ArticlePDF Available

Abstract and Figures

Wireless Sensor Networks (WSNs) is a promising technology with enormous applications in almost every walk of life. One of the crucial applications of WSNs is intrusion detection and surveillance at the border areas and in the defense establishments. The border areas are stretched in hundreds to thousands of miles, hence, it is not possible to patrol the entire border region. As a result, an enemy may enter from any point absence of surveillance and cause the loss of lives or destroy the military establishments. WSNs can be a feasible solution for the problem of intrusion detection and surveillance at the border areas. Detection of an enemy at the border areas and nearby critical areas such as military cantonments is a time-sensitive task as a delay of few seconds may have disastrous consequences. Therefore, it becomes imperative to design systems that are able to identify and detect the enemy as soon as it comes in the range of the deployed system. In this paper, we have proposed a deep learning architecture based on a fully connected feed-forward Artificial Neural Network (ANN) for the accurate prediction of the number of k-barriers for fast intrusion detection and prevention. We have trained and evaluated the feed-forward ANN model using four potential features, namely area of the circular region, sensing range of sensors, the transmission range of sensors, and the number of sensor for Gaussian and uniform sensor distribution. These features are extracted through Monte Carlo simulation. In doing so, we found that the model accurately predicts the number of k-barriers for both Gaussian and uniform sensor distribution with correlation coefficient (R = 0.78) and Root Mean Square Error (RMSE = 41.15) for the former and R = 0.79 and RMSE = 48.36 for the latter. Further, the proposed approach outperforms the other benchmark algorithms in terms of accuracy and computational time complexity.
Content may be subject to copyright.
Expert Systems With Applications 211 (2023) 118588
Available online 19 August 2022
0957-4174/© 2022 Elsevier Ltd. All rights reserved.
Contents lists available at ScienceDirect
Expert Systems With Applications
journal homepage: www.elsevier.com/locate/eswa
A deep learning approach to predict the number of 𝑘-barriers for intrusion
detection over a circular region using wireless sensor networks
Abhilash Singh a, J. Amutha b, Jaiprakash Nagar c, Sandeep Sharma d,
aFluvial Geomorphology and Remote Sensing Laboratory, Indian Institute of Science Education and Research Bhopal, 462066, India
bUniversity School of ICT, Gautam Buddha University, Greater Noida, 201312, India
cSubir Chowdhury School of Quality and Reliability, Indian Institute of Technology Kharagpur, 721302, India
dDepartment of Electronics Engineering, Madhav Institute of Technology and Science, Gwalior, Madhya Pradesh, 474005, India
ARTICLE INFO
Keywords:
WSNs
Binary sensing model
Gaussian distribution
Uniform distribution
Barrier coverage
Deep learning
ABSTRACT
Wireless Sensor Networks (WSNs) is a promising technology with enormous applications in almost every walk
of life. One of the crucial applications of WSNs is intrusion detection and surveillance at border areas and in
the defence establishments. The border areas are stretched over hundreds to thousands of miles, hence, it is
not possible to patrol the entire border region. As a result, an enemy may enter from any point absence of
surveillance and cause the loss of lives or destroy the military establishments. WSNs can be a feasible solution
for the problem of intrusion detection and surveillance at the border areas. Detection of an enemy at the border
areas and nearby critical areas such as military cantonments is a time-sensitive task as a delay of a few seconds
may have disastrous consequences. Therefore, it becomes imperative to design systems that can identify and
detect the enemy as soon as it comes within the range of the deployed system. In this paper, we have proposed
a deep learning architecture based on a fully connected feed-forward Artificial Neural Network (ANN) for the
accurate prediction of the number of 𝑘-barriers for fast intrusion detection and prevention. We have trained
and evaluated the feed-forward ANN model using four potential features, namely area of the circular region,
sensing range of sensors, transmission range of sensors, and number of sensor for Gaussian and uniform sensor
distribution. These features are extracted through Monte Carlo simulation. In doing so, we found that the
model accurately predicts the number of 𝑘-barriers for both Gaussian and uniform sensor distribution with
correlation coefficient (R = 0.78) and Root Mean Square Error (RMSE = 41.15) for the former and R = 0.79
and RMSE = 48.36 for the latter. Further, the proposed approach outperforms the other benchmark algorithms
in terms of accuracy and computational time complexity.
1. Introduction
The unquenchable thirst of people for political and military power
is compelling them to extend their boundaries and grab other people’s
natural resources. In order to achieve this goal, they may try different
techniques such as gaining information about the military establish-
ments, the number of military personnel at a given place, regions of
natural resources, and the vulnerabilities of authorities that they can
exploit. In addition, illegal immigration, smuggling of drugs, and other
banned commodities across the boundaries are immediate concerns
that must be dealt with immediately. Therefore, it is crucial to identify
an intruder or an unauthorised activity accurately in a timely manner
as they all are time-sensitive issues that may result in havoc if not
prevented in time. Furthermore, encroachment in the border areas and
Corresponding author.
E-mail addresses: sabhilash@iiserb.ac.in (A. Singh), roniamutha@gmail.com (J. Amutha), jpnagar@iitkgp.ac.in (J. Nagar), sandeepsvce@mitsgwalior.in
(S. Sharma).
unauthorised entry into the prohibited regions is a serious issue making
border surveillance mandatory.
It is a well-known fact that most real-life problems can be solved
with the help of suitable technologies. Fortunately, WSNs are widely
used and is a popular technology that can resolve the concerned
issue (Huang et al.,2018;Keung, Li, & Zhang,2012;Singh, Sharma and
Singh,2021;Singh, Sharma, Singh, & Kumar,2019). WSNs are widely
deployed for various military applications such as intrusion detection in
border areas, combat monitoring, an unauthorised access to prohibited
areas, land mines detection, battlefield surveillance, reconnaissance,
and so on (Amutha, Sharma, & Nagar,2020;Kandris, Nakas, Vomvas,
& Koulouras,2020;Sharma & Nagar,2020;Si, Ma, Tao, Fu, & Shu,
2020).
https://doi.org/10.1016/j.eswa.2022.118588
Received 1 July 2021; Received in revised form 7 August 2022; Accepted 13 August 2022
Expert Systems With Applications 211 (2023) 118588
2
A. Singh et al.
Researchers have proposed various border surveillance and intru-
sion detection techniques using WSNs (Amutha, Nagar and Sharma,
2021;Arjun, Indukala, & Menon,2019;Aseeri, Ahmed, Shakib, Ghor-
bel, & Shaman,2017;Benahmed & Benahmed,2019;Gavel, Raghu-
vanshi, & Tiwari,2021;Karanja & Badru,2021;Karthick, Prabaharan,
& Selvaprasanth,2019;Mostafaei, Chowdhury, & Obaidat,2018). The
proposed techniques use either simulations or Internet of Things (IoT)
methods to validate their techniques. However, simulation methods for
the validation of proposed techniques have high time complexity, i.e.,
the time taken to obtain a single output at a given value of sensor and
sensing range is in several hours. Also, IoT devices are very expensive
and require a huge financial investment. The high time complexity
and financial issues can be minimised to a negligible level using ma-
chine learning approaches to validate and predict the performance
of WSNs before their actual deployment in a given region (Kotiyal,
Singh, Sharma, Nagar, & Lee,2021;Mishra, Varadharajan, Tupakula,
& Pilli,2018;Singh, Nagar, Sharma and Kotiyal,2021). However,
the accurate and timely detection and prevention of intrusion through
machine learning approaches is still an ill-posed problem that has been
insufficiently investigated. To address this issue, we propose a deep
learning architecture for accurate and timely intrusion detection and
prevention.
In this paper, we proposed a fully connected feed-forward ANN
architecture to predict the number of 𝑘-barriers for accurate intrusion
detection and prevention in WSNs using potential features. We have
extracted four features, namely area of the circular region, sensing
range of sensors, the transmission range of sensors, and the number
of sensors through the Monte Carlo simulation approach. Afterward,
we used these features to train and evaluate the performance of the
feed-forward ANN model using R, RMSE, bias, and computational time
complexity as performance metrics.
Further, the rest of the paper is divided into seven sections. In
Section 2, we have discussed the related works. In Section 3, we have
presented the system model. In this section, we have discussed the
sensor distribution models and the sensing model. In Section 4, we have
discussed the simulation experiment. In Section 5, we have discussed
the machine learning model. In this section, we have discussed the
feature importance, feature sensitivity, and model setup. In Section 6,
we have presented the results. Lastly, in Sections 7and 8, we have
presented the discussion and conclusion of this study, respectively.
2. Related works
Deep learning is a subset of machine learning algorithms which
has been applied for intrusion detection using WSNs (Amutha, Sharma
and Sharma,2021;Lee et al.,2021;Singh, Amutha, Nagar, Sharma, &
Lee,2022a,2022b;Sood, Prakash, Sharma, Singh, & Choubey,2022).
It is also employed for pattern matching and network security where
it identifies the malicious activities occurring in the network and is
termed as Network Intrusion Detection System (NIDS). The accuracy of
the NIDS can be improved with the help of Recurrent Neural Network
IDS (RNNIDS) (Sohi, Seifert, & Ganji,2021) which is capable of identi-
fying the complex patterns resulting in an enhanced intrusion detection
rate. Authors in Yin, Zhu, Fei, and He (2017) have proposed an RNNs
approach which examines the system behaviour, type of intrusion, and
the impact of intrusion on the intrusion detection accuracy with the
help of learning rate and the number of neurons. The major limitation
of RNNIDS is that it fails to minimise the false positives; thus, not
able to achieve the maximum detection rate. This issue has been dealt
with the help of a deep learning architecture that combines classifiers
with Convolution Neural Networks (CNN) having Long Short Term
Memory (LSTM), thus, offering maximum detection rate (Pektaş &
Acarman,2019). Here, the CNN acquires spatial information, and the
LSTM acquires temporal features from the received packets in the
network. The optimal parameters in the feature space are obtained by
employing an estimator known as tree-structured Parzen. This method
focuses on generating flow-based statistical features rather than data-
set features. Although the abnormal traffic can be detected with an
accuracy rate of 99.09% and a false alarm rate of 0.0227, it fails to
compute the computational time complexity of flow-based intrusion
detection method.
A deep learning approach that uses an auto-encoder strategy ex-
hibits an improvement in the time complexity by 18.12% (Abbasi,
Bashir, Qureshi, ul Islam, & Jeon,2021). This strategy uses multi-
layer perception to replace the detection of hierarchical features and
unsupervised feature learning. Thus, it enhances the pattern matching
mechanism for intrusion detection with the help of Deep Learning-
based Feature Extraction (DLFE), which is an Optimisation of Pattern
Matching (OPM) approaches. Another approach called Scale-Hybrid-
IDS-AlertNet (SHIA) (Vinayakumar et al.,2019) performs data com-
putation at the network and the host-level for the intrusion detection
and delivers the relevant alert notifications to the controller auto-
matically. A Deep Neural Network (DNN) can render an effective
IDS that can detect and classify the intruders crossing the Region of
Interest (RoI). Here, the performance of SHIA is measured in terms
of multi-class classification of the DNN, accuracy, True Positive Rate
(TPR), and False-Positive Rate (FPR). Although SHIA is scalable and
shows improved performance in handling large data-sets of real-time
systems, it does not compute the number of 𝑘-barriers and incurs high
computational costs.
Despite having a high computational complexity, ensemble-based
techniques have a high level of accuracy as compared to the base mod-
els. An ensemble based DNN framework for the continuous analysis of
intrusion detection along with the ability to learn hierarchical data-sets
automatically has been proposed in Folino, Folino, Guarascio, Pisani,
and Pontieri (2021). Here, a log-stream of an intrusion detection system
maintains an ensemble that contains classifiers trained on discrete
chunks of the data-set instance and a combiner model. The combiner
model performs reasoning on the instance parameters and classifier
predictions; after that, the efficacy is computed in terms of data scarcity
and accuracy that allows to analyse the diverse ensemble aggregation
strategies. This ensemble approach utilises the unstructured data that
does not comply with the transfer learning strategies and is found in
the log-stream of NIDS.
Barrier coverage using WSNs plays a crucial role in the detection
of an unauthorised personnel or object trying to enter the prohibited
region. Barriers are formed by the sensors deployed over the entire
given RoI. The early depletion of sensor’s energy leads to the failure of
the sensors causing blind spots in the barriers. To overcome this issue,
authors in Saraereh, Ali, Al-Tarawneh, and Khan (2021) have pro-
posed a set-based max-flow procedure for the re-deployment of mobile
sensors. This method determines the vulnerable locations and deploys
the mobile sensors which strengthens and prolongs the longevity of
the barrier. Although the algorithm exhibits higher efficiency in terms
of the computation time, it fails to predict the number of 𝑘-barriers
for the IDS. In Nagar and Sharma (2018), the authors have derived
an analytical closed-form expression using mobile sensors for the 𝑘-
barrier coverage probability of a WSN. Here, they have calculated
the total area covered by an intruder travelling at a given angle to
cross the RoI, Then, this total area is utilised to obtain the closed-
form expression for the 𝑘-barrier coverage probability of the WSN. In
another work presented in Singh, Nagar et al. (2021), the authors have
employed three machine learning approaches, namely Gaussian Process
Regression (GPR), Scaling GPR (S-GPR), and Center mean GPR (C-GPR)
to predict the 𝑘-barrier coverage probability of a WSN. The proposed
GPR technique quickly detects and prevents any intrusion from taking
place at any location in the RoI. A three-level hierarchy scheme to
detect a mobile intruder in distributed WSNs is proposed in Sharma
and Chauhan (2020) improving the precision of intrusion detection. To
maximise the probability of intrusion detection, the authors have used
diverse sensing techniques, 𝑘-mean clustering, and the Likelihood Ratio
Test (LRT) methods. The LRT fusion rule used in this scheme is efficient
Expert Systems With Applications 211 (2023) 118588
3
A. Singh et al.
Fig. 1. Illustration of (a) Gaussian sensor distribution and (b) Uniform sensor distribution. The blue-filled circles represent sensors.
in terms of metrics like detection probability and false alarm rate at a
given number of sensors and the speed of the mobile intruder.
The algorithms and strategies proposed in the literature fail to
accurately predict the number of 𝑘-barriers for intrusion detection.
Hence, the overall aim of this study is to overcome the limitation
of the previous studies in terms of accuracy and computational time
complexity using a deep learning approach.
3. System model
This section briefly discusses sensor distribution models, sensing
range model, and the performance metric, namely the number of
𝑘-barriers.
3.1. Sensor distribution model
The choice of sensor distribution model depends on the required
application. In this work, we consider two-sensor distribution models;
(i) Gaussian distribution model and (ii) uniformly distribution model.
3.1.1. Gaussian sensor distribution model
In this model, a finite number of sensors are installed in a finite
circular region of radius 𝑅meters following a 2D Gaussian distribution,
also known as a normal distribution (Fig. 1a). Thus, the Probability
Density Function (PDF) for a location (𝑥, 𝑦)to be installed with a sensor
is given by Wang, Xie, and Agrawal (2008)
𝑓(𝑥, 𝑦) = 1
2𝜋𝜎𝑥𝜎𝑦
𝑒
−( (𝑥𝑥𝑐)2
2𝜎2
𝑥
+(𝑦𝑦𝑐)2
2𝜎2
𝑦
)(1)
where (𝑥𝑐, 𝑦𝑐)represents the centre of the circular region, 𝜎𝑥and 𝜎𝑦
are the standard deviations of 𝑥and 𝑦location coordinates respec-
tively. Furthermore, the location of a sensor inside the circular region
represented by (𝛾, 𝜙)can also be modelled with the help of position
coordinates (𝑥, 𝑦)if
𝑓(𝑥, 𝑦) (𝑥𝑥𝑐)2+ (𝑦𝑦𝑐)2𝑅=𝑓(𝛾, 𝜙) 𝛾𝑅(2)
3.1.2. Uniform sensor distribution model
In this model, a finite number of sensors are distributed uniformly
and randomly inside a finite circular region (denoted by ) of radius
𝑅meters (Fig. 1b). The position of a random sensor within the circular
region is represented by 𝑃𝑐= (𝛾, 𝜙), where, 𝛾 [0, 𝑅], denotes the
distance of the sensor from the centre of the circular region, and
𝜙 [0,2𝜋], denotes the angle between the 𝑥-axis and the line that
passes through the sensor location. The resulting sensor distribution
probability density function is given by
𝑓𝑝() = {1, 𝑖𝑓 𝑃𝑐
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 (3)
Furthermore, the probability that a sensor is located at an arbitrary
position 𝑃𝑐= (𝛾, 𝜙)inside the circular region is given by
𝑓(𝑃𝑐) = 1
𝜋𝑅2(4)
3.2. Binary sensing range model
The binary sensing range model is one of the most widely employed
sensing range models for estimating the performance of WSNs (Laran-
jeira & Rodrigues,2014;Nagar, Chaturvedi, & Soh,2020). A point 𝑖
denoted by 𝑃𝑖(𝑥𝑖, 𝑦𝑖)inside the 2D circular region will be covered by a
sensor 𝑗located at 𝑆𝑗(𝑥𝑗, 𝑦𝑗), if the Euclidean distance of point 𝑖from
the sensor 𝑗is less than or equal to the sensing range 𝑟𝑠of the sensor.
Mathematically, it can be represented as
𝑃(𝑆𝑖) = {1, 𝑖𝑓 𝑑 (𝑆𝑖, 𝑃𝑖)𝑟s
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 (5)
3.3. Coverage graph
Coverage measures how well sensors monitor the RoI in which they
are deployed. The 𝑘-coverage ensures that each point in the target RoI
is covered by at least 𝑘distinct sensors, where 𝑘is a positive integer
having a typical value greater than one. A connected 𝑘-coverage is
achieved when each point in the RoI is covered by at least 𝑘distinct
sensors and each sensor is able to communicate with each other. An
optimal 𝑘-coverage of the network is obtained when each point within
the RoI is covered by at least 𝑘distinct sensor without any overlapping
area. A Coverage Graph (CG) is denoted by CG(N) =(V, E), where V
denotes the number of vertices and Edenotes the number of edges.
Here, each vertex represents a sensor and an edge represents a link
between any two sensors if and only if they fall within the coverage of
each other. To construct a CG for a circular RoI, we have built sensor-
disjoint cycles within the entire RoI. A WSN rendering two-barrier
coverage inside a circular RoI is shown in Fig. 2a, and its respective
CG is depicted in Fig. 2b.
3.4. Barrier and barrier path
A barrier along the boundaries of a circular RoI can be constructed
by taking the union of distinct sensor coverage. Any arbitrary point
from where an intruder may enter the RoI is known as the point of
Expert Systems With Applications 211 (2023) 118588
4
A. Singh et al.
Fig. 2. Illustration of (a) 2-Barrier coverage and (b) Coverage graph.
Table 1
Simulation parameters.
Parameters Values
Simulator NS-2.35
Radius of the circular region (R) 40 to 127 m
Number of sensors (N) 100 to 400
Sensing range (Rs) 15 to 40 m
Transmission range (Rtx) 30 to 80 m
Mac type IEEE 802.11
Sensor’s deployment type (a) Gaussian Distribution
(b) Uniform Distribution
Sensing model Binary sensing model
intrusion and any possible path that an intruder may follow to reach the
target point is called an intrusion path. In order to provide a guaranteed
barrier coverage, there must exist at least one barrier for every possible
intrusion path. In this way, any intrusion attempt can be identified and
prevented in a timely manner. The maximum number of barrier paths
𝐵𝑃𝑚𝑎𝑥 that can be constructed for a given intrusion path without any
overlapping coverage is given by Eq. (6).
𝐵𝑃𝑚𝑎𝑥 =𝑁
𝑘(6)
where Nrepresents the total number of sensors and 𝑘represents the
number of sensors required to ensure 𝑘barrier coverage for a possible
intrusion path.
4. Simulation experiment
We have obtained the simulation results using network simulator
NS-2.35, one of the widely used network simulators to obtain the per-
formance metrics of WSNs. Table 1 shows different network parameters
and their values used to get the simulation results for the number
of 𝑘-barrier paths. Here, we have assumed that any two sensors can
communicate with each other iff the transmission range of sensors (Rtx)
is at least twice the sensing range of sensors (Rs), i.e.,Rtx 2Rs.
5. Machine learning model
Machine learning algorithms are broadly classified into supervised
and unsupervised learning algorithms. In supervised machine learning,
we work with labelled data sets. It is mainly used to solve either
regression or classification problems. In contrast, unsupervised machine
learning algorithms deal with unlabelled data sets and are mainly used
to perform clustering and dimension reduction tasks.
In this study, our objective is to assess the potential of fully con-
nected feed-forward ANN to map the number of 𝑘-barriers using rele-
vant features. To evaluate the relevancy of the selected features, we
have calculated the feature importance score and performed feature
sensitivity analysis.
5.1. Feature importance
This study used the area of the RoI, sensing range, transmission
range, and the number of sensors as the potential features and 𝑘-
barriers as the predictand. In machine learning, the selection of input
features significantly affects its performance (Singh, Kotiyal, Sharma,
Nagar, & Lee,2020). Hence, before training the machine learning
model, we have evaluated the relative importance of each selected
feature on the predictand. In doing so, we opted regression tree ensem-
ble technique (Torres-Barrán, Alonso, & Dorronsoro,2019). We have
first trained a regression tree ensemble model by boosting hundred
regression trees using the Least Squares gradient Boosting (LSBoost)
ensemble aggregation method (i.e., 𝑟=100), each with a learning rate
of one (i.e., 𝛼=1), and the classical decision tree (i.e., decision stumps)
has been considered as a weak learner. The LBoost algorithm trains
one weak learner at a time and also detects its weak points. Based on
such weak points, it generates a new weak learner (li) and evaluates
its weight (i.e., 𝑤𝑖). According to (Eq. (7)), the algorithm improves the
current model (𝑀𝑖) by emphasising on the prior weak learner’s (𝑀𝑖−1)
weak point. After it has been trained, it integrates the weak learner
into the existing model, and creates a single strong learner (𝑀𝑟,i.e., an
ensemble of weak learners) iteratively.
𝑀𝑖=𝑀𝑖−1 +𝛼 . 𝑤𝑖. 𝑙𝑖(𝑖= 1,2,3,, 𝑟)(7)
In addition, we determine the relative feature importance score by
evaluating the overall variations in the node risk (𝛥𝑅) due to the split
on each feature, and then normalising it by the total number of branch
nodes (𝑅𝑏𝑛). Mathematically, it is represented as in (Eq. (8));
𝛥𝑅 =
𝑅𝑝 (𝑅𝑐ℎ1+𝑅𝑐ℎ2)
𝑅𝑏𝑛
(8)
where 𝑅𝑝denotes the node risk of the parent and 𝑅𝑐ℎ1&𝑅𝑐ℎ2denotes
the node risk of two children. The node risk at an individual node (Ri)
is mathematically represented as in (Eq. (9));
𝑅𝑖=𝑃𝑖. 𝐸𝑖(9)
where 𝑃𝑖denotes the probability of node iand 𝐸𝑖denotes the mean
square error of the 𝑖𝑡ℎ node.
5.2. Feature sensitivity
Estimating the feature importance score only tells us about the
relative importance of each feature. However, it does not convey
how the features are associated with the predictand, i.e., whether the
predictand value increases with feature (positive impact) or decreases
with features (negative impact). To evaluate this, we have performed
the sensitivity analysis of the features using the Partial Dependence Plot
(PDP) (Friedman,2001;Singh, Nagar et al.,2021). PDP measures the
average effect of a single or more features by marginalising the effect of
all other features taken into consideration. We considered the combined
Expert Systems With Applications 211 (2023) 118588
5
A. Singh et al.
Fig. 3. Structure of the fully connected feed-forward ANN model having four inputs, two hidden layers having 20 neurons each, and one output (i.e., 4:20:20:1).
impact of two features simultaneously from the input feature set (i.e.,
𝜗) on the predictand by marginalising the impact of the remaining
features. To do so, a subset 𝜗sand a complimentary set (𝜗c) of 𝜗sare
extracted from the feature set (𝜗= {k1, k2, .. . , kn}) where, nrepresents
the total features. Using (Eq. (10)), we can compute any prediction on
𝜗.
𝑓(𝜗) = 𝑓(𝜗𝑠, 𝜗𝑐)(10)
The partial dependence of the feature in 𝜗scan be determined by
calculating the expectation (Ec) of Eq. (11).
𝑓𝑠(𝜗𝑠) = 𝐸𝑐[𝑓(𝜗𝑠, 𝜗𝑐)]
=𝑓(𝜗𝑠, 𝜗𝑐). 𝜌𝑐(𝜗𝑐). 𝑑𝜗𝑐(11)
where 𝜌𝑐(𝜗𝑐)indicates the marginal probability of 𝜗𝑐, which is repre-
sented in Eq. (12).
𝜌𝑐(𝜗𝑐) 𝑝(𝜗𝑠, 𝜗𝑐). 𝑑𝜗𝑠(12)
Then, the partial dependency of the feature in 𝜗𝑠can be determined by
:
𝑓𝑠(𝜗𝑠) 1
𝑇
𝑇
𝑖=1
𝑓(𝜗𝑠, 𝜗𝑐
𝑖)(13)
where 𝑇represents the total number of observations.
5.3. Model setup
In this subsection, we have discussed the architecture of the fully
connected feed-forward ANN including its working, activation function,
and the training algorithm.
5.3.1. Feed forward ANN
A Neural Network (NN) is a model which mimics the oversim-
plification of the brain performance that operates under a particular
specific function of interest. The primary objective of the NN model is
to discover a mapping function (f) that predicts a target function (f*)
through training the NN using labelled training data sets. During the
training phase, the network learns the range of parameters from the
training data. Once we trained the model, we need to validate and test
its performance using the unseen data.
A network can be subdivided into basic information-processing ele-
ments called neurons, which are the building blocks of ANN. Layers are
groups of neurons, and the network is comprised of interconnections
between these layers. There are diverse perspectives of linking layers
together that lead to several other forms of NNs like feed-forward
neural networks, recurrent neural networks, and convolutional neural
networks. In general, the optimisation problem of a neural network can
be represented by Eq. (14), where lrepresents the loss function, and W
is the learnable parameter. The main objective is to learn Wso that the
variance between the output of fcan be minimised, when the input x0
and the actual output yare provided.
𝑚𝑖𝑛 𝑙(𝑓(𝑥0, 𝑊 ), 𝑦)(14)
In this study, we have trained a fully connected feed-forward ANN.
It is a category of NN formed by organising neurons so that all neurons
in each layer are linked to every other neuron in the adjacent forward
layer. The data flows mainly in one direction, i.e., forward, from the
input neurons to the output, passing through the hidden layers (if any).
Since each neuron has one activation function, the total activation func-
tion for a layer equals the total output value. The training algorithm
utilises the output findings to calibrate the sensor connection weight
value (Moldovan, Caţaron, & Andonie,2020;Novickis, Justs, Ozols, &
Greit¯
ans,2020). Any continuous function can be approximated by a
Expert Systems With Applications 211 (2023) 118588
6
A. Singh et al.
feed-forward ANN with one hidden layer. However, the desired hidden
size might be high, making learning unfeasible. Feed-forward ANN is
well-suited for unstructured data, such as data that is not sequential or
time-dependent.
In this study, we structured a feed-forward ANN that consists of
two hidden layers and one output layer, as illustrated in Fig. 3. Each
hidden layers consist of twenty neurons. A common bias value is added
to each neuron in the hidden layers, which is followed by an activation
function.
5.3.2. Activation function
In feed-forward ANN, the activation function is one of the essential
key elements of the neuron (Duch & Jankowski,1999). It significantly
impacts the performance of the neural networks by modifying the
neuron output. The Universal Approximation Theorem states that a
feed-forward ANN with one hidden layer and an arbitrary sigmoidal
function with adequate sensors can estimate any continuous function
with no restrictions on the number of sensors or the size of the weights.
In this study, we have used the hyperbolic tangent sigmoid transfer
function at each layer because it is a bipolar sigmoid function that
has a positive response for a positive input and a negative response
for a negative input. Hence, it eliminates the problem of negative
responses for positive values. Further, as the complexity and non-
linear of the problems increases (when we increase the number of
sensors and the monitoring area in WSNs), the advantage of using the
hyperbolic tangent sigmoid transfer function becomes more apparent.
The mathematical model is expressed as:
𝑎=2
(1 + 𝑒(−2𝑛))−1 (15)
This expression is mathematically equivalent to tanh(n). However,
the computational time complexity of Eq. (15) is lower than tanh(n).
5.3.3. Training algorithm
To minimise the error in the output, the values of weights and
biases need to be updated. This is done with the help of a backpropa-
gation training algorithm. In the backpropagation algorithm, the input
is transmitted to the hidden layer, which then perpetuates back the
sensitivity in order to minimise the error rate by updating the weights
and the bias during the process. This algorithm results in low conver-
gence, and in some cases, it also leads to over-fitting. To address these
issues, and for quick convergence without over-fitting, approaches like
Levenberg–Marquardt backpropagation (LM) and Bayesian Regular-
ization (BR) backpropagation, and Scaled Conjugate Gradient (SCG)
backpropagation have been developed. To minimise the sum of squares
at every iteration, LM utilises the conjugate gradient backpropagation
method. LM is used for curve fitting problems, while SCG is used for
pattern recognition problems. Bayesian regularization backpropagation
algorithm uses an objective function that incorporates the sum of
squared weights and the residual sum of squares in order to reduce
the prediction errors for attaining the desired model.
In this study, we have used the LM backpropagation algorithm. It is
a non-linear optimisation-based approach for training ANNs, by which
it uses second-order derivatives for improved convergence behaviour.
The LM algorithm offers the features of the steepest descent approach
along with the Gauss–Newton method, which provides an invertible
matrix, named Hessian matrix (H) which is shown in Eq. (16):
𝐻(𝑥) 𝐽𝑇(𝑥)𝐽(𝑥) + 𝜇𝐼 (16)
where, 𝜇represents the combination co-efficient, Jand Irepresents
the Jacobian and identity matrix respectively. The LM modification to
the Gauss–Newton algorithm (Hagan & Menhaj,1994) is represented
in Eq. (17):
𝛥𝑥 = [𝐽𝑇(𝑥)𝐽(𝑥) + 𝜇𝐼 ]−1𝐽𝑇(𝑥)𝑒(𝑥)(17)
The algorithm seems to be the steepest descent when the value of 𝜇
becomes high, whereas the algorithm is Gauss–Newton when the value
of 𝜇is minimal. The LM algorithm for the weight update rule (Mathew,
Griffin, Alamaniotis, Kanarachos, & Fitzpatrick,2018) is determined as
a function of Jacobian matrix and error vector (e), which is represented
in Eq. (18):
𝑤(𝑡+ 1) = 𝑤(𝑡)−(𝐽𝑇
𝑡𝐽𝑡+𝜇𝐼 )−1 +𝐽𝑡𝑒𝑡(18)
We have randomly divided (using the Mersenne Twister generator)
the complete data set (182 ×5) in a 55:15:30 ratio for training, vali-
dation, and testing of the feed-forward ANN algorithm. The complete
methodology is shown in Fig. 4.
6. Results
This section presents the results of feature importance analysis,
feature sensitivity analysis, and feed-forward ANN.
6.1. Feature importance of each features
Once we evaluated the feature importance of each feature for both
Gaussian and uniform sensor distribution, we plotted the relative im-
portance graph for both scenarios (Fig. 5). The higher the importance
score, the more important the feature is. We found that the relative
importance score of the sensing range, transmission range, and sensors
is the same and the maximum. In contrast, the area has the least relative
importance amongst all the features for both scenarios.
6.2. Feature sensitivity curve
To analyse the feature sensitivity, we have plotted the surface plot
of PDP considering two features at a time along with the corresponding
two-dimensional plot with an axis-aligned histogram representing the
distribution of the features. For four features, we have six possible
sensitivity plots. Again, we have plotted it for both scenarios i.e., for
Gaussian and uniform sensor distribution in Figs. 6 and 7, respectively.
For both scenarios, we observed that the area of the circular region
has a negative impact on the number of barriers. In contrast, sensing
range, transmission range, and the number of sensors positively impact
the number of barriers. In a nutshell, we observed a similar trend with
slight variation in the values for both scenarios.
6.3. Performance of the fully connected feed-forward ANN model
Once we trained the feed-forward ANN model for Gaussian and Uni-
form distribution scenarios, we evaluated its performance by plotting
a linear regression curve between the predicted and observed barriers.
We have used R, RMSE, and bias as the performance metrics. A high
value of R represents that the predicted values are well in accord with
the observed value. A low value of RMSE represents a more accurate
model. A positive value of bias shows overestimation, and a negative
value of bias shows underestimation. Afterward, we performed error
and residual analysis for a robust conclusion.
6.3.1. For Gaussian sensor distribution
To evaluate the performance of the trained feed-forward ANN for
Gaussian sensor distribution, we have reported the training, validation,
testing, and overall accuracy. For training accuracy, we fed the training
data set as input to the trained feed-forward ANN and evaluated its
performance. We found that the trained model work reasonably well on
the training data set with R =0.82, RMSE =43.38, and bias = 5.11
(Fig. 8a). The presence of a small negative bias indicates that the values
are slightly getting underestimated. Evaluating the model performance
using the test data results in bias study and hence its performance needs
to be evaluated using the unseen/new data sets. In doing so, we have
Expert Systems With Applications 211 (2023) 118588
7
A. Singh et al.
Fig. 4. Flowchart of the methodology.
Fig. 5. Feature importance graph.
Expert Systems With Applications 211 (2023) 118588
8
A. Singh et al.
Fig. 6. Partial dependency plot for circular region considering Gaussian sensor distribution.
Fig. 7. Partial dependency plot for circular region considering uniform sensor distribution.
Expert Systems With Applications 211 (2023) 118588
9
A. Singh et al.
Fig. 8. Performance of the feed-forward ANN for Gaussian sensor distribution case (a) training accuracy, (b) validation accuracy, (c) testing accuracy, and (d) overall accuracy.
first validated the trained feed-forward ANN model through validation
data set (Fig. 8b). We found that the trained model performs well and
results in a good fit (with R =0.76, RMSE =34.79, and bias =3.99)
while tuning the hyper-parameters. Afterward, we used the test data
for unbiased evaluation of the trained model (Fig. 8c). We observed
that the trained model performed well on the test data (with R =0.76,
RMSE =29.86, and bias =17.31). The predicted value accord well
with the observed value with slight scattering. The presence of positive
bias represents that the values are slightly overestimated during the
testing phase. Finally, we combined all the data sets together (training,
validation, and testing) and fed it into the trained feed-forward ANN
model to calculate the model’s overall accuracy. We found that the
trained model performs reasonably well on the complete data sets with
R=0.78, RMSE =41.15, and bias =3.02 (Fig. 8d).
Further, to analyse the error distribution during the training, valida-
tion, and testing phase, we have performed error analysis and plotted
the combined error histogram using twenty bins (Fig. 9). The combined
error from the trained feed-forward ANN model ranges from 87.94
(leftmost bin) to 150.8 (rightmost bin) and follows a slightly right-
skewed Gaussian distribution. The peak of the distribution lies near
the zero error line indicating a more accurate model. The region left
to the zero error line indicates overestimated region, and the one on
the right represents an underestimated region. Overall, the number of
instances in the overestimated region is higher than the underestimated
region, results in the overestimation of predicted values by the trained
deep learning model. This statement is validated by the presence of a
positive bias of 3.02 (Fig. 8d).
Furthermore, to evaluate the appropriateness of the trained model,
we have performed the residual analysis and plotted the time series
plot of the test data along with the corresponding residual plot. We
have plotted the observed values (in blue) and predicted values (in
red) along with their 95% Confidence Interval (C.I). The dashed line
represents the RMSE value of the testing phase. The residuals are well
scattered and do not follows any specific pattern indicating a good fit
(Fig. 10).
6.3.2. For uniform sensor distribution
Similar to the Gaussian sensor distribution scenario, we have also
evaluated the performance of the deep learning model for uniform
sensor distribution. We reported the training, validation, testing, and
overall accuracy of the feed-forward ANN model that we trained for
uniform sensor distribution. For training accuracy, we have evaluated
the model over the training data sets. In doing so, we observed that the
model performs quite well over the training data sets (Fig. 11a). The
predicted values are close to the observed values (with R =0.79, RMSE
=55, and bias =3.65). However, the RMSE is high, and R is slightly
low as compared to the training accuracy of the Gaussian sensor distri-
bution. Afterward, we feed the validation data sets to the model input
to report the validation accuracy. The predicted values (while tuning
the hyper-parameters) are in agreement with the observed values with
R=0.81, RMSE =34.81, and bias =10.35 (Fig. 11b). For unbiased
evaluation, we have evaluated the trained model performance over the
test data set. We observed that the trained model performs well over
the test data with R =0.78, RMSE =34.58, and bias =19.89 (Fig. 11c).
Expert Systems With Applications 211 (2023) 118588
10
A. Singh et al.
Fig. 9. Error analysis with error histogram with 20 bins for Gaussian sensor distribution case.
Fig. 10. Residual analysis of the feed-forward ANN output for Gaussian sensor distribution case.
Finally, we feed the combined data into the model to report the overall
accuracy.
We found that the trained model also performs well over the com-
plete data set with R =0.79, RMSE =48.36, and bias 9.55 (Fig. 11d).
Further, we perform the error analysis to understand the distribu-
tion of error in a uniform sensor distribution scenario. In doing so, we
plotted the combined error histogram using twenty bins (Fig. 12). We
observed a similar trend as for the Gaussian sensor distribution scenario
despite the high error range. The total error ranges from 111.1
(leftmost bin) to 169 (rightmost bin) and follows a slightly right-skewed
Gaussian distribution. Here also, the peak of the distribution lies near
the zero error line indicating an accurate model. The total number of
instances in the overestimated region is higher than the underestimated
region. Due to this, a positive bias is present in the model.
Furthermore, we have performed the residual analysis and plotted
the time series plot of observed–predicted values for the testing phase
along with the corresponding residual plot (Fig. 13). Here, the residuals
are well scattered and do not follow any particular path or pattern,
indicating that the linear plot is a good fit.
7. Discussion
This study uses fully connected feed-forward ANN to predict the
number of 𝑘-barriers for intrusion detection in WSNs. We trained two
Expert Systems With Applications 211 (2023) 118588
11
A. Singh et al.
Fig. 11. Performance of the feed-forward ANN for uniform sensor distribution case (a) training accuracy, (b) validation accuracy, (c) testing accuracy, and (d) overall accuracy.
Fig. 12. Error analysis with error histogram with 20 bins for uniform sensor distribution case.
Expert Systems With Applications 211 (2023) 118588
12
A. Singh et al.
Fig. 13. Residual analysis of the feed-forward ANN output for uniform sensor distribution case.
separate feed-forward ANN models for Gaussian and uniform sensor
distribution scenarios. We observed that the proposed architecture of
a fully connected feed-forward ANN model gives promising results for
both scenarios. Although the correlation coefficient value is nearly
equal for both scenarios, the RMSE and bias value for the Gaussian
sensor distribution scenario is better.
7.1. Comparing with different variant of feed-forward ANN
For an unbiased evaluation, we have generated different scenarios
of the feed-forward ANN model based on the number of hidden layers
used, ranging from shallow to deep feed-forward ANN model (Table 2).
To do so, we have selected six different scenarios corresponding to
1, 2, 3, 4, 5, and 10 hidden layers. A feed-forward ANN model with
more than ten hidden layers will result in high time complexity and
eventually not an optimal solution for intrusion detection, which is
a time-sensitive application. We have reported the training, valida-
tion, testing, and overall accuracy for all six scenarios. Based on the
overall performance, we have categorised the performance of each sce-
nario into poor, fair, and satisfactory. We found that the feed-forward
ANN with 2, 3, and 10 layers resulted in satisfactory performance.
Out of these three scenarios, feed-forward ANN with two layers (i.e.,
4:20:20:1) shows the best performance.
7.2. Comparison with the benchmark algorithms
Various other findings have been reported for high intrusion detec-
tion accuracy through the machine learning approach (Nancy et al.,
2020;Safaldin, Otair, & Abualigah,2021). Hence, any conclusion
based on comparing different scenarios of a single algorithm may
result in a biased conclusion. To ensure a fair evaluation of the pro-
posed approach, we have compared the results of feed-forward ANN
with the other benchmark algorithms using R, RMSE, and bias as the
performance metrics. We have selected Radial Basis Neural Network
(RBN), Exact Radial Basis Neural Network (ERBN), Recurrent Neu-
ral Network (RNN), LSTM, Boosting (Least-square boosting) Ensemble
Learning (EL), Bagging EL (Random Forest), Binary Decision Tree
(BDT), General Regression Neural Network (GRNN), Gaussian Process
Regression (GPR), and Support Vector Regression (SVR) as potential
benchmark algorithms because these algorithms are amongst best per-
forming algorithms in the black-box and explainable based machine
learning category (Belle & Papantonis,2021;Elias et al.,2020;Lin
et al.,2020;Lundberg et al.,2018;Mansor et al.,2020;Roscher, Bohn,
Duarte, & Garcke,2020;Singh, Gaurav, Rai and Beg,2021;Zhang,
Zhang, Ye, Fang, & Han,2020). On comparing, we found that the
feed-forward ANN outperforms all the benchmark algorithms in terms
of R, RMSE and bias (Table 3). Other than the feed-forward ANN,
SVR performs well and ranks second among the benchmark algorithms.
Interestingly, we found that some benchmark algorithms show a strong
correlation (i.e., high R value); however, they produce biased results.
The bias values are very high, indicating that these models strongly
overestimate the response variable.
7.3. Comparison of the computational time complexity
Further, we have also compared the performance of the algorithms
mentioned above in terms of computational efficiency. We have es-
timated and plotted the computational time complexity graph of all
the three algorithms for both scenarios (Fig. 14). We observed that
the feed-forward ANN algorithm exhibits the lowest, and the LSTM
exhibits the highest computational time complexity. Apart from this,
we have also plotted the time complexity of the Monte Carlo simulation
for three different scenarios. We have estimated the time-complexity
for sensors as 100, 200, and 300. We have kept the other features
constant (area 5000, sensing range =15, and transmission range =
30). We observed that the computation time-complexity of the Monte
Carlo simulation increases with the number of sensors. This shows the
efficacy and need of the proposed deep learning architecture to cut
down the computational cost during the network setup time.
8. Conclusion
This study presented a fully connected feed-forward ANN architec-
ture for the accurate mapping of the number of 𝑘-barriers for intrusion
detection using WSNs. In doing so, we have trained two separate
feed-forward ANN models for both Gaussian and uniform sensor dis-
tribution using four potential features extracted through simulations.
While evaluating the feature importance and feature sensitivity for
Expert Systems With Applications 211 (2023) 118588
13
A. Singh et al.
Table 2
Comparison of the performance of different scenarios of feed-forward ANN.
Scenarios Training Validation Testing Overall Performance
R RMSE Bias R RMSE Bias R RMSE Bias R RMSE Bias
4:20:1 (Single layer) Gaussian 0.40 60.66 11.69 0.28 53.27 17.26 0.55 61.46 19.05 0.44 59.69 3.23 Fair
Uniform 0.51 67.08 6.44 0.53 52.25 8.81 0.53 52.25 8.81 0.56 65.06 17.31 Fair
4:20:20:1 (Two layers) Gaussian 0.82 43.38 5.11 0.76 34.79 3.99 0.76 29.86 17.31 0.78 41.15 3.02 Satisfactory
Uniform 0.79 55.00 3.65 0.81 34.81 10.35 0.78 34.58 19.89 0.79 48.36 9.55 Satisfactory
4:20:20:20:1 (Three layers) Gaussian 0.81 42.94 2.91 0.75 44.94 13.66 0.72 37.51 2.08 0.76 42.93 1.06 Satisfactory
Uniform 0.82 49.54 14.69 0.80 48.62 18.31 0.66 47.90 22.24 0.77 50.17 17.51 Satisfactory
4:20:20:20:20:1 (Four layers) Gaussian 0.44 59.93 10.18 0.30 58.59 20.99 0.54 57.61 12.55 0.47 58.74 12.50 Fair
Uniform 0.08 78.32 55.49 0.01 70.94 43.85 0.12 81.12 84.73 0.09 78.08 62.60 Poor
4:20:20:20:20:20:1 (Five layers) Gaussian 0.46 62.44 24.53 0.28 67.11 17.13 0.42 50.32 11.17 0.43 60.00 19.40 Fair
Uniform 0.11 83.11 102.25 0.06 84.13 97.75 0.08 62.41 70.25 0.07 78.22 91.91 Poor
4:20:20: ... :20:20:1 (Ten layers) Gaussian 0.64 48.35 25.09 0.82 46.20 9.78 0.75 45.01 22.25 0.69 47.72 21.96 Satisfactory
Uniform 0.75 47.90 0.44 0.78 59.69 3.47 0.70 59.17 5.20 0.73 53.23 1.30 Satisfactory
Table 3
Comparison with the benchmark algorithms.
Performance
metrics
Methods
Feed forward ANN RBN ERBN RNN LSTM Boosting EL Bagging EL (RF) BDT GRNN GPR SVR
Gaussian
Uniform
Gaussian
Uniform
Gaussian
Uniform
Gaussian
Uniform
Gaussian
Uniform
Gaussian
Uniform
Gaussian
Uniform
Gaussian
Uniform
Gaussian
Uniform
Gaussian
Uniform
Gaussian
Uniform
R 0.78 0.79 0.68 0.34 0.69 0.68 0.97 0.96 0.08 0.07 0.90 0.89 0.99 0.99 0.93 0.93 0.89 0.88 0.98 0.98 0.63 0.66
RMSE 41.15 48.36 65.67 168.51 75.12 89.68 15.71 22.27 1.16 1.42 37.31 46.00 11.37 12.63 29.29 32.90 39.66 48.80 11.49 13.72 48.44 52.22
Bias 3.02 9.55 39.19 137.51 60.11 71.11 61.45 65.36 17.10 20.9 59.06 69.96 53.90 62.35 57.42 63.22 60.70 71.69 53.18 61.07 18.24 20.26
Fig. 14. Comparison of the computational time complexity of the feed-forward ANN with the benchmark algorithms with three different scenarios of Monte Carlo simulations. A
log scale is used for the vertical axis.
Gaussian and uniform sensor distribution scenarios, we observed that
sensing range, transmission range, and number of sensors are the most
relevant features amongst all, which positively impact the predictand.
In contrast, the area is the least important feature, and it is negatively
related to the predictand. After training the models, we evaluated their
performance over the unseen data and found that both models provide
promising results.
Further, we have compared the performance of the feed-forward
ANN with the benchmark algorithms. We found that the proposed feed-
forward architecture outperforms the benchmark algorithms in terms of
accuracy and computational time complexity.
This study is a step towards the accurate and time-efficient predic-
tion of the number of 𝑘-barriers for intrusion detection using WSNs. Our
methodology can be used to cut down the time and cost requirements
during practical network setup.
Expert Systems With Applications 211 (2023) 118588
14
A. Singh et al.
CRediT authorship contribution statement
Abhilash Singh: Conceptualization, Methodology, Software, Data
curation, Validation, Writing original draft, Visualization, Investi-
gation, Writing review & editing. J. Amutha: Conceptualization,
Methodology, Software, Data curation, Validation, Writing original
draft, Visualization, Writing review & editing. Jaiprakash Nagar:
Conceptualization, Methodology, Data curation, Visualization, Writing
original draft, Writing review & editing. Sandeep Sharma: Method-
ology, Data curation, Visualization, Investigation, Writing review &
editing, Supervision, Project Administration.
Declaration of competing interest
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared to
influence the work reported in this paper.
Data and code availability
The datasets generated during and/or analysed during the cur-
rent study can be downloaded from https://www.kaggle.com/datasets/
abhilashdata/ffannid-intrusion-detection-in-wsnshere (Data) and the
code can be downloaded from https://in.mathworks.com/matlabcentral/
fileexchange/116670-a-deep-learning-approach-to-predict-the- number-
of-k-barriers?s_tid=prof_contriblnkhere (Code) (accessed on 25 August
2022).
Acknowledgements
The authors would like to acknowledge IISER Bhopal, MITS Gwalior,
IIT Kharagpur, and Gautam Buddha University Greater Noida for
providing institutional support. They would like to thank to the editor
and all the anonymous reviewers for providing helpful comments and
suggestions.
References
Abbasi, J. S., Bashir, F., Qureshi, K. N., ul Islam, M. N., & Jeon, G. (2021). Deep
learning-based feature extraction and optimizing pattern matching for intrusion
detection using finite state machine. Computers and Electrical Engineering,92, Article
107094.
Amutha, J., Nagar, J., & Sharma, S. (2021). A distributed border surveillance (dbs)
system for rectangular and circular region of interest with wireless sensor networks
in shadowed environments. Wireless Personal Communications,117(3), 2135–2155.
Amutha, J., Sharma, S., & Nagar, J. (2020). WSN strategies based on sensors,
deployment, sensing models, coverage and energy efficiency: Review, approaches
and open issues. Wireless Personal Communications,111(2), 1089–1115.
Amutha, J., Sharma, S., & Sharma, S. K. (2021). Strategies based on various aspects
of clustering in wireless sensor networks using classical, optimization and machine
learning techniques: Review, taxonomy, research findings, challenges and future
directions. Computer Science Review,40, Article 100376.
Arjun, D., Indukala, P., & Menon, K. U. (2019). PANCHENDRIYA: A multi-sensing
framework through wireless sensor networks for advanced border surveillance and
human intruder detection. In 2019 international conference on communication and
electronics systems (ICCES) (pp. 295–298). IEEE.
Aseeri, M., Ahmed, M., Shakib, M., Ghorbel, O., & Shaman, H. (2017). Detection of
attacker and location in wireless sensor network as an application for border
surveillance. International Journal of Distributed Sensor Networks,13(11), Article
1550147717740072.
Belle, V., & Papantonis, I. (2021). Principles and practice of explainable machine
learning. Frontiers in Big Data, 39.
Benahmed, T., & Benahmed, K. (2019). Optimal barrier coverage for critical area
surveillance using wireless sensor networks. International Journal of Communication
Systems,32(10), Article e3955.
Duch, W., & Jankowski, N. (1999). Survey of neural transfer functions. Neural Computing
Surveys,2(1), 163–212.
Elias, I., Rubio, J. d. J., Martinez, D. I., Vargas, T. M., Garcia, V., Mujica-Vargas, D., et
al. (2020). Genetic algorithm with radial basis mapping network for the electricity
consumption modeling. Applied Sciences,10(12), 4239.
Folino, F., Folino, G., Guarascio, M., Pisani, F. S., & Pontieri, L. (2021). On learning
effective ensembles of deep neural networks for intrusion detection. Information
Fusion,72, 48–69.
Friedman, J. H. (2001). Greedy function approximation: a gradient boosting machine.
The Annals of Statistics, 1189–1232.
Gavel, S., Raghuvanshi, A. S., & Tiwari, S. (2021). Maximum correlation based mutual
information scheme for intrusion detection in the data networks. Expert Systems
with Applications, Article 116089.
Hagan, M. T., & Menhaj, M. B. (1994). Training feedforward networks with the
marquardt algorithm. IEEE Transactions on Neural Networks,5(6), 989–993.
Huang, H., Gong, T., Zhang, R., Yang, L.-L., Zhang, J., & Xiao, F. (2018). Intrusion de-
tection based on 𝑘-coverage in mobile sensor networks with empowered intruders.
IEEE Transactions on Vehicular Technology,67(12), 12109–12123.
Kandris, D., Nakas, C., Vomvas, D., & Koulouras, G. (2020). Applications of wireless
sensor networks: an up-to-date survey. Applied System Innovation,3(1), 14.
Karanja, S., & Badru, R. (2021). Development of a low cost wireless sensor network
for surveillance along Kenya-Somalia border. Preprint.
Karthick, R., Prabaharan, A. M., & Selvaprasanth, P. (2019). Internet of things based
high security border surveillance strategy. Asian Journal of Applied Science and
Technology (AJAST),3, 94–100.
Keung, G. Y., Li, B., & Zhang, Q. (2012). The intrusion detection in mobile sensor
network. IEEE/ACM Transactions on Networking,20(4), 1152–1161.
Kotiyal, V., Singh, A., Sharma, S., Nagar, J., & Lee, C.-C. (2021). ECS-NL: An enhanced
cuckoo search algorithm for node localisation in wireless sensor networks. Sensors,
21(11), 3576.
Laranjeira, L. A., & Rodrigues, G. N. (2014). Border effect analysis for reliability
assurance and continuous connectivity of wireless sensor networks in the presence
of sensor failures. IEEE Transactions on Wireless Communication,13(8), 4232–4246.
Lee, S.-W., Mohammadi, M., Rashidi, S., Rahmani, A. M., Masdari, M., Hossein-
zadeh, M., et al. (2021). Towards secure intrusion detection systems using deep
learning techniques: Comprehensive analysis and review. Journal of Network and
Computer Applications, Article 103111.
Lin, H., Dai, Q., Zheng, L., Hong, H., Deng, W., & Wu, F. (2020). Radial basis function
artificial neural network able to accurately predict disinfection by-product levels
in tap water: Taking haloacetic acids as a case study. Chemosphere,248, Article
125999.
Lundberg, S. M., Nair, B., Vavilala, M. S., Horibe, M., Eisses, M. J., Adams, T., et al.
(2018). Explainable machine-learning predictions for the prevention of hypoxaemia
during surgery. Nature Biomedical Engineering,2(10), 749–760.
Mansor, M., Mohd Jamaludin, S. Z., Mohd Kasihmuddin, M. S., Alzaeemi, S. A.,
Md Basir, M. F., Sathasivam, S., et al. (2020). Systematic Boolean satisfiability
programming in radial basis function neural network. Processes,8(2), 214.
Mathew, J., Griffin, J., Alamaniotis, M., Kanarachos, S., & Fitzpatrick, M. E. (2018).
Prediction of welding residual stresses using machine learning: comparison between
neural networks and neuro-fuzzy systems. Applied Soft Computing,70, 131–146.
Mishra, P., Varadharajan, V., Tupakula, U., & Pilli, E. S. (2018). A detailed investigation
and analysis of using machine learning techniques for intrusion detection. IEEE
Communications Surveys & Tutorials,21(1), 686–728.
Moldovan, A., Caţaron, A., & Andonie, R. (2020). Learning in feedforward neural
networks accelerated by transfer entropy. Entropy,22(1), 102.
Mostafaei, H., Chowdhury, M. U., & Obaidat, M. S. (2018). Border surveillance with
WSN systems in a distributed manner. IEEE Systems Journal,12(4), 3703–3712.
Nagar, J., Chaturvedi, S. K., & Soh, S. (2020). An analytical model to estimate the
performance metrics of a finite multihop network deployed in a rectangular region.
Journal of Network and Computer Applications,149, Article 102466.
Nagar, J., & Sharma, S. (2018). K-barrier coverage-based intrusion detection for wireless
sensor networks. In Cyber security (pp. 373–385). Springer.
Nancy, P., Muthurajkumar, S., Ganapathy, S., Kumar, S. S., Selvi, M., & Arputharaj, K.
(2020). Intrusion detection using dynamic feature selection and fuzzy temporal
decision tree classification for wireless sensor networks. IET Communications,14(5),
888–895.
Novickis, R., Justs, D. J., Ozols, K., & Greit¯
ans, M. (2020). An approach of feed-forward
neural network throughput-optimized implementation in FPGA. Electronics,9(12),
2193.
Pektaş, A., & Acarman, T. (2019). A deep learning method to detect network intrusion
through flow-based features. International Journal of Network Management,29(3),
Article e2050.
Roscher, R., Bohn, B., Duarte, M. F., & Garcke, J. (2020). Explainable machine learning
for scientific insights and discoveries. IEEE Access,8, 42200–42216.
Safaldin, M., Otair, M., & Abualigah, L. (2021). Improved binary gray wolf optimizer
and SVM for intrusion detection system in wireless sensor networks. Journal of
Ambient Intelligence and Humanized Computing,12(2), 1559–1576.
Saraereh, O. A., Ali, A., Al-Tarawneh, L., & Khan, I. (2021). A robust approach for
barrier-reinforcing in wireless sensor networks. Journal of Parallel and Distributed
Computing,149, 186–192.
Sharma, A., & Chauhan, S. (2020). Sensor fusion for distributed detection of mobile
intruders in surveillance wireless sensor networks. IEEE Sensors Journal,20(24),
15224–15231.
Sharma, S., & Nagar, J. (2020). Intrusion detection in mobile sensor networks: A case
study for different intrusion paths. Wireless Personal Communications, 1–21.
Si, P., Ma, J., Tao, F., Fu, Z., & Shu, L. (2020). Energy-efficient barrier coverage
with probabilistic sensors in wireless sensor networks. IEEE Sensors Journal,20(10),
5624–5633.
Expert Systems With Applications 211 (2023) 118588
15
A. Singh et al.
Singh, A., Amutha, J., Nagar, J., Sharma, S., & Lee, C.-C. (2022a). AutoML-ID:
automated machine learning model for intrusion detection using wireless sensor
network. Scientific Reports,12(1), 1–14.
Singh, A., Amutha, J., Nagar, J., Sharma, S., & Lee, C.-C. (2022b). Lt-fs-id: Log-
transformed feature learning and feature-scaling-based machine learning algorithms
to predict the k-barriers for intrusion detection using wireless sensor network.
Sensors,22(3), 1070.
Singh, A., Gaurav, K., Rai, A. K., & Beg, Z. (2021). Machine learning to estimate surface
roughness from satellite images. Remote Sensing,13(19), 3794.
Singh, A., Kotiyal, V., Sharma, S., Nagar, J., & Lee, C.-C. (2020). A machine learning
approach to predict the average localization error with applications to wireless
sensor networks. IEEE Access,8, 208253–208263.
Singh, A., Nagar, J., Sharma, S., & Kotiyal, V. (2021). A Gaussian process regression
approach to predict the k-barrier coverage probability for intrusion detection in
wireless sensor networks. Expert Systems with Applications,172, Article 114603.
Singh, A., Sharma, S., & Singh, J. (2021). Nature-inspired algorithms for wireless sensor
networks: A comprehensive survey. Computer Science Review,39, Article 100342.
Singh, A., Sharma, S., Singh, J., & Kumar, R. (2019). Mathematical modelling for
reducing the sensing of redundant information in WSNs based on biologically
inspired techniques. Journal of Intelligent & Fuzzy Systems,37(5), 6829–6839.
Sohi, S. M., Seifert, J.-P., & Ganji, F. (2021). RNNIDS: Enhancing network intrusion
detection systems through deep learning. Computers & Security,102, Article 102151.
Sood, T., Prakash, S., Sharma, S., Singh, A., & Choubey, H. (2022). Intrusion detection
system in wireless sensor network using conditional generative adversarial network.
Wireless Personal Communications, 1–21.
Torres-Barrán, A., Alonso, A., & Dorronsoro, J. R. (2019). Regression tree ensembles
for wind energy and solar radiation prediction. Neurocomputing,326, 151–160.
Vinayakumar, R., Alazab, M., Soman, K., Poornachandran, P., Al-Nemrat, A., &
Venkatraman, S. (2019). Deep learning approach for intelligent intrusion detection
system. IEEE Access,7, 41525–41550.
Wang, D., Xie, B., & Agrawal, D. P. (2008). Coverage and lifetime optimization of
wireless sensor networks with gaussian distribution. IEEE Transactions on Mobile
Computing,7(12), 1444–1458.
Yin, C., Zhu, Y., Fei, J., & He, X. (2017). A deep learning approach for intrusion
detection using recurrent neural networks. IEEE Access,5, 21954–21961.
Zhang, D., Zhang, N., Ye, N., Fang, J., & Han, X. (2020). Hybrid learning algorithm
of radial basis function networks for reliability analysis. IEEE Transactions on
Reliability.
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Momentous increase in the popularity of explainable machine learning models coupled with the dramatic increase in the use of synthetic data facilitates us to develop a cost-efficient machine learning model for fast intrusion detection and prevention at frontier areas using Wireless Sensor Networks (WSNs). The performance of any explainable machine learning model is driven by its hyperparameters. Several approaches have been developed and implemented successfully for optimising or tuning these hyperparameters for skillful predictions. However, the major drawback of these techniques, including the manual selection of the optimal hyperparameters, is that they depend highly on the problem and demand application-specific expertise. In this paper, we introduced Automated Machine Learning (AutoML) model to automatically select the machine learning model (among support vector regression, Gaussian process regression, binary decision tree, bagging ensemble learning, boosting ensemble learning, kernel regression, and linear regression model) and to automate the hyperparameters optimisation for accurate prediction of numbers of k-barriers for fast intrusion detection and prevention using Bayesian optimisation. To do so, we extracted four synthetic predictors, namely, area of the region, sensing range of the sensor, transmission range of the sensor, and the number of sensors using Monte Carlo simulation. We used 80% of the datasets to train the models and the remaining 20% for testing the performance of the trained model. We found that the Gaussian process regression performs prodigiously and outperforms all the other considered explainable machine learning models with correlation coefficient (R = 1), root mean square error (RMSE = 0.007), and bias = − 0.006. Further, we also tested the AutoML performance on a publicly available intrusion dataset, and we observed a similar performance. This study will help the researchers accurately predict the required number of k-barriers for fast intrusion detection and prevention.
Article
Full-text available
Wireless communication networks have much data to sense, process, and transmit. It tends to develop a security mechanism to care for these needs for such modern-day systems. An intrusion detection system (IDS) is a solution that has recently gained the researcher's attention with the application of deep learning techniques in IDS. In this paper, we propose an IDS model that uses a deep learning algorithm, conditional generative adversarial network (CGAN), enabling unsupervised learning in the model and adding an eXtreme gradient boosting (XGBoost) classifier for faster comparison and visualization of results. The proposed method can reduce the need to deploy extra sensors to generate fake data to fool the intruder 1.2-2.6%, as the proposed system generates this fake data. The parameters were selected to give optimal results to our model without significant alterations and complications. The model learns from its dataset samples with the multiple-layer network for a refined training process. We aimed that the proposed model could improve the accuracy and thus, decrease the false detection rate and obtain good precision in the cases of both the datasets, NSL-KDD and the CICIDS2017, which can be used as a detector for cyber intrusions. The false alarm rate of the proposed model decreases by about 1.827%.
Article
Full-text available
The dramatic increase in the computational facilities integrated with the explainable machine learning algorithms allows us to do fast intrusion detection and prevention at border areas using Wireless Sensor Networks (WSNs). This study proposed a novel approach to accurately predict the number of barriers required for fast intrusion detection and prevention. To do so, we extracted four features through Monte Carlo simulation: area of the Region of Interest (RoI), sensing range of the sensors, transmission range of the sensor, and the number of sensors. We evaluated feature importance and feature sensitivity to measure the relevancy and riskiness of the selected features. We applied log transformation and feature scaling on the feature set and trained the tuned Support Vector Regression (SVR) model (i.e., LT-FS-SVR model). We found that the model accurately predicts the number of barriers with a correlation coefficient (R) = 0.98, Root Mean Square Error (RMSE) = 6.47, and bias = 12.35. For a fair evaluation, we compared the performance of the proposed approach with the benchmark algorithms, namely, Gaussian Process Regression (GPR), Generalised Regression Neural Network (GRNN), Artificial Neural Network (ANN), and Random Forest (RF). We found that the proposed model outperforms all the benchmark algorithms
Article
Full-text available
We apply the Support Vector Regression (SVR) machine learning model to estimate surface roughness on a large alluvial fan of the Kosi River in the Himalayan Foreland from satellite images. To train the model, we used input features such as radar backscatter values in Vertical–Vertical (VV) and Vertical–Horizontal (VH) polarisation, incidence angle from Sentinel-1, Normalised Difference Vegetation Index (NDVI) from Sentinel-2, and surface elevation from Shuttle Radar Topographic Mission (SRTM). We generated additional features (VH/VV and VH–VV) through a linear data fusion of the existing features. For the training and validation of our model, we conducted a field campaign during 11–20 December 2019. We measured surface roughness at 78 different locations over the entire fan surface using an in-house-developed mechanical pin-profiler. We used the regression tree ensemble approach to assess the relative importance of individual input feature to predict the surface soil roughness from SVR model. We eliminated the irrelevant input features using an iterative backward elimination approach. We then performed feature sensitivity to evaluate the riskiness of the selected features. Finally, we applied the dimension reduction and scaling to minimise the data redundancy and bring them to a similar level. Based on these, we proposed five SVR methods (PCA-NS-SVR, PCA-CM-SVR, PCA-ZM-SVR, PCA-MM-SVR, and PCA-S-SVR). We trained and evaluated the performance of all variants of SVR with a 60:40 ratio using the input features and the in-situ surface roughness. We compared the performance of SVR models with six different benchmark machine learning models (i.e., Gaussian Process Regression (GPR), Generalised Regression Neural Network (GRNN), Binary Decision Tree (BDT), Bragging Ensemble Learning, Boosting Ensemble Learning, and Automated Machine Learning (AutoML)). We observed that the PCA-MM-SVR perform better with a coefficient of correlation (R = 0.74), Root Mean Square Error (RMSE = 0.16 cm), and Mean Square Error (MSE = 0.025 cm2). To ensure a fair selection of the machine learning model, we evaluated the Akaike’s Information Criterion (AIC), corrected AIC (AICc), and Bayesian Information Criterion (BIC). We observed that SVR exhibits the lowest values of AIC, corrected AIC, and BIC of all the other methods; this indicates the best goodness-of-fit. Eventually, we also compared the result of PCA-MM-SVR with the surface roughness estimated from different empirical and semi-empirical radar backscatter models. The accuracy of the PCA-MM-SVR model is better than the backscatter models. This study provides a robust approach to measure surface roughness at high spatial and temporal resolutions solely from the satellite data. Keywords: surface roughness; Sentinel-1; Sentinel-2; machine learning models; AutoML; backscatter models
Article
Full-text available
Artificial intelligence (AI) provides many opportunities to improve private and public life. Discovering patterns and structures in large troves of data in an automated manner is a core component of data science, and currently drives applications in diverse areas such as computational biology, law and finance. However, such a highly positive impact is coupled with a significant challenge: how do we understand the decisions suggested by these systems in order that we can trust them? In this report, we focus specifically on data-driven methods—machine learning (ML) and pattern recognition models in particular—so as to survey and distill the results and observations from the literature. The purpose of this report can be especially appreciated by noting that ML models are increasingly deployed in a wide range of businesses. However, with the increasing prevalence and complexity of methods, business stakeholders in the very least have a growing number of concerns about the drawbacks of models, data-specific biases, and so on. Analogously, data science practitioners are often not aware about approaches emerging from the academic literature or may struggle to appreciate the differences between different methods, so end up using industry standards such as SHAP. Here, we have undertaken a survey to help industry practitioners (but also data scientists more broadly) understand the field of explainable machine learning better and apply the right tools. Our latter sections build a narrative around a putative data scientist, and discuss how she might go about explaining her models by asking the right questions. From an organization viewpoint, after motivating the area broadly, we discuss the main developments, including the principles that allow us to study transparent models vs. opaque models, as well as model-specific or model-agnostic post-hoc explainability approaches. We also briefly reflect on deep learning models, and conclude with a discussion about future research directions.
Article
Full-text available
Node localisation plays a critical role in setting up Wireless Sensor Networks (WSNs). A sensor in WSNs senses, processes and transmits the sensed information simultaneously. Along with the sensed information, it is crucial to have the positional information associated with the information source. A promising method to localise these randomly deployed sensors is to use bio-inspired meta-heuristic algorithms. In this way, a node localisation problem is converted to an optimisation problem. Afterwards, the optimisation problem is solved for an optimal solution by minimising the errors. Various bio-inspired algorithms, including the conventional Cuckoo Search (CS) and modified CS algorithm, have already been explored. However, these algorithms demand a predetermined number of iterations to reach the optimal solution, even when not required. In this way, they unnecessarily exploit the limited resources of the sensors resulting in a slow search process. This paper proposes an Enhanced Cuckoo Search (ECS) algorithm to minimise the Average Localisation Error (ALE) and the time taken to localise an unknown node. In this algorithm, we have implemented an Early Stopping (ES) mechanism, which improves the search process significantly by exiting the search loop whenever the optimal solution is reached. Further, we have evaluated the ECS algorithm and compared it with the modified CS algorithm. While doing so, note that the proposed algorithm localised all the localisable nodes in the network with an ALE of 0.5-0.8 m. In addition, the proposed algorithm also shows an 80% decrease in the average time taken to localise all the localisable nodes. Consequently, the performance of the proposed ECS algorithm makes it desirable to implement in practical scenarios for node localisation.
Article
For an efficient detection of intrusion in the network, the system requires sensitive information. This information contains various features depending upon the network. The selection of useful features and relevant information from the data set has always been difficult for the user. These features bound the performance of the system and also restrict them in making accurate decisions. This irrelevant feature plays a major role in making confusion between the normal data pattern against the attack data pattern. This article proposes a novel Maximum correlation-based mutual information technique for efficient feature selection (MCMIFS) in a data network. The proposed method is utilized with Kernel Extreme Learning Machine (KELM) based multiclass classifier for effective intrusion detection. The performance of the proposed scheme is evaluated using standard intrusion detection datasets with feature ranging from 42 to 155. The hybrid of MCMIFS with KELM reveals the effectiveness of the proposed scheme by improving the accuracy of detection, i.e., 99.97%, decreasing the FPR, i.e., 0.19, by decreasing the computational complexity. The proposed Intrusion Detection Scheme (IDS) results are analyzed and compared with the other existing techniques.
Article
Providing a high-performance Intrusion Detection System (IDS) can be very effective in controlling malicious behaviors and cyber-attacks. Regarding the ever-growing negative impacts of the security attacks on computer systems and networks, various Artificial Intelligence (AI)-based techniques have been used to introduce versatile IDS approaches. Deep learning is a branch of AI techniques, mainly based on multi-layer artificial neural networks. Recently, deep learning techniques have gained momentum in the intrusion detection domain and several IDS approaches are provided in the literature using various deep neural networks to deal with privacy concerns and security threats. For this purpose, this article focuses on the deep IDS approaches and investigates how deep learning networks are employed by different approaches in different steps of the intrusion detection process to achieve better results. It classifies the studied IDS schemes regarding the deep learning networks utilized in them and describes their main contributions and capabilities. Besides, in each category, their main features such as evaluated metrics, datasets, simulators, and environments are compared. Also, a comparison of the deep IDS approaches main properties are provided to illuminate the main techniques applied in them as well as the area less focused in the literature. Finally, the concluding remarks in the deep IDS context are provided and possible directions at the subsequent studies are listed.
Article
Deep learning has gained popularity for feature extraction in the field of Network Intrusion Detection and Prevention System (NIDPS) to extract the patterns matching and secure the networks by detecting the unknown and malicious activities. The malicious activities and security attacks are disturbing the normal operations of networks. The new attacks are difficult to monitor due to their new features and pattern types. Different types of methods have been adopted for feature extraction and pattern matching. Deep learning is one of them and subfield of machine learning where it solves the optimization issues layer-wise by looking at the deep structure. On the other hand, pattern matching is another considerable method for intrusion detection due to its variety of applications. However, pattern matching methods are consuming more than 70% of the total running time and cause overhead. In this paper, we propose two methods including Deep Learning-based Feature Extraction (DLFE) and Optimization of Pattern Matching (OPM) for NIDPS systems to optimizes the pattern matching engine in intrusion detection. The experiments are performed by using the snort ruleset for pattern matching and obtained the results. The experiment results show the better performance of proposed methods in terms of time, throughput, and memory.
Article
Classification-oriented Machine Learning methods are a precious tool, in modern Intrusion Detection Systems (IDSs), for discriminating between suspected intrusion attacks and normal behaviors. Many recent proposals in this field leveraged Deep Neural Network (DNN) methods, capable of learning effective hierarchical data representations automatically. However, many of these solutions were validated on data featuring stationary distributions and/or large amounts of training examples. By contrast, in real IDS applications different kinds of attack tend to occur over time, and only a small fraction of the data instances is labeled (usually with far fewer examples of attacks than of normal behavior). A novel ensemble-based Deep Learning framework is proposed here that tries to face the challenging issues above. Basically, the non-stationary nature of IDS log data is faced by maintaining an ensemble consisting of a number of specialized base DNN classifiers, trained on disjoint chunks of the data instances’ stream, plus a combiner model (reasoning on both the base classifiers predictions and original instance features). In order to learn deep base classifiers effectively from small training samples, an ad-hoc shared DNN architecture is adopted, featuring a combination of dropout capabilities, skip-connections, along with a cost-sensitive loss (for dealing with unbalanced data). Tests results, conducted on two benchmark IDS datasets and involving several competitors, confirmed the effectiveness of our proposal (in terms of both classification accuracy and robustness to data scarcity), and allowed us to evaluate different ensemble combination schemes.