main behaviors are implemented in the network: one
is cooperative and benevolent, so that a node always provides
good services and recommendations. The other one is a
malevolent behavior, where a node tries to disrupt the network
by implementing one of the trust attacks presented in Section
II-C. Table III shows the optimal configuration of the simu-
lation parameters for the proposed system, and the different
weights used for the model. For simplicity, we suppose that
the service requester is able to perfectly rate the received
service providing binary feedback: 1 for satisfactory services
and 0 otherwise. Finally, Table IV presents the values for the
relations created by the objects and for their computation ca-
pabilities. Between two objects that belong to the same owner
and then are linked by an OOR, the relationship factor has been
assigned with the highest value. CLORs have been set with
only a slightly lower value since they are established between
domestic objects and objects of the same workplace. SORs are
relationships established between objects that are encountered
occasionally (then owned by acquaintances) and for this reason
a smaller value is given. Finally, the PORs are the riskiest,
since they are created between objects of the same brand but
that never met and depend only on the model object. If two
nodes are tied by two or more relationships, the strongest
relation with the highest factor is considered. Computation
capabilities are divided into two classes: Class1 is assigned
to objects with only sensing capabilities, that is, an object
just capable of providing a measure of the environment status
and to the RFID-tagged objects. Class2 is assigned to objects
with great computational and communication capabilities; to
this class belong objects such as smartphones, tablets, vehicle
control units, displays, set top boxes, smart video cameras.
To find the optimal setting for the residual weight, i.e.
1−δLij , we analyze the model’s response at varying the other
weights, namely α,βand γ. Table V displays the transaction
success rate when the system has reached the steady-state
phase. As expected, the external opinion has more impact than
TABLE III
SIMULATION PARA ME TER S
Parameter Description Value
αResidual weight of the
Computation Capabilities 0.3
βResidual weight of the
Relationship Factor 0.3
γResidual weight of the
External Opinion 0.4
Ltr Number of transactions to
train the Dynamic Knowledge 5
TH Threshold to consider
a service as ”good” 0.5
LsTemporal window to compute
Usefulness and Perseverance Score 10
ρParameter of the
geometric distribution 0.4
uBasic increase of the
Perseverance Score 0.1
dBasic decrease of the
Perseverance Score 0.2
I Number of nodes in the network 791
Percentage of malicious nodes 25%
TABLE IV
PARA MET ER S FOR RELATIONSHIP FACTO R AN D COMP UTATIO N
CAPABILITIES
Relationship Factor
Relationship OOR C-LOR SOR POR
Rij 1 0.9 0.6 0.5
Computation Capabilities
Capabilities Class 1 Class 2
Cj1 0.4
the static characteristics, since it can help to identify malicious
behaviors, however, since we are considering the startup phase,
they are still useful when there is no information available.
Ltr is selected based on the machine learning algorithm
validation. As shown in the next section, the selected value
ensures a sufficient initialisation for the iSVM algorithm and
an efficient prediction in the classifications. The ρparameters
guarantees a compromise in the evaluation of the feedback: a
value close to 1 only considers the newest feedback, while
a value close to 0 considers all the feedback as equally
important. Finally, uand dare picked asymmetric in order to
encourage benevolent behaviours and punish malicious nodes.
TABLE V
PARAMETERS SETTINGS
α= 0.1 β= 0.1 γ= 0.8 SR = 0.83
α= 0.1 β= 0.8 γ= 0.1 SR = 0.82
α= 0.8 β= 0.1 γ= 0.1 SR = 0.81
α= 0.3 β= 0.3 γ= 0.4 SR = 0.85

8
B. Simulation Results for ML algorithms
This Section aims to validate the performance of the incre-
mental SVM (iSVM) algorithm and to compare it with other
incremental machine learning algorithms.
In order to validate the performance of the algorithms
we have used the Receiver Operating Characteristic (ROC)
curve and Area Under the ROC (AUC) curve as performance
metrics. The ROC represents the diagnostic ability of a binary
classifier system, i.e. the true positive rate versus the false
positive rate at different classification thresholds. Lowering
the classification threshold classifies more items as positive,
thus increasing both False Positives and True Positives. The
measure of performance between the algorithms is provided
by the AUC, which indicates how much a model is capable
of distinguishing between classes: a model whose predictions
are 100% wrong has an AUC of 0.0; one whose predictions
are 100% correct has an AUC of 1.0. We compare the
performance of the iSVM with two well-known incremental
algorithms, the incremental Logistic Regression (iLR) [37] and
an incremental artificial neural network, the incremental Radial
Basis Function network (iRBF) [38]. The testing network used
for the validation is composed by a requester interacting with
nodes, as providers, that implement each a different behaviour,
from benevolent to all of the seven possible attacks. We vary
the number of total transactions among the requester and all
the providers to study the ability to learn of the algorithms;
we consider that out of all the transactions, 70% of them are
used to train the incremental models while the remaining 30%
are used for the validation. Figure 2 shows the trend of the
ROC curve for 4 experiments based on 160, 650, 1600 and
3200 transactions of the requester. Considering all the possible
providers, this means that the number of transactions used for
validation with each node is 6, 25, 60 and 120 transactions.
The Figure shows how the iSVM is able to outperform the
other two algorithms: except for the first set of simulations,
with only 6 transactions per node used for validation, the
iSVM has the best values of AUC: the system continuously
learns from the processed data so that the iSVM increases
its percentage of correct predictions with the growth of the
dataset of transactions. Moreover, even if the accuracy of the
iSVM is low when considering few transactions per node, the
proposed model is able to mediate it thanks to the training
phase, which makes use of other parameters to obtain higher
accuracy in selecting trustworthy nodes.
C. Simulation Results for Trust Management Model
We evaluate the performance of the proposed system by
analyzing the success rate, i.e. the ratio between the number
of successful transactions and the total number of transactions,
or by directly calculating the level of trust computed by a node.
We compare the performance of the proposed model with
two well known models by the research community that,
similar to our model, are designed for the same scenario,
i.e. Social IoT scenario, namely the model proposed by Nitti
et al. [18] and the one presented in [19] by Chen et al.
Both these models make use of a subjective approach where
every node has its own vision of the network and relies
0 0.2 0.4 0.6 0.8 1
False Positive Rate
0
0.5
1
True Positive Rate
ROC curve - 6 transactions/node
iSVM - AUC: 0.94162
iLR - AUC:0.96667
iRBF - AUC:0.96441
0 0.2 0.4 0.6 0.8 1
False Positive Rate
0
0.5
1
True Positive Rate
ROC curve - 25 transactions/node
iSVM - AUC: 0.94162
iLR - AUC:0.94707
iRBF - AUC:0.88862
0 0.2 0.4 0.6 0.8 1
False Positive Rate
0
0.5
1
True Positive Rate
ROC curve - 60 transactions/node
iSVM - AUC: 0.95964
iLR - AUC:0.94443
iRBF - AUC:0.88926
0 0.2 0.4 0.6 0.8 1
False Positive Rate
0
0.5
1
True Positive Rate
ROC curve - 120 transactions/node
iSVM - AUC: 0.96817
iLR - AUC:0.95635
iRBF - AUC:0.91658
Fig. 2. ROC curves for the machine learning algorithms for 4 experiments
based on 6, 25, 60 and 120 transactions per node.
2500 3000 3500 4000 4500 5000 5500 6000 6500 7000 7500 8000
Transaction Number
0.75
0.8
0.85
0.9
0.95
1
Success Rate
ME - Proposed Approach
DA - Proposed Approach
ME - Nitti et al.
DA - Nitti et al.
ME - Chen et al.
DA - Chen et al.
Fig. 3. Transaction success rate for two classes of trust attacks.
on the recommendations from its friends to speed up the
evaluation of trust. Differences in the performance of the
models can depend on the structure of the social network
considered and on the types of service/information requested.
To this, we did not consider our ad-hoc social network but we
have adopted the Social IoT dataset described in the previous
subsection, opportunistically re-scaled to a size comparable
to their experiments. Moreover, we have considered the same
requests for all the three models, so we are confident that
the obtained results are consistent with those obtained by the
authors.
These comparisons are aimed at analyzing the improve-
ments we obtain with respect to the state of the art in the
specific reference SIoT scenario. We tested all the different
types of attacks, except for the SA and the SPA, which are
avoided by default in our system: even if a node creates
multiple identities or provides good recommendations for
itself, the computed trust can not be influenced.
Figure 3 shows the transaction success rate when malicious
nodes implement two trust-related attacks, ME and DA. We
consider that 25% of the nodes are malicious and in the case
of the DA, they only act maliciously with nodes that they meet
occasionally or they have never met, i.e. with nodes they have
a weak relation with, such as POR and SOR. All the models

9
TABLE VI
PERFORMANCE COMPARISON OF THE THREE MODELS AGAINST THE
DIS CRI MI NATORY ATTACK .
Proposed Chen et al. Nitti et al.
n1T14 = 0.97 T14 = 0.81 T14 = 0.83
Requester n2T24 = 0.97 T24 = 0.8T24 = 0.8
n3T34 = 0.02 T34 = 0.31 T34 = 0.42
1000 2000 3000 4000 5000 6000 7000 8000
Transaction Number
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Success Rate
10 % 20 % 30 % 40 % 50 % 60 % 70 %
Fig. 4. Transaction success rate at increasing values of %of malicious nodes.
have a good reaction to these two attacks and are able to
achieve a high success rate, ranging from 88% to 94%. This is
not a surprising result, since both these attacks are usually the
ones used to test trustworthiness models. All the implemented
algorithms have a better performance to the Discriminatory
Attack w.r.t. to the ME, even if devices implementing ME
are easier to be identified since they do not behave differently
according to the requester: this can be explained considering
that due to the changing behavior of the DA, the total number
of transactions in which a node acts as malicious are only a
subset of all its transaction.
To better understand how the three models react to the
DA, we set up a small network of 4 nodes fully connected,
where 3 benevolent nodes, n1,n2and n3, have 15 interactions
each with one malevolent DA node n4. Only the relation
{n3, n4}is weak, so n4only behaves maliciously with n3
and benevolent with n1and n2. The results are shown in
Table VI: as expected, in all the models, both n1and n2
have a high trust value for n4while, despite n3is able to
identify n4as a malevolent node in all the models, the trust
value obtained is highly variable. Only our proposed model
assigns a really low trust value to n4, while the other two
models compute higher values due to the strong influence of
the common friends within their algorithms.
We now want to analyze the results at varying percentage
of the malicious nodes. Figure 4 refers to a scenario where all
the malicious nodes implement ME: it shows that even with
70% of malicious nodes the success rate is over 50% and the
algorithm is still able to converge. This happens since every
node has its own vision of the network based on the acquired
Dynamic Knowledge, however, the accuracy decreases, since it
increases the possibility that all the available service providers
are malicious. We need more than 75% of malicious nodes for
20 40 60 80 100 120 140 160 180 200
Transaction Number
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Trust
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Service Feedback
Trust
Service Feedback
Fig. 5. Trust value of a malicious node that performs an On-Off Attack.
the success rate to drop below 0.5: we have run a similar test
also for the other two algorithms: Chen’s algorithm is able to
resist over 80% of malicious nodes while Nitti’s performance
is similar to our with 75% of malicious nodes. This result is
related to the subjective approach of these two models, where
each node takes its own decisions.
The focus of the next set of simulations is to test how
the proposed model works with the dynamic behavior of
the nodes, i.e. against the OOA. We suppose that after 40
transactions, a malicious node starts to change its behavior
from benevolent to malicious and vice versa every 20 transac-
tions. Figure 5 illustrates the trust value of a node performing
such attack and shows how the algorithm is able to quickly
adapt to the changes in the node behavior: only 3 transactions
are needed to modify the trust value of the malicious node,
both when the node is exploiting its good reputation and
when it is trying to build up its trust. Table VII presents a
comparison with the other two models in terms of the number
of transactions needed to change the trust value past 0.5 and
highlighting the initial and final trust, Tiand Tfrespectively,
computed before and after the changing behavior. We note how
our model is the fastest one to recognize the dynamic behavior
so that only a node changing its behavior every 2 transactions
is able to successfully being undetected. Moreover, we also
observe that the final trust values Tfassigned by our model
are rather confident, since they are closer to the trust limits,
i.e. 0 for malicious nodes and 1 for benevolent nodes, while
the other two models compute a trust value of around 0.5, thus
indicating uncertainty in the evaluation of the node.
The next set of simulations focus on the reaction of the
models against BSA (solid lines) and WA (dotted lines), as
shown in Figure 6. In the BSA case, the requester node
receives high recommendation values concerning a malicious
provider from two common malicious friends. To tackle this
attack is important to understand how each model manages
the recommendations received by the common friends: in our
model, such recommendations are used only in the startup
phase and their weight decreases with the number of transac-
tions as the Dynamic Knowledge acquires more experience
(see Eq. 2). Chen’s and Nitti’s algorithms share a similar
approach: the indirect opinion has always a certain relevance

10
TABLE VII
PERFORMANCE COMPARISON OF THE THREE MO DE LS AGA IN ST TH E ON-OFF ATTACK.
Proposed Chen et al. Nitti et al.
# trans. TiTf# trans. TiTf# trans. TiTf
ON →OFF 3 0.99 0.09 4 0.81 0.49 5 0.82 0.45
OFF →ON 3 0.05 0.77 5 0.41 0.5 5 0.17 0.5
0 2 4 6 8 10 12 14 16 18 20
Transaction Number
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Success Rate
BSA - Proposed Approach
WA - Proposed Approach
BSA - Nitti et al.
WA - Nitti et al.
BSA - Chen et al.
WA - Chen et al.
Fig. 6. Trust value of a malicious node that performs a Ballot Stuffing Attack
and a Whitewashing Attack.
0 2 4 6 8 10 12 14 16 18 20
Transaction Number
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Trustworthiness
BMA - Proposed Approach
Benevolent node - Proposed Approach
BMA - Nitti et al.
Benevolent node - Nitti et al.
BMA - Chen et al.
Benevolent node - Chen et al.
Fig. 7. Trust value of benevolent node with or without BMA.
in the trust score computation, however, its weight is different
in the two models (0.15 in Chen’s algorithm and 0.3 in Nitti’s).
From the Figure, we can see the trust value of the nodes
implementing BSA and we can notice how our proposed
model is almost non affected by the BSA (low trust values
after only a few transactions), while the trust value computed
by the other two models is definitely higher but still under the
0.5 threshold, thus marking the BSA nodes as malicious. In
the case of WA, a malicious node with a bad reputation after
10 transaction leaves and re-joins the network to reset its trust
to the default value. All the models are able to identify the
node with few transactions and to label it again as malicious.
However, Nitti’s and Chen’s algorithms are more robust to
this attack, since the gain in the trust value of the WA node
is lower w.r.t. our model.
The next set of experiments tests the BMA, where a
malicious node provides false recommendations to decrease
TABLE VIII
PER CEN TAGE OF P OS ITI VE T RAN SAC TI ONS F OR AN OP PO RTUN IS TIC
SERVICE ATTACK OV ER 100 TRANSACTIONS
% of Positive Transactions
Trust Percentile Proposed Nitti Chen
10% 100 86 82
20% 93 68 67
30% 86 61 57
the trust of benevolent nodes. We first test if this attack
could lead a requester to choose a malevolent node over a
benevolent one: all the models select the malevolent node only
once and are then able to select the benevolent node. This
is due to the higher importance given by the models to the
direct experiences w.r.t. indirect recommendations. Moreover,
the number of nodes implementing BMA does not affect this
result. Then we investigate how the trust value changes in a
scenario where a benevolent node is attacked by bad-mouthing
nodes w.r.t. a benevolent node with no attackers. Figure 7
shows how our model is only affected by the BMA in the
startup phase and it is then able to achieve the same trust
values for the two benevolent nodes; the other two models
present a lower trust value, which does not increase with the
number of transactions, due to fixed parameters external to the
requester experience, such as the centrality or the computation
capabilities. Moreover, it clearly appears how BMA nodes
can confuse the network, especially in Nitti’s algorithm which
gives a higher weight to the indirect opinion than Chen’s.
The next set of simulations examines the OSA, where a
node changes its behavior so that its trust value computed by
the requester maintains an acceptable level. However, a node’s
goal is not to have a high trust value but rather to have a value
higher than the other providers in order to be chosen (and
then have a chance to behave maliciously). To test this attack,
we consider only a service requester and a malicious service
provider performing the attack. We suppose that the provider
is perfectly aware of its trust reputation and act maliciously
only when its trust value is among the 10%, 20% and 30%
percentile of the most trustworthy nodes. Considering 100
transactions between the two nodes, Table VIII shows the
percentage of positive transactions for the three models. As
expected, a larger percentile enables the malicious node to
perform more attacks, however the node could not be selected
as a provider if there are other possible providers for the same
service. If the malicious node wants to be sure to be selected
and set a stringent percentile, the number of opportunities to
behave maliciously reduces. However, our approach is able
to compel the malicious node to perform the highest number

11
TABLE IX
AVERA GE OF TRU ST F OR TH E BE NEV OLE NT NO DE S WIT H ER ROR P ERC EN TAGE AN D THE M AL ICI OU S NOD ES .
Benevolent Node - Error percentage Malicious Node
0% 10% 20% 30% 40% 50% ME/DA WA OOA OSA
Trust 0.99 0.86 0.79 0.72 0.57 0.28 0.02 0.05 0.12 0.88
2000 3000 4000 5000 6000 7000 8000
Transaction Number
0.75
0.8
0.85
0.9
0.95
1
Success Rate
Proposed Approach
Nitti et al.
Chen et al.
Fig. 8. Transaction success rate with all types of malicious attacks.
of positive transactions w.r.t. to the other two models, thus
indicating the ability of the model to cope with this attack.
In particular, if a malicious node wants to stay in the 10%
percentile, it has to always perform benevolent.
We want now to show how the three models respond to a
network with a mix of all the attacks analyzed. The result,
in terms of transaction success rate, is shown in Figure 8,
considering 5% of malicious nodes for each type of attack, for
a total of 30% malicious nodes. Our model is able to converge
faster and to outperform the other two with a success rate of
over 95%. By analyzing which attacks had a higher impact, we
see how simple attacks are better managed by Chen and Nitti’s
algorithms, however, as expected, they highly suffer smart
attacks, such as OSA and OOA, which are not sufficiently
tackled by them.
Finally, the last set of simulations is aimed to understand
how our system reacts when benevolent nodes offer poor
services due to errors related to several reasons. We then
consider a requester which interacts with benevolent nodes
which have a different probability to respond with an incorrect
service due to some kind of error. For each value of the error
percentage, we simulate 100 transactions between the nodes
and mediate the results over 100 runs. Table IX shows the
resulting trust values for different error rates of the benevolent
nodes and compare them with the trust values of malicious
nodes, without considering the attacks on the recommendation,
i.e. BSA and BMA. Due to the subjective approach of our
model, DA performs similar to ME, since, if it is connected
to the requester by a weak link, it will always provide false
services; nodes implementing WA have a slightly higher trust
value, since they can reset their trust to the default value.
As expected, the results show how increasing the error rate,
the average trust of benevolent nodes decreases. However,
even for nodes with a 50% error rate, their trust is still
higher than nodes implementing OOA, which has a similar
behaviour, i.e. 50% benevolent transactions and 50% malicious
transactions: this is due to the Perseverance Score, which
evaluates negatively the consecutive bad services of the OOA.
Only a node implementing OSA is able to maintain a high
level of trust. In this set of simulations, we consider that an
OSA node acts maliciously only when its trust value is among
the 20% percentile of the most trustworthy nodes. As seen in
Table VIII, this means that the node will have more than 90%
of trustworthy transactions, and thus can be considered as a
node that offers bad services 10% of the time.
V. CONCLUSIONS
In this paper, we have analyzed the possible types of attacks
that can be implemented by nodes to disrupt an IoT system.
We then have proposed a trust management model based on
a Machine Learning algorithm for a Social IoT scenario. The
proposed solution is also applicable to general IoT scenarios,
however, information regarding the type of friendship between
two nodes is able to reduce the uncertainty in the selection
of a trustworthy provider and provide better performance.
The proposed model has been tested against all the different
types of attacks, except for the SA and the SPA, which are
avoided by default. Experiments have shown that our model
is able to overcome all the possible attacks. Furthermore, we
compare our algorithm with two well-acknowledged state-of-
the-art models: simulations show that even if our algorithm
show slightly worse performance when under attack by simple
mechanisms, such as ME, it is able to outperform the other two
models when considering a network with a mix of different
types of attack.
ACK NOW LE DG EM EN TS
This work was supported by Italian Ministry of University
and Research (MIUR), within the PON R&I 2014-2020 frame-
work (Project AIM (Attrazione e Mobilit`
a Internazionale).
REFERENCES
[1] J. Lin, W. Yu, N. Zhang, X. Yang, H. Zhang, and W. Zhao, “A survey
on internet of things: Architecture, enabling technologies, security and
privacy, and applications,” IEEE Internet of Things Journal, vol. 4, no. 5,
pp. 1125–1142, 2017.
[2] D. Gambetta et al., “Can we trust trust,” Trust: Making and breaking
cooperative relations, vol. 13, pp. 213–237, 2000.
[3] K. Li, L. Tian, W. Li, G. Luo, and Z. Cai, “Incorporating social
interaction into three-party game towards privacy protection in iot,”
Computer Networks, vol. 150, pp. 90–101, 2019.
[4] W. Meng, K.-K. R. Choo, S. Furnell, A. V. Vasilakos, and C. W.
Probst, “Towards bayesian-based trust management for insider attacks in
healthcare software-defined networks,” IEEE Transactions on Network
and Service Management, vol. 15, no. 2, pp. 761–773, 2018.

12
[5] L. Atzori, A. Iera, G. Morabito, and M. Nitti, “The social internet of
things (siot)–when social networks meet the internet of things: Concept,
architecture and network characterization,” Computer Networks, vol. 56,
no. 16, pp. 3594–3608, 2012.
[6] H. Xia, C.-q. Hu, F. Xiao, X.-g. Cheng, and Z.-k. Pan, “An efficient
social-like semantic-aware service discovery mechanism for large-scale
internet of things,” Computer Networks, vol. 152, pp. 210–220, 2019.
[7] M. A. Azad, C. Perera, and M. Barhamgi, “Privacy-preserving crowd-
sensed trust aggregation in the user-centeric internet of people networks,”
ACM Transactions on Cyber-Physical Systems, 2020.
[8] R. Chen, F. Bao, and J. Guo, “Trust-based service management for social
internet of things systems,” IEEE transactions on dependable and secure
computing, vol. 13, no. 6, pp. 684–696, 2015.
[9] K. Zhao and L. Pan, “A machine learning based trust evaluation
framework for online social networks,” in 2014 IEEE 13th International
Conference on Trust, Security and Privacy in Computing and Commu-
nications. IEEE, 2014, pp. 69–74.
[10] A. M. Kowshalya and M. Valarmathi, “Trust management for reliable
decision making among social objects in the social internet of things,”
IET Networks, vol. 6, no. 4, pp. 75–80, 2017.
[11] B. Jafarian, N. Yazdani, and M. S. Haghighi, “Discriminative-aware
trust management for social internet of things,” Computer Networks,
p. 107254, 2020.
[12] R. Chen, J. Guo, D.-C. Wang, J. J. Tsai, H. Al-Hamadi, and I. You,
“Trust-based service management for mobile cloud iot systems,” IEEE
transactions on network and service management, vol. 16, no. 1, pp.
246–263, 2018.
[13] B. Gong, Y. Zhang, and Y. Wang, “A remote attestation mechanism for
the sensing layer nodes of the internet of things,” Future Generation
Computer Systems, vol. 78, pp. 867–886, 2018.
[14] P. De Meo, F. Messina, M. N. Postorino, D. Rosaci, and G. M. Sarn´
e, “A
reputation framework to share resources into iot-based environments,”
in 2017 IEEE 14th International Conference on Networking, Sensing
and Control (ICNSC). IEEE, 2017, pp. 513–518.
[15] W. Li, H. Song, and F. Zeng, “Policy-based secure and trustworthy
sensing for internet of things in smart cities,” IEEE Internet of Things
Journal, vol. 5, no. 2, pp. 716–723, 2017.
[16] H. Al-Hamadi and R. Chen, “Trust-based decision making for health iot
systems,” IEEE Internet of Things Journal, vol. 4, no. 5, pp. 1408–1419,
2017.
[17] B. Li, L. Liao, H. Leung, and R. Song, “Phat: A preference and honesty
aware trust model for web services,” IEEE Transactions on network and
service management, vol. 11, no. 3, pp. 363–375, 2014.
[18] M. Nitti, R. Girau, and L. Atzori, “Trustworthiness management in the
social internet of things,” IEEE Transactions on knowledge and data
engineering, vol. 26, no. 5, pp. 1253–1266, 2014.
[19] Z. Chen, R. Ling, C.-M. Huang, and X. Zhu, “A scheme of access
service recommendation for the social internet of things,” International
Journal of Communication Systems, vol. 29, no. 4, pp. 694–706, 2016.
[20] N. B. Truong, H. Lee, B. Askwith, and G. M. Lee, “Toward a trust
evaluation mechanism in the social internet of things,” Sensors, vol. 17,
no. 6, p. 1346, 2017.
[21] D. Airehrour, J. A. Gutierrez, and S. K. Ray, “Sectrust-rpl: A secure
trust-aware rpl routing protocol for internet of things,” Future Genera-
tion Computer Systems, vol. 93, pp. 860–876, 2019.
[22] X. Fan, L. Liu, R. Zhang, Q. Jing, and J. Bi, “Decentralized trust
management: Risk analysis and trust aggregation,” ACM Computing
Surveys (CSUR), vol. 53, no. 1, pp. 1–33, 2020.
[23] A. Altaf, H. Abbas, F. Iqbal, and A. Derhab, “Trust models of internet
of smart things: A survey, open issues, and future directions,” Journal
of Network and Computer Applications, vol. 137, pp. 93–111, 2019.
[24] M. A. Azad, S. Bag, F. Hao, and A. Shalaginov, “Decentralized self-
enforcing trust management system for social internet of things,” IEEE
Internet of Things Journal, vol. 7, no. 4, pp. 2690–2703, 2020.
[25] D. Wang, T. Muller, Y. Liu, and J. Zhang, “Towards robust and
effective trust management for security: A survey,” in 2014 IEEE 13th
International Conference on Trust, Security and Privacy in Computing
and Communications. IEEE, 2014, pp. 511–518.
[26] J. Caminha, A. Perkusich, and M. Perkusich, “A smart trust management
method to detect on-off attacks in the internet of things,” Security and
Communication Networks, vol. 2018, 2018.
[27] A. Jøsang and J. Golbeck, “Challenges for robust trust and reputation
systems,” in Proceedings of the 5th International Workshop on Security
and Trust Management (SMT 2009), Saint Malo, France, vol. 5, no. 9.
Citeseer, 2009.
[28] M. Rashmi and C. V. Raj, “A review on trust models of social internet
of things,” in Emerging Research in Electronics, Computer Science and
Technology. Springer, 2019, pp. 203–209.
[29] J. Guo, R. Chen, and J. J. Tsai, “A survey of trust computation
models for service management in internet of things systems,” Computer
Communications, vol. 97, pp. 1–14, 2017.
[30] R. Chen and J. Guo, “Dynamic hierarchical trust management of mobile
groups and its application to misbehaving node detection,” in 2014 IEEE
28th International Conference on Advanced Information Networking and
Applications. IEEE, 2014, pp. 49–56.
[31] K. Zaidi, M. B. Milojevic, V. Rakocevic, A. Nallanathan, and M. Ra-
jarajan, “Host-based intrusion detection for vanets: a statistical approach
to rogue node detection,” IEEE transactions on vehicular technology,
vol. 65, no. 8, pp. 6703–6714, 2015.
[32] R. Chen, J. Guo, and F. Bao, “Trust management for soa-based iot and
its application to service composition,” IEEE Transactions on Services
Computing, vol. 9, no. 3, pp. 482–495, 2014.
[33] Y. Ma and G. Guo, Support vector machines applications. Springer,
2014, vol. 649.
[34] J. Xu, C. Xu, B. Zou, Y. Y. Tang, J. Peng, and X. You, “New incremental
learning algorithm with support vector machines,” IEEE Transactions on
Systems, Man, and Cybernetics: Systems, vol. 49, no. 11, pp. 2230–2241,
2018.
[35] C. Marche, L. Atzori, and M. Nitti, “A dataset for performance analysis
of the social internet of things,” in 2018 IEEE 29th Annual International
Symposium on Personal, Indoor and Mobile Radio Communications
(PIMRC). IEEE, 2018, pp. 1–5.
[36] C. Marche, L. Atzori, V. Pilloni, and M. Nitti, “How to exploit the social
internet of things: Query generation model and device profiles’ dataset,”
Computer Networks, p. 107248, 2020.
[37] S. Lee and C.-H. Jun, “Fast incremental learning of logistic model tree
using least angle regression,” Expert Systems with Applications, vol. 97,
pp. 137–145, 2018.
[38] P. Reiner and B. M. Wilamowski, “Efficient incremental construction of
rbf networks using quasi-gradient method,” Neurocomputing, vol. 150,
pp. 349–356, 2015.
Claudio Marche received the M.Sc. degree in
telecommunication engineering with full marks in
2018 from the University of Cagliari. Since grad-
uation, he has been working as Researcher in the
Department of Electrical and Electronic Engineering
at the University of Cagliari, in the MCLab research
group. He is currently a Ph.D. student in Elec-
tronic and Computer Engineering at the University
of Cagliari. His current research interests include
Internet of Things (IoT), Social Internet of Things
(SIoT) and Trustworthiness for IoT.
Michele Nitti is an Assistant Professor at the Uni-
versity of Cagliari, Italy since 2015. In 2013, he has
been a visited student at the Department of Manage-
ment, Technology and Economics at ETH Zurich,
Switzerland. He served as a technical program chair
for various international conferences (IEEE BMSB
2017, IEEE IoT V&T Summit 2020) and workshops
(IEEE ICCCS 2019, IEEE GIoTS 2020). Currently,
he is a member of the editorial board for the IEEE
IoT Journal, Elsevier Computer Networks and MDPI
IoT and co-founder of an academic spin-off (Green-
Share s.r.l.) which works in the mobility sector. His main research interests
are in architecture and services for the Internet of Things (IoT), particularly
in the creation of a network infrastructure to allow the objects to organize
themselves according to a social structure (Social Internet of Things - SIoT).