chapter 4 enhanced wireless sensor networks with...
TRANSCRIPT
59
CHAPTER 4
ENHANCED WIRELESS SENSOR NETWORKS WITH
TRUST AND REPUTATION MODELS
4.1 Introduction
This Chapter deals with the impact of malicious servers over different trust and reputation
models for static, dynamic and oscillatory behavior of wireless sensor networks. Trust and
reputation models in wireless sensor network have attracted wider appreciation among
global community. Trust can be referred as a specific level of probability with which a
particular node will perform an action in a domain in which it affects its own function
before it can monitor its capacity. Reputation may be defined as an expectation about a
node's behavior based on its present information or past observations. Trustworthy nodes
can only be identified with trust and reputation models [Römer et al., 2004; Marsh, 1994;
Marti et al., 2006]. Many researchers have proposed trust and reputation models for
guaranteeing a specific security and accuracy level. Still, there is a dire need to give more
emphasis in this domain to enhance the coverage area of trust and reputation models in the
wireless sensor network applications.
4.2 Trust and Reputation Models in Wireless Sensor Networks
The description about the five major trust and reputation model of our concern is given in
the subsequent subsections.
60
4.2.1 Eigen Trust Model
It is the most frequently used trust and reputation model in the wireless sensor networks
domain. Kamvar et al. [2003] evaluated this model on the basis of the peer's history of
contributions by assigning a unique global trust value in the peer to peer file system for
each peer [Levien, 2000; Douceu, 2002]. For this, authors define Sij as the local trust of
peer i about peer j, in the following manner: Sij = sat(i,j)− unsat(i, j). It is the difference
between satisfactory and unsatisfactory interaction between peers i and j. Further,
normalized local trust value is exhibited in equation (4.1) as given below:
(4.1)
This is ensured that all the value lies in between 0 and 1. Kamvar et al. [2003] also
introduced aggregated local trust values which is defined as tik = ∑ jCijCjk where tik
represents the trust that peer i places in peer k based on asking his friends. Three practical
issues like a priori notions of trust, inactive peers and malicious collectives were also
incorporated by the authors in this model. In the presence of malicious peers, t = (CT) np
will generally converge faster than t = (CT) ne. In the case of inactive peers, Cij can be
refined as follow in equation (4.2):
(4.2)
61
Malicious collectives issue was addressed by the following equation (4.3) in this model.
(4.3)
4.2.2 Peer Trust Model
Xiong and Liu [2004] reported combined aspects related to the trust and reputation
management such as the feedback of a peer receives from other peers; the total number of
transactions of a peer; the credibility of the recommendations given by a peer; the
transaction context factor and the community context factor. The trust value of peer u, T(u)
is represented by the following equation (4.4):
(4.4)
where I(u) represents total number of transactions performed by peer u with all other peers,
p(u,i) represents other participating in ith
transaction peer u, S(u,i) represents normalized
amount of satisfaction peer u receives from p(u, i) in its ith
transaction, CR(v) represents
credibility of the feedback submitted by v, TF(u,i) represents adaptive transaction context
factor in ith
transaction of peer u and CF(u) represents adaptive community context factor
for peer u. On the other hand, the credibility of v from w is computed as follow in equations
(4.5) to (4.6):
(4.5)
62
where
(4.6)
I(u, v) represents the total number of transactions performed by peer u with peer v, IS(v)
represents set of peers that have interacted with peer v , IJS(v, w) denotes common set of
peers interacted with peer v and w for IS(v) ∩ IS(w) computation. The stimulation for the
community for the incentive or rewards is done through the context factor (CF) with the
following expression: CF (u) = F (u)/I (u) where F (u) represents total number of feedback
peer u gives to others.
4.2.3 BTRM-WSN Trust Model
This trust model for wireless sensor networks has been built on the bio-inspired algorithm
of ant colony system [Girao et al., 2006; Gomez et al., 2008; Marmol et al., 2009]. In this
model, most trustworthy path leads to the most reputable service provider in a network.
WSN launched a set of artificial agents while searching for the most reputable service
provider. In order to carry out a decision about next sensor, a probability is given to each
arc by the following equation (4.7):
(4.7)
63
Modification about the ants [Mármol and Perez, 2010] pheromone trace (distributed
decision making mechanism) for two sensors (s1, s2) is done in the following manner in
equation (4.8):
(4.8)
Where denoted the convergence value of
and φ represents a parameter controlling the amount of pheromone. Equation (4.9)
depicts the best path found by all ants.
(4.9)
Where Q(SGlobal_Best) denotes path quality. The quality of the Sk paths is measured as the
average of all the edges belongs to that path in equation (4.10):
(4.10)
Where % Ak denotes the percentage of trustworthy paths. The punishment or rewards of the
path leading to the selected peer is given by equation (4.11):
(4.11)
Equation (4.12) represents the distance factor ( joining the link between sensor r and
s.
64
(4.12)
4.2.4 Power Trust Model
Zhou and Hwang [2007] proposed this model for scalable and robust peer to peer
reputation system, specifically applicable in dynamically growing networks. These peer to
peer networks may be either structured or unstructured. This model builds a trust overlay
network (TON) on top of all nodes in a peer to peer system. In order to calculate the global
reputation score of each participating peer, it is mandatory to aggregate the local trust
scores of all the nodes. All global scores form a reputation vector V = {v1, v2, ., vn}
fulfilling that ∑Vi = 1. In order to compute vector V, consider the trust matrix R = (rij)
defined over an n-node TON, where rij belongs to [0, 1] is the normalized local trust score
defined by
is the most recent feedback score that node i rates node j.
Next, an initial reputation vector V(0) is set for instance, vi = 1/n and while |V(t) - V(t-1)| > e,
the successive reputation vectors are recursively computed as V(t+1) = RT ×V(t) ..After k
iterations, the global reputation vector will provide converge to the Eigenvector of the trust
matrix R. Finally, power nodes update their global reputation scores.
4.2.5 Linguistic Fuzzy Trust Model (LFTM) Model
This model uses the concept of fuzzy logic for the data reasoning [Marmol et al., 2010]. It
uses the representation power of linguistically labeled fuzzy sets for the satisfaction of a
client or the goodness of a server. Also, it remains exploited by the inference power of
fuzzy logic as in the imprecise dependencies between the originally requested service and
the actual received one or the punishment to apply in case of fraud. The expected resultant
will be an easily interpretable system with adequate performance. In this model, a set of
65
linguistic labels describing several levels of a variable or concept could be associated with a
fuzzy set. The resultant set constitutes linguistic labels such as: “Very Low”, “Low”,
“Medium”, “High” and “Very High”. This defined fuzzy set associated with above said
labels specifies the level of client satisfaction.
4.3 Malicious Server Investigations
We implemented our proposed model with Java-based simulator to test trust and reputation
models for wireless sensor networks [Marmol, 2009]. We considered ten networks
composed of fifty sensor nodes, each for ten scenarios in a two dimensional fields focusing
on severe malicious conditions as shown in figure 4.1.
Figure 4.1: Malicious servers scenario
66
Table 4.1 displays the parameters of the five trust and reputation models. We evaluated
three parametric aspects namely: accuracy, path length and energy consumption for
malicious servers in wireless sensor networks. We have developed a whole scenario
focused on three main targets. First, we are interested in finding the accuracy of our model
in the presence of malicious servers. In other words, we want to know the tolerance limit of
the system in contrast with malicious servers. Since our model has a strong basis on
random or probabilistic decisions, it is considered that it would also be quite interesting to
take care about the accuracy and path length in terms of current and average value for the
entire system. Shorter path is always given due consideration as it consumes fewer
resources.
Table 4.1: Trust and Reputation models parameters
Eigen Trust
Model Parameters
BTRM-WSN
Model
Parameters
LFTM Trust
Model Parameters
Epsilon = 0.1
Pre Trusted Peers
Percentage = 0.3
Pre Trusted
Peers Weight=0.25
Zero Trust Node Selection
Probability = 0.2
Path Length
Factor=0.71
Alpha=1.0
Phi=0.01
Initial
Pheromone=0.85
q0=0.45
Num
Iterations=0.59
Punishment
Threshold=0.48
Rho=0.87
Beta=1.0
Num Ants=0.35
Transition
Threshold=0.66
Phi=0.01
Rho=0.87
q0=0.45
Num Ants=0.35
Num Iterations=0.59
Alpha=1.0
Beta=1.0
Initial Pheromone=0.85
Path Length
Factor=0.71
Transition
Threshold=0.66
Punishment
Threshold=0.48
U_MIN=0.0
U_MAX=1.0
VH_A=0.8
VH_B=0.9
VH_C=1.0
VH_D=1.0
H_A=0.55
H_B=0.7
H_C=0.8
H_D=0.9
M_A=0.3
M_B=0.45
M_C=0.55
M_D=0.7
L_A=0.1
L_B=0.2
L_C=0.3
L_D=0.45
VL_A=0.0
VL_B=0.0
VL_C=0.1
VL_D=0.2
Peer Trust
Model Parameters
Alpha = 1.0
Beta=0.0
Power Trust
Model Parameters
Epsilon=0.0001
Power Nodes
Percentage=0.01
Power Nodes Weight = 0.15
67
In order to measure of the adaptability of our model, we gathered the energy consumption
of all the models in the concerned scenario. The simulations have carried with the
following structure. We deployed our model ten times (i.e. each client applied for a service
ten times) over ten dynamic wireless sensor networks randomly generated, each one
composed of fifty sensors. On each network, the percentage of sensors acting as clients is
always 70%. Rests of the 30% sensors are acting as servers. We simulated with ten random
WSNs having a 10% to 50% of malicious servers with relay value 5 % and a radio range of
15 m. Table 4.2 displays the summary of parameters deployed for malicious server
evaluation in this model
Table 4.2: Scenario parameters
Scenario Options Value
Client
Relay Server
Malicious Server
Radio Range
Delay
Number Execution
Number of Network
WSN Area
Minimum Number of Sensors
Maximum Number of Sensors
WSN Orientation
70 %
5 %
10%, 20%, 30%, 40%, 50%
15 m
0 s
10
10
100 m × 100 m
50
50
Dynamic
4.3.1 Accuracy Concerns
The term accuracy in the context of trust and reputation models may be defined as the
selection percentage of trustworthy nodes. The accuracy evaluated from two viewpoints
namely: current and average. Current accuracy denotes the trustworthiness value calculated
for the last node, whereas average accuracy presents the value of all nodes available in the
said framework. Figure 4.2 depicts the current accuracy versus malicious servers for five
68
different trust and reputation models in the context of dynamic wireless sensor networks.
All the models exhibit decrement in their performance with an increase in malicious servers
percentage except LFTM. For certainty, models show steady state performance of
decrements. Power trust model shows more tolerance with respect to malicious server
ranging from 10 to 50 %. We can rank them in order of their significance in the proposed
scenario as (i) Power trust (ii) BTRM-WSN (iii) Eigen trust (iv) Peer trust (v) LFTM
model.
Figure 4.2: Current accuracy with malicious servers for dynamic wireless sensor networks
According to figure 4.3, approximately same behavior is observed for average accuracy
except LFTM model. LFTM model shows some non linearity in behavior corresponding to
10 15 20 25 30 35 40 45 5030
40
50
60
70
80
90
100
Malicious Servers (%)
Curr
ent
Accura
cy
(%)
BTRM-WSN
EIGEN TRUST
PEER TRUST
POWER TRUST
LFTM
69
malicious servers due to fixed set of outcomes in this model. This shows a good agreement
with the results reported by Marmol and Perez [2011] where the authors emphasized
towards the sensor nodes value in a specific region. We enhanced this contribution to a
certain extent by incorporating malicious servers and five trust models on an individual
wireless sensor network framework. Additionally, we estimated the statistical validity
analytically for current accuracy (CA) and average accuracy (AA) by calculating mean,
Figure 4.3: Average accuracy with malicious servers for dynamic wireless sensor network
standard deviation (SD), variance, population standard deviation (PSD) and variance with
PSD. Table 4.3 depicts the statistical validity analysis of the accuracy parameter of
proposal.
10 15 20 25 30 35 40 45 5020
30
40
50
60
70
80
90
100
Malicious Servers (%)
Avera
ge A
ccura
cy (
%)
BTRM-WSN
EIGEN TRUST
PEER TRUST
POWER TRUST
LFTM
70
Table 4.3: Statistical validity analysis for accuracy (%)
T&R
Models
BTRM-WSN Eigen Trust Peer Trust Power Trust LFTM
Accuracy CA AA CA AA CA AA CA AA CA AA
Mean 94.3 94.53 91.81 92.04 77.79 83.1 97.5 97.7 57.59 55.1
SD 3.9 3.51 6.25 4.67 17.39 10.1 2.2 1.1 23.79 19.5
Variance
(SD)
15.3 12.35 39.18 21.82 302.4 101.8 4.7 1.3 566 381.3
PSD 3.5 3.14 5.59 4.17 15.55 9.1 1.9 1.0 21.28 17.5
Variance
(PSD)
12.3 9.88 31.34 17.46 241.0 81.5 3.8 1.0 452.8 305.1
4.3.2 Path Length Estimations
The path length is defined as the number of resources in a particular network utilizes with a
particular trust and reputation model. Figure 4.4 represented the current path length
Figure 4.4: Current path length with malicious servers for dynamic wireless sensor network
10 15 20 25 30 35 40 45 502.5
3
3.5
4
4.5
5
5.5
6
6.5
7
7.5
Malicious servers (%)
Curr
ent
Path
Length
BTRM-WSN
EIGEN TRUST
PEER TRUST
POWER TRUST
LFTM
71
variation with malicious servers in different trust and reputation models of dynamic WSN.
We evaluated the current and average path length on the similar pattern of accuracy for
both of the trust and reputation models. Current path length depicts the resource utilization
value calculated for the last node whereas average path length exhibits the value of all
nodes present in the scenario. We calculated accuracy and path length both in terms of its
current and average value. Figure 4.4 and figure 4.5 show the current and average path
length utilized by different trust and reputation models in presence of malicious servers.
Marmol et al. [2011] reported only the values of the average path length with a number of
sensors in static wireless sensor networks. We extended the work towards the both average
and current path lengths for the proposed scenarios of trust and reputation models in
Figure 4.5: Average path length with malicious servers dynamic wireless sensor network
10 15 20 25 30 35 40 45 503.5
4
4.5
5
5.5
6
6.5
7
7.5
8
8.5
Malicious Severs (%)
Avera
ge P
ath
Length
BTRM-WSN
EIGEN TRUST
PEER TRUST
POWER TRUST
LFTM
72
presence of malicious servers. We have already presented the scalability impact on the
wireless sensor network in Chapter 2. We enhanced this evaluation by incorporating
malicious servers, resource utilization and energy evaluation aspect on a single platform.
Further, we investigated the statistical validity analytically for current path length (CPL)
and the average path length (APL) by calculating mean, SD, variance, population standard
deviation (PSD) and variance with PSD. Table 4.4 reflects the statistical validity analysis of
path length parameter in our framework.
Table 4.4: Statistical validity analysis for path length
T&R Models BTRM-
WSN
Eigen Trust Peer Trust Power Trust LFTM
Accuracy CPL APL CPL APL CPL APL CPL APL CPL APL
Mean 4.59 4.41 4.71 4.75 4.39 4.80 4.39 4.47 5.29 6.42
SD 1.61 0.60 0.65 0.28 0.44 0.17 0.76 0.28 1.24 1.29
Variance
(SD)
2.59 0.36 0.42 0.08 0.19 0.03 0.58 0.08 1.53 1.68
PSD 1.43 0.54 0.58 0.25 0.39 0.15 0.68 0.25 1.11 1.16
Variance
(PSD)
2.07 0.33 0.33 0.06 0.15 0.02 0.46 0.62 1.22 1.34
4.3.3 Energy Consumption
One of the major issues, when dealing with the wireless sensor network is energy
consumption. Lastly, we emphasized on the average energy consumption by static,
dynamic, oscillatory and combination of dynamic and oscillatory modes of wireless sensor
network. As per the references [Li et al., 2001; Sánchez et al., 2006], power requirements
of a sensor node can be analyzed as a function of distance. For most of the models, energy
consumption E by a message at distance d is given by E (d) = dα + C. Where α represents
73
the attenuation factor and C is constant used for radio signal and dimensionless [Li et al.,
2007; Dorigo et al., 2006]. Table 4.5 compares the five trust and reputation models from
the energy consumption aspect as shown below.
Table 4.5: Energy consumption analysis for malicious servers
Malicious
servers (%)
BTRM-
WSN (μJ)
Eigen Trust
(μJ)
Peer Trust
(μJ)
Power Trust
(μJ)
LFTM
(μJ)
10 8.6 × 10 13
3.1 × 10 18
1.4 × 10 15
1.5 × 10 14
1.9 × 1016
20 1.4× 10 13
3.1 × 10 18
2.3 × 10 15
2.1 × 10 15
3.1× 10 15
30 1.1× 10 14
3.8 × 10 18
1.1 × 10 15
1.1 × 10 15
1.8× 10 15
40 1.6× 10 11
3.4 × 10 18
2.5 × 10 15
1.5 × 10 15
5.2× 10 14
50 6.0× 10 13
1.7 × 10 18
3.2 × 10 15
2.7 × 10 15
8.0× 10 14
We extended this work of reference [Marmol and Perez, 2011] towards the dynamic
wireless sensor networks by assessing different trust and reputation models corresponding
to malicious servers. Our analysis shows that BTRM-WSN model consumes minimum
power as compare to the rest of the models in presence of malicious server percentage
ranging from 10 to 50. We also observed that LFTM model shows decrement in energy
consumption with the increase in malicious servers due to certain nodes become halted in
their behavior with the increase in malicious servers. This indicates the significant
improvement over the results as reported in the reference [Marmol and Perez, 2011].
Moreover, we investigated the confidence interval of the entire framework with the
confidence level in contrast with malicious servers. We found that the confidence interval
increases with the increase in malicious server’s percentage in our proposal as highlighted
in figure 4.6. Overall, we have investigated the entire framework hundred times for
different sensor values with respect to different trust and reputation models. We have
74
observed that more complexity in any trust and reputations model attract more resources
utilization and power consumption.
Figure 4.6: Confidence Interval analysis
We added a variety of evaluation strategies build on accuracy, path length, satisfaction and
power consumption for sensor node operations in our proposed framework which makes
over scenario more robust as compared to the approach reported in reference [Pan et al.,
2013]. We extended the concept by adding more robust constraints like malicious servers,
path length, energy consumption and a combination of all these aspects on a single platform
for the trust and reputation models realization in wireless sensor networks. Our analysis
shows that there remains always significant impact of malicious sensors over wireless
sensor networks, resulting in the overall system performance degradation. Next, Section
present the more deeper evaluation of different WSN modes in trust and reputation models.
0 5 10 15 20 25 30 35 40 45 500
5
10
15
20
25
30
35
40
45
50
Malacious Servers (%)
Confidence I
nte
rval
Confidence level (95%)
Confidence level (99%)
75
4.4 Evaluation of Static, Dynamic and Oscillatory Wireless Sensor
Networks
We launched our model 100 times (i.e. each client applied for a service 100 times) over 100
static, dynamic, oscillating and combination of dynamic and oscillatory WSNs randomly
generated, each one composed of fifty sensors. We have also considered a special case
integrating both the dynamic and oscillating networks. On each network, the percentage of
sensors acting as clients is always a 75%. Therefore, the 25% left are sensors acting as
servers. We tried over 100 random WSNs for static, dynamic, oscillatory and combination
of dynamic and oscillatory WSN modes with five trust and reputation models.
Figure 4.7: Comprehensive scenario setup
76
Additionally, the sensor belonging to our developed networks are spread over the area of
100 m × 100 m with relay value 5% and radio range 15 m as shown in figure 4.7.
4.4.1 Accuracy Analysis
In our evaluation, first we have judged accuracy from two viewpoints namely: (i) current
accuracy (ii) average accuracy as depicted in figure 4.8 and figure 4.9. The outcomes of our
experiment for five different trust and reputation models, in context of different WSN
modes are shown in figure 4.8.
Figure 4.8: Graph of current accuracy versus WSN modes
We assigned static mode value of 1, dynamic mode value of 2, oscillatory mode value of 3
and combination of dynamic and oscillatory mode value of 4 in our simulation. All the
models show decrements in their performance as we switch from WSN mode value 1 to 4.
1 1.5 2 2.5 3 3.5 40
10
20
30
40
50
60
70
80
90
100
WSN Modes
1
Curr
ent
Accura
cy %
BTRM-WSN
EIGEN TRUST
PEER TRUST
POWER TRUST
LFTM
77
Even though upto certain extent power trust model shows the steady behaviors. We can
rank them in order of their significance in the proposed scenario as (i) Power Trust (ii)
BTRM-WSN (iii) Eigen Trust (iv) Peer Trust (v) LFTM model. As far as the accuracy of
LFTM model is concerned, it deviates itself in both the cases of current and average
accuracy from the rest of models, because of the additional computation involved to
calculate five levels of satisfaction (i) very low (ii) low (iii) medium (iv) high (v) very high.
In our scenario, power trust model outperforms rest of the models in static, dynamic and
oscillatory WSNs mode whereas peer trust seems to be best for the combined measure of
dynamic and oscillatory mode. Approximately, similar behaviors are shown in figure 4.9
Figure 4.9: Graph of average accuracy versus WSN modes for different trust and reputation
models
1 1.5 2 2.5 3 3.5 410
20
30
40
50
60
70
80
90
100
WSN Modes
Avera
ge A
ccura
cy (
%)
WSN Modes
1 = Static, 2 = Dynamic, 3 = Oscillatory, 4 = Dynamic & Oscullatory
BTRM-WSN
EIGEN TRUST
PEER TRUST
POWER TRUST
LFTM
78
for all trust and reputation models for average accuracy except in the LFTM model. It
shows some non linearity in behaviour corresponds to WSN modes. Marmol et al. [2008]
reported some similar scenario for static WSNs and emphasis on the value of the number of
sensor in a specific region. We enhanced the contributions to certain extent by
incorporating different WSN modes with respect to five trust and reputation models. Also,
it can be checked from figure 4.8 and figure 4.9 that the current and average accuracy
obtained in static WSN mode remains always greater than rest of other modes. This is due
to the static state of the WSN nodes where the positions of all the nodes are fixed and
predetermined. On the contrary side, the positions of all the nodes in rest of WSN modes
are variable. This shows a good agreement with the results reported in reference [Marmol et
al., 2008].
4.4.2 Path Length Analysis
Marmol et al. [2008] reported the values of the average path length for static WSNs with
specific number of sensors in the BTRM-WSN model. We extended this work for both of
the average and current path lengths in our proposed scenarios with five mentioned trust
and reputation models. In the current path length analysis, BTRM-WSN consumes fewer
resources in static and oscillatory WSNs mode and peer trust model consume fewer
resources in dynamic and combined resultant of dynamic and oscillatory WSNs mode as
depicted in figure 4.10.
79
Figure 4.10: Graph of current path length versus WSN modes for different trust and
reputation models
Whereas for average path length analysis, again BTRM-WSN consumes fewer resources in
static and oscillatory WSNs mode and power trust model consume less resources in
dynamic and combined resultant of dynamic and oscillatory WSNs mode as shown in
figure 4.11.
4.4.3 Energy Consumption Analysis
Table 4.6 compares the five trust and reputation models from energy consumption aspect.
A comparative analysis about the energy consumption with respect to sensors value
increment was reported in reference [Marmol et al., 2008].
1 1.5 2 2.5 3 3.5 42
4
6
8
10
12
14
WSN Modes
Curr
ent
Path
Length
BTRM-WSN
EIGEN TRUST
PEER TRUST
POWER TRUST
LFTM
80
Figure 4.11: Graph of average path length versus WSN modes for different trust and
reputation models
In our proposal, we extended this concept towards the different WSN modes by correlating
these modes with different trust and reputation models. Analysis shows that BTRM-WSN
model consumes minimum power in case of static WSN, power trust in case of dynamic
WSN, peer trust in case both oscillatory WSN as well as a combination of dynamic and
oscillatory WSN. As far as the maximum power consumption is concerned, Eigen trust
consumes maximum power in all the cases. Next section highlights the concept of collusion
with wireless sensor networks.
1 1.5 2 2.5 3 3.5 43
4
5
6
7
8
9
10
11
WSN Modes
Avera
ge P
ath
Length
BTRM-WSN
EIGEN TRUST
PEER TRUST
POWER TRUST
LFTM
81
Table 4.6: Energy consumption (μJ) analysis for different WSN modes
T&R
Models
Static
WSN
Dynamic
WSN
Oscillatory
WSN
Dynamic & Oscillatory
WSN
BTRM 1.1 × 1015
1.4 × 1014
1.7 × 10 15
2.2 × 10 15
Eigen Trust 1.7 × 1020
3.3 × 1018
1.9 × 1020
4 × 1018
Peer Trust 1.4 × 10 15
4.8 × 1014
1.2 × 10 15
2.6 × 1014
Power Trust 1.2 × 1018
4.1 × 1013
1.2 × 1018
4.1 × 1014
LFTM 1.2 × 1017
2.2 × 1016
1.1 × 1018
2.2 × 1016
4.5 Collusion Based Realizations of WSN
We designed a wireless sensor network template using the following parameters. The 20%
of all nodes in a randomly created WSN acted as clients and the rest of 80% nodes acted as
servers as shown in figure 4.12.
Figure 4.12: Collusion based scenario
82
The 5% of the nodes acted as relay servers not offered any services and acted as relay
nodes. The radio range of the nodes set at ten hops to its neighbours. We consider a
scenario where the percentage of fraudulent servers remained 70% which specifies the
indispensable condition for our WSN framework evaluation. We set the minimum and
maximum number of nodes than can create a WSN equal to 200. Sensor nodes belonging to
our developed networks spread over the area of 100 m × 100 m. A total of ten networks are
examined and the final results reflect the average value of all the networks. The process of
searching trustworthy server is carried out ten times for each network. Table 4.7 shows the
summary of parameters deployed in our model.
Table 4.7: Scenario parameters
Scenario Options Value
Client
Relay Server
Fraudulent Server
Radio Range
Delay
Number Execution
Number of Network
WSN Area
Minimum Number of Nodes
Maximum Number of Nodes
20 %
5 %
70 %
10 m
0
10
10
100 m × 100 m
200
200
We investigated the comparative analysis of trust and reputation models with static WSN
and dynamic in contrast with and without collusion parameter. We considered four WSN
modes namely (i) static WSN (SW) (ii) static WSN with collusion (SWC) (iii) dynamic
WSN (DW) (iv) dynamic WSN with collusion (DWC). We denote Eigen trust model with
value 1, peer trust model with 2, BTRM-WSN model with value 3, LFTM with value 4 and
83
a trust and reputation infrastructure based proposal (TRIP) model with value 5. The TRIP
model makes separation between selfish nodes spreading wrong information and honest
nodes on the basis of infrastructure [Marmol et al., 2012]. The outcome of the simulations
are described in the following subsections.
4.5.1 Accuracy Realization
The term accuracy in the trust and reputation system may be defined as the selected
percentage of trustworthy nodes. Initially, we calculated average accuracy corresponds to
different trust and reputation models as shown in figure 4.13.
Figure 4.13: Current accuracy of different WSN modes with trust and reputation models
The value of current accuracy remains highest in case of static WSN as compared to rest of
the WSN modes due to the fact that static nodes are less prone to failure than the dynamic
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
10
20
30
40
50
60
70
80
90
100
Eigen Trust Peer Trust BTRM WSN LFTM TRIP
Curr
ent
Accura
cy (
%)
SW
SWC
DW
DWC
84
as well as the combination of static and dynamic WSN with collusion aspect. Next, we
considered the second evaluation for average accuracy with the same WSN framework.
According to figure 4.14, again average accuracy shows the similar behavior with the
current accuracy in the above figure 4.13 as the value of average accuracy remains highest
in case of static WSN than the rest of the WSN modes. For static WSN (SW) and dynamic
WSN (DW) mode, the value of current and average accuracy remains highest in LFTM
model than other models in most fraudulent conditions whereas TRIP model depicts the
minimum value for SW and DW.
Figure 4.14: Average accuracy of different WSN modes with trust and reputation models
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
10
20
30
40
50
60
70
80
90
100
Eigen Trust Peer Trust BTRM WSN LFTM TRIP
Avera
ge A
ccura
cy (
%)
SW
SWC
DW
DWC
85
In case of SWC and DWC mode, the Eigen trust model outperforms rest of the models in
current and average accuracy value whereas peer trust model shows minimum accuracy
value.
4.5.2 Path Length Realization
The path length is defined as the number of resources a particular network utilizes with a
specific trust and reputation model. In the consistent pattern of accuracy evaluation types,
we evaluated the current and average path length on the similar pattern of accuracy for all
the WSN modes. Figure 4.15 and figure 4.16 represent that the value of current and average
path length which remains quiet in case of TRIP model for both the current and average
case viewpoints than other models.
Figure 4.15: Path length of different WSN modes with trust and reputation models
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
1
2
3
4
5
6
7
Eigen Trust Peer Trust BTRM WSN LFTM TRIP
Curr
ent
Path
Length
(%
)
SW
SWC
DW
DWC
86
This is due to the fact that the TRIP model constitutes the fixed infrastructure for its
functionality resulting in lesser path length as compared to other models. Among the rest of
the models, LFTM model consumes lesser path length than the rest of the models in the
case of static WSN (SW) mode and dynamic WSN (DW) mode. For the static WSN with
collusion (SWC) and dynamic WSN with collusion (DWC) mode, BTRM model utilizes
the minimum path length. We also observed that BTRM model utilizes the maximum path
length of all the SW, SWC, DW and DWC WSN modes. This shows the good agreement
with the results reported in reference [Alkalbani et al., 2013]. An initiative towards the
description of energy consumption analysis for different trust and reputation models was
proposed in reference [Alkalbani et al., 2013].
Figure 4.16: Average path length of different WSN modes with trust and reputation models
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
1
2
3
4
5
6
7
Eigen Trust Peer Trust BTRM WSN LFTM TRIP
Avera
ge P
ath
Length
(%
)
SW
SWC
DW
DWC
87
We enhanced this evolution towards a more intricate assessment by incorporating
collusion, satisfaction and energy evaluation aspect in our scenario. We proposed a more
robust framework subsuming different WSN versus collusion scalability within an
individual framework. Xiong et al. [2004] reported peer to peer trust and reputation based
model for structured peer to peer networks including strategies for its implementation and
evaluation in decentralized environmental conditions. Also, Xiong et al. [2004] emphasized
over trust metric in order to assess trustworthiness, feedback and credibility of peer to peer
networks. Chen et al. [2011] proposed parameter estimation based trust model for
unstructured peer to peer networks. We enhanced the contribution to a certain extent by
incorporating collusion, satisfaction and energy consumption parameters for wireless
sensor network. This makes our investigation more robust and real time.
4.5.3 Satisfaction Realization
We calculated the satisfaction level of different WSN modes for LFTM model as shown in
figure 4.17. In the context of trust and reputation model, satisfaction can be defined as a
particular level of subjectivity upto specific degree in which the system can behave as per
desired goal for mentioned probability. The observation shows that in the static mode
satisfaction level is very high as compared to the rest of WSN modes. On the other side,
incorporation of collusion to SW and DW mode decreases its satisfaction value.
88
Figure 4.17: Satisfaction analysis of different WSN modes with LFTM trust and reputation
model
4.5.4 Energy Concerns with WSN
One of the major issues, when dealing with the wireless sensor network is energy
consumption. So, lastly we emphasized on the average energy consumption by SW, SWC,
DW and DWC modes of five trust and reputation models in wireless sensor networks. The
power requirement of a sensor node can be analyzed as a function of distance as per the
references [Li et al., 2001; Sanchez et al., 2006]. For most of the models, energy
consumption E by a message at a distance d is given by [Li et al., 2007; Dorigo et al.,
2006]
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
10
20
30
40
50
60
70
80
90
100
Very High High Medium Low Very Low
Satisfa
ction (
%)
SW
SWC
DW
DWC
89
Where represents attenuation factor and C denotes constant for radio signal and
dimensionless. Table 4.8 compares the five trust and reputation models from the energy
consumption aspect as shown below.
Table 4.8: Energy consumption of trust and reputations models with WSN modes
Trust and Reputation
Models
with WSN modes
Eigen Trust
(mJ)
Peer Trust
(mJ)
BTRM
(mJ)
LFTM
(mJ)
TRIP
(mJ)
Static WSN
(SW) 8.2 × 10
24 8.7 × 10
16 4 × 10
17 2.5 × 10
17 4 × 10
9
Static WSN with
Collusion (SWC) 6.4 × 10
24 1.0 × 10
17 8 × 10
17 1.4 × 10
18 4 × 10
9
Dynamic WSN
(DW) 5.1 × 10
24 3.2 × 10
16 1.1 × 10
17 1.2 × 10
17 4 × 10
9
Dynamic WSN with
Collusion (DWC) 8.6 × 10
24 2.2 × 10
16 6.8 × 10
17 7.6 × 10
17 4 × 10
9
Marmol et al. [2008] reported a comparative analysis of the energy consumption with
respect to sensors value increment. In our proposal, we extended this concept towards the
different WSN modes and simultaneously clubbing these modes with different trust and
reputation models. We observed that the Eigen trust model consume maximum power in all
the SW, SWC, DW and DWC modes whereas TRIP model reported minimum energy
consumption. We observed that more complexity involvement in the Eigen trust model is
the reason for utmost energy consumption and in case of TRIP model energy consumption
is minimum because of the simpler computation involved in trust value computation. We
also extended the mathematical relation as reported in references [Li et al., 2001; Sánchez
et al., 2006] for the energy consumption for trust and reputation models in our framework.
90
Where represents for overall energy consumption, denotes client nodes energy
consumption, depicts server nodes energy consumption, shows energy consumption
for fraudulent node, for relay node energy consumption and denotes energy
consumption used by the simulator. Overall, we investigated the entire framework twenty
times for different WSN modes with five trust and reputation models. One common thing is
observed that more complexity in any trust and reputations model attracts more resources
utilization and power consumption. We added a variety of evaluation strategies based on
accuracy, path length, satisfaction and power consumption for sensor node operations in
our proposed framework which makes over scenario more robust as compared to the
approach reported in reference [Pan et al., 2013]. A new trust and reputation model by
adding additional constraints to BTRM-WSN adopting an interactive multiple ant colony
algorithm was also suggested by Pan et al. [2013]. We extended the concept by adding
more robust constraints like static, dynamic, collusive and a combination of all these
aspects on a single platform for the trust and reputation models investigations in wireless
sensor networks. We enhanced the work of reference [Qureshi et al., 2010] for wireless
sensor networks with collusion and satisfactory aspect evaluation with five trust and
reputation models over wireless sensor networks. Qureshi et al. [2010] presented FIRE
trust and reputation model extension to detect and prevent direct interaction and validate
interaction collusion attacks in wireless networks. Our analysis shows that there remains
always significant impact of collusion over static and dynamic mode of WSN, resulting in
the performance degradation of the overall system.
91
4.6 Summary
This Chapter concluded the impact of malicious sensors on the BTRM-WSN, Eigen trust,
peer trust, power trust and LFTM trust and reputation models in wireless sensor networks.
It is evident from the simulation that there is a strong relationship in between malicious
server’s percentage and resource utilization in trust and reputation model evaluation. We
have estimated accuracy and path length in terms of overall percentage of the functionality
and energy consumption in terms of millijoule specifically for sensor node operations.
Moreover, we evaluated our framework with simulation based experiments for static,
dynamic, oscillatory and a combination of dynamic and oscillatory wireless sensor
networks. After surveying the current state of art in these models, a number of aspects like
accuracy, path length and energy consumption have been discussed and analyzed.
Additionally, collusion based realization over wireless sensor networks are also calculated
on the basis of performance matrices in this Chapter. We focus on three major directions.
Firstly, we evaluated the accuracy, path length, satisfaction and energy consumption for
collusive and non-collusive modes of wireless sensor networks. Secondly, we investigated
the entire framework for comparative evaluation of above discussed trust and reputation
models and lastly the same model is deployed for the mathematical derivation of the energy
equation of a wireless sensor network. Moreover, we proposed a mathematical equation for
overall energy consumption which adds more robustness in our evaluation. We observed
that with the collusion adoption in the WSN modes leads to major performance
degradation. In case of static nodes, collusion affects less to WSN, when it is incorporated
in dynamic mode. Also, node operations remain more in case of collusion than without it.
We can predict that the lesser the collusive nodes, the more the probability of accuracy, the
92
better resource utilization, the adequate satisfaction level and the lesser the energy
consumption of the entire WSN exhibited by the wireless sensor network system.