TRUST MODEL WITH DEFENSE SCHEME IN MANETS

http://www.iaeme.com/IJARET/index.asp 300 editor@iaeme.com
International Journal of Advanced Research in Engineering and Technology (IJARET)
Volume 10, Issue 2, March-April 2019, pp. 300-321. Article ID: IJARET_10_02_030
Available online at http://www.iaeme.com/IJARET/issues.asp?JType=IJARET&VType=10&IType=02
ISSN Print: 0976-6480 and ISSN Online: 0976-6499
© IAEME Publication
TRUST MODEL WITH DEFENSE SCHEME IN
MANETS
Syeda Kausar Fatima
Research Scholar JNTUH, Prof., DCET, Hyderabad, Principal NSAKCET,
Prof., Assoc. Prof. JBIET,India.
Dr. Syeda Gauhar Fatima
Dr. Syed Abdul Sattar
Syed Mohd Ali
ABSTRACT
The reliability of delivering packets through multi-hop intermediate nodes is a
significant issue in the mobile ad hoc networks (MANETs). The distributed mobile nodes
establish connections to form the MANET, which may include selfish and misbehaving
nodes. Recommendation based trust management has been proposed in the literature
as a mechanism to filter out the misbehaving nodes while searching for a packet delivery
route. However, building a trust model that relies on the recommendations from other
nodes in the network is vulnerable to the possible dishonest behaviour, such as bad-
mouthing, ballot-stuffing, and collusion, of the recommending nodes. This paper
investigates the problems of attacks posed by misbehaving nodes while propagating
recommendations in the existing trust models. We propose a recommendation-based
trust model with a defence scheme that utilises clustering technique to dynamically filter
attacks related to dishonest recommendations within certain time based on number of
interactions, compatibility of information and node closeness. The model is empirically
tested in several mobile and disconnected topologies in which nodes experience changes
in their neighbourhoods and consequently face frequent route changes. The empirical
analysis demonstrates robustness and accuracy of the trust model in a dynamic MANET
environment.
Keywords: Ishonest recommendation, filtering algorithm, mobile ad hoc networks,
recommendation attacks, recommendation management, Trust management models

Trust Model with Defense Scheme in Manets
Cite this Article: Syeda Kausar Fatima, Dr. Syeda Gauhar Fatima Dr. Syed Abdul
Sattar and Syed Mohd Ali, Trust Model with Defense Scheme in Manets, International
Journal of Advanced Research in Engineering and Technology, 10(2), 2019, pp. 300-
321.
http://www.iaeme.com/IJARET/issues.asp?JType=IJARET&VType=10&IType=2
1. INTRODUCTION
The lack of infrastructure and central authority such as base stationsto establish and facilitate
communication in the network [1]. It is composed of set of autonomous nodes which work to
agree relay packets for each other and have dynamic topologies, with resource constraints, and
limited physical protection [2]. MANETs’ applications are increasing in future network
paradigms including vehicular and mesh networks. Many civilian and military services are
demanding MANET applications, ranging from emergency rescue services such as hurricane
and earthquake disasters to exchanging critical information on the battlefield or even home and
personal area networking [3]. The formation and sustained existence of MANET services are
mainly based on an individual node’s cooperation in packet forwarding. Due to the unique
characteristics and demanding use, MANETs are vulnerable to attacks launched by
misbehaving nodes [2]. One of the approved mechanisms to improve security in MANETs is
to use trust management techniques to deal with the misbehaving nodes and stimulate them to
cooperate [4].
Trust as a social concept can be defined as the degree of subjective belief about the
behaviour of a particular entity [5]. Existing trust management frameworks for MANETs can
be categorised into two types. The first establishes trust relationships between nodes based on
direct interactions only [7, 8]. The second type is based on direct observations of the node itself
and recommendations provided by other nodes in the network [9, 10]. The use of
recommendation-based trust technique can be advantageous to nodes in discovering
misbehaving nodes prior to interaction, thus avoiding a potential bad experience. Using
recommendations, nodes in MANETs can make more informed decision on the selection of
routing path even if they did not have any direct interactions in the past [9]. Being acquainted
with several distant nodes (not neighbours) can be done sending a single packet to them, and it
could help in saving energy [11]. Together with the advantages comes the challenge of handling
dishonest recommendations in MANETs. In absence of past interactions, a particular node
might not be well informed to make an assessment of trustworthiness of another node. In such
cases, the evaluating node solicits recommendations from the evaluated node's neighbours
(acquaintances) with whom it has a history of interaction. However, to maximise the gain of
individual and their acquaintances, nodes could resort to dishonest behaviours through attacks
such as ballot stuffing, bad- mouthing or colluding. Such attacks could eventually lead to trust
framework malfunction [12]. Solutions proposed to tackle these problems are limited and not
adequately effective [11-16]. For instance, one of the approaches [11] judges the honesty of the
recommending node by referring to their trust values. A recommending node with a high trust
value is preferred and seen as a trustworthy one. However, a node can be trustworthy in terms
of packet forwarding but could be a bad node as a recommending node. Filtering out dishonest
recommending nodes becomes a serious problem when recommending nodes collude with each
other to accomplish a malicious goal [17]. This may result in confusing and misleading the trust
model in judging a nodes’ trustworthiness.
To overcome some of these limitations, this paper proposes a recommendation-based trust
model with a defence scheme to filter out attacks related to dishonest recommendations like
bad-mouthing, ballot-stuffing, and collusion for mobile ad hoc networks. The recommending
node is chosen based on three factors to check its honesty: number of interactions with the

Syeda Kausar Fatima, Dr. Syeda Gauhar Fatima Dr. Syed Abdul Sattar and Syed Mohd Ali
evaluated node, unity of view with the evaluating node for solving the problem of the scarcity
of knowledge, and closeness to the evaluating node. Recommendations are accumulated over a
period of time to ensure the consistency of recommendations provided by a recommending
node regarding the evaluated node. Clustering technique is adopted to dynamically filter out
recommendations between certain timeframe based on a). number of interactions (using
confidence value), b). compatibility of information with the evaluated node (through deviation
test) and c). closeness between the nodes. Different nodes are chosen in the evaluation
procedure to test the performance of the filtering algorithm against various mobile topologies
and neighbourhoods.
2. RELATED WORK
In recent years, different trust and reputation models have been proposed to enhance security
in MANETs to enable nodes to evaluate their neighbours directly or through recommendations
from other nodes in the network. Though the proposed models have paid some attention to the
problem of dishonest recommendations, finding out effective mechanisms to eliminate or
mitigate the influence is still a challenging problem for MANETs.
CONFIDANT [18] uses the personal experience mechanism to deal with the problem of
dishonest recommendation. It applies the deviation test on the received recommendations and
excludes the ones deviating above the threshold value. The reputation value of a recommending
node is updated based on the results from the deviation test. The model cannot prevent the
dissemination of false recommendation and negative recommendation is the only information
exchanged between nodes [19]. Michiardi and Molva [20] propose CORE model, which only
accepts positive recommendation by others. Consequently, this can lead to decreased efficiency
of the system because nodes cannot exchange bad experiences from the misbehaving ones in
the network. Also, CORE cannot be resilient against ballot-stuffing attack as it leaves ways for
misbehaving nodes to collude and gain unfair high ratings. Wang et al. [21] propose a trust-
based incentive model for self-policing mobile ad hoc networks to reduce the impact of false
recommendation on the accuracy of trust value. However, the performance of the model is not
tested against specific attacks such as bad-mouthing. Authors in [22] propose RFSTrust, a trust
model based on fuzzy recommendation similarity, which is presented to quantify and evaluate
the trustworthiness of nodes. They use similarity theory to evaluate the recommendation
relationships between nodes. That is, the higher the degree of similarity between the evaluating
node and the recommending node, the more consistent is the evaluation between the two nodes.
In this model, only one type of situation is considered when selfish nodes attack is present, and
the performance of the model is not tested against other attacks related to recommendation.
Soltanali et al in [23], propose a model of trust to encourage the cooperation between nodes by
using direct observation and recommendation. This model only accepts the last opinion of a
node, which is passed to a reputation manager system at the end of each interval. Considering
only the last opinion is not insightful enough to recognise the fluctuation in node’s behaviour,
like in on-off attack [12]. In an attempt to increase the honesty of utilising recommendations,
Li et al in [10] include a confidence value in their evaluation by combining two values: trust
and confidence into a single value called trustworthiness. They utilise the trustworthiness value
to put weight on recommendations in which a recommending node with higher trustworthiness
value is given more weight. Collusion attack in providing false recommendation is not
considered by this work, and this may cause incorrect evaluation of the received
recommendations [5]. Hermes [13] is a recommendation-based trust model that uses an
additional parameter known as an acceptability threshold (in relation to the confidence level).
The notion of acceptability is used in the computation of recommendation to ensure that
adequate observations of the behaviour of participating node has been obtained. However, the
selection of acceptability is a trade-off between obtaining more accurate trustworthiness value

and the convergence time required to obtain it. A recommendation exchange protocol (REP) is
proposed by Pedro B. et al. [24] to allow nodes to send and receive recommendations from
neighbouring nodes. It introduces the concept of relationship maturity based on how long nodes
have known each other. Recommendations forwarded by long term associates are weighed
higher than that from short term associates. The maturity of relationship is evaluated on the
basis of a single factor by considering only the duration of relationship. Yu et al. in [25] propose
a clustering technique to filter out trustworthy recommendations from untrustworthy ones. They
follow the majority rule by selecting the cluster with the largest number of recommendations
as trustworthy one. They tested their model against some attacks like bad mouthing and ballot
stuffing. However, majority rule could be ineffective as some nodes can collude to perform an
attack, and not provide an honest judgment about other nodes.
The aforementioned discussion highlights limitations of the trust models in their abilities to
shield nodes from malicious behaviour in the network. It can be seen from the literature review
that most of the models relied on single parameter to compute trustworthiness. To address these
limitations, a defence scheme is proposed in this paper using multiple parameters (as specified
in Section 1) to compute the trustworthiness of recommenders. The model underlines the
importance of social properties in evaluating trustworthiness and uses it in investigating the
relation between closeness of nodes and similarity in behaviour. The use of proof of time and
location, missing in the current literature, is considered by the proposed model. False negative
and false positive problems in evaluating the recommendation’s trustworthiness and their
impact on the network performance are thoroughly investigated.
3. ATTACKS RELATED TO RECOMMENDATION MANAGEMENT IN
TRUST AND REPUTATION
3.1. FRAMEWORKS
It is indeed a challenge to safe guard a network against wide range of attacks. Recent focus of
research in this area has been on the problems associated with misbehaving nodes in the context
of packet forwarding, like blackhole or wormhole attack [26]. For quality assurance, it is
important that trust management frameworks be resilient to attacks [10]. Although several
researches have put considerable effort to protect the propagation and aggregation of
recommendations in a trust model, research is still in its early stages [12]. The following attacks,
namely, bad mouthing attack, ballot stuffing attack, selective misbehaviour attack, intelligent
behaviour attack, timedependent attack and location-dependent attack (see Fig. 1 for the
classification of attacks), are targeted at the propagation and aggregation of recommendation
[10, 12, 27]. Location-dependent attack is used for the first time in this paper. The attack
behaviours are summarised below:
• Bad Mouthing Attack (BMA). In this type of attack, conspiring nodes propagate unfairly
negative ratings of good nodes with an ill intent to tarnish their reputation in the
network. Such collusive behaviour may lead to the blocking of valid paths in the
network by confusing the trust and reputation management mechanism.
• Ballot Stuffing Attack (BSA). Propagation of unfairly positive ratings for some poorly
performing nodes by collusive nodes in the network lead to ballot stuffing attack. The
intention of collusive nodes is to mislead the trust mechanism and cause it to
malfunction in accurately reporting the trustworthiness of assessed node.
• Selective Misbehaviour Attack (SMA). This attack victimises some trusted nodes by
propagating false ratings for them, while at the same time acting normal to other nodes.
This type of behaviour can be very difficult to detect for the trust mechanism.

• Intelligent Behaviour Attack (IBA). This attack selectively provides recommendation
with high or low ratings according to the trust threshold. This kind of attack can cause
malfunction to the trust framework by dynamically responding to trust threshold and
behaving based on it.
• Time-dependent Attack (TDA). This attack makes participating nodes to change their
behaviour by time. Nodes can behave normally for a period of time and can misbehave
by providing unfair ratings at other times. This attack also has its roots in the subjective
property of trust.
• Location-dependent Attack (LDA). This attack exploits mobility property of MANETs,
where a node behaves differently according to its location. This attack originates from
the subjective property of trust where behaviours at one location cannot affect
evaluating trustworthiness of nodes at another location.
The summarised attacks belong to two categories: false rating (BMA, BSA, and SMA), and
inconsistent rating based on the trust threshold, time, or location (IBA, TDA, and LDA). Some
of the countermeasures illustrated below can be used for both categories or being specifically
designed for one category. For example, [20] proposes the use of only positive
recommendations, while [18] uses only negative recommendations and this can countermeasure
attacks like ballot stuffing and bad mouthing. This kind of defence can be harmful to trust
information because nodes cannot report their complete experiences. Statistical methods like
Bayesian theory to accurately compute the correctness of recommendations can be a proper
solution to both categories [27]. Proof of sufficient interactions [13] and specifying a certain
threshold of negative and positive recommendation, besides, the majority opinion technique
[25] could also be used to mitigate the effect of false and inconsistent rating. Comparison
between recommendation list and proof of time and location of the recommendation provider
is also a promising solution to time and location-dependent attacks., The method of comparing
time and location is considered first time in the proposed algorithm.
What follows from above discussion is that the recommending nodes' trustworthiness
cannot be assessed by just a single scheme. It should be supported by using many behaviour
and social properties (such as, the closeness between nodes, and proof of time and location)
which are missed in the illustrated literature. In order to improve accuracy and robustness of
the trust model, the influence of the untrustworthy recommendations should be mitigated to
overcome the problem of false negative and false positive.
4. THE PROPOSED MODEL
We propose a recommendation-based trust management model to secure the routing protocol
between source and destination nodes based on the trust value of each node in the path. The
model considers the problem of the attacks discussed earlier due to some misbehaving nodes in
MANETs. We make use of a Bayesian statistical approach similar to that used in [28] for
computing trust values based on the assumption that they follow a beta probability distribution.
Beta distribution is estimated by using two parameters (𝛼, 𝛽). They can be calculated by
accumulating observations of forwarding and dropping packets where 𝛼 represents the
accumulation of positive observations (forwarded packets) and 𝛽 represents the accumulation
of negative observations (dropped packets). The beta distribution can be defined by gamma
function as shown in Eq. (1).
(1)
where 0 ≤ 𝑝 ≤ 1; 𝛼, 𝛽 > 0 with a condition that 𝑝 ≠
0 if 𝛼 < 1 𝑎𝑛𝑑 𝑝 ≠ 1 𝑖𝑓 𝛽 < 1.

Nodes in the network observe each other’s behaviour in order to construct a trust
relationship representing the degree of trustworthiness one node can put on another. These
relationships are useful to help nodes decide whether to forward packets to a specific neighbour
or not. In the proposed model, an initial trust relationship is established between two nodes 𝑖
and 𝑗 as (𝛼𝑖𝑗, 𝛽𝑖𝑗) at time 𝑡, where 𝛼𝑖𝑗 denotes the positive interactions observed by node 𝑖 about
node 𝑗, and 𝛽𝑖𝑗 denotes the negative interactions observed by node 𝑖 about 𝑗. At time 𝑡 = 0, we
start with = 𝛽𝑖𝑗 = 1, which indicates the initial belief held by node 𝑖 about. This value is
translated into complete uncertainty about the distribution of the parameter which means no
observation or evidence has been collected. If the estimated positive and negative interactions
between two nodes 𝑖 and 𝑗 are denoted as 𝜌 and 𝑛 respective-
ly, 𝛼𝑖𝑗 and 𝛽𝑖𝑗 would be calculated as 𝛼𝑖𝑗 = 𝜌 + 1 and 𝛽𝑖𝑗 =
𝑛+ 1 where 𝜌 and 𝑛 ≥ 0. After each observation, the trust metric can be computed and
updated from these parameters as the expectation of beta distribution given by .
The proposed trust model uses clustering technique in order to maximise the consistency of
receiving recommendations. For example, recommendations from a misbehaving node can
have a range of multiple different ratings for the evaluated node. These ratings may be
inconsistent in which they can differ from each other in a short period of time, a malicious act
of the misbehaving node to confuse the trust model. Dynamic clustering of the
recommendations over a period of time can filter out deviated ratings from the list of
recommendations, thus decreasing the influence of false estimations in computing trust value.
The proposed model clusters recommendations based on three different criteria: (a) number of
interactions by the means of using confidence value, (b) compatibility of information with the
evaluated node by the means of deviation test, and (c) closeness between these nodes. The use
of multiple criteria to judge whether a node is dishonest can mitigate the influence of false
negative and false positive ratings.
The model has three components deployed to evaluate trust: (a) Trust Computation
Component that uses direct as well as indirect (second hand) trust information. (b)
Recommendation Manager Component that requests and gathers recommendations for a node
from a list of recommending nodes, and (c) Cluster Manager Component which filters out
dishonest recommendations from the list and sends out a list of trustworthy recommendations
to the manager component. Fig.2 shows the model’s components and their interaction process.
The recommendation manager and cluster manager components are described in section 5.
The trust computation component obtains direct trust value from two nodes that have
already initiated a trust relationship. These two nodes can continue to interact with each other
at least for a period of time they are within the range. Direct trust value is considered to be
accurate and its computation invulnerable to dishonest recommendations.

Figure 1 Attacks Related to Misbehaviour Problems in Recommendation Management of Trust and
Reputation Frameworks
Direct trust value 𝑇𝑖𝑗 𝑑 of node 𝑖 about 𝑗 is calculated as in Eq. (2).
Influence of past experiences change over time in a dynamic environment. It is thus
important for a trust model to consider this change in influence. The proposed model
incorporates a decay factor (µ) to gradually decrease the influence of past experience over time,
prior to the aggregation with new trust values. Forgetting of past experiences is carried out by
adjusting the time frame of observations while recording the positive or negative experience.
However, trust decays over time even during inactivity periods and it is thus important to
consider the diminishing impact of trust over the time. The first situation is when a node
observes an additional new positive or negative interaction between time 𝑡𝑖 and 𝑡𝑖+1 denoted as
𝜌 𝑛𝑒𝑤 and, then the updated 𝜌 and 𝑛 should be reduced by the decay factor µ before merging
them with the new values. Therefore, at time 𝑡𝑖+1, 𝜌 and 𝑛 is updated respectively according to
the formula in Eq. (3).
𝜌 = 𝜌 𝑜𝑙𝑑 ∗ µ + 𝜌 𝑛𝑒𝑤, 𝑛 = 𝑛 𝑜𝑙𝑑 ∗ µ + 𝑛 𝑛𝑒𝑤 (3)
where 0 ≤ µ ≤ 1, 𝜌 𝑜𝑙𝑑 and 𝑛 𝑜𝑙𝑑 are the old positive and negative experiences observed by
the node. The second situation is when there is no observed new positive and negative
interaction between time 𝑡𝑖 and 𝑡𝑖+1, then, at time 𝑡𝑖+1, 𝜌 and 𝑛 is updated respectively as in Eq.
(4).
𝜌 = 𝜌 𝑜𝑙𝑑 ∗ µ, 𝑛 = 𝑛 𝑜𝑙𝑑 ∗ µ (4)
Indirect trust needs to be considered, when two nodes have not established a previous trust
relationship through exchange of packets or any other form of communication. In such case,
the evaluating node doesn’t have enough experience to judge the trustworthiness of the other
node being evaluated. Indirect trust is also calculated using the beta-function, similarly as the
direct trust was computed earlier. Indirect trust is actually the direct observations obtained by
one node about its neighbours which can be used by another node as second-hand information.
We can say that node ’s direct observations of node 𝑗 could be indirect or second-hand

information to another node 𝑖 (given that node i and j have not interacted in the past). Therefore,
indirect trust value is calculated using (𝛼𝑖𝑗
′,
𝛽𝑖𝑗
′)
and updated by two variables: 𝜌’, describing the
number of positive interactions, and 𝑛’, describing the number of negative interactions. Further,
𝛼𝑖𝑗
′
and ′ are calculated as 𝛼𝑖𝑗
′
= 𝛼𝑖𝑗
′
+ 𝜌′ and 𝛽𝑖𝑗
′
= 𝛽𝑖𝑗
′
+ 𝑛′. If the evaluating node 𝑖 receives 𝛮
recommendations for the evaluated node 𝑗 denoted by 𝑘 = 1, 2, …, 𝛮, indirect trust 𝑇𝑖𝑗 𝑖 of node
𝑖 about 𝑗 is calculated according to the Eq. (5).
Figure 2 Recommendation-based trust model components
While indirect trust information is important to incorporate in a trust model for MANETs,
involving this kind of information can be vulnerable to intentionally generated dishonest
recommendations.
For each node in the network, trust value 𝑇𝑖𝑗 is calculated by combining both direct and
indirect trust values with different weights denoted by 𝑤 𝑑 and respectively.
𝑇𝑖𝑗 is computed according to Eq. (6).
𝑇𝑖𝑗 = 𝑤 𝑑 ∗ 𝑇𝑖𝑗
𝑑 + 𝑤𝑖 ∗ 𝑇𝑖𝑗
𝑖 (6)
where 𝑤 𝑑 + 𝑤𝑖 = 1. The weights are used because of their significant impact on diminishing
the possibility of wrong trustworthiness evaluation of direct and indirect trust information by
nodes. In most of existing models, higher weight is usually given to the direct information as
it is less prone to dishonest recommendation. However, MANETs’ characteristics such as
mobility and frequent change in topology make it difficult to completely trust the source of
information even if it is the nodes selfassessment. The weight in this model is dynamically
calculated based on the quantity and quality of interactions observed by evaluating nodes. If
the evaluating node has enough experience about the evaluated node and the evaluated node is
not compromised or prone to any environmental conditions (e.g. node failure, or low energy
level), it is given equal or more weight than indirect information. While, if the evaluating node
is not able to judge the trustworthiness of the evaluated node, more weight is given to the
indirect trust.

5. CLUSTER-BASED RECOMMENDATION FILTERING
This section analyses the functionalities of recommendation and cluster manager components
and shows how they work together to filter out untrustworthy recommendations. The proposed
filtering technique takes into consideration the dynamic characteristics of MANETs that change
over time. The honesty of recommending nodes is evaluated over a period of time to mitigate
the influence of bad behaviour of the same node over time. Fig. 3 shows the dynamic topology
of MANETs. Consider that, a node 𝑖 wants to evaluate another node 𝑗 by requesting
recommendations from its neighbours. The evaluating node 𝑖 receives a list of recommending
nodes referred as {𝑘1, 𝑘2, 𝑘3, …, 𝑘𝛮). At time 𝑡𝑖 (refer Fig. 3(a)), the location and number of
recommending nodes differ from the recommending nodes at time 𝑡𝑖+1 as shown in Fig. 3(b).
Recommendation manager in the proposed model works as an intermediate component
between indirect trust computation and cluster manager components. It helps in detecting and
eliminating false recommendations. Recommendation manager has three important roles: (1)
send recommendation request to the evaluating node’s neighbours; (2) collect received
recommendation and send it to the cluster manger which runs the filtering procedure; (3) receive
the filtered recommendation and send it back to the trust computation component.
Recommendation manager requests and gathers recommendation list for an evaluating node 𝑖
about node 𝑗 from a list of recommending nodes {𝑘1, 𝑘2, 𝑘3, …., 𝑘𝑁} between time 𝑡𝑖 and
𝑡𝑖+1and send it to the cluster manager to run the filtering algorithm. After filtering, it receives
the trustworthy clusters as a list of honest recommendations denoted as {𝑘1𝑇𝑟, 𝑘2𝑇𝑟, 𝑘3𝑇𝑟,
𝑘𝛮𝑇𝑟}. The final task is to send the trustworthy cluster 𝐶𝑇𝑟𝑢𝑠𝑡𝑤𝑜𝑟𝑡ℎ𝑦 to the requesting node.
Algorithm 1 illustrates the recommendation manager algorithm.
Figure 3 Recommendation by time
1. Algorithm 1: Recommendation Manager Algorithm
2. For each recommendation request Do
3. Send request to neighbours
4. Collect received recommendations
5. Construct 𝐿 = {𝑘1,2, 𝑘3, 𝑘𝛮}
6. Send 𝐿 to the cluster manager for processing
7. Receive trustworthy cluster 𝐶𝑇𝑟𝑢𝑠𝑡𝑤𝑜𝑟𝑡ℎ𝑦 = {𝑘1𝑇𝑟, 𝑘2𝑇𝑟, 𝑘3𝑇𝑟, …, 𝑘𝛮𝑇𝑟}
8. Send 𝐶𝑇𝑟𝑢𝑠𝑡𝑤𝑜𝑟𝑡ℎ𝑦 to the requesting node
9. End For

Nodes are clustered based on three values, namely: confidence value, deviation value, and
closeness value. The following subsections will explain these values and give an overview of
the clustering process and its algorithm.
A. Confidence Value 𝑉𝑖𝑘 𝑐𝑜𝑛𝑓
The notion of confidence was introduced in [29] where confidence value and trust value are
combined together to derive a single trustworthiness value of a node. Following that, trust
models in [10, 13, 30] have also considered the confidence value as a desired parameter to
achieve a single trust value to represent the trustworthiness of nodes. Confidence value can be
used to solve the problem of short-term and long-term observations. That is, nodes may have
the same level of trust with different number of observations. For example, the trust value of a
node at the initial time with 𝛼 = 𝛽 = 1 is 0.5, and after a sequence of positive and negative
interactions in which 𝛼 = 𝛽 = 50, the node has the same trust value of 0.5 about the evaluated
node (see Table 1 for more information). Confidence value starts from 0 in case of no
observations between nodes and increases gradually with the number of recorded observations.
Relying only on the trust value can raise the problem of short-term and long-term observations.
Nodes in the network can have nearly the same level of trust though they may have different
levels of observations. Consequently, this can lead to wrong estimation in judging the ability of
nodes to be honest recommending node.
The proposed filtering algorithm clusters recommending nodes based on the level of
confidence for two reasons. Firstly, the nodes with higher confidence value (those having
sufficient interactions with evaluated node) are desirable because the higher number of
interactions will offer rich information that would help in choosing better recommending nodes.
Secondly, the recommending nodes with very high confidence value in the early rounds in the
network (when there are no enough interactions) are more likely to be attackers. Consequently,
it may lead to exclusion of dishonest nodes from the recommendations list in early stages. The
confidence value is computed as the variance of the beta distribution with some modifications
as in [10] and [13]. Nodes use the confidence value to make a correct decision about the
trustworthiness of recommending nodes taking into account the number of observations
accumulated by each node. Suppose that 𝑖 is an evaluating node that received recommendations
from a recommending node 𝑘, confidence value is calculated as in Eq. (7).
where is the beta distribuation varience between 𝑖 and 𝑘, 𝛼𝑖𝑘 and 𝛽𝑖𝑘 is the accumulated
positive and negative interactions between 𝑖 and 𝑘?
Using this formula the value of confidence falls between the interval of [0, 1], where 0
means that no previous interactions are recorded between the evaluating and evaluated node
while 1 means complete confidence in the evaluated node. The rational of using and computing
the confidence value is shown in Fig. 4. We compare the confidence value computed using the
Table 1. Levels Of Confidence For The Proposed Model And Tmuc Model With The Same
Trust Levels
Proposed method with that in [31] (we call it TMUC for short), which computes the
confidence value using only the standard deviation. The proposed computation method of
confidence value can effectively reflect the knowledge held by nodes based on the number of
interactions better than the calculation in TMUC. For example, when 𝛼 = 𝛽 = 1 which means

there is no previous interaction between two nodes, the proposed method of computing
confidence value is 0 while in TMUC, it is nearly 0.91 which is a high value close to 1. Starting
with high confidence value in case of no interactions can confuse the trust mechanism and
prevent it from making good judgement about behaviour of the evaluated node. Table 1 shows
the values of positive and negative interactions and the confidence value for each level of
interaction for both the proposed model and the work in TMUC. Fig. 4 explains the relationship
between interactions and the level of confidence when the trust levels are the same.
From Fig. 4, it can be seen that the proposed method of computing confidence offers a better
range for the confidence value as compared to that by TMUC. This variation reflects better
accumulated interactions when the trust values (refer Table 1) are same. When there are no
interactions, confidence value from the proposed model is 0 and it progresses with the
increasing number of interactions. Whereas with TMUC, the confidence value is already at 0.91
in case of no interactions and thus is nearly at saturation level when number of interactions
more than 19.
B. Deviation Value 𝑉𝑖𝑗 𝑑𝑒𝑣
Deviation value represents to what extent the received recommendation is compatible with the
personal experience of evaluating node.
Figure 4 Relationships between Interactions and Confidence for the proposed model and TMUC
model
This value has been used by the means of the deviation test in [18] to ensure the unity of
view with the receiving node. Each node compares received recommendation with its own first-
hand information and accepts only those not deviating too much from self-observations. In the
proposed model the deviation value is used as an additional parameter in the clustering
algorithm to filter out any recommendations deviating beyond a predefined deviation threshold.
A problem that could arise here is when the evaluating node lacks historical information for
interactions with the evaluated node, thus not providing a base value for comparison. In order
to overcome this problem, the proposed method compares the confidence level of the evaluating
node with that of the recommending node.
The confidence value is calculated using Eq. (7). The deviation test is only applied if both
nodes have similar level of confidence. Assume that there are three nodes (𝑖, 𝑗 and 𝑘), and node
𝑖 attempts to calculate the trust value of its neighbour node 𝑗 using recommendation provided
by node 𝑘. In this scenario, node 𝑖 first compares its confidence level which denoted as
𝐶𝑜𝑛𝑓_𝐿𝑒𝑣𝑒𝑙 with the recommending node as in Eq. (8). If the confidence difference is less than
a threshold value denoted as 𝐶𝑜𝑛𝑓 𝑇ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑, then node 𝑖 calculates the deviation value as a

difference between the receiving recommendation and direct observations of the evaluated node
as held by the evaluating node as in Eq. (9). The resulting value is compared to a predefined
deviation threshold 𝑑 and we exclude any recommendations that differ widely from the
evaluating node’s own information.
𝐶𝑜𝑛𝑓_𝐿𝑒𝑣𝑒𝑙 = |𝐶𝑉𝑖𝑗 − 𝐶𝑉 𝑘𝑗| ≤ 𝐶𝑜𝑛𝑓_𝑇ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 (8)
where is the confidence value of 𝑖 about 𝑗, and 𝐶𝑉 𝑘𝑗 is the confidence value of 𝑘 about 𝑗? If
the Eq. (8) is successful, deviation value 𝑉𝑖𝑗 𝑑𝑒𝑣 is calculated as follows.
where 𝑇𝑖𝑗 𝑑 is the direct trust value of 𝑖 about 𝑗, and 𝑇 𝑘𝑗 𝑟 is the received trust value of 𝑘 about 𝑗.
C. Closeness Centrality Value
Trust is a social concept and it is thus possible to apply the perceptions of social life in trust
computation and recommendation propagation. An interesting direction of trust research in
MANETs is to utilise social relationships in evaluating trust among nodes in a group setting by
employing the concept of social structures [5]. The proposed model uses the concept of
closeness centrality between the evaluating nodes and the recommending node from the social
trust. Closeness centrality [32] measures the distance between the evaluated node and the
recommending node in terms of physical distance, number of hops, or delays. In the proposed
model closeness centrality is a measure of the distance between the evaluating node and the
recommending node. The use of the closeness centrality enhances the filtering algorithm as
close nodes are likely to possess same nature and counter nearly same environmental and
operational conditions over a period of time in the network. Furthermore, close friends may
have more interactions in the time of friendship. Consequently, trust values for the close
neighbours converge to nearly same level. This may help in recognising the untrustworthy
recommending node whose recommendation is much different from the close recommending
nodes. Closeness value
𝑉𝑖𝑗 𝑐𝑙𝑜𝑠𝑒 refers to the degree of node 𝑖’s closeness to a recommending node 𝑘 at time 𝑡 and is
calculated by Eq. (10).
where ) are the positions of node 𝑖 and node 𝑘 at time 𝑡 and 𝑑 𝑑𝑖𝑠 is a
predefined distance threshold between node 𝑖 and node 𝑘 which should be less than the
transmission range.
D. Cluster procedure
The cluster manager in the proposed model receives a list of recommendations from the
recommendation manager and processes it using a clustering technique. The clustering
algorithm is run by the evaluating node on all the recommendations in the list 𝐿 = {𝑘1, 𝑘2, 𝑘3,
…, 𝑘 𝑛}. A vector of
three values (𝑉𝑖𝑗 𝑐𝑜𝑛𝑓, 𝑉𝑖𝑗 𝑑𝑒𝑣, 𝑉𝑖𝑗 𝑐𝑙𝑜𝑠𝑒) is provided by a recommending node for the clustering
operation. The clustering algorithm divides the vectors from the recommending nodes into a
predefined number of clusters denoted as K. Initially each vector is considered as a cluster, and
then two clusters with the shortest Euclidean distance are merged together to produce a new
cluster. The clustering process is repeated by merging two clusters from the previous iteration
until the predefined number of clusters K is reached. The first step of the clustering process
aims to merge vectors with the closest similarity. In the second step, it selects the trustworthy
clusters if all the recommending nodes in a specified cluster satisfy the following rules:

𝐶 𝑇𝑟𝑢𝑠𝑡𝑤𝑜𝑟𝑡ℎ𝑦 = { if (𝑉𝑐𝑜𝑛𝑓𝑖𝑗 ≥ 𝑑𝑐𝑜𝑛𝑓𝑚𝑖𝑛) and (𝑉𝑐𝑜𝑛𝑓𝑖𝑗≤ ) 𝑖𝑓 (𝑉𝑑𝑒𝑣𝑖𝑗≤ 𝑣) and
(𝑉𝑐𝑙𝑜𝑠𝑒𝑖𝑗≤ 𝑑𝑑𝑖𝑠)
𝑈𝑛𝑡𝑟𝑢𝑠𝑡𝑤𝑜𝑟𝑡ℎ𝑦
𝑅𝑖j 𝑜𝑡ℎ𝑒𝑟 𝑤𝑖𝑠𝑒
where, is the trustworthy recommendation, dconfmin is the minmum confidence
threshold, is the maximum confidence threshold.
The next step is to apply majority rule to select the cluster with largest number of
members. In the final step, trustworthy clusters are returned to the recommendation manager
and to the evaluating node to update its indirect trust of the evaluated node. The proposed
cluster process works as shown in Algorithm 2.
1. Algorithm 2: Cluster Manager Algorithm
2. For each recommendation list 𝐿 Do
3. For each rating vector in the list (𝛼 𝑟, 𝛽 𝑟) Do
4. Calculate confidence value 𝑉𝑖𝑗 𝑐𝑜𝑛𝑓 as in Equ. 7
5. Calculate deviation value 𝑉𝑖𝑗 𝑑𝑒𝑣 as in Equ. 8, 9
6. Calculate closeness value 𝑉𝑖𝑗 𝑐𝑙𝑜𝑠𝑒 as in Equ. 10
7. Construct data vector as (𝑉𝑖𝑗 𝑐𝑜𝑛𝑓, 𝑉𝑖𝑗𝑑𝑒𝑣, 𝑉𝑖𝑗𝑐𝑙𝑜𝑠𝑒)
8. End For
9. Initialize each vector as a unique cluster
10. Repeat
11. For each vector Do
12. Merge two clusters with the shortest Euclidean distance
13. End For
14. Until number of clusters = 𝐾
15. For each cluster that appeared in the previous iteration Do
16. If Then
17. If (𝑉𝑖𝑗 𝑑𝑒𝑣 ≤ 𝑑𝑑𝑒𝑣) and (𝑉𝑖𝑗 𝑐𝑙𝑜𝑠𝑒 ≤ 𝑑 𝑑𝑖𝑠) Then
18. Select trustworthy cluster
19. End If
20. End If
21. End For
22. For each chosen trustworthy cluster Do
23. Apply the majority rule
24. Return trustworthy cluster 𝐶 𝑇𝑟𝑢𝑠𝑡𝑤𝑜𝑟𝑡ℎ𝑦
25. End For
6. SIMULATION AND RESULTS
The simulation is conducted using NS2 simulator [33], an open-source discrete event simulator
designed to support research in computer networking. It involves various modules to help test
several network components such as packet, node, routing, application and transport layer
protocol. NS2 features permit us to extend the DSR routing protocol that supports MANETs
architecture. The proposed trust model components are added to the simulator to test the validity
of the proposed model. A network with 50 mobile nodes is simulated randomly moving in an

area of 700×700 square meter. Several nodes are randomly selected to provide false rating
information in the form of bad-mouthing and ballot-stuffing attacks. There are 15 source-
destination pairs and each source transmit 2 packets per second with a Constant Bit Rate (CBR),
and pause time 60s, which is the time nodes need to pause to begin travelling to the next
destination with a speed of 10 m/s, the packet size is 512 bytes and the simulation time is 500s.
The mobility model utilised in this paper is the random way point which is the most commonly
used model in ad hoc networking research [34] It is easy to use, and movement could be
considered as realistic which is very similar to the real-world movement [35]. However, the
proposed model can fit any other type of mobility models like RPGM model [36]. The
maximum bad-mouthing and ballot-stuffing attacks percentage used in the simulation scenario
is 80% misbehaving nodes, which is enough percentage to test these attacks. An optimistic
scheme is used in choosing trust threshold value at 0.2 in which all nodes are initially expected
to be trusted and normally behaving [10]. Table 2 shows the parameters used in configuring the
network for the experiment. Bad-mouthing and ballot-stuffing attacks with additional
permission to collude in both attacks are used in order to evaluate the defense scheme against
dishonest recommendation. Number of dishonest nodes range from 0% to 80% and the
dishonest recommendations provided deviate 50% from the honest recommendations. Badly
behaving nodes (selfish nodes) counting to 20% always existed in the network and were
responsible for collusion and jamming. Results from the experiment are based on multiple runs
and a negligible variation is noticed.
Table 2 Network Configuration Parameters
Parameter Value
Nodes 50
Area 700 m X 700 m
Speed 10 m/s
Radio Range 250 m
Movement
Random waypoint
model
Routing Protocol DSR
MAC 802.11
Source-destination
pairs
15
Transmitting
capacity
2 Kbps
Application CBR
Packet size 512 B
Simulation time 500 s
Trust threshold 0.4
Publication timer 30 s
Fading timer µ 10 s
Deviation threshold
ddev 0.5
Conf_Threshold 0.4
𝑑𝑚𝑖𝑛𝑐𝑜𝑛𝑓 0.5
𝑑𝑚𝑎𝑥𝑐𝑜𝑛𝑓 0.9
𝑑𝑑𝑖𝑠 200 m

A. Performance Evaluation
The flow of the simulation is as follows. The performance of the entire network is represented
by two parameters: Network throughput and Packet loss in the presence of bad-mouthing,
ballot-stuffing and selfish nodes. The trust value of a good node (not misbehaving) is evaluated
against bad-mouthing attack to see the influence of such attack with and without incorporating
the proposed defence scheme. The trust value of a bad node (misbehaving) is also evaluated
against ballot-stuffing attack to see how such attackers can distort the trust value of this node.
The performance of the proposed model in terms of recognised dishonest recommendations,
false negative and false positive in the presence of bad-mouthing attacks with and without the
defence scheme is examined. Similar experiment is conducted for ballot-stuffing attack.
Finally, a comparative study is conducted with the maturity model [24] proposed in the
literature.
Fig. 5 demonstrates the effect of dishonest recommendation on two performance metrics;
throughput and packet loss for the whole network. The y-axis in Fig. 5(a) shows the percentage
of throughput, both with and without the defence scheme, in the presence of dishonest nodes
varying from 0% to 80% of the total population of nodes. It is observed that the network
throughput without a defence falls from nearly 80% when the dishonest nodes are not present
to nearly 30% when their population increases to 80%. Slight decrease and then increase is
noticed in the throughput (Fig. 5a) for the network with defence when the percentage of
dishonest recommendation nodes increases from 40% to 80%. This may be due to the fact that
the throughput not only depends on the number of misbehaving nodes but is also affected by
the degree of connectivity (number of neighbours),ability of the nodes to classify their
neighbours and the time required to achieve the classification (which are different in each
simulation due to network topology and mobility). However, the proposed defence mechanism
was able to keep the value of throughput at nearly 80% even in case of higher population of the
dishonest nodes. This can be translated as the defence scheme’s ability in mitigating the
negative effect of dishonest recommendations on the throughput performance. The impact of
dishonest nodes on packet loss is shown in the Fig. 5(b). The percentage of packet loss rises
with increasing the percentage of dishonest nodes from 20% to over 60% when no defence
incorporated in the network. While only 20% packet loss can be maintained using the proposed
defence scheme in the presence of dishonest recommending nodes that vary from 0% to 80%
of the nodes in the network. Similarly, the percentage of packet loss decreases slightly when
the percentage of dishonest recommendation nodes increases from 70% to 80% for the same
reasons as discussed in the analysis of Fig. 5(a). It can be seen from the above analysis that
dishonest recommendations can significantly impact the throughput and packet loss metrics by
confusing the trust model. The proposed technique can keep those metrics at an acceptable level
even when the population of dishonest nodes is high.
Fig. 6 demonstrates the average of the indirect trust held by other nodes in the network for
a good node (node 12 in this case) and a bad node (node 4 in this case). The x-axis in Fig. 6(a)
displays the range for the population of bad-mouthing nodes from 0% to 80%. The y-axis shows
the average of the indirect trust value for a good node (node 12 in this case) as held by all the
nodes that have interacted with it in the past. A comparison has been made between three
different parameters as follows. First, the indirect trust value when there are no dishonest nodes,
called expected value. Second, the indirect trust

Figure 5 Network performance in the Presence of Dishonest Recommending nodes for a) Network
Throughput; b) Network Packet Loss
value when dishonest nodes are present and the defense scheme is working, with defense.
Third, the indirect trust value when the dishonest nodes are present and the defense technique
is not working, no defense. It can be seen that with increasing population of badmouthing
attackers, the average trust value of node 12 declines in case of no defense, whereas, the trust
value remains the same as the expected value in case of with defense.
The effects of ballot-stuffing attack are shown in Fig. 6(b). In the x-axis is the percentage
of ballot-stuffing attack that varies between 0% to 80% and y-axis shows the values for the
indirect trust compared against the same three parameters i.e. expected value, with defense and
no defense cases. From the figure, it can be seen that the attacking nodes have propagated
unfairly positive rating for the dishonest node (node 4) thereby raising its trust value to above
0.9 while the attacker population was 80%. The results here show that the defense algorithm is
capable of mitigating the influence of dishonest nodes by filtering out unfair ratings.
To test the proposed defense scheme further, we define three additional metrics: (a)
recognized proportion, representing the number of dishonest recommendations identified by
node , (b) false negative proportion, indicating the number of dishonest recommendations
identified as honest by node ,(c) false positive proportion, indicating the number of honest
recommendations identified as dishonest by node . Fig.7 and 8 show the results for these three
metrics in the presence of bad-mouthing and attack. The x-axis in Fig.7(a) shows the
percentage of bad-mouthing attack while y-axis shows the proportion of the recognized
dishonest recommendation, false negative and false positive with the defense scheme in action.
It can be observed that the defense algorithm can effectively mitigate the dishonest
recommendation propagated by the bad-mouthing attackers regarding the recognition and false
negative metrics. While it keeps the false positive proportion at a very low level (about 2%)
when the attack percentage is more than 50%. Fig. 7 (b) shows the case when the defense
scheme is not in action. It can be seen that the proportion of recognized dishonest
recommendation drops to less than 10% when the percentage of dishonest nodes increase to
80% and consequently the proportion of false negative increases with the increase in dishonest
recommending nodes. As the defense scheme is not in action here, it accepts all the
recommendations propagated in the network and updates the indirect trust value based on these
recommendations. Therefore, the proportion of false positive remains at zero (Fig 7(b)).
Fig. 8 (a) shows results for ballot-stuffing attack. The proposed defence scheme is seen to
be identifying dishonest recommendations and eliminating false negative effectively. The
proportion of false positive is maintained at a reasonable level. The effect of dishonest
(a) Network throughput (b) Packet Loss

recommendation in Fig. 8(b) is obvious. When there is no defence incorporated the proportion
of recognition drops from about 0.9 to nearly 0.1 with variation of the ballot-stuffing attackers
from 0.1 to 0.8. The false negative proportion also increases to nearly 0.9 with the increasing
percentage of the dishonest recommending nodes.
Finally, the performance of the proposed model is compared with the maturity model
proposed in [24] in terms of two metrics: trust level error (TLE) which represents the proportion
of error in evaluating the trust level of a node 𝑖 (node 8 in this case); and trust level evaluation
of a good node (node 1 in this case) by another node 𝑗 in the network. We follow the same
network configuration and node selection which is provided in the maturity model (see [24] for
details) to conduct this experiment. In this configuration, a high-speed network is presented
with high node mobility, which is different from our first configuration. This configuration of
the test network allows us to show the effectiveness of the proposed scheme. Fig. 9 shows the
results of this experiment. Fig. 9(a) displays the trust level error over the simulation time. It can
be seen that the proposed model can keep the TLE smaller than the error reported by the
maturity model. The TLE in case of the proposed model is stable for the entire time of
evaluation and converges to very small value nearly 0.01 towards the later phase. While for the
maturity model, the TLE value is high initially (0.35) as compared to that of the proposed model
and this only converged to 0.1 towards the end (time unit 3000). Fig. 9(b) shows the
effectiveness of the proposed defence scheme in evaluating the trust value of
Figure 6 Trust Evaluation for a) Good-Node 12's Trust Value in the Presence of Bad-mouthing
Attack; b) Bad-Node 4's Trust Value in the Presence of Ballot-stuffing Attack
Figure 7 Recognised, False Negative, and False Positive Proportion in The Presence of Bad-mouthing
Attack for a) With Defence; b) Without Defence
(a) Good-Node 12's Trust Value (b) Bad-Node 4's Trust Value
(a) With Defence (b) Without Defence

Figure 8 Recognised, False Negative, and False Positive Proportion in The Presence of
Ballot-stuffing Attack for a) With Defence; b) Without Defence

Figure 9 Comparative study with maturity model for a) Trust Level Error; and b) Good-Node 1’ Trust
Level
A good node (node 1) from the network. It considers the following scenarios: the expected
trust value when there is no dishonest recommendation (TLNDR), and the same when there is
35% dishonest recommendation (TL35DR) both for the proposed model and the maturity
model. The results show that the proposed model with the defence scheme can manage to avoid
the dishonest recommendation and keep the trust value of node 1 near to the expected value and
slightly higher than the results of the maturity model.
B. Cost Of The Defence Scheme
Mobile Adhoc networks are characterised by constrained resources in terms of communication,
memory usage and computational complexity requirements. Any proposed model or defence
scheme must reflect the trade-offs between accuracy of trustworthiness and network
performance. As gathering and propagating trust information among distributed node can
consume more resources of energy and time, it can enhance the decision making. Dynamic and
highly mobile networks which suffer from several points of failure require techniques to
enhance the decision making on nodes trustworthiness. However, the proposed defence scheme
is lightweight in several aspects. In terms of communication, the proposed model is suitable for
MANETs because only recommendation request and reply packets are used to send and receive
a list of recommendations. The packets of recommendations are exchanged between a single
source of information which is represented in the recommendation manager to and from the
evaluating node and the recommending nodes. The data size and length is very small as every
recommending node provides just three parameters of accumulated positive and negative
observations and its current position. The communication is also enhanced by on-demand
scheme in which recommendation is requested whenever needed. Therefore, the defence
scheme is conducted without network flooding and acquisition delay. The defence scheme is
characterised with the advantage of a role-based management scheme for filtering dishonest
recommendation in which three different components are interoperated to accomplish the task.
The use of clustering in distributed networks can facilitate the data aggregation and reduce the
computational power by each node to evaluate the trustworthiness of other nodes. One of the
costs put on the proposed defence is the complexity that can be countered in maintaining the
cluster and selecting the most trustworthy cluster. Another cost is the memory consumption in
which the defence scheme consumes more memory to store recommendation for a period of

time for conducting the filtering algorithm by the recommendation and clustering managers
which is run by the evaluating node but no memory consumption on the side of the evaluated
node. An additional cost is the time consumption which is more than the traditional defence
which uses single recommender information to update the trustworthiness of the evaluated
node. These costs can be reduced in the proposed defence scheme by using only the very last
recommendations to be including in the clustering filtering computation. Dynamic selection of
the number of recommendations based on a period of time can have many advantages, (1)
reduce complexity and memory usage, (2) exclude any old recommendation from the
calculation, (3) reduce the time that is used to select the trustworthy cluster.
7. CONCLUSIONS
A recommendation-based trust model with a defence scheme is developed and analysed to filter
attacks related to dishonest recommendation exchanged by nodes in the MANET. The use of
recommendation can efficiently allow nodes to acquaint with each other without previous
interactions, but it exposes nodes to dishonest and unfair recommendation. Therefore, the
proposed defence scheme utilises the clustering technique to filter out unfair recommendations
exchanged by nodes in the network based on three values: (a) the level of confidence held by a
node about others, (b) deviation threshold which ensures the unity of views between evaluating
node and the evaluated node, and (c) closeness centrality value to ensure that recommending
node is a close friend to the evaluating node for a period of time. The proposed model is tested
by extensive simulation in terms of throughput and packet loss, against both bad-mouthing and
ballotstuffing attacks, and also compared with other proposal. The simulation results indicate
that the proposed defence scheme can safely incorporate correct indirect trust evidences
received by recommendations and eliminate untrustworthy ones. Moreover, it reduces the effect
of false negative and false positive problems in selecting recommending nodes. The proposed
model can be extended by weighting recommendations based on time and location to mitigate
the influence of location and time dependent attacks.
REFERENCES
[1] H. Deng, W. Li, and D. Agrawal, “Routing security in wireless ad hoc networks,’’
Communications Magazine, IEEE, 40, (10), pp. 70-75, 2002.
[2] B. Wu, J. Chen, J. Wu, and M.Cardei, “A survey of attacks and countermeasures in mobile
ad hoc networks,’’ Wireless Network Security (Springer), pp. 103-135, 2007.
[3] I. Chlamtac, M. Conti, and J.J.-N Liu, “Mobile ad hoc networking: imperatives and
challenges,’’ Ad Hoc Networks, 1, (1), pp. 13-64, 2003.
[4] W. Li, J. Parker, and A. Joshi, “Security through collaboration and trust in manets,’’ Mobile
Networks and Applications, 17, (3), pp. 342-352, 2012.
[5] J.-H Cho, A. Swami, and R. Chen, “A survey on trust management for mobile ad hoc
networks,’’ Communications Surveys & Tutorials, IEEE, 13, (4), pp. 562-583, 2011.
[6] Z. Yan, P. Zhang, and T. Virtanen, ‘’Trust evaluation-based security solution in ad hoc
networks,’’ Proc. Seventh Nordic Workshop on Secure IT Systems, 14, 2003.
[7] A.A. Pirzada, and C. McDonald, ‘’Establishing trust in pure ad-hoc networks,’’, Proc. 27th
Australasian conference on Computer science, 26, pp. 47-54, 2004.
[8] N. Pissinou, T. Ghosh, and K. Makki, ‘‘Collaborative trust-based secure routing in multihop
ad hoc networks,’’ NETWORKING 2004. Networking Technologies, Services, and
Protocols; Performance of Computer and Communication Networks; Mobile and Wireless
Communications (Springer, 2004), pp. 1446-1451, 2004.
[9] S. Buchegger, and J.Y. Le Boudee, ‘’Self-policing mobile ad hoc networks by reputation
systems,’’ Communications Magazine, IEEE, 43, (7), pp. 101-107, 2005.

[10] R. Li, J. Li, P. Liu, and J. Kato, ‘’A Novel Hybrid Trust Management Framework for
MANETs,’’ Distributed Computing Systems Workshops, 2009. ICDCS Workshops' 09. 29th
IEEE International Conference on, pp. 251-256, 2009.
[11] S. Ganeriwal, L.K. Balzano, and M.B. Srivastava, ‘’Reputation-based framework for high
integrity sensor networks,’’ ACM Transactions on Sensor Networks (TOSN), 4, (3), pp. 15,
2008.
[12] Y. Yu, K. Li, W. Zhou, and P. Li, ‘’Trust mechanisms in wireless sensor networks: Attack
analysis and countermeasures,’’ Journal of network and computer applications, 35, (3), pp.
867-880, 2012.
[13] C. Zouridaki, B.L. Mark, M. Hejmo, and R.K. Thomas, ‘‘E-Hermes: A robust cooperative
trust establishment scheme for mobile ad hoc networks,’’ Ad Hoc Networks, 7, (6), pp.
1156-1168, 2009.
[14] S.R. Zakhary, and M. Radenkovic, ‘‘Reputation-based security protocol for manets in
highly mobile disconnection-prone environments,’’ Wireless On-demand Network Systems
and Services (WONS), 2010 Seventh International Conference on, pp. 161-167, 2010.
[15] H. Xia, Z. Jia, L. Ju, X. Li, and Y. Zhu, ‘‘A subjective trust management model with
multiple decision factors for MANET based on AHP and fuzzy logic rules,’’ Green
Computing and Communications (GreenCom), 2011 IEEE/ACM International Conference
on, pp. 124-130, 2011.
[16] K. S.G.a.C. ‘‘A Novel Defence Scheme Against Selfish Node attack in
MANET,’’International Journal on Computational Sciences & Applications (IJCSA, 3,
(3)), 2013.
[17] S. Chen, Y. Zhang, Q. Liu, and J. Feng, ‘‘Dealing with dishonest recommendation: The
trials in reputation management court,’’ Ad Hoc Networks, 10, (8), pp. 1603-1618, 2012.
[18] S. Buchegger and J.Y. Le Boudec, ‘’A Robust Reputation System for P2P and Mobile Ad
hoc Networks,’’ Proc. P2PEcon’04, 2004.
[19] H. Li and M. Singhal, ‘’Trust management in distributed systems,’’Computer, 40, (2), pp.
45-53, 2007.
[20] P. Michiardi, and R. Molva, ‘‘Core: A COllaborative REputation mechanism to enforce
node cooperation in Mobile Ad Hoc Networks,’’ Communications and Multimedia Security
Conference, pp. 107–121, 2002.
[21] K. Wang, M. Wu, and S. Shen, ‘‘A trust evaluation method for node cooperation in mobile
ad hoc networks,’’ Information Technology: New Generations, 2008. ITNG 2008. Fifth
International Conference on, pp. 10001005, 2008.
[22] J. Luo, X. Liu, and M. Fan, ‘’A trust model based on fuzzy recommendation for mobile ad-
hoc networks,’’ Computer Networks, 53, (14), pp. 2396-2407, 2009.
[23] S. Soltanali, S. Pirahesh, S. Niksefat, and M. Sabaei, ‘‘An efficient scheme to motivate
cooperation in mobile ad hoc networks,’’ Networking and Services, 2007. ICNS. Third
International Conference on, pp. 98-98, 2007.
[24] P.B. Velloso, R.P. Laufer, D. de O Cunha, O.C.M. Duarte, and G. Pujolle, ‘’Trust
management in mobile ad hoc networks using a scalable maturity-based model,’’ Network
and Service Management, IEEE Transactions on, 7, (3), pp. 172-185, 2010.
[25] H. Yu, S. Liu, A.C. Kot, C. Miao, and C. Leung, ‘’Dynamic witness selection for
trustworthy distributed cooperative sensing in cognitive radio networks’’, Communication
Technology (ICCT), 2011 IEEE 13th International Conference on, pp. 1-6, 2011.
[26] J. Cai, P. Yi, J. Chen, Z. Wang, and N. Liu, ‘‘An adaptive approach to detecting black and
gray hole attacks in ad hoc network,’’Advanced Information Networking and Applications
(AINA), 2010 24th IEEE International Conference on, pp. 775-780, 2010.
[27] K. Hoffman, D. Zage, and C. Nita-Rotaru, ‘’A survey of attack and defense techniques for
reputation system,s’’ ACM Computing Surveys (CSUR, 42, (1), pp. 1), 2009.

[28] A. Jøsang, R. Ismail, and C. Boyd, ‘’A survey of trust and reputation systems for online
service provision,’’ Decision support systems, 43, (2), pp. 618-644, 2007.
[29] G. Theodorakopoulos and J.S. Baras, ‘’Trust evaluation in ad-hoc networks,’’ Proc. 3rd
ACM workshop on Wireless security, pp. 1-10, 2004.
[30] G.V. Crosby, L. Hesterand, N. Pissinou, ‘‘Location-aware, Trust-based Detection and
Isolation of Compromised Nodes in Wireless Sensor Networks,’’ International Journal of
Network Security, 12, (2), pp. 107-117, 2011.
[31] M.K. Denko, T. Sun, and I. Woungang, ‘’Trust management in ubiquitous computing: A
Bayesian approach,’’ Computer Communications, 34, (3), pp. 398-406, 2011.
[32] E.M. Daly, and M. Haahr, ‘’Social network analysis for information flow in disconnected
delay-tolerant MANETs,’’ Mobile Computing, IEEE Transactions on, 8, (5), pp. 606-621,
2009.
[33] T. Issariyakul and E. Hossain, ‘‘Introduction to network simulator NS2,’’ Springer, 2011.
[34] C. Bettstetter, G. Resta, and P. Santi, ‘‘The node distribution of the random waypoint
mobility model for wireless ad hoc networks,’’ Mobile Computing, IEEE Transactions on,
2, (3), pp. 257-269, 2003.
[35] H. Simaremare, A. Syarif, A. Abouaissa, R.F. Sari, and P. Lorenz, ‘‘Performance
comparison of modified AODV in reference point group mobility and random waypoint
mobility models,’’ Communications (ICC), 2013 IEEE International Conference on, pp.
3542-3546, 2013.
[36] J.-L. Huang and M.-S. Chen, ‘’On the effect of group mobility to data replication in ad hoc
networks,’’Mobile Computing IEEE Transactions on, 5, (5), pp. 492-507, 2006.

TRUST MODEL WITH DEFENSE SCHEME IN MANETS

More Related Content

What's hot (13)

Similar to TRUST MODEL WITH DEFENSE SCHEME IN MANETS (20)

More from IAEME Publication (20)

Recently uploaded (20)

TRUST MODEL WITH DEFENSE SCHEME IN MANETS