Malicious_Node_Detection_Using_Machine_Learning_and_Distributed_Data_Storage_Using_Blockchain_in_WSNs
Malicious_Node_Detection_Using_Machine_Learning_and_Distributed_Data_Storage_Using_Blockchain_in_WSNs
ABSTRACT In the proposed work, blockchain is implemented on the Base Stations (BSs) and Cluster
Heads (CHs) to register the nodes using their credentials and also to tackle various security issues. Moreover,
a Machine Learning (ML) classifier, termed as Histogram Gradient Boost (HGB), is employed on the BSs
to classify the nodes as malicious or legitimate. In case, the node is found to be malicious, its registration
is revoked from the network. Whereas, if a node is found to be legitimate, then its data is stored in an
Interplanetary File System (IPFS). IPFS stores the data in the form of chunks and generates hash for the
data, which is then stored in blockchain. In addition, Verifiable Byzantine Fault Tolerance (VBFT) is used
instead of Proof of Work (PoW) to perform consensus and validate transactions. Also, extensive simulations
are performed using the Wireless Sensor Network (WSN) dataset, referred as WSN-DS. The proposed model
is evaluated both on the original dataset and the balanced dataset. Furthermore, HGB is compared with other
existing classifiers, Adaptive Boost (AdaBoost), Gradient Boost (GB), Linear Discriminant Analysis (LDA),
Extreme Gradient Boost (XGB) and ridge, using different performance metrics like accuracy, precision,
recall, micro-F1 score and macro-F1 score. The performance evaluation of HGB shows that it outperforms
GB, AdaBoost, LDA, XGB and Ridge by 2-4%, 8-10%, 12-14%, 3-5% and 14-16%, respectively. Moreover,
the results with balanced dataset are better than those with original dataset. Also, VBFT performs 20-30%
better than PoW. Overall, the proposed model performs efficiently in terms of malicious node detection and
secure data storage.
INDEX TERMS Blockchain, histogram gradient boost, IPFS, malicious node detection, VBFT, WSN.
I. INTRODUCTION monitoring, etc., [1]. Sensor Nodes (SNs) are used to monitor
A Wireless Sensor Network (WSN), comprising thousands and gather environmental data. Besides, in crowd sensing
of nodes, is widely used in several applications like supply networks, SNs send massive amounts of the collected data
chain management, military surveillance, environmental to the nearby nodes and Cluster Heads (CHs). This process
decreases the cost of different types of equipment and con-
ventional methods for data collection. However, some nodes
The associate editor coordinating the review of this manuscript and do not participate in crowd sensing networks due to privacy
approving it for publication was Nitin Nitin . issues.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
6106 VOLUME 11, 2023
M. Nouman et al.: Malicious Node Detection Using ML and Distributed Data Storage Using Blockchain in WSNs
Moreover, in the absence of a security mechanism, WSNs Moreover, most of the researchers propose the Interplan-
become vulnerable to malicious nodes that modify the data etary File System (IPFS) for data storage, which was intro-
for their own interest. Furthermore, the SNs are resource duced by Juan Benet [14]. IPFS shares many of the same
constrained and do not perform efficient resource utilization. characteristics as blockchain. It uses a P2P, decentralized,
In addition, traditional methods are unable to detect malicious and distributed file storage system. Besides, IPFS nodes are
nodes. Whenever an attack is performed by a malicious node, the machines that execute the IPFS software to store and
the network is compromised, and malicious nodes perform retrieve files from the IPFS network. IPFS nodes use content
malicious activities that affect the entire network. To prevent addressing to store and retrieve the files. All IPFS nodes
the nodes from acting maliciously, many authors propose store the files in the form of chunks, similar to a BitTorrent
authentication schemes that allow only the authentic nodes network. There is no effect on the network if one node
to join the network [2]. However, the existing authentication fails. Furthermore, it uses two types of data structures to
schemes depend upon centralized entities, which are vulner- distribute the file. One is Distributed Hash Table (DHT),
able to cyber-attacks. and the second is Merkle Directed Acyclic Graph (DAG).
In WSNs, SNs are either randomly or statically deployed When nodes send a file to the IPFS for storage, then SHA256
depending upon network topology. SNs gather environmental algorithm is executed on the file, and the hash value for each
data and transfer it to their destination. However, some SNs stored file is generated. The hash value is called a content
do not store the location information because their topology identifier, which is used to retrieve the stored files from
is frequently changed, and the usage of a large number of the IPFS.
sensing nodes may cause network information congestion.
To solve this issue, a WSN is split into sub-networks that CHs A. MOTIVATION AND CONTRIBUTIONS
manage. CHs get data from SNs and send it to Base Stations The motivation of this work comes from the fact that in
(BSs) [3]. Moreover, SNs are resource constrained in terms WSNs, nodes are are randomly deployed. This random
of low storage and computational power. Also, SNs are prone deployment leads to various issues like loss of data, security
to different types of attacks and are easily compromised by risk, etc. In WSNs, data is collected from the surrounding
malicious nodes. Many researchers propose different tech- environment. WSNs are easily accessible, and any node can
niques to avoid malicious attacks and detect malicious nodes join them. As a result, malicious nodes enter the network and
[4], [5]. However, detection of the malicious node in WSNs perform malicious activities that affect the entire network.
depends on a third party, which can easily be compromised. The authors in [15] propose a centralized authentication
Therefore, blockchain is introduced to overcome the prob- mechanism that registers the nodes and protects confidential
lems associated with centralization and the involvement of node identification from an unauthorized node in WSNs.
third parties [6], [7], [8], [9], [10]. However, the centralized system causes the issue of a single
WSN nodes produce vast amount of data and store them point of failure. Moreover, SNs have resource constraints and
on a centralized system. However, security breaches and do not efficiently detect malicious behavior in the network.
failures might destabilize the WSNs. Therefore, a Peer-to- Also, malicious nodes can easily damage and compromise
Peer (P2P) network is proposed to overcome centralization the WSNs [16]. Furthermore, malicious nodes collect false
issues related to data storage [11]. In a P2P network, nodes data and deliver it to destination nodes where blockchain
directly transfer the data from the source to the destina- is deployed to store the data [17]. However, storing huge
tion without the assistance of a third party. With the rapid volumes of data in a blockchain is very expensive. In addition,
expansion of WSN nodes, P2P architecture faces security blockchain uses the PoW consensus mechanism for block
and privacy challenges. Therefore, blockchain technology is generation, which consumes a huge amount of computation
introduced to address the security issues of WSNs through a power during block generation [13]. Further motivation can
distributed, decentralized, and immutable ledger [12]. Once be taken from [18]. The results are provided in Section IV in
data is added to the blockchain, it will never be tampered the manuscript.
by any malicious party due to the distributive nature of The proposed model’s key contributions include the
the blockchain. Furthermore, the idea of integrating WSN following:
and blockchain has attracted much attention from the pub- • in a WSN, a blockchain based decentralized authenti-
lic. However, blockchain consumes a lot of computational cation mechanism is used to protect disclosure of node
resources, whereas, SNs have limited resources. Also, when identities by external nodes,
incorporating the new blockchain design into the WSNs, • for data storage in a WSN, IPFS is deployed that inte-
some other issues may arise. Besides, the Proof of Work grates blockchain technology. The cost of storing data in
(PoW) consensus mechanism is widely used in blockchain the blockchain is minimized when storing data on IPFS.
that effectively reduces the number of malicious nodes The data is stored in chunks in IPFS, and the hashes are
and verifies the transaction. However, the PoW consensus created that are recorded in the blockchain,
mechanism requires a large amount of computational power • the proposed blockchain based network uses the Ver-
to confirm a transaction and add it to the block in the ifiable Byzantine Fault Tolerance (VBFT) consen-
blockchain [13]. sus mechanism [19], which reduces the blockchain
VOLUME 11, 2023 6107
M. Nouman et al.: Malicious Node Detection Using ML and Distributed Data Storage Using Blockchain in WSNs
transaction cost and increases the throughput as com- the blockchain. If the IoT device is successfully registered
pared to the existing consensus mechanisms like PoW and authenticated, the activity is performed according to its
and capability. Similarly, users need to be authenticated in the
• the comparative analysis of the proposed classifier, i.e., blockchain network to be able to control and manage IoT
Histogram Gradient Boost (HGB), with Adaptive Boost devices. It restricts the malicious nodes from becoming a
(AdaBoost), Gradient Boost (GB), Extreme Gradient part of the network and stores all evidences on a blockchain.
Boost (XGB), Linear Discriminant Analysis (LDA), and In [26], the modified version of the station-to-station (STS)
ridge classifiers is performed. The analysis is done on protocol is presented. It first authenticates the user and then
the basis of numerous performance metrics, including establishes a secret exchange session key that ensures user
accuracy, precision, recall, micro F1-score, and macro anonymity inside a group.
F1-score. In [29], a blockchain based data structure model is used for
The remainder of the manuscript is organized as follows. malicious node detection. WSN nodes have limited memory
The related work is presented in Section II, while the problem and computational power, and are unable to detect mali-
statement and proposed system model are given in Section III cious nodes. Whenever an attack is performed on a node,
and Section IV, respectively. Section VI presents the out- it is compromised by a malicious node. In [30], the authors
comes of the simulations performed to verify the accuracy propose the three layered SDN architecture that monitors
of the proposed model. Section VI provides the feasibility and analyzes the traffic in the IoT environment. Another
of the proposed model. In Section 7, the conclusion of the pertinent point is that a blockchain is used for decentralized
manuscript is provided. attack detection. As a result, fog computing and mobile edge
computing provide attack detection, reducing the number
II. RELATED WORK of attacks that occur at the edge layer. In [31], a secure
In WSNs, SNs share information and communicate with each and privacy-preserving model is proposed for the smart city.
other. WSNs are easily accessible, and any node can join Three modules make up the proposed model. The first module
them. Malicious nodes acquire legitimate node identities, is trustworthiness, where authors use the blockchain among
which makes it easy for them to become part of the network. the IoT devices to maintain trust. The second module is two-
The authors propose a lightweight blockchain IoT authen- level privacy, where enhanced PoW is used in blockchain
tication scheme in [20]. This scheme ensures integrity and to achieve confidentiality and prevent the poisoning attack.
non-repudiation in the network. Whenever IoT nodes com- The third module is the intrusion detection system, which is
municate with one another, they must first authenticate each used for malicious node detection. XGB classifier is utilized
other, which is done using a lightweight blockchain. In [21], in the process of identifying malicious nodes. In [32], the
the authors develop a hybrid blockchain model for IoT nodes authors propose the secure privacy-preserving framework.
to prevent malicious or fake data packets from spreading The presented model has two major components: two-level of
throughout the network. Public and private blockchain make privacy and an intrusion detection mechanism. Blockchain is
up a hybrid blockchain. Between CHs and BSs, the pub- utilized in two-level privacy to securely transmit data among
lic blockchain is implemented, while the private blockchain IoT nodes. The two-level privacy uses principal component
is implemented between CHs and SNs. SNs are authenti- analysis (PCA) to transform data into a new form to protect
cated on CH using a smart contract, and CHs are authenti- it against inference attacks. The authors use gradient boost
cated on BS. In [22], blockchain and reinforcement learning anomaly detection (GBAD) for the intrusion detection system
based model is proposed for efficient and secure routing based on light gradient boost model (LGBM). GBAD is
in WSNs. The reinforcement learning algorithm selects the deployed in a smart city that can proficiently classify nor-
best possible routing path. It avoids the malicious routing mal and malicious observations. In [33], a blockchain based
links that might send data through compromised nodes, while automatic (AutoML) model is proposed for customer services
blockchain is used for node authentication and managing all to overcome the third parties’ challenges. IoT devices are
routing information. In [23], blockchain based key manage- used to collect data, and blockchain is used for secure data
ment is presented to tackle the issue of certificate-less key exchange in an open environment. Furthermore, AutoML is
management. The blockchain performs node authentication, designed to process data and reduce expert costs. In [34],
registration, and joining or quitting of nodes. In addition, the authors propose an ensemble learning technique that uses
it provides the mechanism for the detection of the compro- multiple ML techniques to classify data. The final classifica-
mised node. In [24], a data structure based on blockchain tion report is obtained based on all classifiers’ votes.
is used to hold nodes’ authentication and trust informa- In [35], secure routing with multi-layered IoT architec-
tion. Blockchain authentication consists of three aspects: ture is proposed, where light blockchain and cloud are used.
public keys, block mining, and mutual influence, while the Light blockchain is used for security and privacy, while the
blockchain trust model consists of two aspects: knowledge cloud is used for data storage. In [36], two different kinds of
based trust and trust evaluation. In [25], blockchain is used blockchain are used in a WSN: one for storing data and the
to overcome IoT issues. IoT devices register themselves on other for managing how users can access data. A verifiable
data possession consensus mechanism is also used to reduce earn incentives on the blockchain network. In [42], a scal-
the cost of computing. In [37], a decentralized blockchain able and secure blockchain is proposed for the IIoT system.
mechanism is proposed for Internet of Things (IoT) mon- First, a self-adaptive PoW consensus mechanism is proposed,
itoring and controlling, in which each entity can track and which adjusts the difficulty for nodes as per their behaviours.
communicate. The data controller manager receives the data The self-adaptive PoW consensus mechanism efficiently
and filters out specific data stored in the blockchain. In [38], reduces computational power. Moreover, asymmetric cryp-
the multiple synergistic proofs green consensus method is tography is used for access control, giving users more options
proposed to address the issue of limited data storage. As a for managing data authority. Furthermore, a blockchain based
result, less space is being allocated to blockchain data. When directed acyclic network is used to improve throughput and
peer nodes verify a transaction or a block, they frequently transaction time. In [43], a hybrid model based on an SDN
send the same information. This is not favourable and dete- and blockchain is proposed for the smart city. Two types of
riorates network performance. In [39], a blockchain based smart city nodes exist in the network: edge node and core
aggregation scheme is proposed that decreases the device’s node. The core nodes are provided with the data from the edge
duty cycle. Moreover, the reduction in the risk of transmitting nodes after the edge nodes receive the data from the sensors.
a large amount of data at the risk of increasing data delay These edge nodes act as centralized entities because edge
at the IoT device is achieved. The selection of gathered data nodes also use SDN technology, and their computing power
is based on the channel’s quality, and the most recent data and storage are less than that of core nodes. Whereas, core
structure statistics. The expense is incurred in sending the nodes are the powerful nodes that receive the sensors’ data
Merkle-Patricia tree data structures as evidence of inclusion and perform mining. Core nodes, also called miner nodes, use
for the most recent data. In [40], an optimized sampling rate blockchain technology for mining transactions and enhancing
strategy is proposed for IoT sensors that transmit data using security.
blockchain and Tangle technologies, which decrease the age In [44], a lightweight exclusive OR (XOR) hash algorithm
of information, and make efficient end-user processing and is used, which provides secure and reliable data routing
networking resources. using blockchain technology. In [45], rolling blockchain is
In [41], blockchain is used to resolve IoT networks’ secu- used for WSNs to ensure that the WSN nodes and data are
rity and privacy issues, providing a secure distributed and secured from attackers. In [46], a blockchain system with
immutable ledger. Some nodes in a blockchain network are mobile edge computing allows mobile miner nodes to per-
responsible for mining, which entails validating transactions form computationally intensive tasks on surrounding edge
and adding new blocks to the blockchain. Mining nodes nodes. As a result of this strategy, backhaul and latency are
require a high computing capacity as compared to conven- minimized. In [47], a trust-aware localization routing proto-
tional nodes. IoT devices have limited power, battery life, etc. col with class based dynamic encryption is proposed. This
This research encourages IoT devices to purchase process- proposed method first searches the secure path from source to
ing power through edge servers, participate in mining, and destination and then forwards the data packet. The selection
of a secure path is made using the trust value. Moreover, B. FORMULATION PROBLEM
blockchain based encryption is used for data integrity. In [48], In WSNs, nodes are randomly deployed, and data is collected
a trust based range-free safe localization method is proposed. from the surrounding environment. WSNs are easily accessi-
The trust values of beacon nodes are communicated through ble, and any node can join them. As a result, malicious nodes
the blockchain with their nearby nodes. Trustworthy beacon enter the network and perform malicious activities that affect
nodes are selected as miners for block mining. However, it is the entire network. The authors in [15] propose a centralized
a time consuming process. Table 1 presents the summarized authentication mechanism that registers the nodes and pro-
related work. tects confidential node identification from an unauthorized
node in WSNs. However, the centralized system causes the
III. PROPOSED SYSTEM MODEL issue of a single point of failure. Moreover, SNs have resource
In this section, the WSN network, developed in the proposed constraints and do not efficiently detect malicious behaviour
work, is discussed. Also, this section introduces the reason- in the network. Also, malicious nodes can easily damage and
able assumptions that are used to propose the network for compromise the WSNs [16]. Furthermore, malicious nodes
WSNs. Figure 1 presents the proposed system model. collect false data and deliver it to destination nodes where
blockchain is deployed to store the data [17]. However, stor-
A. ASSUMPTIONS
ing huge volumes of data in a blockchain is very expensive.
In addition, blockchain uses the PoW consensus mechanism
• SNs and CHs are resource constrained and each node has
for block generation, which consumes a huge amount of
a unique identity.
computation power during block generation [13].
• BSs provide a certain amount of data storage and com-
putational power for processing the data sent by the SNs.
• BSs are resource enriched and trustworthy nodes for the C. SYSTEM MODEL DESCRIPTION
CHs and SNs. Our proposed network model’s nodes are divided into SNs,
• There is a possibility of the occurrence of energy nodes CHs, and BSs. SNs are randomly deployed. Whereas, CHs
in WSNs, as discussed in [49]. However, in the proposed are chosen based on their high residual energy relative to SNs
work, it is assumed that no energy holes exist. and their closeness to the BSs. SNs and CHs are registered
D. REGISTRATION
SNs and CHs play an important part in the overall system
model. SNs and CHs must create accounts on the blockchain
and send registration requests to BSs. After getting their
registration responses, the SNs and CHs send data to BSs.
E. SENSOR NODES
In the proposed model, we adopted the WSN model, which is
also used by [64]. SNs are randomly deployed according to
their functionalities. SNs collect data from their surrounding
area and transmit it to the CH node. Each SN is directly
connected to a CH in the network and shares its data and cre-
dentials after it is fully registered. The credentials are stored
on the blockchain, which ensures that the SNs are secured.
The data once stored in the blockchain can not be forged as
blockchain becomes data integrity and tamper-proof nature.
The security of SNs is necessary because if they are secured,
only then the data coming from them can be regarded as
authentic, as in [65].
FIGURE 2. Proposed model’s workflow.
F. CLUSTER HEADS
CHs are intermediate nodes that receive data from SNs, pro-
on the blockchain, which is deployed on BSs. Following cess it, and then pass it to the BSs. CHs’ storage capacity and
the registration, the blockchain authenticates SNs and CHs computational power are higher than those of SNs and lower
using Node_ID. SNs collect data and send it to CHs. While than those of BSs. Each CH is directly connected to a BS to
CHs process data and transmit it to BSs. BSs utilize an ML shares its information and credentials with it.
classifier, HGB, to determine whether data is transmitted to a
malicious node or a normal node. When data is transferred to G. BASE STATION
a malicious node, the HGB classifier quickly recognizes that
A BS is a powerful node that has the highest computational
the data belongs to which malicious class, classifies the attack
power and storage capacity in the proposed network. It is
to that class, and reports to the blockchain. Blockchain then
also considered the core node of a network. In the pro-
revokes the malicious node’s registration. Otherwise, data is
posed network, BSs receive data from CHs, perform some
stored in an IPFS database. IPFS generates a unique identifier
complex operations, and verify if the data is being trans-
(hash) for the data and sends it back to the BSs, where it
ferred to a malicious mode or a normal node. The malicious
is stored on the blockchain. Moreover, public blockchain
node’s registration is revoked if data is delivered to it. If not,
is implemented on the BSs. The public blockchain is cus-
it keeps information in the IPFS database. BSs store the net-
tomized to allow nodes to add, remove and validate transac-
work nodes’ credentials and monitor the whole network. BSs
tions. Furthermore, a VBFT consensus mechanism is used in
serve as trusted nodes for other nodes or subnetworks in the
the blockchain to verify and store transaction nodes’ creden-
network.
tials and cryptography hashes [19]. Figure 2 is a representa-
tion of the workflow associated with the proposed model. The
H. CUSTOMERS
steps involved in the system model are given in Algorithm 1.
In the proposed model, shown in Figure 1, customers rely on
IPFS and blockchain. Customers need to access the data that
Algorithm 1 Pseudo Code of the Proposed Model the SNs gathered. Therefore, customers are initially regis-
Step-1: All nodes are registered on the BSs, where tered on the blockchain. The blockchain confirms whether the
blockchain is implemented. customer identity exists or not. If the customer is validated,
Step-2: SNs collect data from the surrounding area and relay it is allowed to join the network. A customer enters the
it to CHs, while CHs forward the data to the BSs [64]. network and requests the hash of the desired data, already
Step-3: The BSs are trained on an ML classifier, HGB, that recorded on the blockchain. The customer provides the hash
classifies the data and sends it either to a malicious node or a value to the IPFS database to retrieve the data associated with
normal node. the hash.
Step-4: Data is stored on the IPFS if the BS classifies the
node as a normal node. Otherwise, the BS revokes the node’s I. MALICIOUS NODE DETECTION USING MACHINE
registration. LEARNING CLASSIFIERS
Step-5: The IPFS provides hashes that are stored on the Some malicious nodes enter the network as legitimate nodes
blockchain implemented on the BSs. to complete their registration process. After the registration,
these malicious nodes change their behaviour and act mali- DT learned from the mistakes of a prior DT. These DTs are
ciously to attack the network nodes. In our proposed system, sequentially connected to each other, and each DT minimizes
the ML classifier is deployed on BSs to classify the normal the error of the previous DT. Furthermore, the additive model
and malicious nodes. We conduct comparative analyses using combines the outcomes of each step, given the strong learner.
six different ML classifiers for malicious node detection. The time complexity of GB is O(ftnlogn) [53]. Here, logn
These classifiers are AdaBoost, GB, XGB, LDA, ridge, and represents the depth of the weak learners. The GB algorithm
HGB, which are used to classify malicious and legitimate is presented below.
nodes. After nodes’ classification, BSs revoke the malicious
nodes from the network. The classifiers are further discussed Algorithm 3 Gradient Boost Algorithm
below. InitializeP model with constant value: F0 x =
Freund and Schapire invented AdaBoost in 1996 [50]. argminr ni=1 L(yi , r).
It was the first ensemble boosting technique, which aims to for m = 1 to M
combine multiple weak learners and make a strong model Compute residual rim = −[ ∂L(y i ,F(xi )
∂F(xi ) ]F( x)=Fm−1 (x) for
because a single weak learner is not able to predict an accurate i = 1, . . . , n
class. The combination of weak classifiers makes a new Train regression tree with feature x against r and create
strong classifier after the voting mechanism, and AdaBoost terminal nodes reasons Rjm for
is one of them. Because of less time complexity, fast per- P j = 1, . . . , m
Compute rjm = argminr xi εRjm L(yi , F(m−1)(xi ) + r) for
formance, and no difficulty in implementation, AdaBoost is j = 1, . . . , m
the most efficient and effectively used classifier in computer Pjm
Update the model: Fm (x) = Fm−1 (x) + v j=1 rjm 1
vision. Also, boosting methods are considered greedy in
(x ∈ Rjm )
terms of dealing with the exponential error function. The
endfor
usage of AdaBoost improves the accuracy of weak classifiers.
Output f(x) = Fm (x).
This algorithm initially assigns equal weights to all samples
and passes them to the first weak learner. The weak classifier
is trained, giving the output in the form of 1, -1. After that, LDA is a broader variant of Fisher discriminate analysis
weights are assigned in the second round to each observation. (FDA), also known as normal discriminate. LDA is a super-
This process is repeated several times, creating a set of weak vised ML technique used for classification and dimension-
classifiers. The time complexity of AdaBoost is O(ftn) [51]. ality reduction, invented by Ronald A. Fisher in 1936 [54].
Here, f represents the features, t represents the weak learners The primary goal of LDA is to reduce higher dimension
while the number of dataset samples is presented by n. The data into lower dimension data to prevent losing important
AdaBoost algorithm is presented below. information and reduce the consumption of computational
resources. However, the number of features surpassing the
Algorithm 2 AdaBoost Algorithm number of samples along with the nonlinearity of the data
Initialize the observation weights wi = 1/N , i = 1, 2, . . . , N. points cause the LDA to fail. In dimensionality reduction,
for m=1 to M three steps are involved. In the first step, separability between
(a) Fit a classifier Gm (x) to training data using weight wi . the classes, known as between the class matrix or between
PN
i=1 wi I (yi ̸ =Gm (xi )) the class variance, is calculated. The goal is to maximize the
(b) Compute errm = PN .
i=1 wi separation between the two classes. The difference between
(c) Compute αm = log((1 − errm )/errm ). the mean of class and the data point of a class, known as
(d) Set wi ← wi .exp[αm .I (yi ̸ = Gm )], i = 1, 2, . . . , N. within class matrix or within class variance, is calculated in
endfor the second step. The aim of this calculation is minimizing
Output Gm (x)= sign[ M m=1 αm Gm (x)].
P
the within class matrix or within class variance. In the third
step, the new lower dimensional space is built and the new
GB is a supervised ML algorithm invented by Friedman, data point is projected onto it. The time complexity of LDA
2001 [52]. It is an ensemble technique that is used for classi- is O(Mf 2 ) [55] in the case where the number of instances
fication and regression. It is different from AdaBoost, and it exceeds the number of features. M represents the mean of
comprises three parts: loss function, weak learners, and addi- instances. The LDA algorithm is presented below.
tive model. The loss function is used to minimize the residual XGB was created by Tianqi Chen in 2014 in order to
and converge the final output. While the weak learners are improve the performance and speed of ML models [56].
used to make predictions. Initially, GB uses two models to Despite being scalable and a highly accurate extension of GB
start with a base model and find the residual passes to the first for boosted tree algorithms, XGB needs heaps of computing
weak learner. It combines several weak learners and makes power. It refers to engineers’ goal of pushing the limit of
a single strong learner using the additive model. Individual computation resources for the GB technique. This technique
weak learners act as decision trees (DT) in GB. These DTs are sequentially generates DTs. Weights are very important in
constructed so that each new tree fits within the residuals of XGB. All independent variable are assigned weights and
the previous step, allowing the model to reduce error. A new subsequently fed into the DT predicting results. When the
Algorithm 4 Linear Discriminant Analysis Algorithm this algorithm more efficient as compared to XGB, LGB,
Given a set of N samples [xi ]Ni=1 ].
and GB in terms of memory consumption and training speed.
Compute the mean of each class µ(1 × M ). It is also used to convert continuous or numerical variables
Compute the total mean of all data µ(1 × M ). into categorical features and deal with noisy data. The time
Calculate between class matrix SB (M × M ). complexity of HGB is O(ft(n_ bins)) [60]. Here, n_bins rep-
for Class i= 1, 2, . . . , c resents the data instances in the current block to generate the
Compute within-class matrix of each class Swi (M × M ), histogram. The HGB algorithm is presented below.
as follows:
Swj = xi ϵwj (xi − µ)(xi − µ)T
P
Algorithm 6 Histogram Gradient Boost Algorithm
Construct a transformation matrix for each class (Wi ) as
InitializeP model with constant value: F0 x =
follows:
argminr ni=1 L(yi , r).
Wi = Sw−1 i SB for m=1 to M
The eigenvalues (V ) and eigenvectors (ω) of each transfor-
Compute residual rim = −[ ∂L(y i ,F(xi )
∂F(xi ) ]F( x)=Fm−1 (x)
mation matrix (Wi ) are calculated, the calculated eigenvector
for i=1,. . . ,n
and eigenvalues of the i-th class.
Apply binning technique.
Sort the eigenvectors in descending order according to their
Sort the data features.
corresponding eigenvalues.
Distribute the data feature in bins.
Project the samples of each class onto their lower dimen-
Train regression tree with feature x against r and create
sional space.
terminal nodes region Rjm for Pj=1, . . . , m
Y=X.T
Compute rjm = argminr xi εRjm L(yi , F( m − 1)(xi ) + r)
endfor
for j=1, . . . , m Pjm
Update the model: Fm (x) = Fm−1 (x) + v j=1 rjm 1
tree wrongly predicts a variable, the weight of the variables (x ∈ Rjm )
is increased, and these variables are provided in the second endfor
DT. These various predictors are combined to form a more Output f( x) = Fm (x).
robust and precise model. Three steps are performed in XGB.
Firstly, it reduces overfitting by using regularization. Second,
it optimizes sorting with parallel execution, which increases A ridge classifier is a type of ridge regressor. The ridge
runtime speed. Finally, it prunes the tree using the maximum classifier first converts the target variable into binary form
depth of the DT as a parameter, minimizing the total runtime. (1,−1) and then treats it as a ridge regressor. Hoerl and
The time complexity of XGB is O(tdxlogn) [57]. Here, d rep- Kennard introduced the ridge regressor in 1970 as a regular-
resents the height of the tree, and x represents the missing ization method for reducing model complexity [61], [62]. The
values. The XGB algorithm is presented below. time complexity of the ridge classifier is 0(n3 ) [63]. It uses
the coefficient estimator for variables that are not linearly
Algorithm 5 Extreme Gradient Boost Algorithm independent but are highly correlated. The ridge estimator
Data: Dataset and hyperparameters Initialize f0 x; shrinks the coefficient value and produces a new value close
for k=1 to M Calculate gk = ∂L(y,f )
∂f ;
to the actual population. Furthermore, it involves plenty of
coefficient mechanisms, meaning that no coefficient is left
Calculate hk = ∂ L(y,f
2 )
∂f 2
; when the model is built. Due to the penalty mechanism, the
Determine the structure by choosing splits with maximized loss function is minimized. The ridge classification algorithm
gain is presented below.
G2 G2 2
A = 21 [ HLL + HRR + GH ];
G
Determine the leaf weights w∗ = − H ;
PT
Determine the base learner b(x)= j=1 wI ; Algorithm 7 Ridge Classification Algorithm
Add trees fk (x) = fk − 1(x) + b(x); Step-1: Input data matrix X holds training dataset and data
endfor matrix X-test holds the test dataset.
Result f (x) = M
P
k=0 fk (x) Step-2: For each test data xϵ X-test, calculate the regression
parameter vector ά as ά = arga min||x − Xi α||22 + λ||α||22
A histogram is used to count the frequency of data across where, λ represents the regularization parameter and i class.
a specific time period. It is also known as binning or Step-3: Perform projection of the new test sample X onto the
bucket [58]. Instead of calculating the split points on the subspace of each class i using ά as x́i = Xi άi .
sorted feature values, HGB applies the binning method to Step-4: Calculate distance between the test sample x and the
the DT [59]. The binning method is applied to data for pre- class-specifics sub-sample x́i .
processing, which sorts the feature values and then divides the Step-5: Test sample x is assigned to that class whose distance
sorted feature values into numerous buckets or bins. It makes is minimum.
J. DATASET DESCRIPTION model used in [64] with a slight modification. The modifi-
The WSN dataset (WSN-DS) used in this study was pub- cation is that in the proposed model, two BSs are used, while
lished in [64]. According to this research, the WSN-DS in [64], only one BS is used. The reason is that for implement-
was developed with the help of the Low Energy Adaptive ing blockchain in our scenario, at least two BSs are required,
Clustering Hierarchy (LEACH) routing protocol. SNs are while the model in [64] is a centralized model. Moreover,
used to collect the data and deliver it to the CH. The CH in this research, we use Google Collaboratory and the Python
receives data from the SNs and transmits it to the BS. The programming language to evaluate the performance of the
BS then aggregates the data from all CHs and generates proposed model using ML classifiers on both datasets: origi-
the dataset. This dataset has 18 features and five classes. nal WSN-DS and balanced WSN-DS. Furthermore, Solidity
The dataset’s features are Node ID, Is CH, RSSI, Distance language is used to develop Smart contracts, implemented in
to CH, Average distance to CH, Max distance to CH, Current Remix IDE. Also, the Remix web3 environment is integrated
energy, ADV_CH send, ADV_CH receives, Distance CH to with the Ethereum wallet using Metamask. Then, virtual
BS, Data send, Data received, etc. More details about the currency is transferred to the Ethereum wallet accounts using
features are given in [64]. This dataset is divided into five Ontology and Ganache to evaluate VBFT and PoW test net-
different classes. The first class is normal, while the remain- works’ transaction costs. Furthermore, IPFS is installed on
ing classes are concerned with the DoS attacks. DoS attacks Windows, and Visual Studio code is used to upload and down-
include Grayhole attacks, Blackhole attacks, Time division load files using IPFS. The hardware specifications are an
multiple access (TDMA) attacks, and Flooding attacks. The Intel(R) Core(TM) i5-4200U processor running at 1.60GHz
details about the classes and distribution of instances are with 12GB of RAM and a 64-bit operating system.
given in Table 2. Furthermore, the WSN-DS consists of
374661 instances, with 340066 instances belong to the nor- A. BLOCKCHAIN RESULTS’ DISCUSSION
mal class, and 34595 instances belong to the malicious class. The results of the blockchain experiment are analyzed in this
It indicates that WSN-DS is highly imbalanced, which could section. The proposed model is evaluated using two consen-
lead to a problem of weak generalization by classifiers. sus mechanisms. Figure 3 compares the average transaction
costs of the VBFT and PoW consensus mechanisms. There
TABLE 2. Details of WSN-DS. are three types of functions that are used in the proposed
model: Reg(), Auth(), and Revoke(). WSN nodes are reg-
istered using the Reg() function, while the Auth() function
confirms whether the node’s identity exists or not. If the WSN
nodes are validated, they are allowed to join the network.
When the ML model discovers a malicious node in the net-
work, the Revoke() function is called. We observe that the
cost of a VBFT transaction is lower than the cost of a PoW
K. DATA SAMPLING transaction. This is because VBFT selects random verifiers
The WSN-DS is highly imbalanced, as mentioned above in in each round, reducing the probability of malicious nodes.
Table 2. When data from the majority and minority classes In contrast, PoW selects mining nodes with a large number
is not balanced, it indicates a biasness in favour of the of processing resources in each round, leading to a high PoW
majority class. As a result, classification accuracy decreases, transaction cost. Moreover, the transaction costs of Auth()
and the classifiers’ performance degrades. In our case, the and Revoke() in both PoW and VBFT are nearly the same.
number of normal class instances is greater than the number The reason is that both Auth() and Revoke() use a single
of malicious class instances. Therefore, it is necessary to attribute for each node. However, the cost of Reg() is differ-
balance the data before giving it to the classification model. ent. The reason is that Reg() function uses three attributes for
In literature, two types of balancing techniques are used to each node to complete the registration. PoW transaction is
deal with imbalanced data. One is oversampling, and the more costly than VBFT because it chooses the miner with
second is undersampling. The oversampling increases the the highest computational power to validate the transaction.
number of instances, whereas, the undersampling decreases In the case of VBFT, it selects the random verifiable for the
the number of instances. Both are used to solve the prob- miner who validates the transaction.
lem of data imbalance. Both have their own sets of benefits Figures 4(a) and 4(b) show how much time it takes to
and drawbacks. This research uses the Synthetic Minority upload and download a file using IPFS. IPFS is used to upload
Oversampling Technique (SMOTE) to handle the imbalanced and download five different files with sizes of 10kB, 100kB,
data [66]. It duplicates the minority class instances by using 1MB, 10MB, and 100MB. The upload time of 10kB, 100kB,
an existing instance to make new instances. 1MB, 10MB, and 100MB files is 2.93s, 3.03s, 4.00s, 4.52s,
and 5.01s, respectively. At the same time, the download time
IV. RESULTS AND DISCUSSION of the files is 2.22s, 2.44s, 3.25s, 3.7s, and 4.0s, respectively.
The proposed model’s simulation results are the content of The computing time increases as the size of files increases.
this section. In the proposed model, we adopted the WSN We notice that the file’s upload time is higher than its
samples out of all positive samples. The reason is that the TABLE 3. Computational complexity of classifiers.
HGB classifier organizes data features before equally dis-
tributing them. Then it employs the weak learners sequen-
tially for prediction.
F1-score is the harmonic mean of precision and recall.
It determines the single value that balances both precision and
recall. The F1-scores of the classifiers differ for the balanced
and original datasets, as shown in Figures 8(a) and 8(b).
HGB classifier achieves the highest F1-score as compared
to GB, AdaBoost, LDA, XGB, and ridge. HGB achieves a
99.5% F1-score on the balanced dataset, while other clas-
sifiers achieve 98%, 93%, 88%, 98%, and 87%, F1-score 0.5 means it performs poorly. In Figure 9, the point closer
respectively. HGB achieves a 97% F1-score for the original to the top left corner, which is equal to 1, denotes higher
dataset, while the other classifiers achieve 97%, 84%, 79%, classification results.
97%, and 75% F1-score, respectively. The HGB classifier The time complexity of six classifiers in terms of training
obtains a high F1-score because precision and recall are both and prediction time is shown in Table 3. We execute six
high on balanced and original datasets. classifiers on the original and balanced datasets, and record
Figure 9 shows the receiver operating characteristic (ROC) the training and prediction times in seconds (s). Table 3
curves of WSN-DS for multiclass using HGB classifier. The shows that the ridge and LDA classifiers achieve the best time
ROC area under the curve (AUC) is utilized for overall assess- complexity on the original and balanced datasets, while GB
ment. The ROC curves show all possible differences between and XGB obtain the worst time complexity. The reason is
true positive rate (TPR) and false positive rate (FPR) across that ridge and LDA are weak classifiers. Whereas, GB and
multiple decision thresholds. The AUC evaluation metric XGB use many DTs, which perform better than LDA and
converts this curve into a value between [0.5, 1]. A value ridge. The proposed HGB classifier takes more time than
of 1 indicates that the classifier performs efficiently, while LDA and ridge, but it takes less time than boosting classifiers
FIGURE 7. (a) Recall of classifiers on the balanced dataset. (b) Recall of FIGURE 8. (a) F1-score of classifiers on the balanced dataset. (b) F1-score
classifiers on the original dataset. of classifiers on the original dataset.
C. COMPARATIVE ANALYSIS OF THE PROPOSED MODEL original dataset. The reason is that it is an ensemble learning
This section evaluates the proposed model’s efficiency in classifier that uses multiple weak learners. Secondly, it uses
comparison to state-of-the-art ML classifiers with respect to the binning method, which distributes the data features into
attacks, and micro and macro F1- scores. The proposed model bins. After the data feature distribution in bins, HGB trains
is evaluated using six different classifiers: AdaBoost, GB, the weak learners on distributed data features, which gives
XGB, LDA, ridge, and HGB. SMOTE is used to balance the them higher scores than others. Therefore, we use the HGB
WSN-DS, and each classifier is trained on the original and classifier combined with the SMOTE method in our model
balanced datasets. The performance metrics show that the for malicious node detection.
HGB multi-label classifier gives the highest result in terms We compute F1-scores for the detection performance of
of micro and macro F1-score among the six classifiers. HGB six classifiers using the original and balanced datasets. The
performs more efficiently on the balanced dataset than on the results are presented in Tables 4 and 5. When computing
TABLE 4. The detection performance of six classifiers using the dataset classifying the nodes as malicious and legitimate. Once the
balanced through SMOTE.
model is trained, it will check for every incoming node and
remove it if it is malicious before it becomes a part of the
network. Furthermore, when the proposed model is applied
on a similar network, it helps the network to perform in the
presence of legitimate nodes only. The WSN is operated in
an uncontrolled environment. With the application of ML,
the nodes acting maliciously in the network are identified.
TABLE 5. The detection performance of six classifiers on the original
The identified malicious nodes are removed from the network
dataset. and only the legitimate nodes are used in routing activities.
Thereby, making the system to operate in a controlled envi-
ronment. Besides, the SNs do not have any heavy compu-
tational requirements. Regular SNs send their data to the
CHs. CHs’ storage capacity and computational power are
higher than those of SNs. The CHs receive the data from
the neighboring nodes and forward it to the BS. This task is
conventional in the WSN and can be implemented easily in
micro-F1, the average metric is calculated by averaging all practice. The ML classification to detect the malicious nodes
classes’ contributions. In contrast, in macro-F1, the metrics is performed at the BSs and not at the CHs. The BS has access
are computed separately for each class and then they are to the electricity grid, has high computing power and can
averaged. As a result, each class is treated equally, regardless be equipped with the necessary hardware to accomplish the
of the number of samples in each class. Tables 4 and 5 classification task. In other words, we should clarify that the
show that HGB gives the highest micro F1-score and macro SNs do not incur heavy computational tasks, and all parts of
F1-score among all classifiers. Table 5 shows that both micro- the network can be implemented easily in practice including
F1 and macro-F1 values are equal in the balanced dataset blockchain, etc.
because the number of test samples for each class is equal.
However, in Table 5, the micro-F1 values of each class for VI. CONCLUSION
the balanced dataset are higher than the macro-F1 values This study proposes a network model to detect malicious
for the original dataset. The reason is that the number of nodes in WSNs. SNs and CHs are registered on BSs that
test samples for each class is large, and the F1-score values are responsible for monitoring the whole network and storing
are high. the credentials of the network nodes. In addition to this,
In conclusion, we use numerous performance metrics for blockchain technology is deployed on BSs. Both the verifi-
evaluating the proposed model. On the balanced and orig- cation and registration of nodes are done through blockchain.
inal datasets, the proposed model achieves high accuracy, Moreover, a consensus mechanism, VBFT, is used to validate
precision, recall, and F1-score. There are many reasons for the transactions, which reduces transaction costs. Moreover,
this enhanced performance. First, it is an ensemble learn- the network nodes’ credentials and the hash values that IPFS
ing boosting technique that uses many DTs consecutively produces are stored in the blockchain. Furthermore, the ML
to train the classifier. Second, it uses the binning method classifier, referred to as HGB, is utilized to identify malicious
on the dataset, which makes HGB classifier different from nodes. The simulation results show that the HGB classifier
other boosting techniques. The binning method organizes outperforms AdaBoost, GB, LDA, XGB, and ridge classifiers
data features and uniformly distributes them. After organizing in terms of accuracy, precision, recall, micro-F1 score, and
the data into bins, HGB uses several DTs to train the classi- macro-F1 score.
fiers. This strategy of the HGB classifier gives better results In the proposed work, individual classifiers are used for
than other boosting classifiers. In the proposed work, the classification, which does not provide enhanced efficiency.
simulations are performed using the WSN-DS. The proposed Moreover, the proposed work lacks in providing the vul-
work is scalable and efficient enough to be deployed using nerability analysis of the smart contracts, which deteriorates
the Empirical Dataset Generation Framework (EDGF) for the practicality of the work in the real world. Moreover, the
WSNs [67]. monitoring of the WSN/IoT systems is beyond our scope
at the current instant. In the future, to tackle the mentioned
V. FEASIBILITY OF THE PROPOSED MODEL issues, a stacking model will be used in WSNs for perform-
In the proposed work, ML and blockchain are used in the ing more efficient malicious node detection. Furthermore,
WSN. The blockchain brings the benefit of tackling the secu- the Oyente tool will be used to assess the smart contracts’
rity issues in WSNs. As WSNs comprise a large number of vulnerabilities. Moreover, this research will be conducted in
SNs, so, there exists an issue of efficiently distinguishing various sectors using real-world networks. Besides, we aim
the malicious nodes from the legitimate nodes. For this pur- to perform WSN/IoT system monitoring, as in [68], in the
pose, ML techniques are employed. The techniques help in future.
[38] Y. Liu, K. Wang, S. Member, Y. Lin, W. Xu, and S. Member, ‘‘LightChain: [59] A. Guryanov, ‘‘Histogram-based algorithm for building gradient boost-
A lightweight blockchain system for industrial Internet of Things,’’ IEEE ing ensembles of piecewise linear decision trees,’’ in Proc. Int. Conf.
Trans. Ind. Informat., vol. 15, no. 6, pp. 3571–3581, Jun. 2019, doi: Anal. Images, Social Netw. Texts. Cham, Switzerland: Springer, Jul. 2019,
10.1109/TII.2019.2904049. pp. 39–50, doi: 10.1007/978-3-030-37334-4.
[39] P. Danzi, S. Member, A. E. Kalør, and S. Member, ‘‘Delay and commu- [60] J. Garcia. Histogram for Efficient Gradient Boosting.
nication tradeoffs for blockchain systems with lightweight IoT clients,’’ Robotenique.github.io. Accessed: Jan. 17, 2023. [Online]. Available:
IEEE Internet Things J., vol. 6, no. 2, pp. 2354–2365, Apr. 2019, doi: https://robotenique.github.io/posts/gbm-histogram
10.1109/JIOT.2019.2906615. [61] B. Behera and G. Kumaravelan, ‘‘Performance evaluation of
[40] A. Rovira-Sugranes and A. Razi, ‘‘Optimizing the age of informa- machine learning algorithms in biomedical document classification,’’
tion for blockchain technology with applications to IoT sensors,’’ Performance Evaluation, vol. 29, no. 5, pp. 5704–5716, 2020, doi:
IEEE Commun. Lett., vol. 24, no. 1, pp. 183–187, Jan. 2020, doi: 10.1002/9781118150238.ch8.
10.1109/LCOMM.2019.2949557. [62] G. C. McDonald, Ridge Regression. 2011, pp. 173–188, doi:
[41] X. Ding, J. Guo, D. Li, and W. Wu, ‘‘An incentive mechanism for building a 10.1002/9781118150238.ch8.
secure blockchain-based Internet of Things,’’ IEEE Trans. Netw. Sci. Eng., [63] R. Yin, Y. Liu, W. Wang, and D. Meng, ‘‘Sketch kernel ridge regres-
vol. 8, no. 1, pp. 477–487, Jan. 2021. sion using circulant matrix: Algorithm and theory,’’ IEEE Trans. Neu-
[42] J. Huang, L. Kong, G. Chen, M.-Y. Wu, X. Liu, and P. Zeng, ‘‘Towards ral Netw. Learn. Syst., vol. 31, no. 9, pp. 3512–3524, Sep. 2020, doi:
secure industrial IoT: Blockchain system with credit-based consensus 10.1109/TNNLS.2019.2944959.
mechanism,’’ IEEE Trans. Ind. Informat., vol. 15, no. 6, pp. 3680–3689, [64] I. Almomani, B. Al-Kasasbeh, and M. Al-Akhras, ‘‘WSN-DS: A dataset
Jun. 2019. for intrusion detection systems in wireless sensor networks,’’ J. Sensors,
[43] P. K. Sharma and J. H. Park, ‘‘Blockchain based hybrid network architec- vol. 2016, pp. 1–16, 2016, doi: 10.1155/2016/4731953.
ture for the smart city,’’ Future Gener. Comput. Syst. vol. 86, pp. 650–655, [65] C. Dehury, S. N. Srirama, P. K. Donta, and S. Dustdar, ‘‘Securing clustered
Sep. 2018, doi: 10.1016/j.future.2018.04.060. edge intelligence with blockchain,’’ IEEE Consum. Electron. Mag., early
[44] K. Haseeb, N. Islam, A. Almogren, and I. U. Din, ‘‘Intrusion pre- access, Apr. 4, 2022, doi: 10.1109/MCE.2022.3164529.
vention framework for secure routing in WSN-based mobile Inter- [66] D. Rosadi, D. Arisanty, W. Andriyani, S. Peiris, D. Agustina, D. Dowe, and
net of Things,’’ IEEE Access, vol. 7, pp. 185496–185505, 2019, doi: Z. Fang, ‘‘Improving machine learning prediction of peatlands fire occur-
10.1109/ACCESS.2019.2960633. rence for unbalanced data using SMOTE approach,’’ in Proc. Int. Conf.
[45] S. Kushch and F. Prieto-Castrillo, ‘‘A rolling blockchain for a dynamic Data Sci., Artif. Intell., Bus. Anal. (DATABIA), Nov. 2021, pp. 160–163.
WSNs in a smart city,’’ 2018, arXiv:1806.11399. [67] D. K. Sah, K. Cengiz, P. K. Donta, V. N. Inikollu, and T. Amgoth, ‘‘EDGF:
[46] M. Liu, F. R. Yu, Y. Teng, V. C. M. Leung, and M. Song, ‘‘Computa- Empirical dataset generation framework for wireless sensor networks,’’
tion offloading and content caching in wireless blockchain networks with Comput. Commun. J., vol. 180, pp. 48–56, Dec. 2021.
mobile edge computing,’’ IEEE Trans. Veh. Technol., vol. 67, no. 11, [68] P. K. Donta and S. Dustdar, ‘‘The promising role of representation learning
pp. 11008–11021, Nov. 2018, doi: 10.1109/TVT.2018.2866365. for distributed computing continuum systems,’’ in Proc. IEEE Int. Conf.
[47] M. H. Kumar, V. M. Y. Suresh, and J. S. G. Nagalalli, ‘‘RETRACTED Service-Oriented Syst. Eng. (SOSE), Aug. 2022, pp. 126–132.
ARTICLE: Trust aware localized routing and class based dynamic block
chain encryption scheme for improved security in WSN,’’ J. Ambient
Intell. Humanized Comput., vol. 12, pp. 5287–5295, Apr. 2020, doi:
10.1007/s12652-020-02007-w.
[48] R. Goyat, G. Kumar, M. K. Rai, R. Saha, R. Thomas, and T. H. Kim, MUHAMMAD NOUMAN received the bache-
‘‘Blockchain powered secure range-free localization in wireless sensor lor’s degree in computer science from the Federal
networks,’’ Arabian J. Sci. Eng., vol. 45, no. 8, pp. 6139–6155, Aug. 2020, Urdu University of Arts, Science and Technol-
doi: 10.1007/s13369-020-04493-8.
[49] Y. Chang, X. Yuan, B. Li, D. Niyato, and N. Al-Dhahir, ‘‘Machine-
ogy (FUUAST), Islamabad, Pakistan, in 2019.
learning-based parallel genetic algorithms for multi-objective optimization He is currently pursuing the master’s degree in
in ultra-reliable low-latency WSNs,’’ IEEE Access, vol. 7, pp. 4913–4926, computer science with the Communication Over
2019. Sensors (ComSens) Research Laboratory, Depart-
[50] B. Schölkopf, Z. Luo, and V. Vovk, Empirical Inference: Festschrift in ment of Computer Science, COMSATS Univer-
Honor of Vladimir N. Vapnik. New York, NY, USA: Springer, 2013, sity Islamabad, Pakistan, under the supervision of
pp. 1–287, doi: 10.1007/978-3-642-41136-6. Dr. Nadeem Javaid. His research interests include
[51] S. Chaudhury. Tuning of Adaboost With Computational Complexity. computer networks, wireless sensor networks, blockchain, and machine
Medium.com. Accessed: Jan. 17, 2023. [Online]. Available: learning.
https://medium.com/@chaudhurysrijani/tuning-of-adaboost-with-
computational-complexity-8727d01a9d20
[52] A. Natekin and A. Knoll, ‘‘Gradient boosting machines, a tutorial,’’ Fron-
tiers Neurorobot., vol. 7, p. 21, Dec. 2013, doi: 10.3389/fnbot.2013.00021.
[53] Prashant. Computational Complexity of Machine Learning Algorithm. UMAR QASIM received the B.S. degree in com-
Medium.com. Accessed: Jan. 17, 2023. [Online]. Available: puter science and the M.B.A. degree from Ham-
https://medium.com/analytics-vidhya/computational-complexity-of-
dard University and the M.S. and Ph.D. degrees
ml-algorithms-1bdc88af1c7a
[54] A. Tharwat, T. Gaber, A. Ibrahim, and A. E. Hassanien, ‘‘Linear dis- in information systems from the New Jersey Insti-
criminant analysis: A detailed tutorial,’’ AI Commun., vol. 30, no. 2, tute of Technology, USA. He has been in the IT
pp. 169–190, 2017, doi: 10.3233/AIC-170729. field for over 20 years and has taught at vari-
[55] M. Krause. Computational Complexity of Linear Discrimination Analysis. ous post-secondary institutions in North America
Stackexchange.com. Accessed: Jan. 17, 2023. [Online]. Available: and Pakistan. He has extensive experience in the
https://stats.stackexchange.com/questions/142565/computational- field of information technology. He worked at Dal-
complexity-for-linear-discriminant-analysis housie University, McMaster University, and vari-
[56] D.-K. Choi, ‘‘Data-driven materials modeling with XGBoost algorithm and ous other software development companies in USA, for more than ten years.
statistical inference analysis for prediction of fatigue strength of steels,’’ He headed the Digital Preservation Program with the University of Alberta
Int. J. Precis. Eng. Manuf., vol. 20, no. 1, pp. 129–138, Jan. 2019, doi: for eight years and was responsible for internal operations and external
10.1007/s12541-019-00048-6.
preservation partnerships. He is currently working as an Associate Professor
[57] M. Virgolin. Time Complexity for Different Machine Learning Algorithm.
Marcovirgolin.github.io. Accessed: Jan. 17, 2023. [Online]. Available: with the University of Engineering and Technology (UET) Lahore, where he
https://marcovirgolin.github.io/extras/details_time_complexity_machine is involved in teaching and research and serving on various committees both
_learning_algorithms at the department and university level. He maintained and shared expertise on
[58] S. Premanand. Histogram Boosting Gradient Classifier. digital preservation with the university community and the professional com-
Analyticsvidhya.com. Accessed: Jan. 17, 2023. [Online]. Available: munity of practice at large. His research interests include the various areas of
https://www.analyticsvidhya.com/blog/2022/01/histogram-boosting- information systems, including but not limited to wireless sensor networks,
gradient-classifier databases, digital preservation, data science, and information management.
6120 VOLUME 11, 2023
M. Nouman et al.: Malicious Node Detection Using ML and Distributed Data Storage Using Blockchain in WSNs
HINA NASIR received the bachelor’s degree in MUHAMMAD IMRAN (Member, IEEE) worked
information and communication systems engi- with King Saud University (KSU), Saudi Arabia,
neering from the NUST School of Electrical Engi- as an Associate Professor. He is currently work-
neering and Computer Science (SEECS), in 2008, ing as a Senior Lecturer with the School of
the M.S. degree in computer science from Interna- Science, Engineering and Information Technol-
tional Islamic University (IIUI), in 2015, under the ogy, Federation University, Australia. He is also
supervision of Dr. Nadeem Javaid, and the Ph.D. the Founding Leader of the WIreless Networks
degree in computer science under the supervision and Security (WINS) Research Group, KSU,
of Dr. Nadeem Javaid. She worked as an Assistant from 2013 to 2021. His research interests include
Professor with the Department of Computer Sci- mobile and wireless networks, the Internet of
ence, Air University, Islamabad. Currently, she is doing postdoctoral research Things, big data analytics, cloud/edge computing, and information security.
at Leeds University, U.K. Her research interests include wireless sensor His research is financially supported by several national and international
networks, underwater wireless sensor networks, cooperative communication, grants. He has completed a number of international collaborative research
cooperative routing, buffer-aided cooperative communication, energy har- projects with reputable universities. He has published more than 300 research
vesting in wireless networks, 5G networks, and the Internet of Things. articles in peer-reviewed, highly-reputable international conferences (90),
journals (198), editorials (15), book chapter (1), and two edited books. Many
of his research articles are among the highly cited and most downloaded.
His research has been cited more than 11,500 with H-index of 55 and i-10
index of 175 (Google Scholar). He has received a number of awards and
fellowships.
He served as the Editor-in-Chief for European Alliance for Innovation
(EAI) Transactions on Pervasive Health and Technology and an Associate
Editor for IEEE Communications Magazine. He is serving as an Associate
Editor for top ranked international journals, such as IEEE Network, Future
Generation Computer Systems, and IEEE ACCESS. He served/serving as
a Guest Editor for about two dozen special issues in journals, such as
IEEE Communications Magazine, IEEE Wireless Communications Maga-
zine, Future Generation Computer Systems, IEEE ACCESS, and Computer
Networks. He has been involved in about 100 peer-reviewed international
conferences and workshops in various capacities, such as the chair, the
co-chair, and a technical program committee member. He has been consec-
utively awarded with an Outstanding Associate Editor of IEEE ACCESS, in
2018 and 2019, besides many others.