Content uploaded by Nadeem Javaid

Author content

All content in this area was uploaded by Nadeem Javaid on Jan 12, 2021

Content may be subject to copyright.

Int. J. of Web and Grid Services, Vol. x, No. x, 202X 1

Attack Resistance based Topology Robustness of

Scale-free Internet of Things for Smart Cities

Talha Naeem Qureshi1, Nadeem Javaid∗,1,

Ahmad Almogren∗,2, Zain Abubaker1,

Hisham Almajed2, Irfan Mohiuddin2

1Department of Computer Science, COMSATS University Islamabad, Islamabad

44000, Pakistan

2Department of Computer Science, College of Computer and Information

Sciences, King Saud University, Riyadh 11633, Saudi Arabia

∗Correspondence:

Nadeem Javaid; (nadeemjavaidqau@gmail.com, www.njavaid.com),

Ahmad Almogren; (ahalmogren@ksu.edu.sa)

Abstract: In Internet of Things (IoTs), the increase in the number of devices is

directly proportional to the number of applications. The exponential growth of devices

increases both the network complexity and risk against topology robustness. Moreover,

the network is also prone to targeted and malicious attacks. In this paper, Enhanced Angle

Sum Operation ROSE (EASO-ROSE), Enhanced ROSE, Adaptive Genetic Algorithm

(AGA) and Cluster Adaptive Genetic Algorithm (CAGA) are proposed to cater the

topology robustness issue for IoT enabled smart cities. In addition, the proposed solutions

keep the nodes’ initial degree distribution unchanged by maintaining the scale-free

nature of the topology. Enhanced ROSE and EASO-ROSE signiﬁcantly improve the

topology robustness by calculating the diﬀerence in nodes’ degree while rearranging

the surrounding angles according to the highest degree node. CAGA and AGA also

signiﬁcantly improve the topology robustness by using adaptive probabilities of crossover

and mutation that guide both algorithms to converge towards global optimum solution.

Extensive simulations are preformed to evaluate the performance of the proposed

strategies. Schneider Ris used as a performance metric in the simulations. The results

depict that the proposed algorithms perform 61.3%, 48.3%, 45.5% and 34.95%, better

than simulating annealing algorithm.

Keywords: Internet of Things, topology robustness, malicious attacks, data driven, scale-

free.

1 Introduction

In the new era of automation, Internet is a global

network whose primary aim is to connect people and

smart devices together. Stupendous expansion of smart

devices emerges as Internet of Things (IoTs). Due to

massive increase in IoT devices, it is estimated that the

number of IoT devices will reach up to ﬁfty million in

the near future S.E. Collier (2015). Diﬀerent networks

are becoming a part of the IoT network, such as

ﬁfth generation cellular network Tsai et al. (2015),

heterogeneous sensor networks T. N. Qureshi et al.

(2013), ﬂying ad-hoc networks W. Zafar et al. (2016),

hybrid mobile networks Han et al. (2017), N. Javaid et

al. (2018), Waluyo et al. (2012), ad-hoc networks T. Qiu

et al. (2017), Nghiem et al. (2013), J. A and Stankovic

(2014), underwater Wireless Sensor Networks(WSNs) I.

Azam et al. (2017) and IoT-Enabled Smart Homes B.

Hussain et al. (2018). The IoT network becomes more

complex, complicated and vulnerable to fragmentation

by increasing the nodes’ density and network size.

Additionally, IoT nodes process a huge amount of

data that makes the network sensitive and prone to

serious challenges: topology robustness and delay. IoT

has a variety of applications, such as monitoring of

power networks, health care, transportation networks,

industrial and home automation systems, etc. These

applications result in excessive sensor nodes’ deployment

and conversion of a traditional city into a smart city S.E.

Collier (2015), T. Qiu et al. (2017) as described in Fig. 1.

Data is collected and sent to big data servers using the

sensors Taniar et al. (2017). Moreover, smart grids are

also part of the IoT network and they are considered as

one of the biggest applications of IoT S.E. Collier (2015).

Copyright ©201X Inderscience Enterprises Ltd.

2Talha Naeem Qureshi et al.

However, failure in the IoT network of synchronous grids

and power plants causes life threatening damage Taniar

et al. (2008).

IoT networks produce multiple types of data in

extensive quantity. In current era, topology robustness

of IoT networks (based on big data servers) has emerged

as a serious concern Wu et al. (2011). Small world and

scale-free are two commonly used models in complex

network theory. The main purpose of a scale-free model is

to model homogeneous network topologies. It has many

applications that include mobile networks N. Javaid et

al. (2017), cooperative networks N. Javaid et al. (2017),

social networks and transportation system networks

T. Qiu et al. (2017), Waluyo et al. (2009). A small

world model is generally used for heterogeneous network

topologies. It has a relatively high clustering coeﬃcient

and a small average path length.

In IoT networks, the sensors have same

communication range and bandwidth. Therefore, scale-

free models are more suitable and applicable to the IoT

network than the small world model. In a scale-free

model, the degree distribution of nodes is based on power

law distribution resulting in several low degree nodes

T. Qiu et al. (2017). These nodes are often targeted by

random attacks and scale-free networks are proven to

be robust against these attacks J. Melthis et al. (2016),

I. Sohn et al. (2018). Whereas, high degree nodes are

highly vulnerable to intentional and targeted attacks R.

H. Li et al. (2012). Therefore, it is important to enhance

the network robustness and resilience of the scale-free

topologies against malicious attacks M. Mozafari et

al. (2019). Furthermore, scientists and researchers are

continuously ﬁnding ways to make robust scale-free

topologies against malicious, targeted, deliberate and

intentional attacks J. Liu et al. (2019), S. Wang et al.

(2019), S. Wang et al. (2019).

The main contributions of the proposed work are as

follows.

•Multiple data driven strategies are proposed in

order to enhance the robustness of a scale-free IoT

network where deployed nodes are homogeneous in

nature.

•Two enhancements of ROSE are proposed:

Enhanced ROSE and EASO-ROSE T. Qiu et

al. (2017). Both strategies increase topology

robustness of the scale-free IoT networks

without changing the nodes’ degree. Moreover,

the proposed work improves the topology by

making an onion-like structure. Furthermore, the

proposed algorithms outperform the conventional

techniques: ROSE and simulating annealing, in

terms of robustness against malicious attacks.

•Schneider robustness metric Rbased adaptive

probabilities are proposed for crossover and

mutation in CAGA and AGA. Adaptive nature of

the probabilities ensures the convergence towards

a global optimal solution and retention of already

converged solutions.

•CAGA integrates the clustering concept within

a solution space and results in fast convergence

towards global optima.

This work is an extension of T. N. Qureshi et al. (2019).

The rest of the paper is organized as follows. Detailed

related work is presented in Section 2. The proposed

models for the IoT network are given in Sections 3 and

4. Section 5 discusses the topology robustness metrices.

ROSE topology robustness algorithm is explained

in Section 6. The proposed robustness enhancement

schemes: EASO-ROSE, Enhanced ROSE, CAGA and

AGA are presented in Sections 7, 8, 9 and 10,

respectively. Simulation results of the proposed schemes

in comparison to previous schemes are given in Section

11. Finally, the paper is concluded in Section 12.

2 Related Work

Simulating annealing algorithm is used by P. Buesser

et al. (2011). The basic functionality of the simulating

annealing is similar to the Hill Climbing. However,

the algorithm includes edge swapping strategy, which

is based on probability that also records inferior

conﬁgurations in order to improve the topology

robustness. Optimization of a scale-free network is

performed using simulating annealing; however, this

algorithm is based on unnecessary comparison, which

results in decreased computational eﬃciency.

In A.-L. Barabsi et al. (1999), Barabasi-Albert (BA)

model is presented to generate the scale-free topologies.

The generated topologies are based on preferential

attachment concept where the nodes’ degree follows

power law distribution. The criteria for generating the

scale-free topologies in the BA model is given as follows.

•The nodes are added in the network one by one

at each iteration, m≤Nowhere mrepresents the

total number of existing nodes and Norepresents

the total number of nodes allowed in the network.

•Whenever a new node is added to an existing

network, it selects the node iin order to connect.

Node is selected on the basis of probability, which

depends on its degree ki. It means that a node

with maximum edges has a high probability to

be selected as a next hop. This phenomenon is

known as preferential attachment and is described

as Matthew Eﬀect.

Power law distribution is achieved by using the

above mentioned criteria. BA model proves that growing

network converges into a scale-free topology by power

law distribution, which is denoted as p(k)k−γwhere

γ≈3. In the past few years, many researchers have

focused on multiple modeling strategies to enhance the

robustness and resilience of the scale-free networks.

Attack Resistance based Topology Robustness of Scale free IoT for Smart Cities 3

Figure 1: Proposed System Model for Enhanced ROSE and EASO-ROSE

Node

Node

Node

Decision

Processor

Modeling

Agriculture

Industry

Smart

Home

Transport

Security

Smart Cities Cloud

Big Data Server

Health

Care

Topology

Neighbour IDs

Neighbour X axis

Neighbour Y axis

Total Degree

Node List

In the traditional IoT networks, sensors have limited

communication range and some of them have insuﬃcient

neighbors. It results in several preferential attachment

issues in the BA model. Due to these limitations, the

main focus of researchers is on the scale-free network

topologies. In the scale-free networks, there are few

high degree nodes, which are vulnerable to malicious,

intentional and targeted attacks. To tackle the high

degree nodes’ failure issue, a large number of edges are

removed, which causes the network fragmentation. In

order to cater this issue, additional edges are added

between the nodes. However, it requires additional

energy resources, wireless channels and bandwidth. The

structure of the scale-free network is changed that

increases the cost. Thus, there is a need to enhance

topology robustness of the scale-free networks by keeping

same initial degree distribution of the nodes.

In order to measure the robustness of the network

topology, C. M. Schneider et al. (2011), Adhinugraha et

al. (2020) present Schneider Rmetric. The metric refers

to the proportion of the largest connected cluster in

comparison to the total number of nodes. Schneider Ris

used to calculate robustness and resilience of the network

topology against intentional, targeted and malicious

attacks. It also considers the case in which the topology

suﬀers from a major loss without completely collapsing.

Schneider Ris more appropriate for the scale-free IoT

networks rather than previously used metrics.

Hill Climbing algorithm is proposed by H. J.

Herrmann et al. (2011). It uses Schneider Rrobustness

metric for the validation. In the proposed scheme, Monte

Carlo method is used as a base. It includes swapping

of the edges to make an onion-like structure while

maintaining the same degree distribution of nodes. As

a result, robustness of the proposed algorithm increases.

However, the multi-model phenomenon is based on

randomization, which leads algorithm to get stuck into

local optima.

G. Zheng et al. (2012), G. Zheng et al. (2013) propose

two network topologies that are based on a linear growth

evolution concept with accelerated growth evolution

phenomenon for robustness of the scale-free networks. A

practical situation regarding reconstruction of edges, is

considered in both proposed topologies. Energy Aware

BA (EABA) is proposed by Y. Jian et al. (2013). In this

scheme, energy consumption of sensors and connectivity

ranges for scale-free network topology are tuned by

conﬁgurable coeﬃcients. The proposed model considers

both energy consumption and transmission performance;

however, data overhead is increased during modeling

4Talha Naeem Qureshi et al.

process. Similarly, L. Liu et al. (2014) propose a complex

network topology on the basis of both small world and

scale-free networks. As, the IoT nodes have limited

communication range; therefore, the proposed model is

not suitable for the IoT networks. However, model can

be implemented on heterogeneous networks.

V. H. P. Louzada et al. (2013) propose a smart

rewiring solution. A two-fold selection is used to ﬁnd

edges, needed to be compared. Topology robustness

is optimized by converting topology into modular

onion-like structure along with computational eﬃciency.

The model is applied to world air transportation

network. Multi-model phenomenon prevents solution

from escaping the local optima. However, proposed

algorithm does not consider the limited communication

range of the nodes. Thus, it is not applicable to IoT

networks. Memetic algorithm is proposed by M. Zhou

and J. Liu et al. (2014). The proposed algorithm

performs local and global search in solution space to

improve the robustness of topology without changing

degree distribution. Two channels’ operations make the

algorithm complex. The algorithm does not consider the

change in communication range of sensor node, which

makes it complex for the IoT enabled networks.

The authors in W.B. Du et al. (2013) introduce

the shortest path queuing strategy in the BA model.

The strategy is designed for the scale-free networks

and is applied on the transportation system by the

authors. T. Qiu et al. (2016) propose a scheme to

enhance robustness of topological structure of the IoT

networks. The importance is deﬁned by two greedy

factors. Additionally, the concept of local importance

of nodes is deﬁned. Moreover, topology robustness is

enhanced by adding new edges, which change the initial

degree distribution. Thus, it increases the cost of the

system. In addition, the designed scheme does not

consider the limited range of sensor nodes.

ROSE algorithm is proposed in T. Qiu et al. (2017).

The proposed strategy is applied on the scale-free WSNs.

As a result, network topology robustness is increased

using various operations to convert the topology into

onion-like structure. In this scheme, the ﬁrst node

is used as a reference node in angle sum operation.

Nevertheless, in real scenario, onion-like structure is

referred by the highest degree node. The correct selection

of reference node makes the topology more robust.

In T. Qiu et al. (2017), Genetic Algorithm (GA) is

applied on the IoT network. The proposed scheme

searches for the possible solution sets by Schneider R.

Moreover, the ﬁxed crossover and mutation probabilities

are used, which result in convergence towards local

optima. However, premature and slow convergence lead

to poor performance.

3 Scale-free IoT model for Enhanced ROSE

and EASO-ROSE

In the IoT network, sensors have limited energy

resources and communication range. These limitations

have diﬀerent impacts on the construction of the scale-

free topologies. These impacts are as follows.

•Sensors are unable to establish arbitrarily large

distance communication links.

•They have limited connections; hence, they have

limited degree (to keep energy consumption

optimal).

Due to aforementioned constraints, it is diﬃcult to

implement the BA model A.-L. Barabsi et al. (1999).

The proposed system model is motivated from T. Qiu

et al. (2017), T. N. Qureshi et al. (2019), which is

used for topology construction in Enhanced ROSE and

EASO-ROSE, as shown in ﬁg. 1. In order to avoid

aforementioned limitations, we deploy a dense network

topology for the IoT network. Moreover, sensors enable

links with all neighbors within their communication

range. As in the era of developed technology, the

communication ranges of the sensors are increased at

minimum cost. This idea leads the network towards the

dense deployment. Therefore, to support a variety of

dynamic services, a large number of nodes are deployed

in residential areas, smart buildings, smart shopping

centers and smart hospitals to form the IoT network. In

addition, signiﬁcant research is carried out to support

dense network W. P. Tay et al. (2008), M. Hefeeda and

M. Bagheri (2007). The reasons to induce the concept of

a dense network are as follows.

•To make the topology robust.

•To maintain the network coverage where nodes are

likely to be failed.

To achieve the scale-free network properties, sensors

should have communication links with 50% of the nodes

available in the network. To achieve this, sensors must

have enough communication radius; otherwise, high

degree nodes will scatter in the network and lead the

network topology towards fragmentation.

To cater the issues of fragmentation of the network

topology and high degree nodes scattering in the

network, the proposed system model treats every

fragmented segment as a cluster. The scale-free IoT

based topologies are created by transforming each cluster

into an onion-like structure. Afterwards, we fuse all the

clusters in the end. The following are the steps for scale-

free topology construction in the IoT networks.

•There are asynchronous communication links

between nodes, which are in the same range.

•All the edges of the IoT sensor nodes contribute

and form the local world of a respective sensor

node.

Attack Resistance based Topology Robustness of Scale free IoT for Smart Cities 5

Onion Like Structure

Maximum Degree Node

Ring of Second High Degree Nodes

Ring of Third High Degree Nodes

Ring of Forth High Degree Nodes

Ring of Fi h High Degree Nodes

Ring of Low Degree Nodes

The edge having 90 degree or maximum angle

between centroid and mid point. Edge is having

high probability of having equal degree nodes.

Centroid

Figure 2: Onion-like Robust structure against Malicious and Targeted Attacks

•High degree sensor nodes consume high energy

as compared to the low degree nodes. Moreover,

due to the depletion of energy, high degree nodes

fail more quickly than low degree nodes. As a

result, edges between nodes break and topology

gets fragmented.

•Any senor node with the highest edge limit is not

able to create any extra connection or edge more

than a deﬁned limit W. Xiao et al. (2015).

•High degree nodes are more preferable to be joined

by the newly incoming nodes.

•Roulette method is opted by incoming nodes to

deﬁne the degree limit.

A node is connected to the neighbors within its

communication range and form its local world. Whereas,

all nodes form connections on the basis of connection

probability PConnect. Hence, we deﬁne PC onnect for

neighbor node ias follows:

PConnect(i)=Degree(i)/

n

X

j=1

Degree(j),(1)

where the total number of nodes that exists in

a network is denoted by nand Degree(i)refers to

the total number of edges of node i. When a new

node joins the network, it sends its location to big

data servers after obtaining coordinates via Global

Positioning System (GPS). Then, it requests the big data

servers to obtain the data of its neighbors. The big data

server provides the neighbors’ information to all sensor

nodes. After speciﬁc interval of time, all nodes broadcast

their locations to the big data server. Big data server

contains the location information of all nodes belonging

to diﬀerent IoT networks.

4 Scale-free IoT model for CAGA and AGA

The system model for CAGA and AGA algorithms is

motivated from T. Qiu et al. (2017), T. N. Qureshi

et al. (2019), as shown in ﬁg. 3. The sink nodes get

information from topology engine and direct the sensor

nodes regarding that information. Then, data is saved

in big data servers connected to smart cities’ cloud for

further processing. Topology engine and modeling engine

connected to big data servers are responsible to ﬁnalize

the topology of the IoT networks. Afterwards, decision

processor applies optimization algorithm over topology

adjacency matrix, which is obtained through modeling

engine and calculates Schneider R. After that, modeling

engine compares the existing topology with proposed

topology on the basis of Schneider R. Then, topology

engine sends topology change information to all sensors

through sink nodes. Moreover, modeling engine also

computes the adjacency matrix of the topology, as shown

in ﬁg. 4. Then, adjacency matrix is further converted into

chromosomes. Finally, multiple chromosomes are stored

in the population set, referred as solution space.

5 IoT topology robustness metrics

The potential of a network to resist disasters and failures

is termed as robustness T. N. Qureshi et al. (2019). A

network’s robustness is the prime factor that assesses the

resilience of the network against attacks. Some edges or

nodes are destroyed in an IoT network due to energy

depletion or by deliberate malicious attacks. As a result,

network is fragmented. Moreover, preliminary topology

connections are broken down. These attacks are classiﬁed

into two categories.

•Random Attacks: these attacks fragment the

network topology via random selection and

destruction of nodes.

6Talha Naeem Qureshi et al.

Topology

Topology Engine

Topology Extracted

Chromosome

i

j

k

l

i

0

1

1

1

j

1

0

0

1

k

1

0

0

1

l

1

1

1

0

0

1

1

1

0

0

1

0

1

0

Adjacency Matrix

AGA

Population

CAGA Clusters

Cluster 1

Cluster 2

Cluster 3

Cluster 4

Decision

Processor

Big Data Server

Modeling

Agriculture

Industry

Smart

Home

Transport

Security

Smart Cities Cloud

Health

Care

Chromosome

Figure 3: Proposed System Model for CAGA and AGA

IoT Topology Node l

Node k

Node i

Node j

i

j

k

l

i

0

1

1

1

j

1

0

0

1

k

1

0

0

1

l

1

1

1

0

0

1

1

1

0

0

1

0

1

0

Adjacency Matrix

Chromosome

Figure 4: Adjacency Matrix Conversion to Binary Chromosome

•Malicious or Targeted Attacks: these are the

priority based attacks in which all nodes are

prioritized on the basis of their edge density.

Afterwards, nodes having the highest number of

edges are removed, resulting in several connections

breakage at the same time. This process is repeated

over and over again until the whole network is

fragmented. This strategy fragments the complete

topology by attacking few nodes having critical

importance.

All proposed schemes namely, EASO-ROSE, Enhanced

ROSE, CAGA and AGA, prioritize the nodes on the

basis of their degree distribution. The degrees of all

the nodes in IoT topology are extracted from big data

servers of the smart cities. Existing topology robustness

is calculated by malicious or targeted attacks. Prior

to calculate topology robustness, nodes are prioritized

on the basis of their number of edges. Then, nodes

are destroyed by starting from high priority node to

low priority node. The process of prioritized node’s

destruction continues until the complete topology is

fragmented and nodes are isolated. If a network has more

than one high degree nodes, then it randomly elects a

single node and eliminates it from the network.

In topology robustness, it is the proportion of attacks

or percolation threshold resulting in entire structure

crumbling. Besides percolation theory, there are various

robustness benchmarks that employ shortest path Frank

et al. (1970), Latora et al. (2001), Sydney et al. (2008)

and graph spectrum concepts Fiedler et al. (1973). A

threshold value for prioritized destruction of nodes is

denoted by qcthat is the proportion of amount of attacks

required to entirely paralyze and isolate the network

Holme et al. (2002). A condition may arise in which a

network may not completely collapse; however, it may

suﬀer from a huge damage.

Percolation theory based topology robustness metric

Schneider Ris proposed by C. M. Schneider et al. (2011).

Attack Resistance based Topology Robustness of Scale free IoT for Smart Cities 7

If a highest degree node of the network is attacked then

it is removed from the network. Then, Schneider Ris

used to calculate robustness. It is relatively simple in

terms of computations and bit complex in terms of the

highest degree node elimination. Schneider Rcertiﬁes the

resilience of topology that after how many destructive

attacks on high degree nodes, the network still remains

operative. We employ this enhanced standard to evaluate

the proposed algorithms against intentional, deliberate,

malicious and targeted attacks. This metric calculates

the maximum degree of the nodes, sub graphs in

the topology and also estimates the magnitude of

impairment formed due to it. Schneider Ris described

as per equation below:

R= 1/N

N

X

Q=1

s(Q),(2)

where total number of nodes in the topology is

represented by N, the proportion of nodes to be deleted

is denoted by Q, fraction of sensor nodes in the largest

connecting portion after failing QN high degree sensor

nodes is denoted by s(Q). In order to evaluate the

networks of diﬀerent sizes, normalization factor 1/N is

used. Schneider Rvaries from limit 1/N to 0.5. Where

the initial limit (1/N) represents the robustness of an

isolated graph and 0.5 represents the robustness of a full

mesh topology.

6 ROSE

ROSE T. Qiu et al. (2017) is designed to work in a

centralized system. All nodes share their coordinates

with centralized system. Then, neighbor list is calculated

and shared with all nodes. ROSE optimizes the

network topology by converging towards an onion-like

structure. According to C. M. Schneider et al. (2011),

nodes exhibiting onion-like formation are more resilient

against targeted and intentional attacks. The important

characteristic of an onion-like network is that it has high

degree nodes at the center. These nodes form a core

surrounded by low degree nodes attached together in

the form of a ring. The degrees of the nodes gradually

decrease from inner to outer side of onion. ROSE

transforms the topology into onion-like structure as

explained by Schneider (through tri-phase operation,

based on onion-like structure). In which, ﬁrst phase is

the detection of independent edges followed by second

phase of degree diﬀerence operation and ﬁnally the third

stage of angle sum operation.

6.1 Independent edges

We relate the graph as G= (V , E) having Vvertices

and Eedges and form a scale-free network. The whole

network consists of Nnumber of nodes and Eedges

belonging to set M. The following criteria is the baseline

for the selection of independent edges:

•All the nodes i, j, k and lincluding edges eij and

ekl should lie in the communication range of each

other.

•Nodes iand jshould not have any edge with node k

and l, and vice versa. Fig. 5(a) shows the selection

of independent edges eij and ekl.

6.2 Degree diﬀerence operation

C. M. Schneider et al. (2011) and M. Zhou and J. Liu

et al. (2014) found that onion-like structures are robust

against intentional and targeted attacks. It is also noted

that nodes belonging to interconnected rings around the

core are of the same degree. If a high degree node fails

to perform its operation then another high degree node

performs its tasks to make the topology resilient. Then,

to achieve the goal, high degree nodes are connected

with other high degree nodes and low degree nodes are

connected with low degree nodes. As a result, degree

diﬀerence between two nodes in the same ring is reduced.

The basic aim of this operation is to keep the degree

diﬀerence between connecting nodes as low as possible.

After selection of independent edges, ROSE T. Qiu et al.

(2017) calculates the degree of all nodes from both edges

as di, dj, dkand dl. Further, degree diﬀerences SU B0,

SU B1and SU B2are calculated from equations 3 to 5

T. Qiu et al. (2017). All types of swaps between node i,

j,k,lare performed and recorded to calculate topology

robustness using Schneider R.

SU B0=|di−dj|+|dk−dl|,(3)

SU B1=|di−dk|+|dj−dl|,(4)

SU B2=|di−dl|+|dj−dk|.(5)

After the calculation of Rfor all connection options (as

shown in ﬁg. 5), the connection method with maximum

Ris selected. Probability pis calculated to control the

degree diﬀerence operation T. Qiu et al. (2017). Where,

value of plies between range (0,1). Probability pis

calculated as equation 6.

p=max(|di−dk|+|dj−dl|

|di−dj|+|dk−dl|,|di−dl|+|dj−dk|

|di−dj|+|dk−dl|)(6)

pacts as a controlling threshold for swapping. If the

selected edges have low degree diﬀerences as compared

to the previous scenario then the value of Ris checked

for all connection options, as illustrated in ﬁg. 5. The

option meeting the criteria of palong with increasing

Schneider Ris ﬁnally selected.

6.3 Angle sum operation

Onion-like topology is the alignment of nodes and edges

in a ring like pattern. In parallel to the ring pattern

nodes, some horizontal edges also exist with respect

to the core of the onion. ROSE T. Qiu et al. (2017)

takes advantage of these edges and further enhances the

8Talha Naeem Qureshi et al.

Algorithm 1 Degree Diﬀerence Operation

1: Input: V,E,listiand listnode−degr ee

2: Collect neighbor list (listi) and degree distribution list (listnode−degr ee) for nodes (N) from big data servers of

smart city

3: Plot topology G= (V, E) based on data driven from listi

4: Randomly select two edges eij and ekl

5: if eij and ekl are independent edges then

6: Calculate the number of neighbors nifrom listifor node i,j,kand l

7: Degree di→Sum of all neighbors ni

8: Degree dj→Sum of all neighbors nj

9: Degree dk→Sum of all neighbors nk

10: Degree dl→Sum of all neighbors nl

11: SU B0=|di−dj|+|dk−dl|,SU B1=|di−dk|+|dj−dl|,SU B2=|di−dl|+|dj−dk|

12: SU Bmin=min(SU B0,SU B1,S U B2)

13: end if

14: if SU Bmin =SU B0then

15: No change in topology G

16: end if

17: if SU Bmin =SU B1then

18: Remove edges eij,ekl and Add edges eik ,ej l

19: G→G0

20: Calculate R(G0)

21: if R(G0)> R(G), G=G0then

22: end if

23: end if

24: if SU Bmin =SU B2then

25: Remove edges eij,ekl and Add edges eil ,ejk

26: G→G0

27: Calculate R(G0)

28: if R(G0)> R(G)then

29: G=G0

30: end if

31: end if

32: Output: Topology=G

robustness by performing angle sum operation over these

edges. The operation further transforms the topology

into onion-like structure by utilizing the eﬀect of these

horizontal edges. To start with this operation, ROSE

selects the central point of topology as centroid. After

selection of centroid, a random edge eij is elected.

Surrounding angle between centroid and edge eij is

calculated as per equation 7 T. Qiu et al. (2017). The

calculation of angle starts after calculating a center point

of the edge eij and then by drawing a line from center

point to the centroid. The angle formed between drawn

line and edge eij is denoted as αand β. The smaller

angle is recorded as α; whereas, βrepresents the larger

angle. Edge eij is taken as lencfor calculation of αas

shown in ﬁg. 6(a). Virtual lines between centroid to node

iis drawn and recorded as lenb.lenais the line between

centroid and node j. The whole graphical explanation is

shown in ﬁg. 6(a).

α=arccos(len2

b+len2

c−len2

a

2∗lenb∗lenc

).(7)

Angle αwill be subtracted from 180oto calculate

angle βas per equation 8:

β= 180o−α. (8)

From above equations, αand βare calculated for edge

ekl. All possible swaps are performed as per ﬁg. 6(a),

(b) and (c). Angles αand βare calculated in the same

manner for all possible swaps. For assurance, Schneider

Ris again calculated as per algorithm 1.

7 EASO-ROSE

ROSE T. Qiu et al. (2017) is developed to handle the

centralized systems. However, unlike ROSE, proposed

EASO-ROSE extracts nodes’ geographic location and

calculates neighbor information from big data servers.

All the nodes broadcast their location obtained through

built-in GPS to the big data server in smart cities for

localization purposes.

EASO-ROSE performs the degree diﬀerence

operation the same as ROSE and calculates the degree

Attack Resistance based Topology Robustness of Scale free IoT for Smart Cities 9

Algorithm 2 EASO

1: Randomly select two independent edges eij and ekl

2: if eij and ekl are independent edges then

3: Calculate number of neighbors nifrom listifor node i,j,kand l

4: end if

5: Calculate highest degree node, Highestmax =max(listnode−deg ree)

6: Select Centroid =Highestmax

7: for Edges eij and ekl do

8: Surrounding angle α with respect to Centroid for edge eij

9: Surrounding angle β with respect to Centroid for edge ekl

10: SU M1=α+β

11: end for

12: Remove edges eij ,ekl and Add edges eik,ej l

13: for Edges eik and ejl do

14: Surrounding angle α with respect to Centroid for edge eik

15: Surrounding angle β with respect to Centroid for edge ejl

16: SU M2=α+β

17: end for

18: Remove edges eij,ekl and Add edges eil ,ejk

19: for Edges eil and ejk do

20: Surrounding angle α with respect to Centroid for edge eil

21: Surrounding angle β with respect to Centroid for edge ejk

22: SU M3=α+β

23: end for

24: Calculate Maximum Sum SU Mmax =max(SUM1, SU M2, SU M3)

25: then Swap edges with respect to SU Mmax G→G0

26: Calculate R(G0)

27: if R(G0)> R(G) , G=G0then

28: end if

29: Output: Topology=G

(a) (b) (c)

ij

k

l

ij

k

l

ij

k

l

Figure 5: Exchange Options of Independent Edges, (a) Original Topology, (b) Exchange Option 1, (c) Exchange

Option 2

diﬀerence SU B0, SU B1and SU B2for all possible

swap combinations. Swap combinations with the lowest

degree diﬀerence are selected. After that Schneider Ris

calculated to assure the increase in topology robustness

as per algorithm 2. ROSE considers the very ﬁrst node

of topology as reference and centroid for angle sum

operation. Afterward, it converts the topology to onion-

like structure. According to the C. M. Schneider et al.

(2011) and M. Zhou and J. Liu et al. (2014), core or

high degree nodes are interconnected with each other

and form a ring like structure to be resilient against

network attacks. As we move from inner to outer side of

the onion-like structure, high degree nodes are further

surrounded by rings of low degree nodes. Mostly, the

rings of the nodes are not symmetric. It means that

core of high degree nodes is highly probable, not to

be situated at exact center of the onion. To cater

the mentioned issue, proposed scheme EASO-ROSE

10 Talha Naeem Qureshi et al.

(b) (c) (d)

Len a

(a)

c

c

c c

ij

k

Len b

Len c

i

ii

j

m

m1

m2

m

m2 m2

m1

Figure 6: (a) Geometric Details of Angle Sum Operation (b) First Edge Swap Possibility, (c) Second Edge Swap

Possibility, (d) Third Edge Swap Possibility

requests the node’s degree information from big data

servers and calculates the coordinates of high degree

nodes. Afterwards, these core high degree nodes are

taken as centroid to calculate αand β. Equation 9 is

used for calculation of the highest degree node.

Highesti=max(Listnode−degr ee).(9)

Centroid is elected on the base of the highest degree

node using equation 10:

Centroid =Highesti.(10)

Where, surrounding angles (α) are calculated on the base

of Highestias follows:

α=arccos len2

b(Highesti)+len2

c(Highesti)−len2

a(Highesti)

2∗lenb(Hig hesti)∗lenc(H ighesti)

.(11)

8 Enhanced ROSE

Enhanced ROSE is based on data driven approach. In

which all the nodes send the geographic location to

a big data server using GPS sensor. Big data servers

calculate the neighbor list of each sensor on the base of

their communication range. Afterwards, big data server

calculates the degree of each node and maintain a list.

Initial steps of Enhanced ROSE and ROSE are

same, i.e., “Selection of Independent Edges” and “Degree

Diﬀerence Operation”. SUB0, S UB1and SU B2are

calculated in degree diﬀerence operation and the one

with the lowest value is selected to swapping. Swap

is re-validated by ﬁnding Schneider R. If the value of

Rincreases then the swap is accepted; otherwise, it is

reverted back.

Enhanced ROSE and arccos function calculate

the surrounding angle αsame as ROSE. After that

Enhanced ROSE transforms the topology into onion-

like structure using algorithm 3. At the ﬁnal step, two

independent edges are selected from the topology as evx

and eyz . Then, degree information for nodes v, x, y and z

is extracted from big data server’s database. Afterward,

Enhanced ROSE lists the degrees in descending order

and name them as Enhanced1, E nhanced2, Enhanced3

and Enhanced4. Where, Enhanced ROSE utilizes the

degrees in the formula below to obtain SU BE nhanced.

SU BEnhanced1=|Enhanced1−Enhanced2|+

|Enhanced3−Enhanced4|(12)

Then another degree diﬀerence SU BEnhanced2is based

on following formula:

SU BEnhanced2=|Degreev−Degreex|+

|Degreey−Degreez|(13)

Ratio PEnhanced is based on SU BE nhanced1and

SU BEnhanced2as per formula below.

PEnhanced =SU BE nhanced1

SU BEnhanced2

(14)

We use a swap threshold Sthreshold to limit the number

of swaps. The range of swap threshold lies between

(0,1). Then, we compare the PEnhanced with Sthreshold .

If PEnhanced is less than Sthreshold , then evx and ey z

edges are swaped, accordingly.

9 Adaptive Genetic Algorithm

Dual goals of preserving diversity in population and

maintaining the convergence ability of GA are achieved

through adaptive probabilities of crossover and mutation

M. Srinivas et al. (1994). Ability to locate and

converge towards the global optimum solution in

multi-model landscape is the robustness attribute of

Attack Resistance based Topology Robustness of Scale free IoT for Smart Cities 11

Algorithm 3 Enhanced Convergence Operation

1: Randomly select two independent edges eij and ekl

2: if eij and ekl are independent edges then

3: Calculate number of neighbors (ni) from listifor node i,j,kand l

4: end if

5: highest degree node, Highestmax =max(listnode−deg ree)

6: Select Centroid =Highestmax

7: for Edges eij and ekl do

8: Surrounding angle α with respect to Centroid for edge eij

9: Surrounding angle β with respect to Centroid for edge ekl

10: SU M1=α+β

11: end for

12: Remove edges eij,ekl and add edges eik ,ej l

13: for Edges eik and ejl do

14: Surrounding angle α with respect to Centroid for edge eik

15: Surrounding angle β with respect to Centroid for edge ejl

16: SU M2=α+β

17: end for

18: Remove edges eij,ekl and add edges eil ,ejk

19: for Edges eil and ejk do

20: Surrounding angle α with respect to Centroid for edge eil

21: Surrounding angle β with respect to Centroid for edge ejk

22: SU M3=α+β

23: end for

24: Calculate Maximum Sum SU Mmax =max(SUM1, SU M2, SU M3)

25: then Swap edges with respect to SU Mmax G→G0,Calculate R(G0)

26: if R(G0)> R(G)G=G0then

27: end if

28: Randomly select two independent edges evx and eyz then Calculate degrees of nodes v, x, y and zthrough

listnode−degree

29: Randomly select two independent edges evx and eyz then Calculate Degree of nodes v, x, y and zas through

listnode−degree

30: Arrange Degrees of node v, x, y , z in descending order Enhanced1, E nhanced2, Enhanced3and Enhanced4

31: Calculate SU BEnhanced1=|Enhanced1−Enhanced2|+|Enhanced3−Enhanced4|,SUBE nhanced2=

|Degreev−Degreex|+|Degreey−Degreez|and PE nhanced =SU BEnhanced 1

SU BEnhanced2

32: if PEnhanced < Sthreshold swap edges and new edges are evy and exz then

33: end if

GA. Probabilities of crossover Pcand mutation Pm

are tuned according to ﬁtness values of the result.

The solution with increased robustness is protected

while solutions with low robustness value are entirely

discontinued. Therefore, to locate global optimum

solutions, classical GA uses random search in crossover

and mutation processes. Where crossover occurs with

some probability Pc. AGA operation is described

in algorithm 4. Moreover, adaptive probabilities of

crossover and mutation enable multi-goal achievement

by GA. It enables diversity in the solution space or

population along with enhanced convergence capacity

of GA (towards global optima.) Adaptive Pcand Pm

relief the system from manual and optimized selection of

probabilities. Rmax −Rdecreases with the IoT topology

while converging towards the local optima. Where Rmax

is the maximum ﬁtness value of the population and Ris

the average ﬁtness value before and after the crossover

operation.

In AGA, adaptability varies with the values of Pc

and Pmaccording to the ﬁtness function represented as

Schneider R. When solution stucks into local optima,

proposed algorithm increases the probabilities of Pc

and Pm. Moreover, when solution lies near to global

optima, it decreases the probabilities of Pcand Pm.

Concept is achieved by calculating the ﬁtness function

for ﬁrst population. Then ﬁtness value is compared with

maximum ﬁtness value Rmax of the population that is

0.5. Ris the average ﬁtness value after and before the

crossover operation Pcand Pmwhile converging the

solution towards the local optima. Therefore, Pcand Pm

change inversely when Rmax −Rdecreases. The Pcand

Pmat this stage are calculated using equation 15 and 16.

Both equations are derived by taking motivation from

M. Srinivas et al. (1994).

Pc=C1

Rmax −R,(15)

12 Talha Naeem Qureshi et al.

Algorithm 4 AGA Mechanism

1: Input: V,E,listi,listnode−degr ee and φ∈[0,1]

2: Collect listineighbor list and listnode−degr ee degree distribution list for Nnodes from big data servers of smart

city

3: Plot topology G= (V, E) based on data driven from listi

4: Calculate Adjacency Matrix = Adjmatrix

5: Convert Adjmatrix into chromosome and add to population P opi

6: Select parent chromosomes Chromosomeparent by crossover probability Pc

7: Calculate Schneider R(Chromosomepar ent1&Chromosomeparent2∈[P opi])

8: Calculate exclusive eij |eij ∈Chromosomepar ent1&eij 6∈ Chromosomeparent2

9: Calculate exclusive ekl |ekl ∈Chromosomeparent2&ekl 6∈ C hromosomeparent1

10: if eij and ekl are exclusive then

11: Calculate tempneighbours1, neighbors nl|nl∈Chromosomepar ent1&nl6∈ nk

12: Calculate tempneighbours2, neighbors nj|nj∈Chromosomepar ent1&nj6∈ ni

13: Calculate distances of all tempneighbours1&tempneighbour s2

14: sort according to distances

15: end if

16: Candidate1= least distance node from nk∈tempneighbours1

17: Candidate2= least distance node from nj∈tempneighbours2

18: Calculate tempneighbours3, neighbors Candidate1|Candidate1∈Chromosomeparent1&Candidate16∈ nl

19: Calculate tempneighbours4, neighbors Candidate2|Candidate2∈Chromosomeparent2&Candidate26∈ nj

20: Calculate distances of all tempneighbours3&tempneighbour s4

21: sort according to distances

22: Remove edge el,Candidate1,ej,Candidate2,ek,C andidate3and ei,C andidate4

23: Create edge (ekl &eCandidate1,C andidate3)∈Chromosomepar ent1& (eij &eCandidate2,C andidate4)∈

Chromosomeparent1

24: Randomly select two independent edges eij and ekl

25: if eij and ekl are independent edges then

26: Calculate number of neighbors nifrom listifor node i,j,kand l

27: degree di→Sum of all neighbors ni

28: degree dj→Sum of all neighbors nj

29: degree dk→Sum of all neighbors nk

30: degree dl→Sum of all neighbors nl

31: Sort degree (di, dj, dkand dl) in descending order as (d1< d2< d3< d4)

32: end if

33: Calculate sl= (d1−d2)+(d3−d4), s2= (di−dj)+(dk−dl) and p= (s1/s2)

34: if p<φthen

35: Exchange edges as per (d1, d2, d3and d4)

36: end if

37: Calculate Schneider R0

38: if if R(G0)> R(G)then

39: G=G0

40: end if

41: Note: we have used the symbols “& and |” for “logical AND and OR”.

Pm=C2

Rmax −R.(16)

High values to Pcand Pmeﬀect the already

converged solutions. To deal with this, Pcand Pmshould

vary directly as Rmax −R0where R0is the ﬁtness value

of majority of the chromosomes. Now, equation 17 and

18 are used to ﬁnd values of Pcand Pm, derived by

motivating from M. Srinivas et al. (1994).

Pc=C1Rmax −R0

Rmax −R,(17)

Pm=C2Rmax −R0

Rmax −R,(18)

where, C1and C2are constants.

9.1 Crossover in IoT

In GA, a large population space consists of solutions

obtained by applying crossover and mutation operations

on parent chromosomes (Chromosomeparent ) to form

child chromosomes by retaining a part of parent genes.

Where the best solution is required in large solution

space. In proposed algorithms, degree distribution of

Attack Resistance based Topology Robustness of Scale free IoT for Smart Cities 13

Algorithm 5 CAGA Mechanism

1: Input: V,E,listi,listnode−degr ee,C lustersiz e,P opf iltered and φ∈[0,1]

2: Collect listineighbor list and listnode−degr ee degree distribution list for Nnodes from big data servers of smart

city

3: Plot topology G= (V, E) based on data driven from listi

4: Calculate Adjacency Matrix = Adjmatrix , convert into chromosome and add to population P opi

5: Create clusters Cluster1to Clusternas per C lustersiz e factor deﬁned

6: for each Clusterido

7: Calculate crossover probability Pcon basis of Schneider Rof cluster 1

8: Select parent chromosomes Chromosomeparent cluster i

9: Crossover parent chromosomes Chromosomeparent cluster i

10: Mutation parent chromosomes Chromosomeparent cluster i

11: Calculate Schneider R0If R(G0)> R(G)then G=G0,Best Chromosomecluster i =G

12: end for

13: Elect the Best Chromosome and sent to Filtered population P opf iltered

nodes is kept unchanged as described by T. Qiu et al.

(2017). Parent topologies perform crossover operation

with each other to form new child topologies. Where

parent topologies are selected on the basis of adaptive

crossover probability Pc.

Crossover process starts with the extraction of

exclusive edges from both parents. The term exclusive

edges represent the edges, which are only present in

one parent. Topology of Chromosomepar ent1forms an

exclusive edge with Chromosomepar ent2and similarly

Chromosomeparent1as described by T. Qiu et al. (2017).

Complete operation is given in algorithm 4 T. Qiu et al.

(2017).

9.2 Mutation in IoT

The goal of mutation is to introduce new individuals in

GA and to reproduce lost ﬁtness of GA. It also avoids

the solution to stuck into local optima and to stop

from premature convergence. It also helps to ﬁnd the

global ﬁttest solution from solution space and to move

towards the global optima. In proposed AGA, mutation

is performed with adaptive probability Pmand rest of

the process occurs as described by T. Qiu et al. (2017);

however, instead of static probability, it uses adaptive

probability. Mutation phase starts with the selection of

independent edges as explained in previous section 6.1,

graphically explained in ﬁg. 5 and as per criteria are

given below:

eij =(eij ∈Chromosomeparent1,

eij 6∈ Chromosomeparent2,(19)

ekl =(ekl ∈Chromosomeparent2,

ekl 6∈ Chromosomeparent1.(20)

After selection of independent edges eij and ekl, the

degrees for nodei,nodej,nodekand nodelare sorted

in descending order. We call them node1,node2,node3

and node4. We calculate p,s1and s2as per below

expressions, described by T. Qiu et al. (2017):

s1= (d1−d2)+(d3−d4),(21)

s2= (di−dj)+(dk−dl),(22)

p=s1/s2.(23)

After calculating p, we compare it with threshold φthat

ranges between (0,1). If p<φ then we exchange edges

according to descending order as (d1, d2, d3and d4).

Mutation operation is stated in algorithm 4.

10 Cluster Adaptive Genetic Algorithm

In CAGA, the whole population is divided into optimum

sized clusters J. Zhang et al. (2007). The increase

in the number of clusters leads to increase in both

computational and time complexity. To cater the

aforementioned issues, optimal number of clusters are

found and ﬁxed as per equation 24.

Clustersize =T otal chromosomes in population

4.(24)

All the clusters act independently. Then, values of Pcand

Pmare calculated adaptability using equations 17 and

18. Adaptive probabilities are calculated for each cluster,

which are independent and diﬀerent. After crossover and

mutation process, best chromosome of each cluster is

extracted on the basis of the highest level of Schneider

Rvalue. All the best chromosomes of each cluster are

compared on the basis of Schneider R. The chromosomes

with the highest Schneider Rvalue are ﬁnally exported

to ﬁltered population P opf iltered . Parallel processing in

all clusters leads towards fast convergence of algorithm

and better results. Complete operations of CAGA are

given in algorithm 5 and graphically illustrated in ﬁg. 7.

14 Talha Naeem Qureshi et al.

0

1

1

1

0

0

1

0

1

0

0

1

0

0

0

1

1

1

0

1

1

1

1

0

1

0

1

1

0

0

0

1

1

1

0

0

1

0

1

0

1

0

1

0

1

0

1

0

1

1

0

1

1

1

0

0

1

0

1

0

0

1

0

0

0

1

1

1

0

1

1

1

1

0

1

0

1

1

0

0

0

1

1

1

0

0

1

0

1

0

1

0

1

0

1

0

1

0

1

1

0

1

1

1

0

0

1

0

1

0

0

1

0

0

0

1

1

1

0

1

0

1

1

0

1

0

1

1

0

0

0

1

1

1

0

0

1

0

1

0

0

1

0

0

0

1

1

1

0

1

1

0

1

0

1

0

1

0

1

1

Cluster 1

Cluster 2

Cluster 3

Cluster 4

Best Chromosome ex Cluster 1

Best Chromosome ex Cluster 2

Best Chromosome ex Cluster 3

Best Chromosome ex Cluster 4

0

1

1

0

1

0

1

1

0

1

0

0

1

0

1

1

1

1

0

0

0

1

0

0

1

0

1

1

0

1

0

0

1

0

1

1

1

1

0

0

0

1

1

0

1

0

1

1

0

1

Comparing Fitness Value of

Chromosomes Filtered Population

Figure 7: Population Conversion to Clusters and Election of Best Chromosome

11 Simulations and results

After extracting nodes’ deployment locations from smart

cities, IoT topology is simulated using MATLAB. It

helps the cities to become smart and economical by

adopting a circular design S. Prendeville et al. (2018).

Hence, in order to simulate a smart city, the circular

region is selected with 500 meter diameter. In this

regard, a network of 100 IoT sensors is deployed in the

restricted circular diameter. Moreover, to validate the

concept of scale-free topology, communication range of

each IoT sensor is restricted to 200 meters. Furthermore,

for Enhanced ROSE and EASO-ROSE, edge density is

restricted to 40% of the total nodes in the network.

Whereas, for CAGA and AGA, edge density is ﬁxed to

60%.

11.1 Comparison of Enhanced ROSE and

EASO-ROSE with Existing Algorithms

Enhanced ROSE and EASO-ROSE are compared

with ROSE and simulating annealing algorithm. 100

iterations are performed for all algorithms under the

same environment. The results are averaged over

20 simulations. Enhanced ROSE and EASO-ROSE

outperformed ROSE and simulating annealing. The

best results (Schneider Rvalue) obtained by Enhanced

ROSE, EASO-ROSE, ROSE and simulating annealing

are 0.150, 0.126, 1.120 and 0.077, respectively, as shown

in ﬁg. 8.

EASO-ROSE performed better than ROSE because

of modiﬁcation in the angle sum operation phase of

ROSE. Original ROSE considers 1st node centroid and

calculates the surrounding angles. However, in original

onion-like topology, high degree nodes interconnected

with each other form a core, which is surrounded by

rings of low degree interconnected nodes. Unlike ROSE,

EASO-ROSE uses the highest degree node centroid.

Enhanced ROSE outperforms its counter schemes in

terms of eﬃciency. The strategy uses the highest degree

node as centroid. Afterwards, similar degree nodes

tend to connect with each other, which converts the

topology into onion-like structure. It leads to the increase

in topology robustness against malicious attacks. All

algorithms are tested over a network of 100 to 300

nodes. Enhanced ROSE has proven to be the most

eﬃcient in converging the topology to robust onion-like

structure. Enhanced ROSE has 17.4%, 22.2% and 61.3%

better eﬃciency as compared to EASO-ROSE, ROSE

and simulating annealing , respectively. Whereas, EASO-

ROSE shows 4.9% and 48.3% better results as compared

to ROSE and simulating annealing. The results are

shown in ﬁg. 9.

0 10 20 30 40 50 60 70 80 90 100

Number of Iterations

0.05

0.1

0.15

Schneider R

Enhanced ROSE

EASO-ROSE

ROSE

Simulating Annealing

Figure 8: Performance comparison between Enhanced

ROSE, EASO-ROSE and other algorithms

100 150 200 250 300

Number of Nodes

0.05

0.1

0.15

0.2

0.25

0.3

Schneider R

Enhanced ROSE

EASO-ROSE

ROSE

Simulating Annealing

Figure 9: Bar comparison between Enhanced ROSE,

EASO-ROSE and other Algorithms with Diﬀerent Sizes

of Network Topologies in IoT

Attack Resistance based Topology Robustness of Scale free IoT for Smart Cities 15

11.2 Comparison of CAGA and AGA with

Existing Algorithms

In the proposed system, the nodes’ geographic

information obtained from big data server is converted

into adjacency matrix for further processing of CAGA

and AGA. Few parameters are adjusted after numerous

independent simulations including population size,

which is based on the number of chromosomes in

a population. 100 iterations are performed after

considering numerous factors including time complexity

and energy consumption. It is observed that the

proposed algorithms are converged to optimal results

within 100 iterations. Few parameters including Pc

and Pmare adaptively set by the proposed algorithms

depending on the ﬁtness of the population.

For comparison, neighourlimit is set to 20. All the

results are the average of 20 independent simulations.

In ﬁg. 10, y-axis shows the improvement in Schneider

Rfor all compared algorithms during each iteration. In

CAGA, the value of Schneider Ris minimum in the start.

Then, till 22nd iteration, there is a rapid improvement

in the value of Schneider Rdue to conversion of the

population into independent clusters and treating each

cluster with diﬀerent Pcand Pm. The best chromosome

of each population is extracted during each iteration

and transferred to ﬁltered population. All transferred

chromosomes from clusters are evaluated using ﬁtness

function and only the best one is retained. Till 22nd

iteration, parallel processing of crossover and mutation

phases results in rapid increase in Schneider Rin

comparison to AGA, simulating annealing and hill

climbing algorithms. Overall, CAGA performs better

as compared to simulating annealing and hill climbing.

After 22nd iteration, the value of Schneider Rincreases

slowly, this is because the value is already increased

to a considerably high level. It means that each

chromosome in each cluster is at considerably good

ﬁtness level. CAGA and AGA are kept on increasing

till 100th iteration. It is important to be noted that

after 50 iterations, the graph of simulating annealing

and hill climbing are depicting a straight line. This is

because both algorithms are trapped in local optima.

Whereas, adaptive mutation probability Pmmakes

CAGA and AGA to come out of local optima and

move towards global optima. AGA performs better

than simulating annealing and hill climbing but not

well in comparison to CAGA. AGA also keeps on

increasing till end and jumps out from local optima

due to adaptive mutation criteria. Simulating annealing

performs better than hill climbing because simulating

annealing also accepts inferior conﬁgurations with some

positive probability that is discarded in hill climbing.

However, performance of simulating annealing and hill

climbing is signiﬁcantly low in comparison to CAGA and

AGA. CAGA shows 10.94%, 36.3% and 45.5% better

performance as compared to AGA, simulating annealing

and hill climbing algorithms. Whereas, AGA performs

25.6% and 34.95% better than simulating annealing and

hill climbing.

0 10 20 30 40 50 60 70 80 90 100

Number of Iterations

0

0.05

0.1

0.15

0.2

0.25

0.3

Schineder R

CAGA

AGA

SIMULATING ANNEALING

HILL CLIMBING

Figure 10: Performance Comparison between CAGA,

AGA and other Algorithms

For comparison, the value of neighourlimit is set to

20 for 100 nodes, it increases by 5 points when 50 new

nodes are added in the network on each iteration. A

scale-free topology is validated by varying the number

of nodes from 100 to 300 while neighourlimit is kept the

same for all algorithms. Fig. 11 shows the performance

comparison of the proposed algorithms CAGA and AGA

with simulating annealing and hill climbing algorithms

by averaging 20 independent simulations. It can be

graphically observed that value of Rdecreases for all

algorithms with increase in IoT enabled nodes’ density.

For large networks fragmentation resistance is more

complicated as compared to small networks. CAGA

and AGA show better robustness as compared to other

two algorithms. Scalability analysis also validates the

eﬃcacy of proposed CAGA and AGA. Moreover, all

IoT topologies remain connected after implementing the

optimization algorithms.

100

150

200

250

300

Number of Nodes

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

Schineder R

CAGA

AGA

SIMULATING ANNEALING

HILL CLIMBING

Figure 11: Bar Comparison between CAGA, AGA

and other Algorithms with Diﬀerent Sizes of Network

Topologies in IoT

12 Conclusion

This paper improves the topology robustness of a

complex IoT network using two proposed models. The

ﬁrst model consists of Enhanced ROSE and EASO-

ROSE strategies while the second model contains

16 Talha Naeem Qureshi et al.

CAGA and AGA algorithms. In the former, the nodes’

geographical location and information are provided by

the smart cities’ cloud server to get the exact coordinates

of the nodes. Also, the selection of the highest degree

node enhances the reliance and topology robustness

by converting the network into an onion-like structure.

On the other hand, in the latter, AGA uses adaptive

probabilities and maintains a balance between local and

global optima. It prevents the optimal solution from

getting stuck in the local optima and moves it towards

the global optima. Whereas, CAGA uses the concept

of multiple clusters and provides fast convergence by

performing parallel operations. Additionally, the initial

degrees of nodes are kept unchanged. To determine

the topology robustness, Schneider Rmetric is used.

Besides, the high value of Schneider Rmeans that the

topology is robust. Extensive simulations are performed

to validate the eﬀectiveness of the proposed models

by comparing with existing models, such as ROSE,

simulating annealing and hill climbing. From the results,

it is obvious that the proposed Enhanced ROSE

and EASO-ROSE outperform ROSE and simulating

annealing while CAGA and AGA perform better than

simulating annealing and hill-climbing in terms of high

Schneider R.

Acknowledgement

This work was supported by the Deanship of Scientiﬁc

Research at King Saud University under Grant RG-1437-

035.

References

S.E. Collier, “The Emerging Enernet: Convergence of

the Smart Grid with the Internet of Things”, Rural

Electric Power Conference (REPC), 2015 IEEE, 2015

Tsai, C. Wei, H. H. Cho, T. K. Shih, J. S. Pan, and J.

J. Rodrigues. “Metaheuristics for the Deployment of

5G.” IEEE Wireless Communications 22, no. 6 (2015):

40-46.

T. N. Qureshi, N. Javaid, A. H. Khan, A. Iqbal, E.

Akhtar and M. Ishfaq, “Balanced Energy Eﬃcient

Network Integrated Super Heterogenous Protocol for

Wireless Sensor Networks”, International Workshop

on Body Area Sensor Networks (BASNet), 2013

W. Zafar, B.M. Khan, “Flying Ad-Hoc Networks:

Technological and Social Implications”, IEEE

Technology and Society Magazine, 2016.

Han, Guangjie, L. Liu, S. Chan, R. Yu, and Y. Yang.

“HySense: A hybrid mobile crowdsensing framework

for sensing opportunities compensation under

dynamic coverage constraint.” IEEE Communications

Magazine 55, no. 3 (2017): 93-99.

N. Javaid, A. Sher, H. Nasir, and N. Guizani,

“Intelligence in IoT based 5G Networks: Opportunities

and Challenges”, IEEE Communications Magazine,

2018.

Waluyo, Agustinus Borgy, David Taniar, Wenny

Rahayu, Ailixier Aikebaier, Makoto Takizawa, and

Bala Srinivasan. “Trustworthy-based eﬃcient data

broadcast model for P2P interaction in resource-

constrained wireless environments.” Journal of

Computer and System Sciences 78, no. 6 (2012):

1716-1736.

T. Qiu et al., “Heterogeneous Ad Hoc Networks:

Architectures, Advances and Challenges,” Ad Hoc

Networks, vol. 55, 2, Feb. 2017, pp. 14352.

Nghiem, Thao P., Agustinus Borgy Waluyo, and David

Taniar. “A pure peer-to-peer approach for kNN query

processing in mobile ad hoc networks.” Personal and

Ubiquitous Computing 17, no. 5 (2013): 973-985.

J. A and Stankovic, “Research Directions for the Internet

of Things”, IEEE Internet of Things J., vol. 1, no. 1,

Feb. 2014, pp. 39.

I. Azam, N. Javaid, A. Ahmad, W. Abdul, A. Almogren

and A. Alamri, 2017. “Balanced load distribution with

energy hole avoidance in underwater WSNs.” IEEE

Access, 5, pp.15206-15221.

B. Hussain, Q.U. Hasan, N. Javaid, M. Guizani, A.

Almogren and A. Alamri, 2018. “An Innovative

Heuristic Algorithm for IoT-Enabled Smart Homes

for Developing Countries.” IEEE Access, 6, pp.15550-

15575.

T. Qiu, J. Liu, W. Si, M. Han, H. Ning, M.

Atiquzzaman, “A Data-Driven Robustness Algorithm

for the Internet of Things in Smart Cities”, IEEE

Communications Magazine, 2017.

Taniar, David. “Big data is all about data that we

don’t have.” In 2017 International Conference on

Advanced Computer Science and Information Systems

(ICACSIS), pp. 1-8. IEEE, 2017.

Taniar, David, Clement HC Leung, Wenny Rahayu, and

Sushant Goel. “ High-performance parallel database

processing and grid databases.” Vol. 67. John Wiley

& Sons, 2008.

Wu, Jun, M. Barahona, Y. J. Tan, and H. Z.

Deng. “Spectral Measure of Structural Robustness in

Complex Networks.” IEEE Transactions on Systems,

Man, and Cybernetics-Part A: Systems and Humans

41, no. 6 (2011): 1244-1252.

N. Javaid, M. Ejaz, W. Abdul, A. Alamri, A. Almogren,

I.A. Niaz and N. Guizani, 2017. “Cooperative position

aware mobility pattern of AUVs for avoiding void

zones in underwater WSNs.” Sensors, 17(3), p.580.

Attack Resistance based Topology Robustness of Scale free IoT for Smart Cities 17

N. Javaid, A. Sher, W. Abdul, I.A Niaz, A. Almogren

and A. Alamri, 2017. “Cooperative opportunistic

pressure based routing for underwater wireless sensor

networks.” Sensors, 17(3), p.629.

T. Qiu, A. Zhao, F. Xia, W. Si, D. Wu, “ROSE:

Robustness Strategy for Scale-free Wireless Sensor

Networks”, IEEE/ACM Transactions on Networking,

2017.

Waluyo, Agustinus Borgy, David Taniar, Wenny

Rahayu, and Bala Srinivasan. “Mobile service oriented

architectures for NN-queries.” Journal of Network and

Computer Applications 32, no. 2 (2009): 434-447.

J. Melthis, S. Tang, P. Yang, M. Hanneghan, C. Carter,

“Topologies for Combining the Internet of Things

and Serious Games”, Journal of Intelligent and Fuzzy

Systems, 2016.

I. Sohn, “Robustness Enhancement of Complex

Networks via No-Regret Learning”, ICT Express,

2018.

R. H. Li, J. X. Yu, X. Huang, H. Cheng, and Z. Shang,

“Measuring Robustness of Complex Networks under

MVC Attack”, in Proc. 21st ACM Int. Conf. Inf.

Knowl. Manage., New York, NY, USA, Oct./Nov.

2012, pp. 15121516.

M. Mozafari M. Khansari, “Improving the Robustness

of Scale-free Networks by Maintaining Community

Structure”, Journal of Complex Networks, 2019.

J. Liu, H. A. Abbass, K. C. Tan, “Evolving

Robust Networks Using Evolutionary Algorithms”,

Evolutionary Computation and Complex Networks,

Springer, 2019.

S. Wang, J. Liu, “Designing Comprehensively Robust

Networks Against Intentional Attacks and Cascading

Failures”, Information Sciences, 2019.

S. Wang, W. Lv, L. Zhao, S. Nie, H. E. Stanley,

“Structural and Functional Robustness of Networked

Critical Infrastructure Systems under Diﬀerent Failure

Scenarios”, Physica A: Statistical Mechanics and its

Applications, 2019.

T. N. Qureshi, N. Javaid, S. A. Butt, W. Khan, S. A.

Changazi and M. M. Iqbal, “Enhanced Robustness

Strategy for IoT in Smart Cities Based on Data

Driven Approach”, 33nd International Conference on

Advanced Information Networking and Applications

AINA, 2019.

P. Buesser, F. Daolio, and M. Tomassini, “Optimizing

the Robustness of Scale-free Networks with Simulated

Annealing”, in Proc. 10th Int. Conf. Adapt.

Natural Comput. Algorithms ICAN NGA, Ljubljana,

Slovenia, Apr. 2011, pp. 167176.

A.-L. Barabsi and R. Albert, “Emergence of Scaling in

Random Networks.” Science, vol. 286, no. 5439, pp.

509512, 1999.

C. M. Schneider, A. A. Moreira, J. S. Andrade, Jr., S.

Havlin, and H. J. Herrmann, “Mitigation of Malicious

Attacks on Networks,” Proc. Nat. Acad. Sci. USA, vol.

108, no. 10, pp. 38383841, 2011.

Adhinugraha, Kiki, Wenny Rahayu, Takahiro Hara, and

David Taniar. “On Internet-of-Things (IoT) gateway

coverage expansion.” Future Generation Computer

Systems 107 (2020): 578-587.

H. J. Herrmann, C. M. Schneider, A. A. Moreira, J.

S. Andrade, Jr., and S. Havlin, “Onion-like Network

Topology Enhances Robustness Against Malicious

Attacks,” J. Statist. Mech., Theory Experim., vol.

2011, no. 1, pp. 19, 2011.

G. Zheng, S. Liu, and X. Qi, “Scale-free Topology

Evolution for Wireless Sensor Networks with

Reconstruction Mechanism”, Comput. Elect. Eng.,

vol. 38, no. 3, pp. 643651, 2012.

G. Zheng and Q. Liu, “Scale-free Topology Evolution for

Wireless Sensor Network”, Comput. Elect. Eng., vol.

39, no. 6, pp. 17791788, 2013.

Y. Jian, E. Liu, Y. Wang, Z. Zhang, and C. Lin,

“Scale-free Model for Wireless Sensor Networks”, in

Proc. IEEE Wireless Commun. Netw. Conf. (WCNC),

Shanghai, China, Apr. 2013, pp. 23292332.

L. Liu, X. Qi, J. Xue, and M. Xie, “A Topology

Construct and Control Model with Small-world

and Scale-free Concepts for Heterogeneous Sensor

Networks”, Int. J. Distrib. Sens. Netw., vol. 10, no. 3,

2014, Art. no. 374251.

V. H. P. Louzada, F. Daolio, H. J. Herrmann,

and M. Tomassini, “Smart Rewiring for Network

Robustness”, J. Complex Netw., vol. 1, no. 2, pp.

150159, 2013.

M. Zhou and J. Liu, “A Memetic Algorithm for

Enhancing the Robustness of Scale-free Networks

Against Malicious Attacks”, Phys. A, Statist. Mech.

Appl., vol. 410, pp. 131143, Sep. 2014.

W.B. Du, Z.-X. Wu, and K.-Q. Cai, “Eﬀective Usage of

Shortest Paths Promotes Transportation Eﬃciency on

Scale-free Networks”, Phys. A, Statist. Mech. Appl.,

vol. 392, no. 17, pp. 35053512, 2013.

T. Qiu, D. Luo, F. Xia, N. Deonauth, W. Si, A. Tolba,

“Greedy Model with Small World for Improving the

Robustness of Heterogeneous Internet of Things”,

Journal of Computer Networks, Volume 101, 4 June

2016, Pages 127-143

18 Talha Naeem Qureshi et al.

W. P. Tay, J. N. Tsitsiklis, and M. Z.Win, “On

the Impact of node Failures and Unreliable

Communications in Dense Sensor Networks,” IEEE

Trans. Signal Process., vol. 56, no. 6, pp. 25352546,

Jun. 2008.

M. Hefeeda and M. Bagheri, “Randomized k-coverage

Algorithms for Dense Sensor Networks,” in Proc. 26th

IEEE Int. Conf. Comput. Commun. (INFOCOM),

May 2007, pp. 23762380.

W. Xiao, L. Lin, and G. Chen, “Vertex-Degree Sequences

in Complex Networks: New Characteristics and

Applications,” Physica A: Statistical Mechanics and

Its Applications, vol. 437, no. 11, Nov. 2015, pp. 43741.

Frank, Howard, and I. Frisch. “Analysis and Design

of Survivable Networks.” IEEE Transactions on

Communication Technology 18, no. 5 (1970): 501-519.

Latora, Vito, and M. Marchiori. “Eﬃcient Behavior of

Small-world Networks.” Physical review letters 87, no.

19 (2001): 198701.

Sydney, Ali, C. Scoglio, M. Youssef, and P. Schumm.

“Characterizing the Robustness of Complex

Networks.” arXiv preprint arXiv:0811.3272 (2008).

Fiedler, Miroslav. “Algebraic Connectivity of Graphs.”

Czechoslovak mathematical journal 23, no. 2 (1973):

298-305.

Holme, Petter, B. J. K., C. N. Yoon, and S. K. Han.

“Attack Vulnerability of Complex Networks.” Physical

review E 65, no. 5 (2002): 056109.

M. Srinivas, L.M. Patnaik,“Adaptive Probabilities of

Crossover and Mutation in Genetic Algorithms”,

IEEE Transactions on Systems, Man and Cybernetics,

1994.

J. Zhang, H. S. Chung, ,“Clustering-Based Adaptive

Crossover and Mutation Probabilities for Genetic

Algorithms”, IEEE Transactions on Evolutionary

Computation, 2007.

S. Prendeville, E. Cherim, N. Bocken“Circular Cities:

Mapping Six Cities in Transition”, Environmental

Innovation and Societal Transitions, Volume 26,

March 2018, Pages 171-194.