ThesisPDF Available

Towards Attack Resilience of Scale-Free IoT Networks with Topology Modifications (MS Thesis without Source Codes)

Authors:

Abstract and Figures

Nowadays, the Internet of Things (IoT) provides benefits to humans in numerous domains by empowering the projects of smart cities, healthcare, industrial enhancement and so forth. The IoT networks include nodes, which deliver the data towards their destination. However, the removal of nodes due to malicious attacks affects the connectivity of the nodes in the networks. The ideal plan is to construct a topology, which maintains the nodes' connectivity after the attacks and subsequently increases the network robustness. Therefore, in this thesis, werst adopt two different mechanisms for the construction of a robust scale-free network. Initially, a Multi-Population Genetic Algorithm (MPGA) is used to overcome the premature convergence in GA. Then, an entropy based mechanism is used, which replaces the first solution of high entropy population with the best solution of low entropy population to improve the network robustness. Second, two types of edge swap mechanisms are introduced. The Efficiency based Edge Swap Mechanism (EESM) selects the pair of edges with high efficiency to increase the network robustness. The second edge swap mechanism named EESM-Assortativity transforms the network topology into an onion-like structure to achieve maximum connectivity between similar degree nodes in the network. The optimization of the network robustness is performed using Hill Climbing (HC) and Simulated Annealing (SA) methods. The simulation results show that the proposed MPGA Entropy has 9% better network robustness as compared to MPGA. Moreover, the proposed ESMs effectively increase the network robustness with an average of 15% better robustness as compared to HC and SA. Furthermore, they also increase the graph density as well as network's connectivity with high computational cost. Furthermore, we design a robust network to support the nodes' functionality for the topology optimization in the scale-free IoT networks. It is because the computational complexity of an optimization process increases the cost of the network. Therefore, in this thesis, the main objective is to reduce the computational cost of the network with the aim of constructing a robust network topology. Thus, four solutions are presented to reduce the computational cost of the network. First, a Smart Edge Swap Mechanism (SESM) is proposed to overcome the excessive randomness of the standard Random Edge Swap Mechanism (RESM). Second, a threshold based node removal method is introduced to reduce the operation of the edge swap mechanism when an objective function converges at a point. Third, multiple attacks are performed in the network to find the correlation among the measures, which are degree, betweenness and closeness centralities. Fourth, based on the third solution, the Heat Map Centrality (HMC) is introduced that finds the set of most important nodes from the network. The HMC damages the network by utilizing the information of two positively correlated measures. It helps to provide a good attack strategy for robust optimization. The simulation results demonstrate the efficacy of the proposed SESM mechanism. It outperforms the existing RESM mechanism by almost 4% better network robustness and 10% less number of swaps. Moreover, 64% removal of nodes helps to reduce the computational cost of the network. In addition, we also perform topology optimization using a new heuristic algorithm, named as Great Deluge Algorithm (GDA). Afterwards, four rewiring strategies are designed. The first strategy is based on the degree dissortativity, which performs rewiring if maximum connectivity among similar degree nodes is achieved. In second strategy, we propose a degree difference operation using degree dissortativity to make sure that the connected edges possess low dissortativity and degree difference. Whereas the other two strategies consider nodes' load capacity as well as improved GDA to maximize the network robustness. The effectiveness of the proposed rewiring strategies is evaluated through simulations. The results prove that the proposed strategies increase the network robustness up to 25% as compared to HC and SA algorithms. Besides, the strategies are also very effective in increasing the graph density and network connectivity. However, their computational time is high as compared to HC and SA.
Content may be subject to copyright.
Towards Attack Resilience of Scale-Free IoT
Networks with Topology Modifications (MS
Thesis without Source Codes)
By
Muhammad Awais Khan
CIIT/FA18-REE-009/ISB
MS Thesis
In
Electrical Engineering
COMSATS University Islamabad, Islamabad - Pakistan
Spring, 2021
i
COMSATS University Islamabad
Towards Attack Resilience of Scale-Free IoT
Networks with Topology Modifications (MS
Thesis without Source Codes)
A Thesis Presented to
COMSATS University Islamabad
In partial fulfillment
of the requirement for the degree of
MS (Electrical Engineering)
By
Muhammad Awais Khan
CIIT/FA18-REE-009/ISB
Spring, 2021
ii
Towards Attack Resilience of Scale-Free IoT
Networks with Topology Modifications (MS
Thesis without Source Codes)
A Post Graduate Thesis submitted to the Department of Electrical and Computer
Engineering as partial fulfilment of the requirement for the award of Degree of MS
(Electrical Engineering).
Name Registration Number
Muhammad Awais Khan CIIT/FA18-REE-009/ISB
Supervisor:
Dr Nadeem Javaid,
Associate Professor, Department of Computer Science,
COMSATS University Islamabad,
Islamabad, Pakistan
Co-Supervisor:
Dr Sardar Muhammad Gulfam
Assistant Professor, Department of Electrical and Computer Engineering,
COMSATS University Islamabad,
Islamabad, Pakistan
iii
Final Approval
This thesis titled
Towards Attack Resilience of Scale-Free IoT Networks with
Topology Modifications (MS Thesis without Source Codes)
By
Muhammad Awais Khan
CIIT/FA18-REE-009/ISB
has been approved
For the COMSATS University Islamabad, Islamabad
External Examiner:
Dr. Ataul Aziz Ikram
Professor, Department of Electrical Engineering,
FAST-NU, Islamabad, Pakistan
Supervisor:
Dr Nadeem Javaid
Associate Professor, Department of Computer Science,
COMSATS University Islamabad, Islamabad, Pakistan
Co-Supervisor:
Dr Sardar Muhammad Gulfam
Assistant Professor, Department of Electrical and Computer Engineering,
COMSATS University Islamabad, Islamabad, Pakistan
Head of Department:
Dr. Shurjeel Wyne
Associate Professor, Department of Electrical and Computer Engineering,
COMSATS University Islamabad, Islamabad, Pakistan
iv
Declaration
IMuhammad Awais Khan (Registration No. CIIT/FA18-REE-009/ISB) hereby de-
clare that I have produced the work presented in this thesis, during the scheduled
period of study. I also declare that I have not taken any material from any source
except referred to wherever due that amount of plagiarism is within acceptable
range. If a violation of HEC rules on research has occurred in this thesis, I shall
be liable to punishable action under the plagiarism rules of the HEC.
Date: 27 July, 2021
Muhammad Awais Khan
CIIT/FA18-REE-009/ISB
v
Certificate
It is certified that Muhammad Awais Khan (Registration No. CIIT/FA18-REE-009/ISB)
has carried out all the work related to this thesis under my supervision at the
Department of Electrical and Computer Engineering, COMSATS University, Is-
lamabad and the work fulfils the requirement for award of MS degree.
Date: 27 July, 2021
Supervisor:
Dr Nadeem Javaid
Associate Professor, Department of
Computer Science.
Co-Supervisor:
Dr Sardar Muhammad Gulfam
Assistant Professor, Department of
Electrical and Computer Engineering.
Head of Department:
Dr. Shurjeel Wyne
Associate Professor, Department of Electrical
and Computer Engineering
vi
ACKNOWLEDGEMENT
First of all, thanks to Almighty Allah who gave me the strength to complete this
thesis. After that, I will thank my honourable supervisor Dr Nadeem Javaid,
co-supervisor Dr. Sardar Muhammad Gulfam and my parents because without
their support I will not be able to complete my thesis with dignity and respect.
My supervisor Dr Nadeem Javaid helped me in every cause. Whenever I feel
unmotivated or depressed, he keeps pushing me. He gives me the time whenever
I need it the most and I did not see any supervisor giving that much extra time
to his students. Lastly, I am greatly thankful to my ComSens Lab colleagues for
providing me with a warm and friendly atmosphere.
vii
Conference Proceedings
1 Khan, M.A., Javaid, N., Majid, A., Imran, M. and Alnuem, M., 2016, March.
Dual sink efficient balanced energy technique for underwater acoustic sensor
networks. In 2016 30th International Conference on Advanced Information
Networking and Applications Workshops (WAINA) (pp. 551-556). IEEE.
Download
2 Zain-ul-Abidin, M., Khan, M.A., Javaid, N., Khizar, M., Khan, Z.A. and
Qasim, U., 2016, March. Enhanced single chain-based scheme in cylindri-
cal underwater wireless sensor networks. In 2016 30th International Con-
ference on Advanced Information Networking and Applications Workshops
(WAINA) (pp. 343-348). IEEE. Download
3 Khan, M.A., Sher, A., Hameed, A.R., Jan, N., Abassi, J.S. and Javaid, N.,
2016, November. Network lifetime maximization via energy hole allevia-
tion in wireless sensor networks. In International Conference on Broadband
and Wireless Computing, Communication and Applications (pp. 279-290).
Springer, Cham. Download
4 Khan, M.A., Javaid, N., Wadud, Z., Gull, S., Imran, M. and Nasr, K., 2017,
June. Towards energy balancing in heterogeneous wireless sensor networks.
In 2017 13th International Wireless Communications and Mobile Computing
Conference (IWCMC) (pp. 786-791). IEEE. Download
5 Khan, M.A., Javaid, N., Javaid, S., Khalid, A., Nasser, N. and Imran, M.,
2020, June. A novel cooperative link selection mechanism for enhancing
the robustness in scale-free IoT networks. In 2020 International Wireless
Communications and Mobile Computing (IWCMC) (pp. 2222-2227). IEEE.
Download
viii
Journal Publications
1 Wadud, Z., Javaid, N., Khan, M.A., Alrajeh, N., Alabed, M.S. and Guizani,
N., 2017. Lifetime maximization via hole alleviation in iot enabling hetero-
geneous wireless sensor networks. Sensors, 17(7), p.1677. Download
ix
ABSTRACT
Towards Attack Resilience of Scale-Free IoT Networks with
Topology Modifications (MS Thesis without Source Codes)
Nowadays, the Internet of Things (IoT) provides benefits to humans in numerous
domains by empowering the projects of smart cities, healthcare, industrial en-
hancement and so forth. The IoT networks include nodes, which deliver the data
towards their destination. However, the removal of nodes due to the malicious
attacks effects the connectivity of the nodes in the networks. The ideal plan is to
construct a topology, which maintains the nodes’ connectivity after the attacks and
subsequently increases the network robustness. Therefore, in this thesis, we first
adopt two different mechanisms for the construction of a robust scale-free network.
Initially, a Multi-Population Genetic Algorithm (MPGA) is used to overcome the
premature convergence in GA. Then, an entropy based mechanism is used, which
replaces the worst solution of high entropy population with the best solution of
low entropy population to improve the network robustness. Second, two types of
edge swap mechanisms are introduced. The Efficiency based Edge Swap Mech-
anism (EESM) selects the pair of edges with high efficiency to increase the net-
work robustness. The second edge swap mechanism named EESM-Assortativity
transforms the network topology into an onion-like structure to achieve maximum
connectivity between similar degree nodes in the network. The optimization of
the network robustness is performed using Hill Climbing (HC) and Simulated An-
nealing (SA) methods. The simulation results show that the proposed MPGA
Entropy has 9% better network robustness as compared to MPGA. Moreover, the
proposed ESMs effectively increase the network robustness with an average of 15%
better robustness as compared to HC and SA. Furthermore, they also increase the
graph density as well as network’s connectivity with high computational cost. Fur-
thermore, we design a robust network to support the nodes’ functionality for the
topology optimization in the scale-free IoT networks. It is because the compu-
tational complexity of an optimization process increases the cost of the network.
Therefore, in this thesis, the main objective is to reduce the computational cost of
the network with the aim of constructing a robust network topology. Thus, four
solutions are presented to reduce the computational cost of the network. First,
a Smart Edge Swap Mechanism (SESM) is proposed to overcome the excessive
randomness of the standard Random Edge Swap Mechanism (RESM). Second, a
threshold based node removal method is introduced to reduce the operation of the
edge swap mechanism when an objective function converges at a point. Third,
multiple attacks are performed in the network to find the correlation among the
x
measures, which are degree, betweenness and closeness centralities. Fourth, based
on the third solution, the Heat Map Centrality (HMC) is introduced that finds the
set of most important nodes from the network. The HMC damages the network by
utilizing the information of two positively correlated measures. It helps to provide
a good attack strategy for robust optimization. The simulation results demon-
strate the efficacy of the proposed SESM mechanism. It outperforms the existing
RESM mechanism by almost 4% better network robustness and 10% less number
of swaps. Moreover, 64% removal of nodes helps to reduce the computational cost
of the network. In addition, we also perform topology optimization using a new
heuristic algorithm, named as Great Deluge Algorithm (GDA). Afterwards, four
rewiring strategies are designed. The first strategy is based on the degree dissor-
tativity, which performs rewiring if maximum connectivity among similar degree
nodes is achieved. In second strategy, we propose a degree difference operation
using degree dissortativity to make sure that the connected edges possess low dis-
sortativity and degree difference. Whereas, the other two strategies consider nodes’
load capacity as well as improved GDA to maximize the network robustness. The
effectiveness of the proposed rewiring strategies is evaluated through simulations.
The results prove that the proposed strategies increase the network robustness up
to 25% as compared to HC and SA algorithms. Besides, the strategies are also
very effective in increasing the graph density and network connectivity. However,
their computational time is high as compared to HC and SA.
xi
TABLE OF CONTENTS
Acknowledgements vii
Conference Proceedings 99
Journal Publications 100
Abstract x
List of Figures xv
List of Tables xvii
1 Introduction 3
1.1 Introduction ............................... 4
1.1.1 Research Background ...................... 4
1.1.2 Problem Statement ....................... 5
1.1.3 Thesis Contributions ...................... 7
1.1.4 Organization of thesis ..................... 8
2 Literature Review 9
2.1 Literature Review ............................ 10
3 Modification Strategies 19
3.1 Summary of the Chapter ........................ 20
3.1.1 Construction of a Scale-Free Network ............. 21
3.1.2 Attack Model .......................... 21
3.2 Multi-Population Entropy based Mechanism ............. 22
3.3 Proposed Edge Swap Mechanisms ................... 25
3.3.1 Efficiency based Edge Swap Mechanism ............ 26
3.3.2 Efficiency based Edge Swap Mechanism with Assortativity . 28
3.4 Performance Evaluation ........................ 29
3.4.1 Comparison of Network Robustness of Scale-Free Topologies
under Malicious Attacks .................... 30
3.4.2 Comparison of Graph Density of Scale-Free Topologies un-
der Malicious Attacks ...................... 33
3.4.3 Comparison of Robustness of Scale-Free Topologies with dif-
ferent Network Parameters under Malicious Attacks ..... 35
3.4.4 Comparison of Connectivity of Initial Topology and SAEA
under Malicious and Random Attacks ............. 36
3.4.5 Comparison of Computational Time of Scale-Free Topolo-
gies under Malicious Attacks .................. 36
3.5 Conclusion of the Chapter ....................... 38
4 Computationally Efficient Topology Optimization 39
xii
4.1 Summary of the Chapter ........................ 40
4.2 Scale-Free Network Modeling ..................... 40
4.2.1 Construction of a Scale-Free Network ............. 40
4.2.2 Network Robustness Measure ................. 41
4.2.3 Network Optimization through Selection of Independent Edges 42
4.3 Computationally Efficient Topology Optimization: Overview . . . . 43
4.3.1 Smart Edge Swap Mechanism ................. 44
4.3.2 Threshold based Node Removal ................ 46
4.3.3 Optimization of Network Considering Multiple Attacks . . . 47
4.3.4 Heat Map Centrality ...................... 48
4.4 Simulation Results and Discussion ................... 51
4.4.1 Power Law Distribution .................... 51
4.4.2 Measuring the Extent of Damage Caused by Each Attack . . 52
4.4.3 Execution Time of Different Attacks ............. 53
4.4.4 Convergence of Robustness using Degree based Node Removal 54
4.4.5 Initial Robustness Evaluation considering Multiple Attacks . 55
4.4.6 Robustness Analysis using High Degree Node Attack and
HMC Attack .......................... 56
4.4.7 Swap Cost using High Degree Node Attack and HMC Attack 59
4.4.8 Topology Comparison ..................... 61
4.5 Conclusion of the Chapter ....................... 62
5 Topology Rewiring Strategies 63
5.1 Summary of the Chapter ........................ 64
5.2 Modeling the Topology of the Scale-Free Network .......... 64
5.2.1 Construction of Scale-Free Network Topology ........ 64
5.2.2 Attack Types and Robustness Metric ............. 65
5.3 Topology Optimization ......................... 66
5.3.1 Optimization Algorithms .................... 66
5.3.2 Edge Rewiring Strategies .................... 67
5.3.2.1 Rewiring Strategy using Degree Dissortativity . . . 67
5.3.3 Rewiring Strategy using Degree Difference Operation with
Dissortativity .......................... 69
5.3.4 Rewiring Strategy using Nodes’ Load Capacity ........ 71
5.3.5 Optimized Edge Rewiring Strategy using Nodes’ Load Ca-
pacity .............................. 73
5.4 Performance Evaluation ........................ 75
5.4.1 Comparison of Robustness under Malicious Attacks ..... 75
5.4.2 Comparison of Graph Density under Malicious Attacks . . . 77
5.4.3 Comparison of Connectivity under Malicious and Random
Attacks ............................. 78
5.4.4 Comparison of Computational Time of Different Topologies
under Malicious Attacks .................... 80
5.4.5 Comparison of ROSE-DDO and RS-DDOD ......... 80
5.5 Conclusion of the Chapter ....................... 81
6 Conclusion and Future Work 82
xiii
6.1 Conclusions ............................... 83
6.2 Future Work ............................... 84
7 References 85
Conference Proceedings 99
Journal Publications 100
xiv
LIST OF FIGURES
3.1 Nodes Joining Mechanism. ....................... 20
3.2 Steps Involved in Modeling the Entropy based GA. ......... 23
3.3 Comparison of Network Robustness of MPGA Entropy and MPGA
with Number of Iterations. ....................... 30
3.4 Comparison of Network Robustness of MPGA Entropy and MPGA
with Number of Nodes. ......................... 31
3.5 Comparison of Network Robustness of Scale-Free Topologies under
Malicious Attacks with respect to Number of Iterations. ....... 32
3.6 Comparison of Network Robustness of Scale-Free Topologies under
Malicious Attacks with respect to Number of Nodes. ........ 33
3.7 Comparison of Graph Density of Scale-Free Topologies under Ma-
licious Attacks. ............................. 34
3.8 Comparison of Robustness of Scale-Free Topologies with different
Network Parameters under Malicious Attacks. ............ 35
3.9 Comparison of Connectivity of Initial Topology and SAEA under
Malicious Attacks. ........................... 36
3.10 Comparison of Connectivity of Initial Topology and SAEA under
Random Attacks. ............................ 37
3.11 Comparison of Computational Time of Scale-Free Topologies under
Malicious Attacks. ........................... 37
4.1 (a) Initial Topology (b) Swap 01 (c) Swap 02 ............. 42
4.2 Limitations Identified and their Proposed Solutions ......... 43
4.3 Pearson Correlation Bar Graph (Combined Attacks) ......... 49
4.4 Power Law Distribution ........................ 52
4.5 Extent of Damage ............................ 53
4.6 Execution Time (Single) ........................ 54
4.7 Execution Time (Combined) ...................... 55
4.8 Convergence of Initial Network Topology ............... 56
4.9 Robustness of Initial Network Topology ................ 57
4.10 Robustness Analysis (Degree) ..................... 58
4.11 Robustness Analysis (HMC) ...................... 59
4.12 Number of Swaps (Degree) ....................... 59
4.13 Number of Swaps (HMC) ....................... 60
4.14 (a) Initial Topology (b) Hill-Original (c) Hill-Smart (d) ROSE-
Original (e) ROSE-Smart ........................ 61
5.1 Onion-Like Structure. .......................... 65
5.2 Edge Rewiring Strategy. ........................ 71
5.3 Comparison of Network Robustness under Malicious Attacks with
respect to Number of Iterations. .................... 75
5.4 Comparison of Network Robustness under Malicious Attacks with
respect to Number of Nodes. ...................... 77
5.5 Comparison of Graph Density under Malicious Attacks. ....... 77
xv
5.6 Comparison of Connectivity under Malicious Attacks. ........ 78
5.7 Comparison of Connectivity under Random Attacks. ........ 79
5.8 Comparison of Computational Time under Malicious Attacks for
N=100. ................................. 79
5.9 Comparison of Robustness of ROSE-DDO and RS-DDOD. ..... 80
xvi
LIST OF TABLES
2.1 State of the Art Work ......................... 17
3.1 Network Parameters used in Simulations ............... 30
3.2 Network Parameters .......................... 34
4.1 Limitations Identified, Proposed Solutions and Validations ..... 44
xvii
List of Abbreviations and Mathematical Symbols
ASO Angle Sum Operation
AI Artificial Intelligence
BA Barabasi-Albert
BP BackPropagation
CC Closeness Centrality
DAO Degree Associativity Operation
DDD Degree Differences using the Dissortativity
DDLP Deep Deterministic Learning Policy
DDO Degree Difference Operation
DD Degree Dissortativity
DDOD Degree Difference Operation with Dissortativity
DDLP Deep Deterministic Learning Policy
EES Efficiency based Edge Swap mechanism
GDA Great Deluge Algorithm
GA Genetic Algorithm
HC Hill Climbing
HDO High Degree Operation
HMC Heat Map Centrality
IoT Internet of Things
MCS Maximum Connected Subgraph
MPGA Multi-Population Genetic Algorithm
MA Multi Agent
ML Machine Learning
NF Node’s Farness
1Thesis by: Muhammad Awais Khan
NNF Node’s Neighbor Farness
ORS-Cap Optimized Rewiring Strategy using nodes’ load Capacity
RESM Random Edge Swap Mechanism
RS Rewiring Strategy
SA Simulated Annealing
SESM Smart Edge Swap Mechanism
VSet of Nodes
WSNs Wireless Sensor Networks
AAdjacency Matrix
rAssortativity
BetiBetweenness Centrality
CiCapacity of a node i
Π(i)Connection Probability of a node i
ki Connection Probability of a node i
kDegree of a node
mEdge Density
GGraph
BLower limit
βNodes’ Capability to process the load
σjk (i) Number of shortest paths between nodes jand kthat pass
through node i
NNumber of Nodes
αPower Law Exponent
RRobustness
sShortest Path
BUpper limit
2Thesis by: Muhammad Awais Khan
Chapter 1
Introduction
3
Chapter 1 Introduction
1.1 Introduction
This section consists of research background,problem statement,contributions
and organization of the thesis. The details are as follows.
1.1.1 Research Background
The Internet of Things (IoT) has become an essential technology nowadays [1
3]. Its integration with the Wireless Sensor Networks (WSNs) [47] provides
good support to the research community. The IoT-WSNs [812] have various ap-
plications including healthcare [1315], industrial [16], intelligent transportation
[17,18], smart environmental monitoring [19], smart cities [20,21], blockchain
[2225] and underwater communications [2630]. An important characteristic of
the IoT-WSNs is that they are operational even in hostile environments [31]. The
nodes in the WSNs are used for efficient data delivery towards the destination
[3234]. However, due to limited energy resources of the nodes [3539], their com-
munication capability, lifetime [4042], etc., are greatly compromised. In the IoT,
the nodes communicate through the Internet [43,44], however, the frequent cyber
attacks [4550] on the network greatly affect their connectivity and reduces the
network robustness. Therefore, the ultimate goal of the researchers is to provide
effective ways to improve the network resiliency against the cyber attacks.
The network topologies [51,52] provide layouts of various communication activities
occurring inside the networks. The resilience of the network topologies against the
attacks depends upon the arrangement of the nodes present in the networks. The
robust network topologies maintain the connections of maximum number of nodes
in a subgraph after the node removal due to the cyber attacks. Different types of
attacks occur in the networks. Two types of attacks are generally considered in
the networks: random and malicious [53]. The random attacks [5456] target any
random node and remove it from the network, while the malicious attacks [57
59] remove the most important node from the network. The nodes’ importance
is measured using degree, betweenness, etc. Thus, the malicious attacks have
greater affect as compared to the random attacks. The attacks split the network
into multiple independent graphs [60] and paralyze the network with time.
A lot of researchers have put their efforts into studying the properties of different
networks such as wireless networks [61,62], social networks [63,64], body area
networks [6567], vehicular networks [6871], robotic networks [72,73]. Most
of these networks are complex networks [7477] and have high importance due to
4Thesis by: Muhammad Awais Khan
Chapter 1 Introduction
their dense nature. A complex network theory [7880] is one of the classic network
theories. It consists of two models, namely, small-world network [81,82] and scale-
free network [83,84] models. The small-world networks have two features, which
differentiate them from other networks. These features include a small average
path length and high clustering coefficient [85]. Moreover, the properties of the
nodes in the networks are heterogeneous [86]. On the other hand, the nodes in
the scale-free networks are homogenous [87] and their degree distribution follows
the power-law [88]. It means that most of the nodes in the network are low
degree nodes while there are few nodes with high degree in the network. The
networks with high number of low degree nodes are robust against the random
attacks. However, these networks show vulnerability against the malicious attacks
due to presence of few number of high degree nodes. Therefore, the researchers
are focusing on the construction of robust scale-free network topology against the
malicious attacks.
Since, the attack on low degree nodes does not affect the performance of the
network; therefore, the main concern of the research work is: how to make the
network robust, when the network faces malicious attacks? In that case, the high
degree nodes are isolated from the network and the performance of the network
tends to decline with the passage of time. Therefore, enhancing the network
robustness against the malicious attacks is a focal point of the research in the
scale-free networks.
The network robustness is measured through a metric proposed by Schneider et al
[89] by analyzing the behavior of the network during node removal. The research
studies [9092] reveal that the network robustness can be improved through dif-
ferent optimization techniques. Therefore, the authors in [93] use edge rewiring
mechanism to construct a robust network topology. The proposed mechanism
achieves success in constructing an onion-like topology [9497], which shows high
robustness against the malicious attacks. However, the proposed mechanism falls
into the local optima and reduces the robustness. Similarly, the edge rewiring
mechanism in [98] involves temperature parameter to modify the network topol-
ogy. However, it uses many redundant operations in the network and limits the
network robustness.
1.1.2 Problem Statement
In this thesis, the problems of constructing a robust scale-free network topology are
addressed so that it can tolerate the attacks. It is known that scale-free network
5Thesis by: Muhammad Awais Khan
Chapter 1 Introduction
topologies are vulnerable to malicious attacks. Therefore, many scale-free topolo-
gies are constructed including the one proposed in [99] where the authors overcome
the premature convergence in the existing Genetic Algorithm (GA) model. Due to
the premature convergence, similar solutions are produced in the network. Thus,
to avoid it, several populations are considered to bring diversity to the network.
However, the proposed mechanism does not consider degree distribution of nodes
in each population. Moreover, in [93,98], the authors use the edge swap mech-
anism to rewire the independent edges for the construction of robust scale-free
network topologies. They have checked the robustness for both alternative con-
nection methods. Still the robustness reduces because the proposed mechanisms
damage the property of the scale-free network by reducing the connections among
similar degree nodes. Besides, the degree to degree correlation is important as it
helps to construct a network with better onion-like topology. However, the previ-
ous mechanisms [93,98] reduce the robustness when nodes with different degrees
are connected with each other, damaging the property of onion-like structure.
Most networks in real world have scale-free nature. Therefore, the nodes, which
act as hubs are considered the most important nodes in these networks and the
removal of these nodes create a serious threat to the connectivity of other nodes.
The cost of the network is important while performing topology optimization us-
ing the algorithms proposed in [93,98]. However, there are some other factors
that increase the cost of the network. For example, it is known that every objec-
tive function has a certain convergence limit and performing further optimization
always provides similar results. For topology optimization case, performing un-
necessary edge swaps may not provide an optimized network topology, except it
increases the cost of the optimization process. Also, the removal of nodes based
on degree and betweenness in [100] has shown success because of a strong positive
correlation between them. However, the betweenness centrality has high compu-
tational cost [101]. Moreover, making the network robust against different types
of malicious attacks is a complex problem in the scale-free networks and the at-
tacker can attack the nodes having different properties like degree, betweenness,
closeness, etc. In addition, the Degree Difference Operation (DDO) in [102] low-
ers the importance of the initial topology by fixing the value of the threshold p,
which is used to limit DDO. It is because the initial topology always maintains
a high degree difference for a fixed value of p. Thus, it can not be considered as
an optimal choice to improve the network robustness. Furthermore, the authors
in [93,98] do not consider nodes’ load capacity during the node removal process.
Thus, when the capacity of the node exceeds the maximum capacity limit, it fails.
The nodes’ failures in the network affect the connectivity of other nodes and reduce
6Thesis by: Muhammad Awais Khan
Chapter 1 Introduction
the network robustness.
1.1.3 Thesis Contributions
The main contributions in this thesis are as follows.
1. A multi-population environment is considered called MPGA to reduce the
network complexity in achieving an optimal solution. The introduction of
an entropy based mechanism for the replacement of one’s bad solution with
other’s good solution increases the network robustness.
2. The fundamental edge swap mechanism has less robustness against the mali-
cious attacks because it does not consider nodes degree for rewiring. There-
fore, the Efficiency based Edge Swap mechanism (EES) is introduced to
increase the robustness of the network against the malicious attacks.
3. Considering the importance of onion-like structure, a new edge swap mecha-
nism is introduced based on the assortativity to make the topology onion-like.
4. Aiming to overcome the randomness of edge swap, a smart edge swap mech-
anism is proposed to evaluate the network robustness against the malicious
attacks.
5. A threshold based node removal method is introduced to reduce the com-
putational complexity of the network. It tackles the problem of performing
unnecessary topology optimization when a convergence is achieved.
6. For the construction of a robust topology, we consider three important cen-
tralities named as degree, betweenness and closeness. The Pearson correla-
tion coefficient is used to find two strong positively correlated measures that
can be used simultaneously.
7. Considering that multiple attacks can occur on the network, the topology is
optimized using a centrality measure named as Heat Map Centrality (HMC).
8. Considering that different types of rewiring can construct an optimized scale-
free topology, the optimization algorithm GDA is introduced to increase the
network robustness.
9. As random edge rewiring increases the network dissortativity, a new edge
rewiring mechanism is proposed to construct a topology with low degree
dissortativity.
7Thesis by: Muhammad Awais Khan
Chapter 1 Introduction
10. Ensuring successful connections among similar degree nodes, we modify
DDO for the construction of robust onion-like topology.
11. Considering that the failures of nodes reduce the network robustness, a topol-
ogy is constructed, which ensures minimum nodes’ failures in the network
using the nodes’ load capacity.
12. Based on the improved performance of edge rewiring using the nodes’ load
capacity, we have redesigned the GDA method to increase the network ro-
bustness.
1.1.4 Organization of thesis
The remaining thesis is organized as follows. Chapter 2presents a detail overview
of the literature. Chapter 3,4and 5show the proposed strategies and their
simulation results. The conclusion and future work are given in Chapter 6.
8Thesis by: Muhammad Awais Khan
Chapter 2
Literature Review
9
Chapter 2 Literature Review
2.1 Literature Review
In this section, we discuss some of the researchers’ work in detail. Sohn et al.
[103] propose a new optimized mechanism rooted from artificial neural networks.
It takes the topologies of the scale-free networks as an input data and gives the
topologies of the hill climbing as an output data. A three-layer mechanism is
designed as a part of network operation, which consists of a single input layer,
multiple hidden layers and a single output layer. The network performs better
in terms of achieving high tolerance against the random and targeted attacks.
Nevertheless, it fails to provide the state of the art understanding for network
reconstruction. The work proposed in [104] selects two vertices randomly from
a given set of vertices, which are paired to form a link. After the establishment
of links, the network attains the shape of a complex network. Then, it picks up
different combinations of vertices based on the probability distribution, which leads
to the emergence of different complex networks. The proposed mechanism achieves
low computational complexity for generating the complex networks. However, this
type of practice cannot be applied to a network that incorporates extremely high
degree vertices, because in that case, it has a weak impact in increasing the network
efficiency for large scale networks. The authors introduce a mechanism in [105] for
the identification of a sinkhole and the malicious attack. In this mechanism, the
responsibility of the data forwarding is done through a clustering mechanism [106].
Nodes’ practice during data routing is observed through a watchdog reputation.
Then the misbehaving nodes can rapidly be recognized and parted away from the
network. Its network efficiency is remarkable with less energy consumption, yet, it
has to prove its worth in dense environment, where the density of the nodes is high;
thus, it becomes difficult for the watchdog reputation mechanism to overcome the
malicious attack.
The heuristic algorithms are used in the research to find a solution to a problem.
The GA based schemes are presented that aims to find an optimal solution in
the complex networks. The MPGA proposed in [107] overwhelm the premature
convergence and drive the network towards the higher robustness. In the MPGA,
the crossover operation and mutation are applied to a limited set of chromosomes
extracted from the adjacency matrix in order to converge the solution from local to
global optima. To further enhance the performance of MPGA, the authors in [108]
introduce the migration operator that selects the worst solution of a population
and replaces it with the good solution of the other population, which reduces
the premature convergence in the MPGA. The proposed methods attain efficient
10 Thesis by: Muhammad Awais Khan
Chapter 2 Literature Review
robustness in expense of low computational efficiency and high data overhead.
The proposed work in [109] includes the two heuristic strategies, CDA and DBA,
that exploit the information of community detection structure and the degree of
nodes severally. After utilizing the necessary information, an attack strategy is
implemented, called the GA-based-Q Attack, where the fitness function is designed
using the modularity Q. The proposed GA-based attack minimizes the attack
probability for small area networks. Nonetheless, the problem for the large area
network is still confusing, which needs to be studied in order to obtain a robust
network.
Aiming to explore the invulnerability of the clustering in the networks to cascad-
ing, FU et al. [110] develop a model for WSNs, where they instigate two new
notions, i.e., “sensing load” and “relay load”. Their focus goes around the analy-
sis of the invulnerability to cascading failure in clustering WSNs, which includes
the random and the scale-free networks. Besides this, the network vulnerability
is improved by the division of the capacity expansion problem for the selection of
eligible nodes and the efficient utilization of the resources. The proposed method-
ology does not consider the energy consumption, which is the main part of the
network enhancement. In contrast to this, the proposed capacity expansion can
only be applied to static WSNs, it could be quite tough for this model to achieve
success in the clustering WSNs with the nodes and sink mobility [111]. The pur-
pose of the work in [112] is to overwhelm the limitations of the Elephant Herding
Optimization (EHO) and raises the performance of the original EHO. For this rea-
son, the authors propose three novel algorithms, namely, “alpha-tuning, cultural-
based, and biased”. The alpha-tuning algorithm enhances the EHO operation by
continuously tuning the alpha value with the increasing number of iterations. The
cultural-based algorithm replaces the worst solution with a novel solution, which
originates randomly from a belief space. The biased algorithm initializes the first
population to determine the good fitness candidate. It forces the first population
to obtain a minimum good fitness population within the given threshold. These al-
gorithms work promise to attain efficient robustness, however, further validations
are required to understand the latent effects of the belief space.
The Ant Colony Optimization (ACO) is another useful technique for finding the
optimal path for the network robustness. A novel ACO based optimization scheme
in the heterogeneous environment is introduced by the Qiu et al. [113] to highlight
the global importance of the optimal paths to efficiently reduce the number of links
that emerged from a node towards the destination. The Small-world topology is
constructed with the advent of the shortest path, thus, reduces the average path
11 Thesis by: Muhammad Awais Khan
Chapter 2 Literature Review
length, which at the same time also reduces the communication delay of the net-
work. The critical node is selected that acts as an endpoint for the added shortest
paths. It increases the nodes’ withstand capability in case of an attack at the
cost of the longer running time. To get an optimize solution for a problem and
improve the network robustness, the authors introduce the concept of ACO [114].
The motivation behind this work is to establish a quantitative model to produce
a feasible solution and increases its convergence speed. The proposed methodol-
ogy enhances the computational efficiency and the robustness of the network in
the sparse case, however, its performance in a large scale network has yet to be
explored.
Many complex networks introduce the edge swap mechanism for increasing the
network robustness. A Simulated Annealing (SA) algorithm proposed by Busser
et al. [98] uses the same concept of the edge swapping mechanism to make the
structure look like an onion keeping the degree distribution of the nodes unchanged
to improve the network robustness. SA optimizes the performance of the scale-
free network, even so, the algorithm causes high computational cost. Su et al.
[115] propose a mechanism to turn the non scale-free networks into the scale-free
networks by highlighting the importance of the preferential attachment in [116].
The strategy proposed in [117] introduces the betweenness centrality for checking
the node’s eligibility as a data forwarder. It calculates the betweenness centrality
of each node and sorts the degree of nodes in the decreasing order. Weights are
assigned to the nodes after finding the shortest path between any two neighbors
of the high degree node. The proposed link adding strategy proved to be efficient
in terms of increasing the network robustness against the random attacks, still, it
becomes fragile when the network faces the malicious attack.
Community structural topology is a useful technique to find the convergence of the
solution towards the global optima. In [118], Yang et al. propose a 3 steps strat-
egy for the community structure to improve the network robustness by keeping
in mind the degree distribution of the nodes constant. They allow each commu-
nity to exhibit an onion like shape through the edges swap, which allows only the
similar characteristics nodes to connect with each other. In addition, during the
malicious attack, the nodes prefer to connect with the high degree node in the
same community, which drives the network towards the higher robustness. The
proposed methodology considers the attack on the vertices, however, its effects on
the edges is still a meaningful task to do. The authors in [119] measure the net-
work robustness by considering the attack on the nodes. A novel robustness index
is designed to check the nodes’ vulnerability to cascading failures, and the network
12 Thesis by: Muhammad Awais Khan
Chapter 2 Literature Review
depletion against the link-based cascading failures. Based on their observed ex-
perimentation, their work provides a better solution for small area networks. Still,
for large area networks, their robustness suffers a downward trend, that is against
our observations. The proposed scheme in [120] considers the interdependent links
as one of the leading causes in the network fragility. Two optimized models are
designed to check the feasibility of a solution towards the higher robustness. These
are the generic edge swapping model and the optimized interdependent network
model. The proposed mechanism optimizes the network robustness against the
targeted links, nevertheless, its performance leans towards negligence when the
number of interdependent links is increased.
To generate the scale-free topology, a new modeling strategy is proposed in [102].
The proposed scheme considers the WSN constraints, i.e., arbitrarily large commu-
nication range of the nodes and the degree of the available nodes in the network.
It improves the robustness of the generated scale-free topology by exploiting the
node’s degree and the position information. Additionally, it rearranges the nodes
into an onion like structure in order to make the topology robust against the
malicious attacks. Meanwhile, it keeps the node’s degree same to keep the topol-
ogy scale-free. A Comparative analysis of the proposed scheme is also performed
against the existing robustness schemes. Simulation results validate the efficacy
of the implemented scheme in counterpart, however, the proposed model is only
valid for the homogeneous scale-free network, its effectiveness in terms of the het-
erogeneous environment is yet to be flourished.
The proposed work in [121] develops a link strategy to mitigate the effects of
a malicious attack on nodes. The authors observe that constructing a limited
cost dependent links are far better than the simple link adding strategy. The
results prove to be effective in developing a robust network, still, the cost of
adding edges is the key problem that is to be considered. In order to enhance
the community robustness, a two-level learning strategy is introduced in [122],
which initializes the population to perform the crossover operation in the hunt for
a better solution in the global area without altering any change into the original
community structure. The prime aspect of the learning strategy is to circumvent
the intercommunity links, in case if these links dissipate, the performance of the
network declines sharply, leaving behind the isolated nodes. The edge swap criteria
assemble those important links to be constructed in the same communities. Thus,
when the community-level learning strategy is applied, the network leans an onion
like structure, which notably improves the network robustness. In spite of that,
the proposed strategy does not assure optimal routes for the case of offspring
13 Thesis by: Muhammad Awais Khan
Chapter 2 Literature Review
production, which can be put into the pitfall of the proposed strategy. In [123],
the authors propose an algorithm that considers the cost during the attack on high
degree nodes. The proposed algorithm finds the combination of nodes with the
minimum cost that helps to optimize the network performance. The algorithm is
based on a redesigned objective function to find a better solution in case of the
malicious attack. It shows good performance in enhancing the network robustness
in case of a malicious attack. However, it does not tell whether the proposed
solution converges to a local optimum or a global, which is considered to be a
major obstacle in overcoming the premature convergence.
Aiming to explore the optimality in the network structure, the authors in [124]
have observed three strategies for checking the network behaviour, namely, ran-
dom rewiring, greedy rewiring and second neighbor rewiring. In random rewiring,
changing the edge of a node affects the network robustness as every node has the
information of all other nodes. Introducing the greedy rewiring results in lim-
iting the connection of nodes, thus, it is not effective in increasing the network
resilience. The second neighbor strategy is a good technique to be applied in the
case of node removal, still, it increases the data overhead of the network, which
is not acceptable. In order to mitigate the effects of a malicious node, the work
presented in [125] holds the data of a malicious node. During the malicious attack,
the network identifies the malicious node and stops forwarding the data of other
nodes towards that node. However, holding the forwarding process may cripple
the network because there might be some nodes that are only associated with this
node. Moreover, holding the data forwarding of these nodes increases the time
complexity of the network. The authors in [126] propose a two-step algorithm
namely: the detection of nodes and its optimization and secondly the nodes rein-
sertion. They also propose two strategies for the network robustness, named as
the static attack and the dynamic attack. However, due to the slow convergence
speed of the proposed mechanism in solving the problem, its performance in the
time complex network is ignored.
The proposed work in [100] is based on a fault-tolerant model, which shows high
network robustness against random node failure. However, it fails to improve the
network robustness against the malicious attacks. In [127], the authors analyze
the threat of malicious attacks on the Artificial Intelligence (AI) community and
inform that many algorithms involving Machine Learning (ML) are fragile against
malicious attacks. However, these algorithms do not ensure a reliable and robust
network. According to [128], the scale-free IoT networks show vulnerability to
malicious attacks. Thus, designing a robust mechanism against the attacks is
14 Thesis by: Muhammad Awais Khan
Chapter 2 Literature Review
challenging. The previous optimization strategies have optimized the network
topology by maintaining the network connectivity; however, the computational
cost of the optimization strategies is relatively high. Due to the increasing demand
for IoT devices, increasing the network robustness against malicious attacks is one
of the challenging issues in the scale-free network [129].
In [130], the authors find that the addition of the links increases the cost of the
network. The conventional Genetic Algorithm (GA) is a good example of an evolu-
tionary algorithm that optimizes the network robustness. However, the premature
convergence in GA reduces the exploration capability and lowers the performance
of the network against the malicious attacks [99]. A similar issue is raised in [131],
where the authors state that improving the topology of the scale-free network
against the malicious attacks is a complex problem. The authors in [132] address
the malicious attacks on a Multi Agent (MA) network. However, the research
focuses only on constructing a robust MA network without considering its cost.
The research in [133] proves that it is necessary to involve both cooperation and
robustness in constructing a robust network. However, this research is only lim-
ited to undirected network. The authors in [134] reveal that node attacks and
link attacks are negatively correlated. Therefore, multi-objective optimization is
a better choice in this case. However, the computational cost for calculating the
robustness of node and link attacks is different due to the more number of links in
the network as compared to the nodes. Another attack strategy is introduced in
[135], which measures the impact of the intentional attacks and analyze that they
are harmful to the stability of the network. However, the optimization strategies
only deal with optimizing the network against one of those attacks.
According to [123], most attack strategies remove all the nodes according to a
specific order. However, in general, the attacker does not consider attacking the
nodes in a specific order. Moreover, removing high degree nodes from the network
costs more than removing low degree nodes. Therefore, controlling the network
robustness by considering the attack cost is a major problem in a network [136].
Furthermore, according to [137], the degree distribution of the nodes during the
attack process is dynamic as it changes with each attack.
The majority of the research studies in [138] test the network robustness against
the random node removal, however, in general, the attacker tries to attack the
most critical nodes in the network. Based on the analysis of [139], measuring the
network robustness against node and link removal is an open issue in the complex
15 Thesis by: Muhammad Awais Khan
Chapter 2 Literature Review
networks. The former measures for network robustness including natural connec-
tivity, controllability robustness, etc., are based on edge and node’s connectivity,
size of the largest connected component, etc. However, these measures have failed
to express the network’s capability in preserving the connectivity of the network.
Due to the growing demand of the complex networks according to [140], there is
also a concern of security issues related to these networks. According to [141], the
link addition strategy increases the cost of the network and changes the degree dis-
tribution of the nodes in the network. To keep the degree distribution unchanged
and reduce the cost of the network, the edge swap mechanism is designed. How-
ever, the random edge swap mechanism increases the computational complexity
of the network and performs many redundant operations during the optimization
process.
According to [142], the assortativity rand the power-law exponent αare important
factors in the scale-free networks, which help to create a strong interaction between
the nodes. The value rclose to 1 tends to make strong interaction between the
nodes of similar degree. However, there is no evidence where the proposed model
highlights the correlation between Rand these measures. The literature work
in [143] reduce the cascading failure by considering the capacity of the nodes.
However, they fail to analyze its effect on network recovery. According to [144],
the degree to degree correlation is an important factor for enhancing the robustness
of the network. However, the Newman’s research reveal that the enhancement of
degree to degree correlation is limited to a certain degree threshold under malicious
attacks.
Schneider et al. in [93] propose an efficient rewiring mechanism based on the
robustness Rto make the network robust against the malicious attack. However,
further study reveals that the random rewiring mechanism uses the series of steps
for finding the network robustness, which increases the computational cost of the
network. The State of the Art Comparisons is described in Table 2.1.
16 Thesis by: Muhammad Awais Khan
Chapter 2 Literature Review
Table 2.1: State of the Art Work
Limitations already
addressed Contribution already done Validation already done Limitations to be
addressed
The acceptance of solution
with the good fitness value
makes the solution stuck in
the local optima [98]
Accepts the worst solution
with a specific probability to
converge the solution towards
the global optima
Network robustness is
increased High computational cost
The fault-tolerant model is
only robust against the
random attacks [100].
The network robustness
against the malicious attacks
is increased using two
operation: HDO and DAO
High robustness is achieved.
The convergence operation
slows down due to redundant
operations.
The degree of nodes and the
communication range of nodes
is not considered [102]
Considers both the degree of
nodes as well as the
communication range of nodes
Improves the network
resilience after the edges
removal
Time and cost complexity is
compromised
Computational complexity of
hill climbing [103] is high
Lower the computational
complexity on the basis of
adjacency matrix
Robust network and high
tolerance in case of the
random and the malicious
attack
Fails to provide state of the
art understanding for network
construction
Isolated nodes are added into
the network that increases the
computational complexity
[104]
Avoiding isolated nodes by
restricting their connection to
only one other vertices
Achieves low computational
complexity in enhancing the
network robustness
Cannot be applied to large
scale networks
Selective forwarding [105]Nodes forward the data to
selected cluster head Enhances network robustness
Fails to perform in dense
environment due to high data
overhead
Single population and
premature convergence in
traditional GA [107]
Multi-population GA is
applied to increase the
diversity in the solution space
Improves the network
robustness and the premature
convergence
High computational time is
involved in finding the desired
solution
Slow convergence of solution
in [108]
Introduces the immigration
operator to further improve
the premature convergence
Robustness is enhanced Computational time is
increased
Link deletion and addition
weaken the network
performance [109]
Link rewiring instead of
deletion and addition
Makes the network robust in
small area networks
Its performance is
compromised in large area
networks
Capacity expansion problem
[110]
Increasing the node capacity
by keeping in mind the
preferential attachment
Network efficiency is increased It can only be applied to static
WSNs
The random replacement of
the worst solution and the
fixed parameters for all
iterations [112]
Replacing the worst solution
with the good one and tuning
the value of the alpha for each
iterations
Robustness of the network is
improved Time complexity is increased
The global optima problem
and the average path length
[113]
The ACO globally finds the
optimal path and reduces the
average path length
Enhances the network
robustness
Increases the computational
complexity of the network
The solution converges
towards local optima [114]
ACO with travelling salesman
problem is introduced to
converge the solution towards
the global optima
Makes the network robust and
decreases the computational
complexity
Fails to find an optimal
solution in dense area network
Optimizing the network
topology is a major problem
[115]
Scale-free network is
constructed from non
scale-free network with the
help of non linear rewiring
method
Network robustness is
increased The data overhead is increased
The link selection with high
robustness is selected without
considering the path emerges
from that link [117]
Nodes betweenness centrality
is considered to minimize the
average path length
Increases the network
robustness against random
attacks
Fragile against the malicious
attack
Nodes deletion decreases the
network performance [118]
Considering the edge swap
without effecting the degree
distribution
Network robustness is
improved High computational cost
The robustness metric Ris
introduced only for nodes
failures [119]
The new robustness metric R
is introduced for link
cascading failure
Makes the network robust in a
small area network
Robustness decreases for large
area networks
A small perturbation in the
network leads to network
failure due to the dependency
between the nodes [120]
Edge rewiring mechanism is
introduced to change the inner
structure of the network
Network robustness is
improved
Performance degrades with the
increase of interdependent
links
Same topology exists for all
types of attack [121]
Changes topology on the bases
of nature of the attack
Improves the network
resilience
Cost is involved in
construction of new edges as
well as the computational time
is increased
Change in the community
structure alters the network
topology and can damage the
network functionality [122]
Propose a new metric Rfor
preserving the community
robustness
Network robustness is
improved
Does not assure the optimal
route
17 Thesis by: Muhammad Awais Khan
Chapter 2 Literature Review
Limitations already
addressed Contribution already done Validation already done Limitations to be
addressed
More nodes are affected due to
the high degree attack in many
heuristic techniques [123]
Uses the high degree attack
with the cost to minimize the
affect of the high degree attack
Makes the network robust
against the attack
Fails to provide convergence of
solution from local to global
optima
Experimental validations of
both the random and the
malicious attack have not done
simultaneously on same
strategies [124]
It validates the effectiveness of
the proposed solution with
both the attacks to compare
the nodes’ performance
Makes the system robust
against the targeted attack Increases the network overhead
Selective attack and reliability
is compromised [125]
Overloaded nodes are
controlled by selective drop
and defective links are
disabled to improve reliability
Enhances the network
robustness Network delay is increased
Targeting the high degree
nodes is a NP hard problem
and its effects do not remain
the same, therefore, it is
difficult to measure [126]
Makes the optimization
problem combinatorial and
solves the issue using the
global information of nodes
Improves the network
robustness
Slow convergence speed and
high computational complexity
Vulnerability of scale-free
network to malicious attack
[128]
Uses back-propagation neural
network to optimize the
network topology
Accuracy of the network is
improved and loss function is
minimized
Time consuming and
additional learning features
are required for training
Rcannot be optimized further
[129]
Optimizing the value of R
through deep neural network
training
The value of Ris optimized Higher network complexity
due to complex training
Cost of the network increases
for high degree node attack
[136]
Modes of attacks are changed
based on attack cost
The validity of the network is
done using s(q) and network
efficiency
Optimization after node
removal is missing
Degree distribution of nodes
during the attacks is dynamic
[137]
Eight attack strategies are
introduced
Network performance is
validated based on the extent
of the attack
Optimization is missing
Network is robust for random
node removal [138]
Construct a robust network
based on high degree node
removal
The performance of the
network is validated by
considering the network
robustness for both in and out
degree nodes
Hard to implement in real
scenario
The cost for evaluating the
robustness for nodes and link
removal is high [139]
Protect the link whose removal
maximizes the RGvalue
The validation is done on
different datasets
Slow convergence of GA
increases the time complexity
of the network
Attack cannot be limited to
nodes [140]
Consider link-attack strategies
on k-core network
The shell-max strategy
validates the effectiveness of
the proposed model
The optimization of network
robustness after the attack is
not done in k-core
decomposition
Removal of nodes in a
community impacts the
structure[141]
Preferential rewiring method
is applied to preserve the
community structure of the
network
The performance of the
network is validated on both
synthetic and real-world
networks
Redundant operations
increases the complexity of the
network
The previous BA model fixed
the value of aand rand
generated a scale-free model
[142]
The proposed scheme focuses
on tuning the value of a and r
for both malicious and random
attacks.
Optimization of the network is
missing.
The effectiveness of the
proposed model is validated
through the malicious and
random attack on both nodes
and links
Not enough evidence to show
the impact of resource
allocation on the performance
of the network [143].
Highest capacity node is
removed from the network and
the recovery process recover
those nodes whose recovery
timing reaches the threshold
Feasible only for static
environment
The resilience loss is used as a
recovery metric to quantify
the recovery process of the
network.
Enhancement of degree to
degree correlation is limited to
a certain degree threshold
under malicious attack [144].
Increases the network’s
robustness by considering the
information of loops in the
network.
Addition of edges increases the
network cost.
The network’s performance is
validated suing R, FVS, and r
18 Thesis by: Muhammad Awais Khan
Chapter 3
Modification Strategies for Scale-Free IoT
Networks
19
Chapter 3 Modification Strategies
3.1 Summary of the Chapter
In this chapter, two different mechanisms are adopted for the construction of a
robust scale-free network. First, a Multi-Population Genetic Algorithm (MPGA)
is considered to overcome the premature convergence in the GA and optimize
the network robustness against the malicious attacks. Then, an entropy based
mechanism is used, which replaces the worst solution of one population with the
best solution of other population to improve the network robustness. Second, two
types of edge swap mechanisms are introduced. The efficiency based edge swap
mechanism selects the pair of edges with high efficiency to increase the network
robustness. The second edge swap mechanism based on assortativity transforms
the network topology into an onion-like structure to achieve maximum connectivity
between the similar degree nodes in the network. The optimization of the network
robustness is performed using Hill Climbing (HC) and Simulated Annealing (SA)
methods. The simulation results show that the proposed MPGA Entropy has
9% better network robustness as compared to MPGA. Moreover, the proposed
edge swap mechanisms have effectively increased the network robustness with an
average of 15% better robustness as compared to HC and SA.
i
jk
l
7
3
5
Figure 3.1: Nodes Joining Mechanism.
20 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
3.1.1 Construction of a Scale-Free Network
Due to the limited communication range of nodes in the WSNs [145], the newly
added node may not have enough neighbors. Thus, the preferential attachment
property [145][146149] is not applicable for the networks having limited commu-
nication range. In the preferential attachment, the high degree nodes have high
probability to become the neighbors of a newly added node. Thus, large num-
ber of connections makes the network more dense and this is the reason why the
importance of scale-free networks is increasing day by day. A scale-free network
consists of an undirected graph G= (V, E ). Where Vdenotes the set of nodes
and Eis the set of edges in the graph [102].
The network modeling starts with a few number of connected nodes, which are
initially deployed in the network. The newly added node selects the highest degree
node as its neighboring node in the network. The connection probability ki of a
neighbor node iis defined in [102] and mathematically, it can be written as:
ki=ki
Pn
q=1 kq
.(3.1)
From Equation (4.1), kiindicates the degree of a node iand kqsignifies the overall
degree of the neighbouring nodes of i, whereas, nis the number of nodes in the
network. Considering the model in [102] where a newly added node iwants to
establish a connection with one of the nodes in its communication range. Suppose,
j,kand lare the nodes in the communication range of a node i. The degrees of
these nodes are 7, 5 and 3, respectively, as shown in Figure 3.1. The connection
probabilities of the nodes are calculated using Equation (3.1), which are 0.3888,
0.2777 and 0.1666, respectively. The selection of the neighbor nodes for a node i
is based on their edge densities represented as m. Roulette wheel method [102] is
adopted for the selection of nodes’ neighbor, which generates random numbers in
the range 0 to 1. The nodes with large number of connections have high probability
of being selected as they occupy large space on the roulette wheel.
3.1.2 Attack Model
The scale-free networks have high resilience against the random attacks, however,
they are vulnerable to malicious attacks. Therefore, the aim of the topology
optimization is to construct a network, which shows high robustness against the
malicious attacks. This reason is that the removal of high degree nodes effects the
connections of most nodes in the network.
21 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
For the evaluation of network robustness, the equation used in [102] is adopted,
which is given as:
R=1
N+ 1
N1
X
n=0
MCSn
N.(3.2)
From Equation (3.2), M CSnsignifies the total number of nodes in the maximal
connected subgraphs in response of eliminating the nth highest degree node from
the network. The factor 1
N+1 is utilized to set the value of the robustness in
the range of 0 to 1. However, the maximum value of the network robustness is
restricted to 0.5.
3.2 Multi-Population Entropy based Mechanism
In this section, an entropy based model is proposed for increasing the robustness
of the scale-free networks. The entropy is the measure of the uncertainty in the
degree information of the nodes in each population. Besides, the major problem in
the conventional GA is the occurrence of the premature convergence, which pro-
duces similar solutions in the networks. The authors in [107] propose MPGA using
several populations to overcome the premature convergence in the network. Each
population in MPGA produces different solutions and brings diversity in the pop-
ulation. However, the proposed MPGA fails to analyze the probability of nodes
having different degrees in each population during the replacement operation. If
the population with high probability of different degree nodes are considered dur-
ing the replacement operation, the network robustness can be increased because
this population produces large varieties of solutions in the networks.
In the topology optimization, the probabilities of different degrees nodes decide
the importance of a topology in the scale-free networks because of having a large
varieties of solutions. Thus, the entropy is used to figure out the probability of
different degrees nodes in different populations. At the beginning, different in-
dividuals are evolved to bring different solutions in the networks. Later on, the
crossover is performed between the two fittest individuals in a population. The
crossover brings diversity by searching the local solutions in the network. Among
the solutions, some are identical and the process of the crossover brings no change
in the network. Therefore, the mutation is performed to lead the solution towards
the global optima. The entropy [150] of the population is calculated after the
mutation and before the replacement operation and its general formula is given in
Equation (3.3):
22 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
Step 1: Scale-Free Topology Step 2: Multi-Population Environment
Step 3: Crossover
Step 4: Mutation
Step 5: Replacement
Step 6: Termination
Ends the operation if
global solution is achieved
If H(X) is greater than H(Y), replace
the worst solution of H(Y) with good
solution of H(X)
Figure 3.2: Steps Involved in Modeling the Entropy based GA.
H(P) = X
k
pPklog2(pPk).(3.3)
23 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
Where pPkdefines the probability of nodes of degrees kin a population Pand
H(P) defines the entropy of a population P.
The entropy of each population is calculated by finding the probability of different
degree nodes in the network. After the evaluation of entropy, the immigration
or replacement operator is used, which replaces the worst solution of low entropy
population with the best solution of high entropy population to increase the solu-
tion diversity. The high entropy population has better solutions as compared to
the low entropy population because the low entropy population provides low di-
versity [151]. The network model of the GA is shown in Figure 4.2. The operation
of MPGA is referred from [108] and the steps involved in implementing the GA in
the multi-population environment are described as follows.
Population Initialization: The first step in GA is the selection of the chro-
mosomes, which are selected from a set of individuals called a population. The
chromosomes are the set of possible solutions for each node and they define the
connections of nodes with other nodes. In GA, the set of solutions is represented by
string 0 and 1 where 1 defines the node’s connection with other node and 0 shows
no connection. For population initialization, several things should be considered
when dealing with the GA. First, the diversity of the population is important as
low diversity might lead to premature convergence. The premature convergence
is a state where the solution converges without bringing optimal solutions in the
network. Second, the size of the population must be kept normal. A large popula-
tion might slow down the performance of the GA while a small population might
not bring optimality in the network.
Selection of Fittest Individual: The selection of the fittest individuals is im-
portant to continue the process of evolution in GA. The fittest individual plays a
vital role in bringing optimality in the network. In GA, the fittest individuals are
selected through roulette wheel method [108].
Crossover Operation: The crossover is applied on the fittest individuals, which
are selected from each population to bring different solutions in the population.
The fittest individuals are selected to generate the topology of the child. The
child retains some part of the topology from both fittest individuals. We consider
an example of crossover from [108] where the parent 1 topology has an exclusive
edge E12, which is not present in the parent 2 topology. Similarly, the parent 2
topology’s exclusive edge E34 does not belong to the topology of the parent 1. The
crossover generates the topology of parent 1 in parent 2 by replacing the edges E38
24 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
and E47 with E34 and E78. The details of the mechanism are already discussed in
[108].
Mutation Operation:In some scenarios, the topologies obtained from the crossover
operation are identical, which bring premature convergence in the network. The
mutation is performed to overcome the premature convergence as it brings diver-
sity in the population. Thus, the topology obtained from the mutation is different
from both parents [108], as shown in Figure 3.2.
Use of Entropy in Replacement Operator: After the implication of GA, the
topologies obtained from two populations are compared based on their entropy.
Considering a population where different degrees nodes are present. Suppose the
minimum degree of a node in a population is 1 and the maximum degree is 5.
Assuming the probabilities of the nodes having degree information in the range
1-5 are given as:
p11=7
14, p12=2
14, p13=3
14, p14=1
14, p15=1
14.(3.4)
The above calculation in Equation (3.4) is performed for a single population
(P= 1). The same steps are followed for P= 2 as well. A population with
a high entropy is selected as a better topology because its population has more
diverse solutions as compared to low entropy population. The replacement oper-
ator is introduced into the network, which replaces the worst solution from the
population of low entropy with the best solution from the population of high en-
tropy. The complete step by step process of GA with the entropy based selection
is described in Figure 3.2 and the total entropy across (P= 1) is calculated in
Equation (3.5):
H(P= 1) = 7
14log2(7
14)2
14log2(2
14)3
14log2(3
14)
1
14log2(1
14)1
14log2(1
14)=1.9213.
(3.5)
3.3 Proposed Edge Swap Mechanisms
In this section, considering the importance of edge swap mechanisms in [93,98],
two edge swap mechanisms are proposed to construct a robust topology against the
malicious attacks. The topologies presented in [93,98] select a random pair of edges
from the network. These pair of edges are swapped in both ways in search of an
optimized scale-free topology. The nodes related to the chosen edges are swapped
to maintain the connectivity of the nodes in the network. However, the proposed
25 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
random edge rewiring produces low robustness. Therefore, this chapter aims to
accomplish high network robustness through more precise selection of edges. The
details of the proposed edge swap mechanisms are discussed underneath.
3.3.1 Efficiency based Edge Swap Mechanism
In [93,98], the authors use the edge swap mechanism to rewire the independent
edges from a topology. The robustness is checked for both alternative connec-
tion methods. However, the mechanism fails to consider nodes’ degrees during
the edge swap, thus, it can remove the connections between similar degree nodes
and damage the property of an onion-like structure. Therefore, an Efficiency
based Edge Swap Mechanism (EESM) is proposed by maintaining the connections
between similar degree nodes to increase the network robustness against the mali-
cious attacks. The entropy is used to analyze the uncertainty in the nodes’ degree
information of the selected pair of independent edges in a chance to increase the
network robustness. To find how accurate the type of rewiring is, the efficiency
of the nodes associated with the edges needs to be evaluated before the swap.
Initially, for calculating the efficiency, the random probability function rand(1, N)
assigns random values to all nodes present in the network. The random function
is used to reduce the computational complexity of the EESM and make the op-
timization process highly rely on the degree information of the nodes. With the
function, only the entropies of the nodes associated with the selected edges are
evaluated. The details of the proposed edge swap mechanism are discussed below.
Suppose the randomly generated probabilities of the nodes i,j,kand lare 0.034,
0.013, 0.026 and 0.004, respectively. The entropy is calculated using HT1, which
can be calculated in Equation (3.6):
HT1=X
q
pqlog2(pq).(3.6)
Here, HT1denotes the uncertainty in the degree of the nodes for the edges (i, j)
and (k, l) in the initial topology and qdefines the node. For the initial topology,
the entropies for the edges (i, j) and (k, l) are 0.2743 and 0.1688, respectively. The
total entropy H(T1) across the edge swap is 0.4431. Similarly, for the edges (i, k)
and (j, l), the entropies are 0.3028 and 0.1133, respectively and the total entropy
H(T2) is 0.4341. Also, for the edges (i, l) and (j, k), the entropies are 0.1977 and
0.2183, respectively and the total entropy H(T3) is 0.4160. The expected degree
length LT1[150] is calculated as:
26 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
LT1=X
q
kqpq.(3.7)
From Equation (3.8), kqis the node’s degrees of the selected independent edges of
the initial topology. The efficiency of the edge η[150] is measured in percentage
and is calculated from Equation (3.8):
ηT1=HT1
LT1
100.(3.8)
Using η, the efficiency of all possible pairs of independent edges are calculated. The
pair of edges having high efficiency among other swap pairs is selected as a robust
swap for improving the network robustness. The process of EES is described in
Algorithm 1. Here, we only describe the process of edge swap mechanisms as the
rest of the processes are already described in [93,98].
Algorithm 1 EESM
Input: Adjacency Matrix (A), Set of edges (E), Graph (G)
Output: A
1: procedure EESM(A)
2: for all edges in Edo
3: Assign the random probabilities to all nodes in N
4: Select two independent edges and mark them as eij and ekl
5: Calculate HT1,LT1and ηT1
6: Remove edges eij and ekl from A. Add eil and ejk in A1
7: G1Gand A1A
8: Remove edges eij and ekl from A. Add eik and ejl in A2
9: G2Gand A2A
10: Calculate the edge efficiency ηT2,ηT3and find ηmax=max(ηT1,ηT2,ηT3)
11: if R(A1)R(A) &&ηT2== ηmax then
12: AA1
13: else if R(A2)R(A) && ηT3== ηmax then
14: AA2
15: end if
16: end for
17: end procedure
In Algorithm 1, the random probabilities are assigned to all nodes in the network.
Two random edges are selected in the network, which are marked as ei,j and ek,l
(Line 4). For the nodes of the selected edges, the entropy H, length Land efficiency
ηfor the initial topology are calculated (Line 5). Moreover, the efficiency of all
the possible pairs of the edges is also calculated (Line 11). A pair of edges with
high efficiency is selected and the topology is updated accordingly (Lines 11-15).
27 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
3.3.2 Efficiency based Edge Swap Mechanism with Assor-
tativity
The assortativity is one of the important factors, which reduces the impacts of
malicious attacks on the scale-free networks. In assortativity, similar degree nodes
are connected with each other and the functionality of the nodes is handled by their
neighboring nodes when they fail. In the scale-free networks where the malicious
attacks occur on high degree nodes, the importance of assortativity is high. It is
because when a high degree node is removed from the network, its neighboring
node replaces its functionality to maintain the connectivity of the nodes. Thus,
the impacts of malicious attacks on the scale-free networks are greatly reduced.
This is the main reason why onion-like structure shows better network robustness
against the malicious attacks because the high degree nodes are connected with
each other in the structure.
Mathematically, the assortativity is denoted with rand is calculated as [100]:
r=PikiPij Aij kikj(Piki2)2
PikiPiki3(Piki2)2.(3.9)
From Equation (3.9), kdenotes the degree of a node and Ai,j is the adjacency
matrix (i, j = 1,2,3, ..., N ) [100]. The value of rranges from -1 to 1. When
r > 0, the similar degree nodes are connected with each other making the network
assortative. For r < 0, different degree nodes are connected with each other
making the network dissortative. For r= 0, the network is neither assortative nor
dissortative.
The proposed schemes in [93,98] consider changing the network topologies to make
them onion-like. However, they do not consider the assortativity while altering the
connections between the nodes. Therefore, in our proposed Efficiency based Edge
Swap Mechanism with Assortativity (EESM-Assortativity), we modify the process
of EES by introducing the assortativity in the network. A pair of independent
edges is selected from the network, namely, eij and ekl and the efficiency of the
selected pair of edges and assortativity of the network topology are calculated. The
edges are swaped by removing eij and ekl and replacing them with eil and ekj . In an
alternate swap, the edges eij and ekl are replaced with eik and ejl . The efficiency
and the assortativity of the topology are evaluated for both swaps. The swap
which improves the robustness as well as increases the efficiency and assortativity
of the network is considered as a robust swap for topology optimization.
28 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
Algorithm 2 EESM-Assortativity
Input: Adjacency Matrix (A), Set of edges (E), Graph (G)
Output: A
1: procedure EESM-Assortativity(A)
2: for all edges in Edo
3: Assign the random probabilities to all nodes in N
4: Select two independent edges and mark them as eij and ekl
5: Calculate rG,HT1,LT1and ηT1
6: Remove eij and ekl from A. Add eil and ejk in A1
7: G1Gand A1A
8: Remove edges eij and ekl from Aand add edges eik and ejl in A2
9: G2Gand A2A
10: Calculate the edge efficiency ηT2,ηT3and find ηmax=max(ηT1,ηT2,ηT3)
11: Calculate rG1and rG2
12: Find r=max(rG, rG1, rG2)
13: if r=rG1&& R(A1)R(A) && ηT2== ηmax then
14: AA1
15: else if r=rG2&& R(A2)R(A) && ηT3== ηmax then
16: AA2
17: end if
18: end for
19: end procedure
In Algorithm 2, the r,H,Land ηare calculated for initial topology (Line 6).
Then the pair of edges with high efficiency and assortativity is selected and the
topology is updated accordingly in (Lines 14-20).
3.4 Performance Evaluation
In this section, first the important parameters are listed, which are used in the
simulation of the scale-free network topologies. The parameters are shown in
Table 4.1. Then, the performances of our proposed topologies are compared with
different scale-free network topologies. The performance evaluation is divided into
two parts. In first part, the performance of MPGA Entropy is compared with
MPGA to show the importance of Entropy in the multi-population environment.
In second part, the performance of our proposed EESM and EESM-Assortativity
is evaluated using HC and SA algorithms. The proposed schemes, which consider
EESM in HC and SA are denoted as HCE and SAE, respectively. On the other
hand, the EESM-Assortativity based schemes are named as HCEA and SAEA,
respectively. MATLAB 2018a is used for the simulations on a laptop having Intel
Core i7 with 4GB RAM.
29 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
Table 3.1: Network Parameters used in Simulations
Parameters Values
Number of Nodes 100
Maximum Degree 25
Network Area 500 ×500 (m2)
Transmission Range 200 m
Edge Density (m) 2
maxiter 100
3.4.1 Comparison of Network Robustness of Scale-Free Topolo-
gies under Malicious Attacks
For robustness analysis, MPGA [107] is used as a comparison scheme for MPGA
Entropy. For EES case, we compare the proposed edge swap mechanisms with HC
and SA in a single-population environment. Initially, 100 nodes are considered in
the network. Later on, the nodes are increased from 100-250 nodes. A maximum
100 iterations are performed for the simulations.
0 20 40 60 80 100
Number of Iterations
0.1
0.15
0.2
0.25
0.3
Robustness
Figure 3.3: Comparison of Network Robustness of MPGA Entropy and
MPGA with Number of Iterations.
In the first scenario, as shown in Figure 3.3, our proposed MPGA Entropy shows
9% higher robustness as compared to MPGA. The use of entropy increases the
diversity of the solutions in the network because the worst solution of low entropy
population is replaced with the best solution of high entropy population. However,
30 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
no such phenomenon exists in MPGA; thus, its robustness is low. At the start,
MPGA has better robustness; however when the number of iterations increases,
the effectiveness of the proposed MPGA Entropy increases. It proves that MPGA
converges too early as compared to proposed MPGA Entropy. Thus, MPGA
Entropy overcomes the premature convergence effectively as it finds more optimal
solution through wide varieties of solutions. Moreover, the results in Figure 3.4
prove the effectiveness of MPGA Entropy for high node density as well.
Figure 3.4: Comparison of Network Robustness of MPGA Entropy and
MPGA with Number of Nodes.
The results in Figure 3.5 show that the proposed HCE and SAE have high robust-
ness against the malicious attacks as compared to HC and SA, respectively. One
of the major issues in HC is the number of redundant operations in the network.
There are some pair of edges, which do not improve the network robustness. Due
to randomness of the edge swap in the HC, these edges are selected again during
the iteration process. Therefore, HC fails to overcome the redundant operations
in the network. On the other hand, SA reduces the redundancy by marking the
pair of edges, which are selected during the optimization process. Thus, it has
high robustness as compared to HC. However, still the optimization of network
robustness in SA is too slow and it does not provide global solutions in the given
time of frame. Moreover, SA only considers improving the network robustness by
changing the connections of edges. Other factors like assortativity, degree differ-
ence, etc., are neglected during the optimization process, thus, results in reduction
31 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
of the network robustness. In addition, SA does not consider nodes’ degree during
the edge swap operation and the edge swap process changes the property of the
onion-like structure. Therefore, the consideration of efficiency in HCE and SAE
for finding the optimal number of swaps in the network increases the network
robustness.
HCE and SAE consider constructing the topology with a better degree distribu-
tion, which helps to make the network robust against the malicious attacks. Over-
all, the proposed mechanisms HCEA and SAEA outperform all other mechanisms
because they consider efficiency and assortativity to make connections between
similar degree nodes. SAEA has high efficacy and is considered the best among
all the proposed schemes for constructing a robust network topology. The use of
assortativity helps to select the topology, where the connections between similar
degree nodes are high. Thus, the effects of malicious attacks are greatly reduced
when high degree nodes are removed from the network. This is because the other
high degree nodes in the network maintain the functionality of the removed nodes
after the malicious attacks. Thus, the topology shows better robustness by main-
taining the connections of most of the nodes in the subgraph. Moreover, the
efficacy of the proposed SAEA is also tested on different nodes’ density and the
results in Figure 3.6 show that SAEA has better network robustness against the
malicious attacks at low and high nodes’ density.
0 20 40 60 80 100
Number of Iterations
0.1
0.15
0.2
0.25
0.3
Robustness
Initial
HC
SA
HCE
SAE
SAEA
HCEA
Figure 3.5: Comparison of Network Robustness of Scale-Free Topologies under
Malicious Attacks with respect to Number of Iterations.
32 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
100 150 200 250
Number of Nodes
0
0.05
0.1
0.15
0.2
Robustness
Initial
HC
SA
HCE
SAE
HCEA
SAEA
Figure 3.6: Comparison of Network Robustness of Scale-Free Topologies under
Malicious Attacks with respect to Number of Nodes.
3.4.2 Comparison of Graph Density of Scale-Free Topolo-
gies under Malicious Attacks
The graph density is the ratio of the number of edges to the maximum possible
number of edges in the graph G. Mathematically, it is written as:
GraphDensity =2E
N(N1).(3.10)
The graph density given in Equation (3.10) is directly related with the network
robustness. The higher graph density shows that a large number of nodes are
connected with each other. A network must maintains high connectivity among
nodes in order to make itself robust against the malicious attacks. Figure 3.8 shows
the graph density comparison of the proposed topologies with HC and SA. We have
performed 100 simulations to get the graph density of the mechanisms. SAEA has
the highest graph density value because it has better maintained the connectivity
between the nodes as compared to other schemes. If the connectivity between
the nodes is high, the effects of the malicious attacks are greatly reduced and the
topology shows high robustness against the attacks. Moreover, for large number
of connections in the topology, the number of nodes in the maximum connected
subgraph also increases. The connections of more nodes in the graph increases the
graph density. The other reason that increases the graph density is the inclusion
33 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
HC
HCE
HCEA
Initial
SA
SAE
SAEA
0
0.01
0.02
0.03
0.04
0.05
Graph Density
Figure 3.7: Comparison of Graph Density of Scale-Free Topologies under
Malicious Attacks.
of assortativity in SAEA, which connects similar degree nodes with each other.
Thus, it constructs a better onion-like topology. Similarly, HCEA, HCE and SAE
outperform the scale-free topologies, which are based on HC and SA. The efficiency
based edge selection in HCE and SAE provides more robust topologies. Also, the
consideration of global information of the network using assortativity in HCEA
and SAEA makes the topology onion-like as it connects similar degree nodes with
each other. On the other hand, HC has lower graph density among other schemes
because of poor selection of edge swap and large number of redundant operations
during the optimization process. HC does not consider the degree information
of nodes while performing edge swap operation. Thus, the nodes tend to loose
connections with each attack and the network damages quickly as compared to
others.
Table 3.2: Network Parameters
Number of
Nodes (N)
Network
Field
(m2)
Maximum
Degree
Communication
Range (m)
50 50 x 50 10 25
100 100 x 100 20 50
150 150 x 150 30 75
34 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
3.4.3 Comparison of Robustness of Scale-Free Topologies
with different Network Parameters under Malicious
Attacks
To verify the performance of our proposed mechanisms, we evaluate the robustness
of the network by considering different network parameters. We vary the number
of nodes, network field, maximum degree of nodes and their communication range
to analyze the robustness of the scale-free network topologies. The network pa-
rameters are given in Table 3.2. From Figure 3.8, it is obvious that the robustness
of the network is high when the number of nodes is set to 50. The small networks
have high robustness against the malicious attacks. This is because the robustness
is inversely related with the number of nodes present in the network. Therefore,
lowering the number of nodes increases the robustness. In all cases, our proposed
SAEA outperforms SA and HC, which shows the effectiveness of the efficiency and
assortativity for making the network robust against the malicious attacks. The
efficiency based mechanism in SA selects pair of nodes with better degree distri-
bution as compared to the initial topology and the assortativity in SAEA is used
to construct an onion-like topology.
50 100 150
Network Parameters
0
0.05
0.1
0.15
0.2
Robustness
Initial
HC
SA
HCE
SAE
HCEA
SAEA
Figure 3.8: Comparison of Robustness of Scale-Free Topologies with different
Network Parameters under Malicious Attacks.
35 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
3.4.4 Comparison of Connectivity of Initial Topology and
SAEA under Malicious and Random Attacks
We analyze the connectivity of our proposed scale-free topology SAEA by conduct-
ing malicious and random attacks. The results are simulated over 100 iterations.
From Figure 3.9, the connectivity of SAEA under malicious attacks is better as
compared to initial topology, which shows the effectiveness of the edge swap mech-
anism performed in SAEA. For the case of random attacks as shown in Figure 3.10,
the initial topology and SAEA show high robustness against the random attacks,
which shows that both topologies are scale-free. Moreover, when the number of
removed nodes increases the subgraph size decreases. It shows that the connec-
tivity of the network is greatly effected when more nodes are removed from the
network.
0 20 40 60 80 100
Number of Malicious Attacks
0
20
40
60
80
100
Maximum Connected Subgraph Size
Initial
SAEA
Figure 3.9: Comparison of Connectivity of Initial Topology and SAEA under
Malicious Attacks.
3.4.5 Comparison of Computational Time of Scale-Free
Topologies under Malicious Attacks
Figure 3.11 shows the evaluation of the proposed SAEA with HC and SA in terms
of computational time. It shows that HC has the lowest computational time
because it does not store the information of the previous solution thus, it does
not perform any extra calculation at each step. SA stores the information of
36 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
0 20 40 60 80 100
Number of Random Attacks
20
40
60
80
100
Maximum Connected Subgraph Size
Initial
SAEA
Figure 3.10: Comparison of Connectivity of Initial Topology and SAEA under
Random Attacks.
previous solution as it accepts the worst solution and tries to make it better thus,
its computational time is high as compared to HC. However, our proposed scheme
SAEA evaluates the efficiency of the edges as well as the network assortativity at
each edge swap. Thus, its computational time is high as compared to HC and SA.
HC SA SAEA
0
2
4
6
8
10
12
14
Computational Time (secs)
Figure 3.11: Comparison of Computational Time of Scale-Free Topologies
under Malicious Attacks.
37 Thesis by: Muhammad Awais Khan
Chapter 3 Modification Strategies
3.5 Conclusion of the Chapter
The study of entropy in WSNs is not new. However, when focusing on improving
the robustness, a good choice is to perform topology optimization with entropy.
Therefore, this chapter has considered the concept of entropy in multi-population
and single population environments for finding the optimal solution in the scale-
free networks. The entropy in multi-population is useful during the replacement
of the worst solution of one population with the best solution of other population
as it helps to increase the network robustness. Moreover, the selection of edges
based on their efficiency makes the network robust. Also, the assortativity helps to
construct an onion-like topology. The network robustness using the proposed edge
swap mechanisms is optimized through HC and SA. The simulations results have
shown the adequacy of the proposed network topologies in terms of improving the
network robustness against the malicious attacks. Therefore, the proposed modifi-
cation strategies are vital in maintaining the connectivity of the nodes during the
malicious attacks, which is proved from the result of the graph density. However,
the proposed edge swap mechanism SAEA has high computational time and it is
a trade-off in this chapter.
38 Thesis by: Muhammad Awais Khan
Chapter 4
Computationally Efficient Topology
Optimization for Scale-Free IoT Networks
39
Chapter 4 Computationally Efficient Topology Optimization
4.1 Summary of the Chapter
In this chapter, four solutions are presented to reduce the computational cost
of the optimization process in the scale-free IoT networks. First, a Smart Edge
Swap Mechanism (SESM) is proposed to overcome the excessive randomness of the
standard Random Edge Swap Mechanism (RESM). Second, a threshold based node
removal method is introduced to reduce the operation of the edge swap mechanism
when an objective function converges at a point. Third, multiple attacks are
performed in the network to find the correlation between the measures, which are
degree, betweenness and closeness centralities. Fourth, based on the third solution,
a Heat Map Centrality (HMC) is used that finds the set of most important nodes
from the network. The HMC damages the network by utilizing the information
of two positively correlated measures. It helps to provide a good attack strategy
for robust optimization. The simulation results demonstrate the efficacy of the
proposed SESM mechanism. It outperforms the existing RESM mechanism by
almost 4% better network robustness and 10% less number of swaps. Moreover,
64% removal of nodes helps to reduce the computational cost of the network.
4.2 Scale-Free Network Modeling
In this section, we discuss the construction of a scale-free network, its robustness
measure and the independent selection of edges from the network.
4.2.1 Construction of a Scale-Free Network
The authors consider a BA model [116] that utilizes the information of the ini-
tially deployed nodes to construct a scale-free network topology. The preferential
attachment property of the BA model allows the newly added nodes to make con-
nections with the high degree nodes in the network. However, due to the limited
transmission range, the newly joined nodes have limited neighbors in their commu-
nication range. Also, it is important for the nodes in the network to have sufficient
neighbors in their communication range due to the growing demand of dense net-
work topologies in the future. The ROSE emphasizes the importance of the dense
WSNs and takes into account the communication range of nodes in the network.
The communication range in ROSE allows the nodes to connect with 50% of the
nodes in the network, making it a dense scale-free network. Moreover, the ROSE
analyzes that for limited transmission range, the division of a network into multi-
ple clusters is a good choice to develop a robust network. Therefore, the following
40 Thesis by: Muhammad Awais Khan
Chapter 4 Computationally Efficient Topology Optimization
aspects of ROSE are considered in the proposed model for the construction of the
scale-free network.
1. The preferential attachment property of a node is limited to its nodes, which
are within its communication range.
2. Considering the limited resources of the nodes in IoT-WSNs, their maximum
degree is limited to a certain threshold.
3. The high degree nodes must be located in the center of the network.
4.2.2 Network Robustness Measure
In the scale-free networks, the attacker can attack the nodes as well as the links
to destroy the connectivity of the nodes in the network. Generally, the attacks
can be random or malicious. The random attacks remove random nodes while
the malicious attacks remove the most important nodes from the network. In
the scale-free networks, we use malicious attacks, which remove the high degree
nodes and damage the connectivity of the network. Initially, the degrees of nodes
are calculated and the node with the highest degree is removed. Also, the edges
connected with the node are also removed. Then, the degree of the nodes is
recalculated and the highest degree node is removed again. The process is repeated
many times until all nodes are removed from the network.
For calculating the network robustness, a metric Rproposed by Schneider et al.
[93] is used based on percolation theory. When a node is removed from the net-
work, the graph is divided into multiple subgraphs. The connectivity of the nodes
is checked and the subgraph where the nodes are maximally connected is con-
sidered for the evaluation of R. We take the mathematical equation from [102]
for evaluating the robustness, which provides the information of the nodes in the
maximal connected subgraphs after removing nth nodes from the network. The
equation for evaluating the robustness in [93] also provides the number of nodes
information in the maximal connected subgraphs, however, it considers the frac-
tion of nodes, which needs to be removed in order to disconnect the entire network.
Both are similar in terms that they both provides the information of the network
connectivity after repeated removal of nodes in the network. The equation for
evaluating the network robustness [102] is given as follows.
R=1
N+ 1
N1
X
n=0
MCSn
N.(4.1)
41 Thesis by: Muhammad Awais Khan
Chapter 4 Computationally Efficient Topology Optimization
i
j
lk
(a)
i
j
lk
(b)
i
j
lk
(c)
Figure 4.1: (a) Initial Topology (b) Swap 01 (c) Swap 02
From Equation (4.1), Ndenotes the total number of nodes and M CSndenotes
the maximal connected subgraphs after nth highest degree node removal from the
network [102].
4.2.3 Network Optimization through Selection of Indepen-
dent Edges
In the scale-free networks, the optimization is performed through edge swap mech-
anism by selecting two independent edges from the graph G= (V, E). Where, V
represents the set of nodes and Erepresents the set of edges. The two selected
edges are said to be independent if they lie within the communication range of
each other and there is no extra connection between these two edges. Figure 1a
shows that ei,j and ek,l are the independent edges. Figure 1b and Figure 1c show
the edge swap performed on these independent edges.
The optimization of the network robustness against the malicious attacks is evalu-
ated by swapping the independent edges in the network. The edges are swapped in
such a way that the updated topology increases the network robustness against the
malicious attacks. If the first swap increases the network robustness, the topology
is updated. If the first swap has low robustness value, the second swap is per-
formed and the topology is updated only if it increases the robustness. If both
swaps fail to optimize the network robustness, the original topology is considered
in the network.
42 Thesis by: Muhammad Awais Khan
Chapter 4 Computationally Efficient Topology Optimization
L1
S1
Independent Edges Selection
Selection of Nodes based on Importance
Edge Swap Process
NodesRemoval Process
Random Selection of Edges Unnecessary Optimization
Smart Selection of Edges S2 Threshold based Node Removal
Robustness Calculation
L3 L4
L2
Multiple Malicious Attacks Betweenness Centrality
increases Computational Cost
S3 S4
Combined Attacks are
considered HMC is considered
i
jlk
i
jlk
i
jk
l
Figure 4.2: Limitations Identified and their Proposed Solutions
4.3 Computationally Efficient Topology Optimiza-
tion: Overview
This section describes our proposed topology optimization mechanism where we
identify four limitations. Each limitation is associated with the optimization of
the network robustness in the scale-free network, as shown in Figure 4.2. The
limitations are denoted as L1, L2, L3 and L4, while their proposed solutions are
provided using S1, S2, S3 and S4, respectively. Table 4.1 shows the mapping
of these limitations with their proposed solutions and validations. L1 and L2
mentioned in Table 4.1 are associated with the limitations in the previous edge
swap mechanism based on redundancy and computational cost of the network.
These limitations are tackled using SESM and threshold based node removal,
respectively. The validation for both these solutions is done using R, number of
swaps, MCS, etc.
For L3, the issue of multiple attacks on the network is tackled using S3, where a
combined attack strategy of the two strongly correlated measures is needed. In
contrast, L4 is associated with finding an attack measure to damage the network’s
connectivity in quick time, as mentioned in Table 4.1. Therefore, S4 introduces a
measure named as Heat Map Centrality (HMC) to overcome L4. We discuss the
solution of each limitation in the given subsections.
43 Thesis by: Muhammad Awais Khan
Chapter 4 Computationally Efficient Topology Optimization
Table 4.1: Limitations Identified, Proposed Solutions and Validations
Limitations Identified Solutions Proposed Validations Done
L1: Random selection of
independent edges increases
the computational cost of
the network
S1:SESM reduces the
randomness through a
smart selection of edges
V1: The performance of the
network is validated
through R(Figure 4.10 and
Figure 4.11) and number of
swaps (Figure 4.12 and
Figure 4.13)
L2: Unnecessary removal of
nodes after the convergence
is achieved increases the
computational overhead
S2: We analyze the
robustness value and set a
threshold for node removal
to reduce the
computational overhead
V2: The efficacy of the
network is validated using
MCS/N for node removal
in the network (Figure 4.5)
L3: Multiple attacks can
happen on the network
simultaneously
S3: Finding the two
strongly correlated
measures to make the
network robust against
multiple attacks
V3: The validation is
provided using Pearson
correlation coefficient
(Figure 4.4), execution time
(Figure 4.6 and Figure 4.7)
and R(Figure 4.9)
L4: Finding the set of most
influential nodes in the
network that can damage
the network in less time is a
challenging task
S4: HMC reduces the
computational cost of the
network by damaging the
network to a greater extent
V4: The performance
parameters used for
validation are R
(Figure 4.11) and number
of swaps (Figure 4.13)
4.3.1 Smart Edge Swap Mechanism
Due to the involvement of randomness in the edge swap mechanism used in HC
and ROSE, many redundant operations are generated in the optimization of the
network robustness. Specifically, in HC, the edge swap mechanism increases the
number of redundant operations in the network because it does not mark the
independent edges after their selection. Thus, these edges are selected again in
the optimization process, which results in increasing the number of redundant
operations in the network. In ROSE, the marking of independent edges reduces the
redundancy of the network. However, the random edge swap mechanism happens
on low degree edges in the network, which results in providing low robustness
against the malicious attacks with the high computational cost. Therefore, a new
selection criteria for independent edges is required to overcome the redundancy
issue and increase the network robustness. It is known that the high degree nodes
are the main target of the attacker and the removal of high degree nodes damages
the topology of the network. Therefore, based on the information of the high degree
nodes, one can protect the connectivity of the nodes by altering their connections.
Furthermore, it is understood that one high degree node replaces other high degree
node in the network. Thus, changes in the connections of these nodes can bring a
44 Thesis by: Muhammad Awais Khan
Chapter 4 Computationally Efficient Topology Optimization
significant improvement in the network. Besides, an onion-like structure has strong
tolerance against the malicious attacks and in the structure, the high degree nodes
are tightly connected with other high degree nodes. Therefore, the edge swap
mechanism based on high degree nodes is a good choice to construct an onion-like
structure.
For L1, as shown in Table 4.1, the problem of edge randomization is controlled
through the selection of high degree nodes. The SESM is proposed to overcome
the random selection of the independent edges in the previous Random Edge Swap
Mechanism (RESM). For the initial topology, a set of high degree nodes is selected
from the network. From the set, two high degree nodes are selected before per-
forming the edge swap mechanism. The information of the neighboring nodes of
these two selected nodes is extracted. The selection process of finding a node
from the neighbors’ set is a complex problem as each node has multiple neighbors.
Therefore, to avoid this problem, a random neighbor is selected from the neigh-
bor’s set. The information of the selected neighbor is utilized for the selection of
independent edges. The independence of the selected edges is checked. If they
are independent, the edge swap operation is performed, else the selection process
continues to try further connections to find the independent edges. The edge swap
mechanism swaps the edges of the network in search for a more optimized network
topology.
Algorithm 3 Smart Edge Swap Mechanism
1: procedure Smart Edge Swap Mechanism(A)
2: Input: A,N,G
3: for all NGdo
4: Find a high degree node from N
5: Calculate neighbors of the high degree node and mark this node as i
6: Pick a neighbor randomly from the neighbors of node iand mark it as
j
7: Perform steps 3-5 again for 2nd high degree node kand its neighbor l
8: Swapcounter = 0
9: if (i, j) and (k, l) are pairs of independent edges in the set Ethen
10: Perform optimization using HC and ROSE
11: if swap is successful then
12: Update Aand G
13: Calculate R
14: Swapcounter=Swapcounter + 1
15: end if
16: end if
17: end for
18: end procedure
45 Thesis by: Muhammad Awais Khan
Chapter 4 Computationally Efficient Topology Optimization
Algorithm 3describes the process of SESM. The high degree node is selected from
graph G, which consists of Nnumber of nodes (Line 4). The neighbor of the
selected node is chosen and the link between them is marked (i, j) (Line 6). The
steps 3-5 are followed for other high degree nodes (Line 7). The swap counter is
initialized (Line 8). If the selected edges (i, j ) and (k, l) are independent, the edge
swap mechanism is performed. For optimization, the operation of HC and ROSE
are used (Line 10). If the swap is successful and the robustness is increased, the
swap counter is incremented (Line 14).
4.3.2 Threshold based Node Removal
Several mechanisms including HC and ROSE increase the network robustness by
focusing on changing the network topology through swapping. Due to the struc-
tural complexity of the network, optimizing the network robustness using an edge
swap becomes a difficult task. Moreover, there is not enough evidence to guide
the network to perform limited edge swaps. Besides, it is understood that the
performance of the network is greatly reduced when a specific optimization task
is performed continuously without any improvement. In the optimization pro-
cess, analyzing the convergence of an objective function is an important factor. It
is useless to perform unnecessary optimization for an objective function after its
maximum value is achieved. In the topology optimization scenario, the objective
is to maximize the network robustness by swapping the edges of the topology un-
til a single node is left in the network. However, this type of process consumes
excess memory and increases the computational cost of the network. Therefore,
based on these problems, a threshold based node removal method is considered
as mentioned in Table 4.1. The method considers removing the nodes one by one
until convergence is achieved. The details of the proposed solution are described
below.
Consider a network whose topology is constructed using the previous BA model.
The preferential attachment property of the scale-free network guides the newly
added node to connect with high degree nodes in the network. These nodes are
added into the network one by one until a topology of the scale-free is generated.
The network robustness is calculated for initial topology by removing the nodes one
by one. For node removal, we have performed several experiments and found out
that almost 60-65% node removal is enough to destroy the network’s connectivity.
The 60-65% node removal guides us to select a threshold value for node removal,
where the robustness value reaches its maximum. Therefore, the node removal is
performed based on the given threshold. For each node removal, the edge swap
46 Thesis by: Muhammad Awais Khan
Chapter 4 Computationally Efficient Topology Optimization
mechanism swaps the independent edges within the given threshold. The topology
that maximizes the Rvalue is considered for the construction of a robust scale-free
network.
Algorithm 4 Threshold Based Node Removal
1: procedure Threshold Based Node Removal(A)
2: Input: A,N,G
3: for all NGdo
4: for k= 1 : N1do
5: Find a high degree node ifrom topology G
6: Remove the node iand update the topology from Gto G2
7: Calculate MCS and evaluate network robustness Ri
8: if Rk> Rk1then
9: Continue the removal process
10: else if Rk==Rk1then
11: Calculate the value for node removal where Rk==Rk1
12: Update the threshold for node removal and perform optimization
using Algorithm 3 with the selected node removal phase
13: end if
14: end for
15: end for
16: end procedure
In Algorithm 4, the threshold based node removal is discussed. The high degree
node is removed from graph G(Line 6). The robustness Ris calculated for each
node removal (Line 7). If the value of Rkis greater than the value of Rk1, the node
removal process is repeated again until all nodes are removed from the network.
If Rkis equal to the previous value Rk1for consecutive node removal steps, the
threshold value is recalculated. The optimization is performed using Algorithm 1
with the node removal at the given threshold (Line 12).
4.3.3 Optimization of Network Considering Multiple At-
tacks
The authors in HC and ROSE have analyzed the network robustness against high
degree node removal. They consider that the degrees of the nodes can provide
more influential nodes from the network. However, the research performed in
[100] has termed betweenness centrality as another metric to measure the most
important nodes from the network. Therefore, in [100], the authors have combined
both measures to find the attack probability of nodes. Still, they have failed to
provide enough evidence about the importance of both these measures in terms
47 Thesis by: Muhammad Awais Khan
Chapter 4 Computationally Efficient Topology Optimization
of computational cost. The work proposed in [101] has discussed both these mea-
sures in terms of computational cost. The authors have considered the degree of
nodes as an excellent parameter to find the local information of nodes. However,
choosing betweenness centrality is not the best option because it increases the
computational cost of the network [101].
To determine the relationship between any two measures, the Pearson correlation
coefficient is used, which is a well-known correlated measure. In our scenario,
initially, three attacks named as degree, betweenness and closeness are induced.
Then, based on these attacks, three robustness measures are calculated. After-
wards, the Pearson correlation coefficient between the robustness of any two cen-
trality measures is evaluated using the following formula.
r=N(PRC1RC2)(PRC1)(PRC2)
p[N(PR2
C1)(PRC1)2][N(PR2
C2)(PRC2)2],(4.2)
From Equation (4.2), RC1and RC2are the evaluated robustness for any two cen-
trality measures. Figure 4.3 shows the correlation comparison of the three central-
ity measures. From Figure 4.3, the strong positive correlation between degree and
closeness attacks shows that both these measures can be considered simultane-
ously in an attack to improve the robustness. The idea of finding the correlation
between the measures is adopted from [130], where the authors use two strong
negatively correlated measures for multi-objective optimization. The optimiza-
tion of both the measures are necessary to increase the robustness. Contrary to
the aforementioned case, the strong positively correlated measures are combined
together to improve the robustness. <