ThesisPDF Available

Enhancing robustness of scale-free IoT networks against random and malicious attacks (MS Thesis without Source Codes)

Authors:

Abstract and Figures

During the past few decades, the Internet of Things (IoT) has made remarkable progress in many real-world applications including healthcare, military, transportation, etc. Multiple sensor nodes are deployed in these _elds to get the required data. Different network topologies are used in IoT and scale-free is one of them. It is mostly preferred due to its robust behavior against random node removal, however, the network collapsed because of malicious attacks. Therefore, in this thesis, robustness of the scale-free networks is enhanced against malicious attacks through optimization. To achieve this, the edge's degree and nodes' distance based edge swap operations are used in the proposed Improved Scale-Free Networks (ISFNs) scheme. In the edge's degree based operation, nodes of similar degrees are linked. Moreover, the connections of the nearest nodes are made in distance based edge swap. These operations help to achieve a better onion-like structure without changing the degree distribution of the network. Therefore, the network becomes robust against malicious attacks. Moreover, no new links or nodes are added in the optimization process, therefore, no extra cost is incurred. Furthermore, to make the network more robust against realistic attacks, the variable attacks are considered. Simulation results of the proposed scheme are compared with ROSE and Simulated Annealing (SA) for different number of nodes. The proposed scheme outperforms the existing techniques for different numbers of nodes and against the low degree, high degree and random attacks. Moreover, ISFNs has 13% and 23% better network robustness as compared to ROSE and SA, respectively. Network Topology Evolution Scheme (NTES) is proposed to prevent the scale-free networks from random and malicious attacks. In this scheme, the network field is divided into two parts with uniformly distributed nodes. After the network's evolution, the nodes are linked with each other through one-to-many correspondence. The division of the network field is made by considering that a network is robust if its size is small. Moreover, to study the hierarchical changes in the degree of nodes, k-core decomposition is used. In addition, nodes' degrees and core based attacks are performed on the network to evaluate the performance of the proposed scheme. Furthermore, the network robustness is analyzed using three optimization techniques: Artificial Bee Colony (ABC), Bacterial Foraging Optimization (BFO) and Genetic Algorithm (GA). The techniques are compared with each other and a technique that efficiently optimizes the network to increase the robustness is selected. In the optimization process, we make use of three edge swap methods. Due to the edge swap, the network robustness is enhanced without changing the degree distribution, so the addition of nodes/links is not required to increase the robustness. Furthermore, NTES is compared with Barabasi Albert (BA) model and Hill Climbing (HC) algorithm against random and malicious attacks. The simulation results show that the proposed NTES optimized using GA outperforms BA and HC by 46.90% and 57.08%, respectively, in terms of robustness. In addition, the network robustness of Scale Free Networks (SFNs) is enhanced against the malicious attacks. For that purpose, initially, a parameterless optimization algorithm JAYA is used because it requires less computational efforts as compared to the heuristic techniques. Then, as the edge swap plays an important role to enhance the robustness of SFNs, therefore, the edge swaps are classified into three categories. For each category, effects on the network's topological parameters such as average shortest path length, assortativity and clustering coefficient are analyzed. Next, the robustness is enhanced with the addition of nodes in the maximum connected subgraphs and the protection of bridge edges maintain the network connectivity. Moreover, optimized network is analyzed for different attack strengths. In simulations, the comparison of JAYA is made with two existing algorithms: ROSE and Simulated Annealing (SA). The network optimized by JAYA has a better robustness against random and malicious attacks, as compared to the existing algorithms. Furthermore, among the edge swap categories, the degree dependent edge swap is better to increase the robustness of SFNs. Moreover, the addition of nodes into the maximum connected subgraphs enhances the robustness and the protection of bridge edges ensures the network connectivity in all the algorithms. Furthermore, the robustness against different attack strengths are analyzed and the results show that high attacks strength paralyzed the network more efficiently.
Content may be subject to copyright.
Enhancing robustness of scale-free IoT networks
against random and malicious attacks (MS Thesis
without Source Codes)
By
Muhammad Usman
CIIT/FA18-REE-016/ISB
MS Thesis
In
Electrical Engineering
COMSATS University Islamabad, Islamabad - Pakistan
Spring, 2021
i
COMSATS University Islamabad
Enhancing robustness of scale-free IoT networks
against random and malicious attacks (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 Usman
CIIT/FA18-REE-016/ISB
Spring, 2021
ii
Enhancing robustness of scale-free IoT networks
against random and malicious attacks (MS Thesis
without Source Codes)
A Post Graduate Thesis submitted to the Department of Electrical and Computer
Engineering as partial fulfillment of the requirement for the award of Degree of
MS (Electrical Engineering).
Name Registration Number
Muhammad Usman CIIT/FA18-REE-016/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
Enhancing robustness of scale-free IoT networks against random
and malicious attacks (MS Thesis without Source Codes)
By
Muhammad Usman,
CIIT/FA18-REE-016/ISB
has been approved
For the COMSATS University Islamabad, Islamabad
External Examiner:
Dr. Ataul-Aziz Ikram
Professor, Department of Electrical Engineering,
FAST-NU, Islamabad
Supervisor:
Dr. Nadeem Javaid
Associate Professor, Department of Computer Science,
COMSATS University Islamabad, Islamabad
Co-Supervisor:
Dr. Sardar Muhammad Gulfam,
Assistant Professor, Department of Electrical and Computer Engineering,
COMSATS University Islamabad, Islamabad
HoD:
Dr. Shurjeel Wyne
Associate Professor, Department of Electrical and Computer Engineering,
COMSATS University Islamabad, Islamabad
iv
Declaration
IMuhammad Usman (Registration No. CIIT/FA18-REE-016/ISB) hereby declare
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: July 09, 2021
Muhammad Usman
CIIT/FA18-REE-016/ISB
v
Certificate
It is certified that Muhammad Usman (Registration No. CIIT/FA18-REE-016/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: July 09, 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
DEDICATION
Dedicated
to my mentor loving Parents, family members and to my Supervisor
Dr. Nadeem Javaid, who motivated and guided me, when I needed.
vii
ACKNOWLEDGEMENT
I am very grateful to Almighty Allah, without His help nothing is possible.
Whenever I feel discouraged, helpless and worried, Allah helps me.
I feel great pleasure in expressing my profound and heartiest gratitude to my super-
visor Dr. Nadeem Javaid and co-supervisor Dr. Sardar Muhammad Gulfam, for
their indispensable guidance, deep consideration, affection and active co-operation
that made possible this work to meet its end successfully well in time.
I would also like to thank all of the members of the ComSens family. Their
company helped me to develop the research attitude and become familiar with the
research.
In the end, I would like to thank all my family members. Their support and trust
helped me to achieve new goals.
viii
ABSTRACT
Enhancing robustness of scale-free IoT networks against
random and malicious attacks (MS Thesis without Source
Codes)
During the past few decades, the Internet of Things (IoT) has made remark-
able progress in many real-world applications including healthcare, military, trans-
portation, etc. Multiple sensor nodes are deployed in these fields to get the re-
quired data. Different network topologies are used in IoT and scale-free is one
of them. It is mostly preferred due to its robust behavior against random node
removal, however, the network collapsed because of malicious attacks. Therefore,
in this thesis, robustness of the scale-free networks is enhanced against malicious
attacks through optimization. To achieve this, the edge’s degree and nodes’ dis-
tance based edge swap operations are used in the proposed Improved Scale-Free
Networks (ISFNs) scheme. In the edge’s degree based operation, nodes of similar
degrees are linked. Moreover, the connections of the nearest nodes are made in
distance based edge swap. These operations help to achieve a better onion-like
structure without changing the degree distribution of the network. Therefore, the
network becomes robust against malicious attacks. Moreover, no new links or
nodes are added in the optimization process, therefore, no extra cost is incurred.
Furthermore, to make the network more robust against realistic attacks, the vari-
able attacks are considered. Simulation results of the proposed scheme are com-
pared with ROSE and Simulated Annealing (SA) for different number of nodes.
The proposed scheme outperforms the existing techniques for different numbers
of nodes and against the low degree, high degree and random attacks. Moreover,
ISFNs has 13% and 23% better network robustness as compared to ROSE and SA,
respectively.Network Topology Evolution Scheme (NTES) is proposed to prevent
the scale-free networks from random and malicious attacks. In this scheme, the
network field is divided into two parts with uniformly distributed nodes. After the
network’s evolution, the nodes are linked with each other through one-to-many
correspondence. The division of the network field is made by considering that a
network is robust if its size is small. Moreover, to study the hierarchical changes
in the degree of nodes, k-core decomposition is used. In addition, nodes’ degrees
and core based attacks are performed on the network to evaluate the performance
of the proposed scheme. Furthermore, the network robustness is analyzed using
three optimization techniques: Artificial Bee Colony (ABC), Bacterial Foraging
Optimization (BFO) and Genetic Algorithm (GA). The techniques are compared
with each other and a technique that efficiently optimizes the network to increase
ix
the robustness is selected. In the optimization process, we make use of three
edge swap methods. Due to the edge swap, the network robustness is enhanced
without changing the degree distribution, so the addition of nodes/links is not re-
quired to increase the robustness. Furthermore, NTES is compared with Barab´asi
Albert (BA) model and Hill Climbing (HC) algorithm against random and ma-
licious attacks. The simulation results show that the proposed NTES optimized
using GA outperforms BA and HC by 46.90% and 57.08%, respectively, in terms
of robustness.In addition, the network robustness of Scale-Free Networks (SFNs)
is enhanced against the malicious attacks. For that purpose, initially, a parame-
terless optimization algorithm JAYA is used because it requires less computational
efforts as compared to the heuristic techniques. Then, as the edge swap plays an
important role to enhance the robustness of SFNs, therefore, the edge swaps are
classified into three categories. For each category, effects on the network’s topolog-
ical parameters such as average shortest path length, assortativity and clustering
coefficient are analyzed. Next, the robustness is enhanced with the addition of
nodes in the maximum connected subgraphs and the protection of bridge edges
maintain the network connectivity. Moreover, optimized network is analyzed for
different attack strengths. In simulations, the comparison of JAYA is made with
two existing algorithms: ROSE and Simulated Annealing (SA). The network op-
timized by JAYA has a better robustness against random and malicious attacks,
as compared to the existing algorithms. Furthermore, among the edge swap cat-
egories, the degree dependent edge swap is better to increase the robustness of
SFNs. Moreover, the addition of nodes into the maximum connected subgraphs
enhances the robustness and the protection of bridge edges ensures the network
connectivity in all the algorithms. Furthermore, the robustness against different
attack strengths are analyzed and the results show that high attacks strength
paralyzed the network more efficiently.
x
Conference Proceedings
1Usman, M., Javaid, N., Abbas, S. M., Javed, M. M., Waseem, M. A., &
Owais, M. (2021, July). A novel approach to network’s topology evolution
and robustness optimization of scale free networks. In Conference on Com-
plex, Intelligent, and Software Intensive Systems (pp. 214-224). Springer,
Cham. Download
2 Abbas, S. M., Javaid, N., Usman, M., Baig, S. M., Malik, A., & Rehman,
A. U. (2021, July). An Efficient Approach to Enhance the Robustness of
Scale-Free Networks. In International Conference on Innovative Mobile and
Internet Services in Ubiquitous Computing (pp. 76-86). Springer, Cham.
Download
3Usman, M., Javaid, N., Khalid, A., Nasser, N., & Imran, M. (2020, June).
Robustness Optimization of Scale-Free IoT Networks. In 2020 International
Wireless Communications and Mobile Computing (IWCMC) (pp. 2240-
2244). IEEE. Download
xi
TABLE OF CONTENTS
Dedication vii
Acknowledgements viii
Abstract ix
Conference Proceedings 95
List of Figures xv
List of Tables xvii
1 Introduction 1
1.1 Introduction ............................... 2
1.1.1 Details of scale-free networks .................. 2
1.1.2 Contributions .......................... 4
1.1.3 Organization of thesis ..................... 5
2 Literature review and problem statement 6
2.1 Literature review ............................ 7
2.2 Problem statement ........................... 18
2.2.1 Details of problem statement ................. 18
3 Edge swap based optimization strategy to enhance the robustness of
Scale-Free Networks (SFNs) 19
3.1 Summary of the chapter ........................ 20
3.2 Generation of scale-free networks ................... 20
3.3 Network model description ....................... 21
3.3.1 Independent edges ....................... 22
3.3.2 Edges’ degree based edge swap ................ 22
3.3.3 Nodes distance based edge swap ................ 23
3.3.4 Measuring network robustness ................. 24
3.3.5 Variable attacks ......................... 24
3.4 Details of ISFNs algorithms ...................... 25
3.5 Simulation results ............................ 29
3.5.1 Construction of scale-free network ............... 29
3.5.2 Comparison of Improved Scale-Free Networks (ISFNs) with
ROSE and Simulated Annealing (SA) ............. 29
3.5.3 Comparison of ISFNs with ROSE and SA against low degree
attacks .............................. 30
3.5.4 Comparison of ISFNs with ROSE and SA against high de-
gree attacks ........................... 31
3.5.5 Comparison of ISFNs with ROSE and SA against random
attacks .............................. 31
xii
3.5.6 Comparison of ISFNs with ROSE and SA against variable
attacks .............................. 32
3.5.7 Network robustness against different number of nodes . . . . 33
3.6 Conclusion of the chapter ....................... 34
4 Network’s topology evolution scheme 35
4.1 Summary of the chapter ........................ 36
4.2 Scale-free model ............................ 36
4.2.1 Construction of Scale-Free Network .............. 36
4.2.2 Details of edge swap ...................... 37
4.2.2.1 Edge swap of randomly selected nodes ....... 38
4.2.2.2 Edge swap of degree based selected nodes ..... 38
4.2.2.3 Edge swap of distance based selected nodes . . . . 39
4.2.3 Metric of robustness ...................... 39
4.3 Network Topology Evolution Scheme overview ............ 40
4.3.1 Evolution of network topology ................ 40
4.3.2 k-core based nodes’ degree distribution ............ 41
4.3.3 Attacks on the proposed topology ............... 41
4.3.4 NTES’s optimization by heuristic algorithms ......... 42
4.4 Details of NTES algorithms ...................... 43
4.5 Simulation results and discussion ................... 46
4.5.1 Network topology evolution .................. 47
4.5.2 Attacks on upper and lower networks ............. 47
4.5.3 Networks connection by one-to-many correspondence . . . . 49
4.5.4 Core based attacks on network ................. 50
4.5.5 Network robustness against random and malicious attacks . 50
4.5.6 Comparison of NTES robustness against different optimiz-
ing techniques .......................... 51
4.6 Comparison of NTES and other algorithms on Scale-Free Topoogy
(SFT) topologies ............................ 52
4.7 Conclusion of the chapter ....................... 54
5 Optimization of scale-free networks 55
5.1 Summary of the chapter ........................ 56
5.2 Overview of JAYA ........................... 56
5.3 JAYA for the scale-free networks ................... 57
5.3.1 Metric of robustness ...................... 58
5.3.2 Scale-free model ......................... 58
5.3.3 Details of the algorithms .................... 59
5.4 Classification of edge swap ....................... 60
5.4.1 Random edge swap ....................... 62
5.4.2 Degree dependent edge swap .................. 63
5.4.3 Distance dependent edge swap ................. 63
5.4.3.1 Global communication efficiency .......... 64
5.4.3.2 Average clustering coefficient ............ 64
5.4.3.3 Average shortest path length ............ 64
5.4.3.4 Assortative coefficient ................ 65
xiii
5.5 Addition of nodes in maximum connected subgraphs ........ 65
5.6 Effect of bridge edge on the network ................. 66
5.7 Effect of different attacks strength ................... 66
5.8 Simulation results ............................ 67
5.8.1 Comparison of JAYA with the existing techniques ..... 67
5.8.2 Analysis of edge swap methods ................ 69
5.8.3 Nodes’ addition in Maximam Connected Subgraph (MCS) . 74
5.8.4 Robustness enhancement by bridge edge protection ..... 75
5.8.5 Network strength against different attacks strength ..... 76
5.9 Conclusion of the chapter ....................... 77
6 Conclusion and future work 79
6.1 Conclusion ................................ 80
6.2 Future work ............................... 81
7 References 82
Conference Proceedings 95
xiv
LIST OF FIGURES
3.1 Scale-free network’s construction ................... 21
3.2 Edge swap mechanism a: First connection method b: Second con-
nection method c: Third connection method ............. 22
3.3 Nodes distance based edge swap a: First connection method b:
Second connection method c: Third connection method ....... 24
3.4 ISFNs evaluation of robustness .................... 30
3.5 Network’s performance evaluation against low degree attacks . . . . 30
3.6 Network’s performance evaluation against high degree attacks . . . 31
3.7 Network’s performance evaluation against random attacks ..... 32
3.8 Network’s performance evaluation against variable attacks ..... 32
3.9 Network’s performance evaluation against different number of nodes 33
4.1 Network evolution by adding edges .................. 37
4.2 Edge swap mechanism ......................... 38
4.3 Nodes’ attack based on core ...................... 42
4.4 Attacks on NMCS of the network ................... 42
4.5 (a) Malicious and random attacks on upper network. (b) Malicious
and random attacks on the lower network ............... 48
4.6 (a) Power-law distribution of the mutual nodes (b) Comparison of
core based attacks and high degree node attacks ........... 49
4.7 (a) Random attacks (b) Malicious attacks .............. 51
4.8 (a) Comparison of optimization algorithms when N = 100. (b)
Comparison between NTES and existing algorithms when N = 100 52
4.9 (a) Comparison of NTES with existing algorithms when random
attacks happen (b) Comparison of NTES with existing algorithms
when malicious attacks happen .................... 53
5.1 Chromosome is obtained from the adjacency matrix ......... 57
5.2 JAYA for the optimization of SFNs (a) Current individual (b) The
best individual (c) Finding the neighbors (d) Selecting the nearest
node (e) The updated individual .................... 58
5.3 Edge swap mechanism (a) Original topology (b) First connection
method (c) Second connection method ................ 63
5.4 Affect of bridge edge on the robustness of SFNs ........... 66
5.5 Effect of edge swap on the robustness of SFNs ............ 67
5.6 Affect of edge swap on the robustness of SFNs ............ 68
5.7 Affect of edge swap on the robustness of SFNs ............ 69
5.8 HDA attack on network ........................ 69
5.9 LDA attack on the network ...................... 70
5.10 Random attack on network ....................... 70
5.11 Assortativity coefficient with different network size and edge swap
methods ................................. 71
5.12 Average clustering coefficient with different network size and edge
swap methods .............................. 72
xv
5.13 Affect of edge swap on average shortest path length ......... 72
5.14 Affect of edge swap on global communication efficiency of network . 73
5.15 Computational complexity of different edge swaps .......... 73
5.16 Effect of edge swap on the robustness of the network ........ 74
5.17 Addition of nodes in MCS ....................... 75
5.18 Bridge edge in the network ....................... 75
5.19 Robustness of the network with attack strength 5 .......... 76
5.20 Robustness of the network with attack strength 10 .......... 77
5.21 Robustness of the network with attack strength 15 .......... 77
xvi
LIST OF TABLES
1.1 List of acronyms ............................ 2
2.1 Problems addressed by previous work. ................ 12
4.1 Mapping the identified limitations, their proposed solutions with
validations ................................ 43
xvii
Chapter 1
Introduction
1
1.1. INTRODUCTION
1.1 Introduction
In this section of thesis, the background and the importance of the SFNs robust-
ness, contributions to improve robustness and organization of thesis have been
presented.
1.1.1 Details of scale-free networks
In this section, first we will discuss the definitions of different terminologies that
we have used secondly, we will discuss about the work in details.
The Internet of Things (IoT) has become an essential part of many real-world
applications due to the availability of a vast range of sensor nodes and the internet.
The IoT is part of healthcare [1], smart grid [2], industry [3], etc., and with the
passage of time, the number of IoT devices is increased, therefore, the networks
become dense. Moreover, due to the ease in the availability of the internet, the
concept of internet of everything has become common.
Table 1.1: List of acronyms
Notations Description
ABC Artificial Bee Colony
BA Barab´asi Albert
BFO Bacterial Foraging Optimization
DDO Degree Difference Operation
GA Genetic Algorithm
HC Hill Climbing
HDA High Degree Adaptive
HDO High Degree Operation
IoT Internet of Things
MA Memetic Algorithm
mEdge density
MCS Maximum Connected Subgraphs
MOO Multi-Objective Optimization
NNumber of Nodes
NC Natural Connectivity
NTES Network Topology Evolution
Scheme
RRobustness
SFNs Scale-Free Networks
WSNs Wireless Sensor Networks
For the effective communication between IoT nodes, different network topologies
are considered [4]. The two widely used topologies are Small World Topology
2Thesis by: Muhammad Usman
1.1. INTRODUCTION
(SWT) [5] and SFT [6,7]. These topologies are part of complex network theory. In
SWT, the heterogeneous nodes are considered that have different communication
ranges, bandwidths and energy. Moreover, the topology has a high clustering
coefficient and average shortest path length. Furthermore, the SFT is formed by
the homogeneous nodes [8] that have similar bandwidth and communication range.
Being part of several critical applications, the IoT networks are subject to cyber
attacks [9]. The attackers mostly affect the networks to take its controllability.
In IoT, two types of attacks are common: random and malicious. In random
attacks, the attackers have less information about the networks, so the nodes
remove randomly. However, in malicious attacks, the attackers have complete
network information and the nodes are removed based on their properties [10]. The
effect of these types of attacks on the IoT topologies is different [11]. The SWT
is robust against malicious attacks, whereas, the SFT is robust against random
attacks. In SFT, the nodes’ degree follows the power-law degree distribution. In
this distribution, the number of low degree nodes is more as compared to the nodes
with high degrees. Therefore, the probability of attacks on low degree nodes is
high against which the SFT is robust. However, the removal of high degree nodes
makes the SFT fragile.
In [6], Barab´asi Albert (BA) model is used to form a SFT based network by follow-
ing the growth and preferential attachment processes. During the network growth
process, a new node is added asynchronously while in preferential attachment,
the local information of the already connected nodes is considered. Based on the
probability of nodes’ degree, the connections are made with nodes that are already
part of a network and have a high degree. Furthermore, the SFT is represented
by graph theory and unweighted and undirected graphs are considered [12]. The
robustness, which is the resilience of a network against attacks [13], is calculated
based on percolation theory based measure proposed by Schneider et al. The
network fragments into multiple subgraphs because of important nodes removal.
The MCS is used to calculate the robustness [14] and a large number of nodes are
required to make the network robust. The sensor nodes have limited energy re-
sources, therefore, the SFNs robustness decreases due to the nodes’ failure. Many
researchers study the methods to increase the lifetime of nodes [1531].
In the network, some nodes are more densely connected internally nodes than the
other nodes in the network and form the community. The community in the SFNs
has an important role in the robustness enhancement. Therefore, in literature, the
3Thesis by: Muhammad Usman
1.1. INTRODUCTION
community structure in analyzed against the random attacks, malicious attacks
and cascading failures [3336,3856].
1.1.2 Contributions
In this section, details of the contributions to enhance the robustness of SFNs is
given. During the optimization the degree is not changed, therefore, no extra cost
is required.
A homogeneous nodes based method is represented for generating a scale-free
network that have the same communication range and energy. These nodes are
robust enough to not only random attacks but also to specific attacks. Moreover
to enhance a SFNs robustness, a new technique, Improved Scale-Free Networks
(ISFNS) is proposed. ISFNs, enhances the robustness of the scale-free network
topology without changing the nodes’ degree distribution. ISFNs consists of two
phases: the edges degree based swap and nodes distance based edges swap to
obtain onion-like structure for the network. The onion-like structure makes the
network robust against the malicious attacks
Moreover, as the size of networks are increasing due to the advancement in the
sensor techniques, the networks are becoming dense. The size of a network has an
effect on network robustness and with the increase in a network size the robustness
decreases. To address the aforementioned problem, a Network Topology Evolu-
tion Scheme (NTES) is proposed. In NTES, the network evolution is started by
dividing the sensor field into two parts. The networks are evolved in each part and
link with each other through one-to-many correspondence. The nodes of network
Aare linked with single or multiple nodes of network Bby following one-to-many
correspondence. To make the network robust, the nodes’ degree follow the power-
law degree distribution. The nodes’ degree changes hierarchically in each ring [59],
therefore, the change is calculated by k-core decomposition. Furthermore, the SFT
is optimized to form long links between the nodes because these links make the
network robust against malicious attacks [60]. In addition, three optimization al-
gorithms, Genetic Algorithm (GA), Bacterial Foraging Optimization (BFO) and
Artificial Bee Colony (ABC), are used to optimize NTES and the one that has
better performance is selected.
Furthermore, to optimize the SFNs the JAYA algorithm is used to find the optimal
solution with less computational cost and without the control parameters. It is
the latest optimization technique used to find the optimal results than the already
existing techniques. According to our knowledge, we are the first that are using
4Thesis by: Muhammad Usman
1.1. INTRODUCTION
JAYA for the optimization of SFNs. Furthermore, to enhance the robustness of
the networks edge swaps are performed. To analyze the effect of edge swap on the
network structure, the edges are classified into three categories: random edge swap,
degree based edge swap and distance based edge swap. The edge swap is performed
by keeping the nodes’ degree distribution constant. In addition, the robustness is
enhanced by adding the nodes in the MCS after the network fragments. Moreover,
the network’s performance is analyzed against different attack strengths.
1.1.3 Organization of thesis
The rest of thesis is organized as: in Chapter 2related work on SFNs is reviewed.
Chapter 3, Chapter 4and Chapter 5present strategy to enhance robustness, net-
work topology evolution scheme and optimization of SFNs, respectively. In Chap-
ter 6, conclusion and future work of the study is presented.
5Thesis by: Muhammad Usman
Chapter 2
Literature review and problem statement
6
2.1. LITERATURE REVIEW
2.1 Literature review
In real-world applications many networks are scale-free. These networks are robust
against random attacks because low degree nodes are present in a large number
and they have a high probability of selection. However, malicious attacks frag-
ment these networks into multiple subgraphs. In malicious attacks, the important
nodes based on their degree, betweenness centrality, closeness centrality, etc., are
removed from the network. Due to the presence of a small number of important
nodes, SFT are prone to malicious attacks. In this thesis, the importance of a
node is calculated based on its degree.
The network topology has a key role in defining the robustness of SFT. In [57],
based on the graph theory, BA model is proposed. In this model, the processes
that are required to construct a network are presented. The model forms the
network topology to an onion-like structure that is robust against the malicious
attacks. However, the limited communication range of sensor nodes in Wireless
Sensor Networks (WSNs) is not considered. Therefore, the nodes may die due to
their excessive usage. Moreover, the constraints of sensor nodes including limited
energy and communication range is considered in [62]. The network is evolved
by considering nodes’ communication range and the threshold value of the nodes’
degree. The robustness is improved through the Degree Difference Operation
(DDO) and Angle Sum Operation (ASO). However, the redundant operations
increase the computational complexity. A similar work has been performed in
the [63]. Furthermore, based on the nodes’ fault probability a network topology
evolution scheme is introduced in [65]. In the network growth process, a new
node joins the network based on the fault probability and communication range of
nodes, simultaneously. The network robustness is enhanced against the malicious
attacks due to the formation of onion-like structure.
The optimization is performed to enhance the robustness of SFT. The optimized
topology has a better network structure and is robust against malicious attacks.
Therefore, a local rewiring based algorithm Hill Climbing (HC) is proposed [57].
In this algorithm, the rewiring is performed by considering the local information of
nodes. Although it provides a better network after the rewiring, however, it traps
into local optima due to the local rewiring process. A similar work has been done
in [58]. The authors in [66] extended the work proposed in [57] by introducing the
global edge swap method. A temperature variable is introduced to get the optimal
solution. Moreover, the authors in [67], use a deluge algorithm to get the optimal
result as compared to SA. The edges are classified with the removal of nodes from
7Thesis by: Muhammad Usman
2.1. LITERATURE REVIEW
the network [68]. The swap operation is performed to increase the number of valid
edges in the network. The increase of nodes in MCS enhance the robustness of
the network.
Different heuristic techniques are used to enhance the robustness of SFT. GA is
mostly used to optimize SFT, however, the classical GA traps to local optima due
to premature convergence. The low population diversity causes the premature
convergence problem and for better results, a global solution is required from the
search space. To solve the less population diversity problem [69,70] introduces the
multi-population based methods. Although the methods provide a high diverse
solution, however, due to the involvement of multi-population the computational
cost is high. Moreover, the involvement of multiple operations makes the method
difficult to implement on real-world networks. Qie et al. [71] solve the premature
convergence problem by introducing the self competition among the individuals
of a population. Moreover, a Memetic Algorithm (MA) is proposed [72] that uses
the local search operator to avoid premature convergence. The optimal solution
is found from the search space by MA. A multi-agents based algorithm [73] is
proposed to get the optimal solution with less computational cost.
Single objective optimization techniques are discussed so far for the optimization
of SFT. However, in some problems, it is difficult to optimize the network with
single objective optimization. Therefore, a Multi-Objective Optimization (MOO)
technique is used to optimize the problems that have more then one objective to
be optimized. A Multi-Objective Evolutionary Algorithm (MOEA) is proposed
in [74] that is used to optimize robustness, when multiple attacks simultaneously
happen on the network. In MOEA, the negatively correlated objective functions
are found by Pearson correlation coefficient and are optimized. Due to MOO,
the computational cost to calculate robustness is high. A similar work has been
done in the literature [7577] to enhance the robustness of SFNs against multiple
objectives. Apart from undirected network, a directed network also define the
important network characteristics. Therefore, a directed network with the emer-
gence of cooperation and controllability of robustness as two important features are
discussed [78]. To find the correlation between network topological features and
their robustness a practical approach is required. Therefore, in [79], an empirical
approach is proposed to better control the network.
Moreover, the IoT based applications and importance of the networks other then
the scale-free is studied in the literature review. In robotic, the controlled move-
ment has great importance to complete the tasks. Therefore, the collision and
8Thesis by: Muhammad Usman
2.1. LITERATURE REVIEW
obstacle avoidance control strategies are studied in the [80,81]. Moreover, the se-
cure networks for the agri-food and the data sharing based on the blockchain are
studied in [82,83]. Using the blockchain network’s and data security is increased.
The WSNs are extensively used in the underwater communications. The effects
of sink mobility on the data gathering are analyzed in [84,85]. The sink nodes are
considered static and the sensor nodes are mobile before. However, the extensive
research prove the importance of sink position for the efficient communication.
In the communication, the delay should be kept minimum for the efficient sharing
of data. Therefore, a routing scheme for the delay sensitive applications pro-
posed [86]. Moreover, the efficient routing is performed in [8789] The routing
is made efficient by managing the energy consumption of the sensors. Further-
more, to increase the network lifetime the protocols are enhanced [90,91]. These
protocols use the proper routing to better utilize the energy. In the same way,
energy utilization and consumption algorithms to increase the network lifetime
are discussed [92,93]. The network attack and repair strategy is analyzed in [94].
The fraction of removed nodes and the average degree of nodes are experimentally
studied.
Some other crossed domain topics are also studied in the literature review to
make use of different concepts. The deep learning is used to detect the brain
tumor [95]. The data set of normal and effected persons are analyzed and based
on the selected features the model is trained to detect the tumor. Furthermore, the
communications related to the body area network based protocol is studied [96].
The faults detection in the WSNs has a great importance. Therefore, base on
the random forest scheme the faults are detected [97]. Moreover, to increase
the lifetime of the networks the cooperation of nodes is very important, which is
analyzed in [98].
In cascading failure and robustness of the network, how they relate with each
other needs to be found [99]. By calculating these objectives comprehensively
normalized robustness used that has high computation cost. Connectivity and
dependence links enhance the robustness [132], however, these links are added
randomly. If both networks follow scale-free topology then these links may change
the network structure. The dependence links are not following the power-law
distribution, therefore, stubborn attacks on these links cost more damage to the
network. Both networks contain high and low degree nodes and attacks on high
degree nodes affect more. Therefore, dependence links need to be added between
these nodes to improve robustness. After a node or edge having high load is
9Thesis by: Muhammad Usman
2.1. LITERATURE REVIEW
removed its load is distributed among other nodes or edges. However, there is a
possibility of simultaneous edge and node removed due to the environmental faults
or extra load then how the network deals with that to minimize the Cascading
Failures (CF) is not discussed [100]. Secondly, due to overloading the robustness
of a network decreases therefore, when an attack or failure happens on one part
of a network that should be removed to protect the other part of the network
to collapse. The proposed scheme better capture the CF, however, no method
is discussed to reduce the CF. The failure of nodes is not mentioned either it is
random or based on degree. In Cyber Physical System (CPS) Distributed Energy
Resources (DER) can help to maintain load that need to be discussed. A similar
work is done in [101103]. Moreover, against the multiple nodes removal from the
network, the cascading failure is calculated in [104]. Using different parameters
the network robustness is analyzed to make a practical approach to enhance the
robustness.
Although [59] gives new perspective of attack based on cores of network, however,
when the fraction of removed nodes is high core based attack is less destructive.
In that case the betweenness centrality attack becomes more vulnerable to the
network. For SFNs onion-like structure is not considered because in that struc-
ture same degree nodes are connected with each other. Shell-min attacks can
be considered as random attacks because they are done on low value of core for
these attacks SFNs are robust. By installing backups network performance in-
creased [105] however, installing backup increase cost. The key nodes usually are
hub nodes so, installing backup for these nodes is not an easy task. Secondly, after
the attack, a key node is removed because there is a high probability that a sec-
ond attack may happen on the backup node. In targeted attacks, the attacker has
complete information of the network [105] how the proposed scheme update the
network information is not discussed. The links are added and removed to imple-
ment the adversarial attack [106], however, the degree distribution is change that
is not considered. The addition of links have a cost that has not been discussed.
Wandelt et al. [107] propose Quick Robustness Estimation (QRE) for estimating
the network robustness. QRE performed better than the betweenness centrality
because it is based on a cheap-to-compute network matrix combination. The pro-
posed system enhanced the robustness of scalable networks, therefore, it is not
for SFNs. The methodology for managing the robustness of the Social Internet of
Things (SIoT) is proposed by the authors in [108]. SIoT is divided into multiple
Enterprise Systems (ESs). The state variable of these ESs are determined and the
interaction between them is anatomized. After that, the nonlinear dynamic model
10 Thesis by: Muhammad Usman
2.1. LITERATURE REVIEW
is developed. However, the applicability of the proposed method is reduced at the
operating stage of SIoT.
The entropy based network robustness optimization technique is proposed in [109].
In this technique, the random failures of nodes are considered and the entropy of
the network degree distribution, scaling factor of power-law and nodes’ connectiv-
ity is studied. However, the malicious attacks against which the SFNs are fragile
is not considered. Apart from that, the network robustness is enhance by adding
links [110]. A similar work is performed in [120,121], that consider the critical net-
work infrastructure and modification of edges, respectively. On the other hand,
the network robustness of SFNs against the removal of links is analyzed in [122].
The malicious attacks are happened on high degree nodes, therefore, the method
is failed to increase the robustness in that case. Moreover, the network robust-
ness in enhance by decreasing the assortativity coefficient against the malicious
attacks [111]. The network robustness against the random attacks by keeping the
nodes’ average degree constant is enhanced in [112,113]. Moreover, the network
robustness is enhanced by modifying the network structural and characteristics
information in [114117].
To enhance the network robustness against the malicious attacks, the parame-
terized networks are made that are robust against both random and malicious
attacks [118]. A similar work has been done in [119]. On the other hand, the lim-
ited energy of sensor nodes are analyzed to deal with sensor nodes’ constraints to
increase the robustness of large-scale WSNs [123]. To increase the network robust-
ness against the attacks on the edges is studied in [124]. The network robustness is
increased against the random and malicious attacks without increasing the num-
ber of driver nodes. Moreover, to study the limited energy of the sensor nodes
in the SFNs, the network is partitioned and a reduced SFN is made [125]. Fur-
thermore, the robustness structure and the effects nodes removal are extensively
studied in [126130]
11 Thesis by: Muhammad Usman
2.1. LITERATURE REVIEW
Table 2.1: Problems addressed by previous work.
Problem identified Proposed solutions Validations Limitations
SFNs survivability and
robustness after cyber
attacks
Topology evolution
based on fault proba-
bility
Considered two op-
erations i.e., HDO and
DAO
Four types of attacks
considered that make
TMSE resilient against
real life attack
Results are validated by
comparing with the exist-
ing algorithms
TMSE outperforms the
existing techniques of dif-
ferent size networks and
different types of attacks
Preferential attachment
may be compromised due
to fault probability
HDO changes connection
of high degree nodes which
have high attack probabil-
ity
DAO not considered the
fault probability that in-
crease the effect of attack
[65]
Multiple attacks on
a network simultane-
ously
Uses multi-objective
optimization to en-
hance the robustness
Effects of attack on topo-
logical features validate the
importance of MOEA
Synthetic and real-life
networks prove the impor-
tance of MOEA
High computation cost
due to calculation of multi-
objectives
Topology improvement
for simultaneous attacks
not discussed
Pareto set generation can
be difficult [74]
Premature conver-
gence of classical
GA
Uses multi-populations
co-evolution to solve
premature convergence
Comparing with exist-
ing algorithms MPGA
enhances the robustness
Optimal population size
should be chosen
Local operator should be
used to make diverse pop-
ulation [69]
Premature conver-
gence of classical
GA
Self-competition
among the individuals
Calculate population
diversity using GPCR
and GDU
Mutation probability
calculated by adaptive
adjustment
Performance is validated
with different algorithms
and attacks
Less time cost as com-
pared to ROCKS which
is multi-population algo-
rithm
Diversity calculation is a
complex process at every
operation
GPCR and GDU have
high computation cost
Uses multiple operations
which make computation
cost high [71]
Link attack based on
betweenness centrality
has high computa-
tional complexity
Link attack based on
shell has better perfor-
mance than between-
ness centrality
Shell based attack
has less computational
complexity
Importance of shell-based
attacks validated for differ-
ent types of network
Shell-min, shell-max and
shell-pro attacks compared
with betweenness central-
ity
Less destructive when the
fraction of removed nodes
is high
Less effective for SFN due
to not following onion-like
structure
[59]
Continued on next page
12 Thesis by: Muhammad Usman
2.1. LITERATURE REVIEW
Table 2.1 Continued from previous page
Problems identified Proposed solutions Validations Limitations
No practical approach
available to under-
stand the relationship
between network
topology features and
network robustness
Empirical approach
to measure robustness
Exhaustive search-
based technique on
small world and real-
life networks
ENC helps to con-
verge by shrinking
search space
RER helps to rectify
ENC
Importance of RER
is validated for different
number of nodes
Network connectedness
improved by increasing
number of RER operations
Exhaustive attack is not
actually considered
Not explain RER for
undirected network
RER changes nodes de-
gree
Due to RER SFN not re-
mains scale-free [79]
The vulnerability of
SFNs due to malicious
attack
Constructed network
considering communi-
cation range of nodes
DDO and ASO
are used to construct
onion-like structure
Through ROSE topol-
ogy converted to onion-like
structure
Network becomes robust
against malicious attack
Target all nodes in the
network may generate re-
dundant operations
DDO and ASO are com-
putationally expensive [62]
High computational
complexity of exist-
ing algorithms is a
hurdle in topology
self-optimization
For the self-
optimization AI based
technique is used
Efficiency and loss func-
tion for training and test-
ing validated
99% efficiency is achieved
Loss function minimized
in 35th iterations
Not suitable for different
size of networks and edge
densities
After different attack
how self-optimization
works not discussed [64]
Due to random edge
swap without consider-
ing the network struc-
ture redundant opera-
tions are performed
After attack edges
are classified into three
types By increasing
the valid edges robust-
ness enhances
For HDA attack algo-
rithm performance is vali-
dated for different size of
networks
The heuristic algorithm
maintains the robustness
Edge swap between in-
valid edges enhances the
computational overhead
When nodes fail the
edges with the nodes also
removed [68]
Network robustness
decreases when the
nodes removed
Optimization algo-
rithms fall into local
optima
NC is used to mea-
sure robustness
Chaotic GA is used
to optimized network
To avoid local optima
logistic maps based
power function carrier
used
Degree distribution
remains unchanged
after optimization
NC increases with in-
creasing iterations [136]
NC and r have positive
correlation that proves en-
hance robustness
Onion-like structure is
achieved through optimiza-
tion
Considering different at-
tacks network is robust
N/A
Continued on next page
13 Thesis by: Muhammad Usman
2.1. LITERATURE REVIEW
Table 2.1 Continued from previous page
Problems identified Proposed solutions Validations Limitations
Community detec-
tion algorithms reveals
individual information
Practical approach
required to protect
these individual’s
privacy
Network is optimized
by using heuristic tech-
nique
CDA and HDA per-
formed to detect com-
munity and high degree
nodes
Q-based attack based
on GA is performed
Addition and removal
of links increase pri-
vacy on individual
Different Pcand Pmare
used to find optimal pa-
rameter for GA
For all networks Q-based
attack outperformed
Decrease in Q value con-
firms the privacy of indi-
vidual
Randomly addition and
removal of links affect SFN
Addition of links have
cost
Onion-like structure is
not follow
Directedness is an
importance feature for
a network
f(c) and Rdifficult
to optimize simultane-
ously
Correlation between
f(c) and Rproves the
negative correlation
MOEA is used for
these objective
r is used to set net-
work structure
PD is used to calcu-
late f(c)
MCS is used to calcu-
late R
Pareto optimal solu-
tion obtain for both ob-
jectives
MOEA optimized both
objectives better than sin-
gle objective algorithm
Different networks are
optimized
Ris assortative whereas
f(x) disassortative there-
fore, onion-like structure is
obtained by shuffling edges
No comparison is made
with other multi-objective
algorithms [78]
Only assortativity is con-
sidered
Other topological param-
eter may prove positive
correlation of these objec-
tives
SFNs are vulnerable
to malicious attacks
Robust network
structure is require
SFNs converter to
onion-like structure
Random edge swap is
made
Degree difference re-
mains same after opti-
mization
Nodes increase in
MCS
Different networks
are used for validations
Network performance im-
proved against malicious
attacks
Swap edge enhances ro-
bustness
Increase in number of
nodes decreases network
robustness
Assortativity and cluster-
ing increase
Random edge swap ef-
fects network structure [57]
Average shortest path
length increased
HC traps into local
optima
Network optimiza-
tion against malicious
attacks required
Network is generated
by using BA mode
Global and local edge
swap performed for ex-
ploration and exploita-
tion
Tis used to avoid lo-
cal optima
αis used to get fast
convergence
Both synthetic and real-
world networks considered
Global edge swap and
local edge swap are com-
pared
Global edge swap has
better robustness due to
exploration
Network optimized by us-
ing αis more robust
No threshold value for
edge swap is set [66]
No real-world network
considered
Continued on next page
14 Thesis by: Muhammad Usman
2.1. LITERATURE REVIEW
Table 2.1 Continued from previous page
Problems identified Proposed solutions Validations Limitations
SFNs are not ro-
bust against malicious
attack
No proper solution
is available to mitigate
cascading failure
Rcand Rare weakly
correlated
Multi-objective opti-
mization is used
High and low objec-
tives value based net-
works generated by SA
Optimal topology is
found by the MAGA by
exploration
Both synthetic and real-
world networks are consid-
ered
Normalized robustness is
better as compared to SA
For different sized net-
works both objective have
improved values
Cascading failure proved
to be more damaging as
compared with intentional
attack
A single measure needs to
be used for these objectives
Cascading failure does
not happens rapidly coun-
termeasures reduce its ef-
fect
A better network struc-
ture required that is robust
against attacks
Robustness of in-
terdependent network
needs to enhance
Connectivity and
dependence links are
added to enhance
robustness
Considering cost
constrain optimal val-
ues of these links need
to be calculated
CPS is more vulnera-
ble and caused cascad-
ing failure in physical
region
One-to-many config-
uration is used to con-
nect these networks
Stubborn and smart
attacker are considered
Defender add links by
calculating intra and
inter degree
Both synthetic and real-
world networks are consid-
ered
Intra and inter de-
gree based attacks are per-
formed on networks
Based on these attacks
links are added
Different intra and inter
degree based networks are
considered
Links are added ran-
domly without considering
degree distribution [132]
Dependence links are not
following power-law distri-
bution
Stubborn attack on de-
pendence links caused net-
work to fail
GA has premature
convergence problem
Population diversity
needs to be high
ROCKS uses
multi-populations
co-evolution to deal
premature convergence
Pmand Pcare dif-
ferent for populations
Populations coordi-
nate using migration
operator
Migration popu-
lation contains best
individuals from all
populations
Degree distribution
remains same after
optimization
Onion-like structure is
obtained after optimiza-
tion
After ROCKS network is
more robust against ran-
dom and malicious attacks
For different size net-
works ROCKS outperform
HC and SA
Due to multi-population
computation cost is
high [70]
After attack self-
optimization is difficult
Difficult to implement
on real-life network due to
complexity
Continued on next page
15 Thesis by: Muhammad Usman
2.1. LITERATURE REVIEW
Table 2.1 Continued from previous page
Problems identified Proposed solutions Validations Limitations
After edge or node re-
moval their respective
edges and links also re-
moved
Multiple nodes and
links removal caused
cascading failure
Node and Edge based
model required to deal
the failure
DB model proposed
to deal cascading fail-
ure
Node importance is
calculated based on its
degree and between-
ness
Removal of impor-
tant node causes other
nodes and edges to be-
come overload
Load is distributed
among node by consid-
ering their initial ca-
pacity
Five metrics used to
evaluate robustness
Initial capacity of nodes
is compared with critical
value of tolerance
Cascading failure of edge
and node of DB model is
less when θ¡ 0.9
DB model has better
value of SGand SCcom-
pared to DN and BN
Load is distributed
among neighboring nodes
which have already high
load
Simultaneous edge and
node removal is not dis-
cussed
Overload edges and
nodes should be removed
to protect remaining
network
Random edge swap
is performed in opti-
mization
Robustness improve
by compromising com-
munity structure
Optimizing per-
formed in a way that
preserve community
structure
Three step strategy
is proposed to preserve
community
Onion-like structure
is introduced in every
community
High degree nodes
are connected with
nodes of their own
community
Considering cost
links can be added
to better preserve
community
Community structure is
compared before and after
optimization
These are preserved with
enhanced robustness
After the node removal
3-steps strategy performed
better compared with sin-
gle step
High degree node of dif-
ferent communities are not
connected with each other
After removal of high de-
gree node its load can not
be distributed by remain-
ing nodes
High degree nodes must
connect to share their load
Removal of impor-
tant node causes de-
crease in robustness
These nodes must
be protected by taking
countermeasures
No method is avail-
able to protect these
nodes
Networks vulnerabil-
ity against key nodes
removal
Key nodes are found
based on MILP
Network is optimized
by heuristic algorithm
Countermeasures in
the form of installing
backup for key nodes
Impact of node’s elimina-
tion on topological param-
eter
Throughput, network de-
lay, and flow is calculate
End-to-end delay is max-
imized by node protection
After removal of a node
there is high probability of
backup node removal
Network information
need to be updated
against smart attacker
Continued on next page
16 Thesis by: Muhammad Usman
2.1. LITERATURE REVIEW
Table 2.1 Continued from previous page
Problems identified Proposed solutions Validations Limitations
Adversarial attack
change network infor-
mation
Network robustness
decrease by these at-
tacks
Adversarial attacks
are considered in this
network
Two new attacks
strategies DILR and
DALR intoduced
RLR is less effec-
tive as compared with
DILR and DALR
SFNs are categorized
into strong to weakest
Networks are generated
by considering Rand α
Each network is attacked
200 times
Average shortest path
length, clustering coeffi-
cient and diagonal distance
is compared with attacks
Degree distribution is
change by edge swaps [106]
Addition of links have
cost
Cascading failures
in interdependent
networks
Novel method to
capture cascading
failure introduced
CPS and PS consid-
ered as interdependent
networks
One-to-One corre-
spondence between
networks
Fraction of survival
nodes is calculated af-
ter each node removal
Removal of a node in
PS caused overloading
in survived nodes
Removal of a node
in CPS decreases MCS
size
Networks are validated
according to asynchronous
failure propagation model
Intra and inter dependen-
cies considered
At even stage nodes re-
moved from CPS and at
odd stage nodes are re-
moves from PS
For load and space of PS
different distributions are
followed
Interdependent network
is more vulnerable to node
removal as compared to
single network
How to reduced cascad-
ing failure is not discussed
[100]
DER impacts to reduce
load needs to be discussed
Importance of SFN ro-
bustness
Memetic algorithm
to optimize network is
used
To enhance robust-
ness global and local
searches used
Population is gener-
ated by swapping edges
of initial network
Crossover is per-
formed by changing
links of parents
Offsprings has same
degree distribution as
parents
Local search operator
is used for exploitation
Optimal solution is
found by 2-tournament
selection
Optimal value of edge
swap
Different sized network
MA RSFM A outper-
formed
Proposed crossover im-
prove network robustness
MARSFMA performed
better against random and
malicious attack
Onion-like structure is
produced
Population has less diver-
sity [72]
Crossover changes degree
of nodes
17 Thesis by: Muhammad Usman
2.2. PROBLEM STATEMENT
2.2 Problem statement
This section presents the problem statement of the thesis. Moreover, the main
problem is divided into three subproblems which are presented below.
2.2.1 Details of problem statement
The SFNs are more suitable for IoT networks because they are resilient to random
attacks. In recent years, a significant attention is paid to enhance the robustness
of these networks against malicious attacks [12,68,131]. For the topology of SFNs,
BA model is proposed in [57], that explains how the nodes are connected to form
a network. Furthermore, for the calculation of robustness, a mathematical equa-
tion based on the percolation theory is proposed in [139]. One way to increase
robustness is by adding edges however, it adds the cost, which is solved by edge
swap. Therefore, by global edge swap, the degree of nodes remains constant and
robustness enhances without increasing cost [62]. In addition, an onion-like struc-
ture is proposed that contains nodes, whose degree decreases hierarchically and
are more robust against malicious attacks. However, the solution falls into local
optima that is solved in [57] with local edge swap. Moreover, due to redundant
operations, it has less efficiency. In [68], the network is constructed based on the
communication range and the threshold of nodes degree. It converts the network
into an onion-like structure, however, due to redundant operations, the network’s
efficiency is reduced.
Critical networks, including healthcare, military, and Internet, etc., have scale-free
nature. These networks should be robust against attacks however, the existing
algorithms [68,70,138] have high cost, therefore, self-optimization is used in [14].
Still, the problem of network robustness is not solved against malicious attacks
and these attacks make the network vulnerable.
18 Thesis by: Muhammad Usman
Chapter 3
Edge swap based optimization strategy to
enhance the robustness of SFNs
19
3.1. SUMMARY OF THE CHAPTER
3.1 Summary of the chapter
During the past few decades, the Internet of Things (IoT) has made remark-
able progress in many real-world applications including healthcare, military, trans-
portation, etc. Multiple sensor nodes are deployed in these fields to get the re-
quired data. Different network topologies are used in IoT and scale-free is one
of them. It is mostly preferred due to its robust behavior against random node
removal, however, the network collapsed because of malicious attacks. Therefore,
in this chapter, the robustness of scale-free networks is improved against malicious
attacks through optimization. To achieve this, the edge’s degree and nodes’ dis-
tances based edge swap operations are used in the proposed Improved Scale-Free
Networks (ISFNs) scheme. In the edge’s degree based operation, nodes of similar
degrees are attached and links between the near neighboring nodes are made in
nodes distance based operation. These operations help to achieve a better onion-
like structure without changing the degree distribution of the network. Therefore,
the network becomes robust against malicious attacks. Moreover, no new links
or nodes are added in the optimization process, therefore, no extra cost is in-
curred. Furthermore, to make the network more robust against realistic attacks,
the variable attacks are considered.
3.2 Generation of scale-free networks
The construction of SFNs is based on the BA model. After the deployment of
nodes in a sensor field, the model considers the growth and preferential attachment
steps. In the first step, at each time interval, a single node is added to the network.
However, in the second step, new nodes prefer to join the network by considering
the degree of the existing neighboring nodes. The new nodes prefer to join the
nodes that initially have high degree in the network. Therefore, due to the limited
resources of the sensor nodes, the chance of network failure is increased.
To deal with that issue, the connection probability based solution is used. The
probability Πdi of a new node iis calculated as follows.
Πdi=di
Pjdj
,(3.1)
where, diand djare the degree of node iand sum of the neighboring nodes degrees,
respectively. The network construction is depicted in Fig. 3.1. There are three
nodes i, j and kthat want to become part of the network. First of all, start
20 Thesis by: Muhammad Usman
3.3. NETWORK MODEL DESCRIPTION
k
k1
k2
k3
k6
k4
k5
i
i1
i2
j1
j3
j4
j
j2
j5
Figure 3.1: Scale-free network’s construction
with the example of node jthat wants to join the network. There are five nodes
in its neighborhood with degrees 1,1,3,1 and 2. Using Eq. 3.1, the connection
probabilities of the neighboring nodes are 0.125, 0.125, 0.375, 0.125 and 0.25. All
the nodes’ connection probabilities are placed into the roulette wheel. On average,
the nodes with a high degree have more area into the roulette wheel as compared
to the low degree nodes. So, there is more probability of a high degree node’s
selection. In this study, the neighbor nodes that have the highest connection
probability are directly selected, however, for the remaining nodes, the roulette
wheel selection is followed. Furthermore, node idetermines the connection with
neighbor nodes according to the edge density m. In the neighbor of node i, their
are two nodes i1and i2with degrees 2. If m= 2 then node imakes connection with
nodes i1and i2directly. Otherwise, depending on the value of m, the connections
are made. For node k, there are six nodes and all these nodes are not connected
with the network. Therefore, when the node kbroadcasts a request message to
make the connections with the nodes in the local community, all the nodes will
receive that message. After that, the nodes reply to node k. At that point,
the node kfollows the First Come First Serve (FCFS) [140] approach because it
provides a simple, efficient and less computational expensive solution. Therefore,
the node that responses early makes the connection with node k. These steps
mentioned above are followed until the network is completely evolved.
3.3 Network model description
The complete details of the proposed ISFNs are given in this section. The de-
scription of the operations as part of the ISFNs is presented in detail. However,
21 Thesis by: Muhammad Usman
3.3. NETWORK MODEL DESCRIPTION
before being familiar with these operations, the knowledge of independent edges
is essential.
3.3.1 Independent edges
The topology of the SFNs is represented as a graph G={N, E }. The set of
nodes and edges are given as N={N1, N2, ..., NN}and E={E1, E2, ..., EN},
respectively. The following conditions should be met to confirm that the edges eij
and ekl are independent.
1. Nodes as part of edge pairs should be in the same communication range.
2. There should be no extra edge except eil and ejk .
In Fig. 3.2a, the original topology with edges eij and ekl are shown. By following
the above mentioned conditions, the selected edges are independent. Moreover,
the first edge swap (eik,ejl) and the second edge swap (eil,ejk ) are represented in
Fig. 3.2b and Fig. 3.2c, respectively.
Figure 3.2: Edge swap mechanism a: First connection method b: Second
connection method c: Third connection method
3.3.2 Edges’ degree based edge swap
For the specific edge eij , the edge degree dij is calculated from the degrees of its
respective nodes. The edge degree is derived from the nodes degree [74] and is
defined as,
dij =pdi×dj,(3.2)
where, degrees diand djare for nodes iand j, respectively. Higher the value of
edge degree, more important is it in the network.
After finding the independent edges from the network, the edge degree is calculated
with Eq. 3.2. After that, based on the Eqs. 3.3,3.4 and 3.5, the degree difference
is calculated against each edge swap. The degree difference of all the edge swaps is
22 Thesis by: Muhammad Usman
3.3. NETWORK MODEL DESCRIPTION
compared and the pair of edges having minimum difference is selected. If a new pair
of edges increases the robustness and the network connectivity is not destroyed,
then the adjacency matrix is updated with the acceptance of edge swap.
DIF0=|dij dkl |,(3.3)
DIF1=|dik dj l|,(3.4)
DIF2=|dil dj k|.(3.5)
The edge swap based on the difference of degree is motivated from [62], where
random pairs are selected from the network. However, in this thesis, the edges are
selected based on the degrees of their respective nodes. The edge degree based swap
helps to connect the similar degree nodes. Moreover, the edge degree difference
operation helps to achieve high robustness by increasing the assortativity. Using
the edges’ degree difference operations, the degree distribution of the original
topology is not changed. Therefore, no extra cost is required to optimize the
network.
3.3.3 Nodes distance based edge swap
It is the second operation of the ISFNs scheme. This edge swap helps to link the
nodes that are near to each other. In the network, the robustness is improved
against malicious attacks when these links exist. As the preferential attachment is
followed by the SFNs in the growth process, the new nodes make the connection
with the existing nodes based on a high degree. Therefore, the low degree nodes
have a chance to connect with the high degree nodes, through this edge swap.
To perform the nodes distance based edge swap, nodes i, j, k and lare selected
from the graph G. There are edges eij and ekl between the nodes (i, j ) and
(k, l), respectively. The average distance is calculated between the nodes of the
independent edges using the Euclidean distance formula. The average distance of
two edges is required for each connection method because there is a possibility of
a distance mismatched between the edges. Therefore, the pair of edges is selected
that has a small distance. Fig. 3.3 shows the nodes distance based edge swap.
D1and D2represent the nodes’ Euclidean distances for the edge eij and ekl,
respectively. In Fig. 3.3a, the original topology is given and the robustness is
calculated against the malicious attacks. Moreover, in Fig. 3.3b, the first edge
swap of the independent edges eik,ejl is performed and D1and D2are calculated.
The same approach is followed in the second edge swap as shown in Fig. 3.3c.
23 Thesis by: Muhammad Usman
3.3. NETWORK MODEL DESCRIPTION
D1
D2D1D2D1D2
ij
kl
ij
kl
ij
kl
ab c
Figure 3.3: Nodes distance based edge swap a: First connection method b:
Second connection method c: Third connection method
The robustness is calculated against these edge swaps and the pair of edges that
provides the highest value of robustness is selected to update the adjacency matrix.
3.3.4 Measuring network robustness
After the malicious attacks on the SFNs, the network is divided into multiple
subgraphs, resulting in the reduction of the network performance. To study the
relationship between robustness and the number of removed nodes, Schneider et
al. [62] and [131] proposed a mathematical equation. According to the Schneider
observation, the robustness Rof a network having Nnodes can be represented
using Eq. 3.6.
R=1
N+ 1
N1
X
n=0
MCSn
N,(3.6)
where, Nis the number of nodes and 1
N+ 1 is the normalization factor and MCS
are formed after the nodes with high degree are removed. Moreover, the robustness
lies in the range of [0, 0.5]. The maximum value of the robustness is less than 0.5,
which means that the maximum number of subgraphs are connected while the
minimum value of the network robustness is approximately zero, which means
that a single high degree node present in the network is affected by the malicious
attacks.
3.3.5 Variable attacks
To study the effect of variable nodes removal from the network in an instant
of time, the variable attacks are performed. In these attacks, the number of
nodes is removed from the network and its connectivity is analyzed. By knowing
these attacks, the defender can easily optimize the network to increase its lifetime.
Variable attacks in this study are performed by randomly selecting the number
of removed nodes in the range of 1 to 10. The number of nodes in the MCS
24 Thesis by: Muhammad Usman
3.4. DETAILS OF ISFNS ALGORITHMS
is calculated after each attack. Due to multiple nodes are randomly removed,
therefore, the effect on network connectivity with multiple nodes removal in a
single instant is analyzed.
3.4 Details of ISFNs algorithms
In this section, details about the algorithms used in the proposed ISFNs scheme
to enhance the robustness of the SFNs are given. The algorithm 7is motivated by
the network generation model of [62]. In this algorithm, the priority is given to the
nodes that reply first to the connection request message. Moreover, a node that
is near to the requesting node has more possibility to make a connection than the
one that is far. Due to this less energy is consumed and the chances of node failure
are reduced. In algorithm 7, the process of SFNs topology modeling is discussed.
The following variables are used in this algorithm.
N: number of nodes present in the network.
r: communication range of nodes.
m: number of edges a new node makes with its neighboring nodes.
thr : threshold value of max degree.
A: adjacency matrix.
The details of the Algorithm 7are as follows.
Initially, the nodes are uniformly random deployed in a sensor field (Line 1). The
network’s evolution starts from the center node Nc. To form a clique, the center
node with its neighboring nodes are found (Lines 3-4). The nodes that are in the
same communication range, form the edge with Ncand the adjacency matrix is
updated (Line 5). After the completion of the clique, the network growth starts
and the node which wants to connect with the network is randomly selected (Line
8). The degree of its neighboring nodes is calculated (Line 9). Depending on the
degree of neighboring nodes the edges are formed. If all the neighboring nodes
are not connected with the network then depending on the value of mthe edges
are formed and the network adjacency matrix is updated (Lines 10-11). However,
if the neighboring nodes are already connected and their degree threshold is not
achieved then the connection probability is calculated (Lines 12-15). Moreover,
for the preferential attachment, the neighboring nodes are selected, based on the
25 Thesis by: Muhammad Usman
3.4. DETAILS OF ISFNS ALGORITHMS
Algorithm 1 SFNs generation
Input: N, r, thr, m
Output: A
1: Randomly deployed N nodes
2: for i = 1 : m do
3: Find Ncand its neighbors
4: Select nodes to form clique
5: Update A
6: end for
7: endfor
8: for j = m : N do
9: Randomly select Nj
10: Calculate djof the neighbors
11: if dj== 0 then
12: Randomly select neighbor
13: else if dj>0 && dj6= thr then
14: Calculate Probability
15: Πdi=di
Pjdj
16: end if
17: endif
18: Use roulette wheel method
19: Select mnodes
20: Update A
21: end for
22: endfor
roulette wheel method and the adjacency matrix is updated (Lines 16-18). The
process is repeated until all the nodes become connected.
Algorithm 8defines the edges’ degree based swap operation. In place of random
selection [62], the edges are made by considering the degree. The degree based
edge swap ensures that the nodes with a similar degree are connected. Hence,
the network structure becomes onion-like. This operation executes after the SFNs
topology generations in IoT. The algorithm uses the following variables.
A: adjacency matrix of initial SFNs.
Aup : adjacency matrix obtained after the swap.
Two nodes Niand Nkare randomly selected from the graph (Line 2). After
that, the neighboring nodes Njand Nlthat have high degrees and are in the
communication range of Niand Nkare selected (Lines 3-6). If the obtain edges
are independent then the degree of each edge is calculated (Lines 7-10). More-
over, for all the possible edges the difference of edge degree is calculated (Lines
26 Thesis by: Muhammad Usman
3.4. DETAILS OF ISFNS ALGORITHMS
Algorithm 2 Edges’ degree based swap
Input: A, N
Output: Aup
1: for all NGdo
2: Find Niand Nk
3: if Ni6=Njthen randomly select
4: Nineighbor Nj
5: Nkneighbor Nl
6: end if
7: end if
8: if eij && ekl are independent then
9: Calculate dij =pdi×djand
10: Calculate dkl =dk×dl
11: end if
12: end if
13: Calculate difference of edge degrees
14: DIF0=|dij dkl |
15: DIF1=|dik dj l|
16: DIF2=|dil dj k|
17: EDD = min(DIF0,DIF1,DI F2)
18: if EDD == DI F1then Swap edges
19: eil &eij
20: Calculate robustness Rup
21: else if EDD == DI F2then Swap edges
22: eik &ejl
23: Calculate robustness Rup
24: end if
25: Endif
26: if Rup > R && Network is connected then
27: Update Aup A
28: end if
29: end if
30: end for
31: end for
11-14) and the minimum degree difference is selected (Line 15). The edges are
swapped according to the minimum difference in edge degree and after each swap,
the robustness is calculated (Lines 16-22). The updated robustness value is com-
pared with the initial network robustness for all the edge swaps and for which the
robustness is high the adjacency matrix is updated (Lines 23-24).
The algorithm 3explains the nodes’ distance based edge swap method. This swap
is performed after the edge’ degree based swap. The following variable is used in
this algorithm.
27 Thesis by: Muhammad Usman
3.4. DETAILS OF ISFNS ALGORITHMS
Algorithm 3 Nodes distance based edge swap
Input: A, N
Output: Aup
1: for all NGdo
2: Randomly select Niand Nk
3: if Ni6=Nkthen
4: Select neighbors Njand Nl
5: else
6: Randomly select Nk
7: Select neighbors Njand Nl
8: end if
9: endif
10: if eij are independent of ekl then calculate
11: D1for Ni, Nj
12: D2for Nk, Nl
13: Avg0= Avg(D1, D2)
14: Repeat Steps 10-12 for edge swap
15: (eik, eij ) and (eil , ej k)
16: end if
17: endif
18: ND = min(Avg0,Avg1,Avg2)
19: if ND == Avg1&& R(Aup)R(A)then
20: Update Aup A
21: else if ND == Avg2&& R(Aup)R(A)then
22: Update Aup A
23: end if
24: endif
25: end for
26: endfor
Aup : Updated adjacency matrix.
In the graph Gof nodes N, randomly select nodes Niand Nk(Line 2). Then
the neighbor nodes of Niand Nkare found (Lines 3-8). If the edges (eij,ekl) are
independent then the distances D1and D2are calculated between the nodes that
form the independence edges (Lines 10-11). The average of D1and D2is calculated
(Line 12). For the other possible edges’ swaps, repeat steps 10-12 (Lines 13-14).
The edge swap is selected against which the average distance is minimum (Line
16). If the original edge pair has the minimum average distance then the original
topology is kept. On the other way, the edges that have the minimum average
distance are selected. If the robustness is increased and the distance between
the nodes is minimum then the adjacency matrix is updated (Lines 17-20). The
process is repeated until all the edges in the graph Gare examined.
28 Thesis by: Muhammad Usman
3.5. SIMULATION RESULTS
3.5 Simulation results
The performance of the proposed ISFNs scheme is evaluated in this section. After
the network evolution, the edges’ degree and nodes’ distance based swaps oper-
ations are used to optimized the network. Furthermore, the network robustness
is assessed against random and malicious attacks. Moreover, the network con-
nectivity against the variable attacks is studied to make the network robust. In
addition, the comparison of ISFNs is made with two existing schemes, ROSE and
SA against different sized networks. Moreover, a system of core i7 having 7th
generation with 16GB RAM and 256GB SSD is used to perform simulations in
MATLAB.
3.5.1 Construction of scale-free network
To validate the performance of the proposed scheme, the synthetic network is
constructed. The sensor field is set to 500×500m2in which the nodes are randomly
deployed. To make sure that each node is connected, the communication range
of nodes is set to 50% of the sensor field. The high communication range enables
the nodes to form a dense network [?]. The maximum and minimum values of
each node’s degree is set to 25 and 2, respectively. Throughout the simulation, the
edge density m= 2. During the network evolution, the BA model is followed and
a single node is added at each interval and the new nodes follow the preferential
attachment.
3.5.2 Comparison of ISFNs with ROSE and SA
The performance of the proposed scheme is evaluated against the existing tech-
niques, ROSE and SA when N= 100. The comparison is shown in Fig. 3.4.
All the techniques improve the robustness of the network as compared to the ini-
tial network. However, the proposed scheme outperforms all the existing and the
initial network. The high robustness of ISFNs is due to the formation of better
onion-like structure. Moreover, the operations as part of the ISFNs reduce the
redundant operations to optimize the network against malicious attacks. From
the results, it can be concluded that the networks are initially robust and the
optimizations enhanced the robustness. The minimum value of SA is due to the
local optima problem. Moreover, due to the possible wrong selection of the center
node and the redundant operations caused by the degree difference and angle sum
operations reduced the performance of ROSE.
29 Thesis by: Muhammad Usman
3.5. SIMULATION RESULTS
Figure 3.4: ISFNs evaluation of robustness
3.5.3 Comparison of ISFNs with ROSE and SA against
low degree attacks
The performance of all techniques is analyzed against the high degree attacks. In
these attacks, the nodes are randomly removed from the network and the effect
of nodes removal on robustness is assessed. The performance of all the optimized
networks against the low degree attack is shown in Fig. 3.5. All the techniques
Figure 3.5: Network’s performance evaluation against low degree attacks
have downward trends due to the low degree nodes removal. As the SFNs are
robust against the low degree nodes attacks. Therefore, the network robustness
decreases gradually. The high robustness against the low degree attacks confirm
that the network remains scale-free after optimization.
30 Thesis by: Muhammad Usman
3.5. SIMULATION RESULTS
3.5.4 Comparison of ISFNs with ROSE and SA against
high degree attacks
SFNs collapse due to the malicious attacks on the high degree nodes. SFNs have
a very small number of high degree nodes that is due to the power-law followed
by these networks. The effect of high degree attacks on the networks that are
optimized by different techniques is shown in Fig. 3.6. The robustness of the
Figure 3.6: Network’s performance evaluation against high degree attacks
network decreases sharply after the high degree nodes are removed. All the tech-
niques have the downward trends, however, due to the change in robustness value,
the schemes could not be compared directly. Therefore, against the number of
removed nodes, the slope is analyzed. As in Fig. 3.6, the existing techniques and
the initial network collapsed after the removal of almost 20 nodes. However, the
proposed ISFNs scheme performs much better and the network fragments after
the removal of 40 nodes. The high robustness value ensures that the network is
optimized with the better formation of an onion-like structure. The two opera-
tions introduced in ISFNs proved their usefulness by providing a robust network
structure against malicious attacks.
3.5.5 Comparison of ISFNs with ROSE and SA against
random attacks
SFNs are robust against the random removal of nodes. The high robustness is
due to the presence of a large number of nodes with a low degree. Therefore,
the networks are analyzed against the random nodes removal. The network’s
performance when the random attacks happen is analyzed, as shown in Fig. 3.7.
After each attack, the techniques have downward trends because the nodes are
31 Thesis by: Muhammad Usman
3.5. SIMULATION RESULTS
Figure 3.7: Network’s performance evaluation against random attacks
removed from the networks. The robustness of ISFNs outperform the existing
schemes and initial network, due to better network structure. Almost 80% of the
nodes are required to be removed from the network to make it collapse. However,
the existing schemes and the initial network resist only against the removal of 60%
of the nodes. The small rise in the ROSE during the 20% and 35% of the nodes
removal is due to the random selection of high degree edges during the degree
difference operation. SA has less robustness as compared to ROSE and ISFNs
because in the network evolution process the constraint of sensor nodes such as
communication range and the threshold value of nodes degrees are not considered,
therefore, the removal of high degree nodes fragment the network.
Figure 3.8: Network’s performance evaluation against variable attacks
3.5.6 Comparison of ISFNs with ROSE and SA against
variable attacks
The network’s connectivity against the removal of different number of nodes is
analyzed in the variable attacks. Moreover, against the multiple nodes’ removal
32 Thesis by: Muhammad Usman
3.5. SIMULATION RESULTS
from the network, robustness is calculated and the number of removed nodes
is randomly selected. The network performance against the variable attacks is
shown in Fig. 3.8. When the number of removed nodes is small, the maximum
number of nodes is present in the MCS. However, as the number of removed
nodes is increased, the nodes in MCS decrease. SA and ROSE performed worst,
however, ISFNs outperform the existing techniques. During the simulations, 3
different numbers of nodes are randomly generated and against all these, ISFNs
has better results. So, these results prove the importance of edges’ degrees and
node’s distance based swaps operations to form the onion-like structure.
Figure 3.9: Network’s performance evaluation against different number of
nodes
3.5.7 Network robustness against different number of nodes
The comparison between ISFNs and the existing techniques for different sized
networks is presented in Fig. 3.9. The increase in the network size causes the ro-
bustness of the network to decrease. It is due to the availability of a large number
of high degree nodes within the same sensor field. Therefore, the removal of high
degree nodes has a severe effect on the network connectivity. All the techniques
have the same downward trends with the increase in the number of nodes. How-
ever, ISFNs outperforms the existing techniques for different number of nodes. It
is due to the better onion-like structure formed by the ISFNs operations. From
the results, it is shown that the proposed scheme is better, when it is implemented
on the dense and complex networks.
33 Thesis by: Muhammad Usman
3.6. CONCLUSION OF THE CHAPTER
3.6 Conclusion of the chapter
Many critical applications have a scale-free nature. These networks have high
robustness against the removal of nodes with a low degree, however, the high degree
nodes removal collapses the network. Therefore, in this chapter, the robustness
of the SFNs is increased by the proposed ISFNs scheme. In this scheme, two
operations, i.e., edges’ degree and nodes distance based edge swaps are used. In
the edges’ degree based operation, the edges are swapped to make the links with
similar degree nodes. Whereas, the nodes’ distance based operation is used to form
the links between the neighboring nodes that are closed to each other. Moreover,
the network is analyzed against the variable attacks. In this attack, different
numbers of nodes are randomly removed from the network and their effect on
the network connectivity is calculated. The degree distribution is not changed by
the ISFNs’s operations, therefore, no extra cost is required. After optimizing the
network, it remains scale-free and is compared with the existing techniques: ROSE
and SA. It is proved that against different network sizes the ISFNs outperforms
the existing techniques.
34 Thesis by: Muhammad Usman
Chapter 4
Network’s topology evolution scheme
35
4.1. SUMMARY OF THE CHAPTER
4.1 Summary of the chapter
In this chapter, against the random and malicious attacks on scale-free networks
a Network Topology Evolution Scheme (NTES) is proposed. In this scheme, the
network field is divided into two parts with uniformly distributed nodes. After the
network’s evolution, the nodes are linked with each other through one-to-many
correspondence. The division of a network field is made by considering that a
network is robust if its size is small. Moreover, to study the hierarchical changes
in the degree of the nodes the k-core decomposition is used. In addition, nodes’
degree and their core based attacks are performed and against these attacks the
proposed scheme is evaluated. Furthermore, the network robustness is analyzed
using three optimization techniques: Artificial Bee Colony (ABC), Bacterial For-
aging Optimization (BFO) and Genetic Algorithm (GA). The results are compared
and a technique that efficiently optimizes the network to increase the robustness
is selected. In the optimization process, we make use of three edge swap methods.
Due to the edge swap, the degree distribution is not changed so, no extra cost for
adding nodes or links is required.
4.2 Scale-free model
For the SFT, the details about the network evolution and the scale-free property
of the network is verified in this section. Initially, the SFT is constructed using
the BA model. Then, to enhance the network robustness, three types of edge
swapping methods are discussed. Moreover, to calculate the robustness of the
SFT, the metric of robustness is studied.
4.2.1 Construction of Scale-Free Network
The operation of dense networks are difficult as they are highly vulnerable to the
attacks that occur on the network links or nodes. The real-world examples of
dense networks are hospitals, military, transportation, etc. The drawback of these
networks is that their operational efficiency deteriorates once the attacks happen.
Therefore, the networks are divided into smaller networks by the graph partition
concept. The small networks’ operations are easy, processing is fast, efficiency is
high and the failures or removal of nodes have less effect on the overall networks.
In this study, we have made a synthetic network by assuming that small-sized
networks are more robust and easy to maintain than large-sized networks. To
make a small-sized network, the network field is divided into two equal parts
36 Thesis by: Muhammad Usman
4.2. SCALE-FREE MODEL
and nodes are randomly deployed. In both parts, the network evolves with equal
number of nodes. The node at the center of the network broadcasts a request
message to its neighboring nodes. Based on the response time, the initial nodes
connect to the center node. After that, the remaining nodes join the network
based on the preferential attachment, as in [57]. The complete process of network
evolution is shown in Fig. 4.1. The dotted line represents the partition of the
NA
NB
NM
CL
ML
C1
C2
C3
C4
Ncu
Ncl
Figure 4.1: Network evolution by adding edges
network, and Ncu and Ncl are the center nodes of the upper and lower networks,
respectively. The network growth starts from the center nodes. After the network
evolves, its both parts are linked by one-to-many correspondence. Moreover, to
increase the network robustness edge swap is performed.
4.2.2 Details of edge swap
The graph theory is used to represent the SFT. With the help of graph theory,
the network is converted to a graph (G) in which the nodes are represented as a
set of vertices V={1,2, ..., N}, whereas, the links between the nodes are shown
as edges E={eij |i, j Vand i6=j}. So, the graph is G= (V, E) and it is
undirected and unweighted graph used to evaluate robustness of SFT. The edge
swap of independent edges is performed to enhance the robustness. Two edges are
independent if all the nodes of these edges are in the same communication range
and they have no extra edge.
In Fig. 4.2, the original topology with the possible first and second edge swaps
are given. As seen in Fig. 4.2(a), the initial topology’s nodes i, j, k and lhave
independent edges eij and ekl. In Fig. 4.2(b) and Fig. 4.2(c), the first and second
possible edge swaps are represented, respectively. Against all the possible edges,
37 Thesis by: Muhammad Usman
4.2. SCALE-FREE MODEL
the robustness is calculated and a pair of edges is selected that gives the highest
robustness.
To enhance the robustness, the edge swap is important because no extra cost is
required to add new node or edges. Therefore, the following types of edge swap
methods are implemented.
1. Edge swap of randomly selected nodes
2. Edge swap of degree based selected nodes
3. Edge swap of distance based selected nodes
4.2.2.1 Edge swap of randomly selected nodes
Nodes iand jare chosen at random from the network to perform a random
edge swap. Then the nodes kand lare selected in the neighborhood of iand
j, respectively. Edges eij and ekl should be independent to make the edge swap.
The robustness is calculated for each edge swap. If the robustness of the network
is increased, then the network is updated. However, new independent edges are
found in case of robustness is not improved. Since, there are many nodes in the
network with a low degree, therefore, in this edge swap mechanism, the probability
of these nodes selection is high.
i
l
j
k
i
l
j
k
i
l
j
k
(a) (b) (c)
Figure 4.2: Edge swap mechanism
4.2.2.2 Edge swap of degree based selected nodes
The degree of the nodes is being used to perform a degree-based edge swap. Ini-
tially, from the network, a high degree node is selected, then a low degree node
from its neighboring region is chosen. The same method is followed for the other
pair of nodes. If the edges are independent, then the edge swap is made, as in
Fig. 4.2. This edge swap connects similar degree nodes. Against all possible
38 Thesis by: Muhammad Usman
4.2. SCALE-FREE MODEL
edge swaps, the robustness is calculated and the edge swap that enhances the net-
work robustness is selected. To reduce the possibility of similar edge selection, the
edges are marked. So, in the next edge swap, these edges are not selected, hence,
computational cost is reduced.
4.2.2.3 Edge swap of distance based selected nodes
In this edge swap method, independent edges in the network are marked and
Euclidean distance is calculated against all the nodes. The edge swap is made
in such a way that longer links are formed between network nodes. Against the
attacks, the existence of long links make the network robust. After making the
long links, the network robustness is calculated. If the network is fully connected
and the robustness is increased the edge swap is accepted and vice versa.
4.2.3 Metric of robustness
When the network is fully connected, it has maximum ability to perform its ser-
vices. The nodes’ failure or removal greatly reduces the network’s performance.
Usually, the importance of a node is calculated based on its degree [14]. In the
network, a node having a high degree is more important than the rest of the nodes.
In this thesis, malicious attacks happen on the nodes based on their degree. To
perform the malicious attacks, the nodes’ degree is calculated and the highest
degree node is removed from the network. For the remaining nodes, the degree
is recalculated and a node having the highest degree is removed. The removal of
nodes is carried out until the network becomes fragmented. These attacks are more
vulnerable for the SFT, therefore, a metric is required to calculate the robustness.
In recent papers, percolation theory based metric is proposed by Schneider et al.
[131], which is extensively used to calculate the robustness of SFT. In this metric,
when the high degree node is removed, the network collapses and is partitioned
into multiple parts known as subgraphs. The network’s robustness is determined
by the number of nodes present in MCS. To calculate robustness R, we use the
following formula.
R=1
N+ 1
N+1
X
n=0
MCSn
N,(4.1)
where, N,1
N+1 and MCSnare network size, the normalization factor and the
number of nodes in MCS when nth high degree node is removed, respectively.
39 Thesis by: Muhammad Usman
4.3. NETWORK TOPOLOGY EVOLUTION SCHEME OVERVIEW
The robustness value lies in the range of (0, 0.5). The minimum value of robustness
is 0, which means the network is fully collapsed. Whereas, the maximum robust-
ness value is 0.5. Due to the limited resources of sensor nodes, the maximum value
is always less than 0.5.
4.3 Network Topology Evolution Scheme overview
In this section, the NTES is proposed to enhance the robustness of SFT. NTES
provides solutions for the decentralized system. The scheme is designed to be
robust against malicious attacks by forming onion-like structure. In this structure,
the center nodes of the network have a high degree. The nodes’ degree decreases
hierarchically when we move away from the center.
Considering the importance of the onion-like structure for the robustness of SFT,
the network topology is constructed. The NTES consists of the following opera-
tions: network topology evolution, networks connection by one-to-many correspon-
dence, SFT attacks, core based attacks and a comparison of heuristic algorithms
to optimize the NTES’s robustness is made.
4.3.1 Evolution of network topology
For malicious attacks, a small sized network is robust. The results can be observed
from [62], [70], [72]. Therefore, the network field is divided into two parts and
nodes are uniformly distributed in it. During the evolution of both parts, the
power-law is followed. The connection of nodes have a major role in network
robustness. The one-to-many correspondence is better as compared to one-to-one
correspondence [132]. Therefore, the connection of both parts is made by one-to-
many correspondence. The high degree nodes of one part connect with low degree
nodes of the other part of the network. Thus, the degree of the edges becomes
smaller, therefore, the effect of the malicious attacks on the links decreases. Two
networks’ topology evolution following the power-law distribution are shown in
Fig. 4.1. The network field division is indicated by the dotted line and the
two portions have the same number of nodes. The blue nodes (NA) and black
nodes (NB) represent the network A and B, respectively. Whereas NMdenotes
the mutual nodes of the network that exists in both parts. The black solid lines
and double dashed lines denote connectivity links CLand the mutual links ML,
respectively. During the network evolution, the nodes are added asynchronously
in both parts.
40 Thesis by: Muhammad Usman
4.3. NETWORK TOPOLOGY EVOLUTION SCHEME OVERVIEW
4.3.2 k-core based nodes’ degree distribution
Different rings based on the degree of the nodes are presented in the onion-like
structure. In each ring, the nodes with the same degree are connected [62]. The
computational cost is high to collapse the network with malicious attacks based
on degree. Therefore, due to the availability of information about a specific node
in the core, less computational cost is incurred. As a result, in each ring, k-
core decomposition determines the nodes’ degree and a node is removed from it
respective ring based on its importance. For these nodes, the core based onion-like
structure is shown in Fig. 4.1.
In k-core decomposition, the cores are created by removing the nodes from the
network. In Fig. 4.1, core C4 contains the IDs of nodes having a low degree,
which are initially removed from the network. Then the other low degree nodes
are removed after recalculating the degrees and their information is stored in the
next core C3. The node removal process is repeated until all of the high degree
nodes have been removed and placed in the internal core C1. Due to the power-
law, a long tail of nodes with low degree presents in SFT. Hence, the removal
of a high degree node from the network causes a specific part of the network to
collapse. Therefore, in that case, less computational cost is required to damage
the network.
4.3.3 Attacks on the proposed topology
It is assumed that the attackers carry complete network topology information and
can execute any attack to collapse the network. Therefore, having the knowledge
about the specific type of attack make the defender capable to manage it. To
increase the effectiveness of the proposed NTES, nodes’s core and degree based
attacks are considered. In core based attack, the nodes are removed from their
respective core. The node of the inner core are removed first then the nodes of the
outer cores are removed. The core based attack is shown in the Fig. 4.3, where
the nodes NRare removed from the inner core. The network is not disturbed by
removing these nodes. However, as the number of removed nodes are increased
the network is fragmented into multiple subgraphs as shown in Fig. 4.4. Three
subgraphs S1, S2 and S3 are made after the core based attacks are happened on
the network. Furthermore, in each subgraph, the high degree nodes NM CS are
present. These nodes are removed to fully collapse the network.
Moreover, the High Degree Adaptive (HDA) attack is considered to remove the
nodes based on the degree. In this attack, the degree of nodes’ present in the
41 Thesis by: Muhammad Usman
4.3. NETWORK TOPOLOGY EVOLUTION SCHEME OVERVIEW
NA
NB
NM
NR
CL
ML
Figure 4.3: Nodes’ attack based on core
NA
NB
NM
NR
NMCS
CL
ML
S1
S2
S3
Figure 4.4: Attacks on NMC S of the network
network is calculated and the highest degree node is removed. Again, the highest
degree node is removed by recalculating the degree. This process is repeated until
all the nodes are removed from the network.
4.3.4 NTES’s optimization by heuristic algorithms
The NTES is optimized by three heuristic algorithms including GA, ABC and
BFO. In GA, the edge swap is performed by considering the exclusive edges [69].
However, in both ABC and BFO, for the better exploration and exploitation
a random position change is required. In the proposed scheme, the nodes are
stationary, therefore, it is not possible to change positions at random. To deal with
this problem, random and degree based edge swaps are used for the exploration
and exploitation, respectively. The exploitation is performed by exploiting local
information of nodes that is the degree of nodes to perform edge swap. However,
42 Thesis by: Muhammad Usman
4.4. DETAILS OF NTES ALGORITHMS
when the solution traps into the local optima the exploration by random edge
swap is performed.
Table 4.1: Mapping the identified limitations, their proposed solutions with
validations
Identified limita-
tions
Solutions proposed Validations done
L1: The effects of
malicious attacks are
severe on large-sized
networks [57,66]
S1: NTES is proposed
in which small-sized
networks evolve
V1: The small-sized
networks are robust to
random and malicious
attacks, as shown in
Fig. 4.8(b)
L2: The links degree
that connect the net-
works do not follow
the power-law [132]
S2: Using the concept
of the interdependent
links, the networks are
connected
V2: The power-law
degree distribution is
validated for the mu-
tual nodes, as shown
in Fig. 4.6(a)
L3: The change
of node’s degree in
each ring considering
onion-like structure is
not known [57]
S3: The same degree
nodes are found using
k-core decomposition
V3: The nodes are re-
moved based on their
degrees, and degree
based cores are cre-
ated in Algorithm 5
L4: Random edge
swap increases the
number of redundant
operations [66]
S4: Long links are
created through dis-
tance based edge swap
V4: Existence of long
links enhances the
network robustness,
as shown in Fig.
4.8(a)
Table 4.1 presents the complete details of the limitations that are identified through
the literature review. Then their proposed solutions and how they are validated
is given.
4.4 Details of NTES algorithms
In this section, the complete details of the algorithms to increase the robustness of
SFT are presented. Algorithm 7describes the network topology evolution process.
The details of the variables used in the algorithm are as follows. N,r,E,m,thr,
N.nei and Arepresent the total number of nodes, communication range, number
of edges, edge density, threshold value of maximum degree, number of neighbors
and adjacency metric, respectively.
This algorithm works as follows. Initially, based on the coordinates, the network
field is divided into two parts and nodes are uniformly random deploy (Lines 2 and
43 Thesis by: Muhammad Usman
4.4. DETAILS OF NTES ALGORITHMS
Algorithm 4 Network topology evolution
Input: N, r, E, m, thr
Output: N.N ei, A
1: procedure Network generation
2: Divide the network field based on coordinates
3: Random deployment of nodes
4: Find Ncu
5: for all niNdo
6: Center node broadcasts request message
7: CReq Neighboring nodes
8: NiCReq
9: if DNi== 0 then
10: Make an edge with the node that replies first
11: else
12: Calculate Pk
13: Pki=ki
Pjkj
14: end if
15: end if
16: Sort Pk
17: if DNH6=thr then
18: Select NH
19: else
20: Select Second NH
21: end if
22: end if
23: if m2then
24: Roulette wheel based node selection
25: end if
26: end if
27: Update: N1.N ei
28: Repeat the process for lower part of the network
29: Update: N2.N ei
30: Connect both parts using Nm
31: Update A
32: end for
33: end for
34: end procedure
35: end procedure
3). After the network field partition and nodes’ deployment, a center node Ncu is
found (Line 4). The Ncu node broadcasts the connection request message to all the
nodes in its neighborhood (Lines 6 and 7). The replies are sent by the neighboring
nodes (Line 8). If all the neighboring nodes have zero degrees, then an edge is
made with the node that replies first (Line 10). Moreover, if the degree of the
neighboring nodes is different, then the connection probability is calculated (Lines
44 Thesis by: Muhammad Usman
4.4. DETAILS OF NTES ALGORITHMS
12 and 13). The addition of nodes in the network is based on the edge density m.
Due to the limited communication range of the nodes, the threshold value of nodes’
degree is fixed. Hence, a new node only connects with a node that has not reached
the maximum degree limit (Lines 15-20). Moreover, if the neighboring nodes have
different degrees, then mnodes are selected by roulette wheel selection (Lines 21
and 22). After that, the list of neighboring nodes is updated (Line 24). The
complete process is repeated for the lower part of the network and the neighbor
list is updated (Lines 25 and 26). Both parts of the network are connected by
the mutual nodes Nmand the links follow the power-law distribution (Line 27).
Finally, the adjacency metric is updated (Line 28).
To improve the robustness of SFT, Algorithm 5performs edge swap. The edge
swap increases the number of nodes in the MCS, which makes the network more
robust. The variables used in this algorithm are N, N.N ei and A. From the
graph G, two different nodes Niand Nkare randomly selected (Lines 3 and 4).
Then the neighboring nodes are found randomly (Lines 5-7). Afterwards, the
initial robustness Riis calculated (Line 8). The interdependency of the edges is
checked and the robustness is calculated after each edge swap (Lines 10-14). The
edge swap that results in the maximum value of robustness is selected and the
adjacency metric is updated (Lines 16 and 17). Furthermore, in an edge swap
based on node degree, two nodes with higher degrees are chosen. Then, the low
degree neighboring nodes are found and the initial robustness is calculated (Lines
18-23). The edges should be independent and the edge swap is made in all the
possible ways and network robustness is calculated (Lines 24-28). The edge swap
resulting in the maximum value of robustness is selected (Line 30). The adjacency
metric is updated by considering the Nnum.
The attacks’ procedure for the robustness of SFT is explained in Algorithm 8.
After the network evolution, low degree nodes are found in the network (Line 3).
These nodes are removed and their indices are stored in the variable C(i) (Line
4). When all nodes are removed, then C(i) is converted into Ncthat contains
the number of nodes based on the cores (Line 5). The nodes are removed based
on core and the robustness is calculated (Line 7). For the degree based attacks,
the degree of nodes in the network is calculated. The robustness calculation is
performed, after the removal of high degree node (Lines 8 and 9). There is a
new calculation of a degree and the nodes with a high degree are removed (Lines
10-14). The process is repeated, until all the nodes are removed from the network.
The robustness is calculated after each high degree node is removed.
45 Thesis by: Muhammad Usman
4.5. SIMULATION RESULTS AND DISCUSSION
Algorithm 5 Edge swap strategies
Input: A, E, N
Output: A
1: procedure Edgeswap()
2: for all NGdo
3: Select Random Node Ni
4: Select Random Node Nk
5: if Ni6=Nkthen
6: Nj=Neighbor of Ni
7: Nl=Neighbor of Nk
8: Calculate Ri
9: end if
10: end if
11: if eij && ekl are independent then swap
12: eik and ejl
13: R1
14: eil and ekj
15: R2
16: end if
17: end if
18: Nnum =max(Ri, R1, R2)
19: Update A according to Nnum
20: For degree based swap
21: Select Ni=NHD1
22: Select Nk=NHD2
23: Nj=Neighbor of Ni
24: Nl=Neighbor of Nk
25: Calculate Ri
26: if eij && ekl are independent then swap
27: eik and ejl
28: R1
29: eil and ekj
30: R2
31: end if
32: end if
33: Nnum =max(Ri, R1, R2)
34: Update A according to Nnum
35: end for
36: end for
37: end procedure
4.5 Simulation results and discussion
In this section, we first make a synthetic network using the proposed NTES. The
networks are evolved by dividing the network field and are connected by one-to-
many correspondence. The robustness is calculated after removing the high degree
46 Thesis by: Muhammad Usman
4.5. SIMULATION RESULTS AND DISCUSSION
Algorithm 6 Nodes’ degree and k-core based attacks strategy
Input: A, E, N , r
Output: RC, RD
1: procedure Attackonnetwork()
2: for all NGdo
3: Find low degree node
4: C(i)DN
5: NcC
6: Remove nodes based on Nc
7: Calculate RC
8: Find DN
9: Sort DN
10: Remove Dmax node
11: if Di6= 0 then
12: Recalculate DN
13: Sort DN
14: Remove Dmax node
15: end if
16: end if
17: end for
18: end for
19: end procedure
nodes from the network. For the optimization of the network, GA, ABC and BFO,
are used and their performance is compared based on robustness. Moreover, the
NTES is compared with BA model and HC algorithm to validate its performance.
The simulations are performed in MATLAB on a core i5, 6th generation system
having 8GB RAM and 512GB HDD.
4.5.1 Network topology evolution
The nodes are randomly deployed in the network field of 500 500m2and the
total number of nodes is 100. The nodes’ communication range is 50% of the
total size of the network field. Since our work is to make small-sized networks,
therefore, based on the coordinates, the network field is divided into two parts.
The nodes are uniformly distributed in both parts. Due to the nodes’ resource
constraints, the minimum and maximum values of the nodes’ degree are set to 2
and 25, respectively.
4.5.2 Attacks on upper and lower networks
After the completion of network topology evolution, on both parts of the net-
work the attacks are performed. In random attacks, nodes are removed randomly,
47 Thesis by: Muhammad Usman
4.5. SIMULATION RESULTS AND DISCUSSION
however, in malicious attacks based on degree the nodes are removed. The effect
of these attacks is shown in Fig. 4.5. To efficiently analyze the effect of nodes’
Figure 4.5: (a) Malicious and random attacks on upper network. (b) Malicious
and random attacks on the lower network
removal on the robustness of the network, two nodes are simultaneously removed
during each attack. Initially, the robustness remains the same, as shown in Fig.
4.5(a). However, as the number of removed nodes is increased, the malicious at-
tacks become severe as compared to random attacks. When 15 pairs of nodes are
removed, the malicious attacks damage the network more. Moreover, against the
random and malicious attacks, the difference in robustness is small that proves the
effectiveness of the proposed technique. Moreover, the value of robustness is not
zero against random and malicious attacks. It is because the attacks are performed
on one part of the network, however, the second part is still connected. The lower
part of the network followed the same approach. The malicious and random at-
tacks effect is different for the lower part of the network, as shown in Fig. 4.5(b)
because of the separate network evolution. Initially, when the small number of
nodes are removed, the effect of random and malicious attacks is the same. How-
ever, as the number of removed nodes is increased, the malicious attacks become
48 Thesis by: Muhammad Usman
4.5. SIMULATION RESULTS AND DISCUSSION
more vulnerable than the random attacks. The robustness value becomes constant
when 88% of the nodes are removed and the network is completely fragmented.
4.5.3 Networks connection by one-to-many correspondence
When the nodes’ degree follow the power-law distribution, the networks become
robust against the random attacks. Therefore, the mutual nodes are connected
by considering the power-law. In Fig. 4.6(a), for the mutual nodes the degree
distribution is shown. dis the degree of node and P(d) represents the probability
of nodes having the degree d. As per the definition of the power-law, the number
of nodes having high degree should be less as compared to the number of low
degree nodes. The results show that the low degree nodes outnumber the high
Figure 4.6: (a) Power-law distribution of the mutual nodes (b) Comparison
of core based attacks and high degree node attacks
degree nodes. So, the power-law is followed by the mutual nodes. Moreover, due
to the predefined limit of the nodes to connect with the other part of the network,
only a few nodes are present in the mutual part of the network.
49 Thesis by: Muhammad Usman
4.5. SIMULATION RESULTS AND DISCUSSION
4.5.4 Core based attacks on network
The core based attacks are performed in Fig. 4.6(b) to study its effect on the
robustness of the SFT. The removal of nodes starts from the inner core that
contains the important nodes based on degree. The robustness is calculated with
the removal of nodes from the outer core. The outer