Research Article International Journal of Advanced Computer Research, Vol 13(65)

ISSN (Print): 2249-7277 ISSN (Online): 2277-7970

[Link]
Design of advanced intrusion detection system for multimodal fuzzy-based
convolutional neural networks (mFCNN)
K. Suresh Kumar1, T. Ananth Kumar2* and R. Nishanth3
Assistant Professor, Information Technology, Sri Krishna College of Technology, Coimbatore, India1
Associate Professor, Computer Science and Engineering, IFET College of Engineering, Villupuram, Tamil Nadu,
India2
Assistant Professor, Department of Electronics and Communication Engineering, Cochin University College of
Engineering Kuttanad, Pullinkunnu, Kerala, India3

Received: 20-June-2023; Revised: 11-November-2023; Accepted: 15-November-2023

©2023 K. Suresh Kumar et al. This is an open access article distributed under the Creative Commons Attribution (CC BY)
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.

Abstract
The Internet of Things (IoT) develops a smart autonomous system capable of controlling and facilitating numerous
human intervention activities such as inventory, electricity, traffic, and health management. It primarily operates by
interconnecting networks and objects in various locations. Communication is carried out between devices to accomplish
diverse missions in industrial, household, and healthcare applications. The involvement of the IoT in these applications
renders the devices susceptible to cyber-attacks. In this work, an intelligent methodology was proposed to defend the
devices against various security-related threats using modified deep learning algorithms. The intrusion detection system
was designed using a hybrid model comprising multimodal fuzzy-based convolutional neural networks (mFCNN) in
association with bidirectional long-short term memory (Bi-LSTM) and the enhanced code element embedding (ECEE)
method. Here, the input for training is derived from real images, videos, and audio information retrieved from various
sensors. Initially, the ECEE in each input frame is converted into suitable vectors. These vectors are then mapped to
corresponding fixed-length vectors for embedding to achieve the most compressed representation. The Bi-LSTM is
utilized for extracting relevant information related to spatial features, thus providing an effective intrusion detection
mechanism. The mFCNN extracts the most critical temporal features by classifying contextual inputs using the captured
videos and images. The experimental results demonstrate that the proposed hybrid model yields better results compared to
existing methodologies, with the accuracy of protection benchmarked at 99.91%, which is higher than other baseline
methods.

Keywords
Internet of Things (IoT), Cybersecurity, Deep learning algorithms, Intrusion detection systems, Hybrid neural networks.

[Link] Despite their advancement, new IoT networks are

The Internet of Things (IoT) is a rapidly expanding vulnerable to a wide range of security threats and
global network that connects both living and vulnerabilities due to the inefficient processing power
nonliving entities. A security flaw within IoT renders of these devices [4]. Due to technological
it susceptible to a wide array of cyberattacks [1]. advancements, IDS are now extremely effective
Users might be reluctant to embrace new technology network security tools. There has been a significant
if the network is perceived as insecure. While IoT is a increase in IoT device usage in recent years. Devices
remarkable concept, its global network of connected with IoT capabilities can connect to the internet and
smart objects makes it particularly prone to transmit data [5], accessible and controllable from
cyberattacks [2]. Intrusion detection systems (IDS) any location due to the internet. IoT is a network
can enhance the security of any network connecting physical and digital devices, not just a
environment. Most modern IDSs employ machine computer network [6]. It has application domains in
learning (ML) algorithms to train and detect network industry, home automation, healthcare, and other
attacks [3]. innovative technologies. ML, particularly deep
learning (DL), is gaining popularity in various
scientific fields, including IoT [7]. DL approaches are
automatically used in IoT security for detecting
*Author for correspondence network intrusions and malware. For example, the
94
 International Journal of Advanced Computer Research, Vol 13(65)

automatic classification of fish based on inherent [Link] works

characteristics is well-known [7]. Contrary to popular The use of ML algorithms for attack detection was
belief, DL-based IoT analysis can be easily proposed by [14], highlighting the need for edge
circumvented; hostile environments are examined for processing in large IoT networks due to the
malicious tasks [8]. Malware classifiers for PDF files limitations of cloud computing. A novel approach
can be deceived into marking malicious files as using a support vector machine (SVM) and a self-
benign. Currently, DL techniques are not secure in taught DL autoencoder for intrusion detection was
IoT environments [9]. However, DL has far-reaching proposed by [15], which improves SVM classifier
implications, which is intriguing. Information performance while saving time and costs. In [16]
retrieval has gained public attention due to its wide contrasts shallow and deep neural networks,
range of applications [10]. While DL-based retrieval suggesting DL's potential to enhance cybersecurity,
of images and text is well-studied, retrieving other with a study [17] showing a three-layer deep neural
formats like graphs is more challenging. The deep network model's superiority. However, the
relevance matching model (DRMM) outperforms application of ML or DL in cybersecurity is
traditional models by addressing key relevance challenging, with many zero-day attacks being
matching factors [11] and can be built on top of long- discovered in IoT due to various protocols used by
term memory. DL-based data filtering systems could devices [18]. A new intrusion detection approach was
become more effective as they start with a predefined developed, employing the intrusion label and decoder
measure of similarity derived from data layer alongside common vulnerabilities and
measurements, but these metrics may not always be exposures (CVE) to prevent unauthorized access
appropriate. DL can help analysts in resolving [18]. Another method [19] predicts security breaches
complex situations. IoT's significance continues to using the Nash equilibrium. He et al. [20] studied
grow, especially for diverse and open mobile IoT- vehicle networks using the two-Rayleigh model,
enabled health networks, which makes large amounts while Feng et al. [21] developed analytical
of healthcare data vulnerable [12]. Healthcare expressions for amplify and forward (AF) relaying
organizations may face network and security issues. networks.
Xu et al. explored cooperative network security using
Weibull fading channels [13], and the 2-Rayleigh In physical layer security, the 2-Rayleigh and
model accurately simulates vehicle-created Nakagami-m models are under scrutiny. Traditional
propagation environments. communications use these models, but mobile IoT
communication is more dynamic and complex [22].
The objective of the paper is to enhance the stability The antenna for mobile vehicular communication
and efficiency of intrusion detection by employing a moves constantly, resulting in multiple scattering
hybrid model that combines multimodal fuzzy-based groups, which classic fading channels like the
convolutional neural networks (mFCNN) with Nakagami-m and 2-Rayleigh [23] do not account for.
bidirectional long-short term memory (Bi-LSTM) Tegos et al. [24] adapted the Nakagami model for
and the enhanced code element embedding (ECEE) mobile environments to assess the performance of
method. This approach aims to classify network mobile secrecy systems. Environmental changes must
traffic, extract critical features, and analyze the be promptly reported for secure device
relationships between data packets to effectively communication. The healthcare sector, particularly in
detect and prevent a wide range of cyber threats. IoT, has seen increased AI application [25]. Kumar et
Additionally, the paper seeks to address the al. [26] employed general regression methods,
challenges in IoT security by proposing a robust, Alhussein et al. [27] proposed an emotion prediction
deep learning-based solution designed to handle system from healthcare big data, Samuel et al. [28]
complex and dynamic network environments. suggested back-propagation networks for hospital
inventory, and Cao et al. [29] discussed using
The remaining text is structured as follows: Section 2 convolutional neural networks (CNNs) for image
discusses previous research. Section 3 describes the processing. Khan et al. [30] compared ML techniques
datasets and pre-processing methods used to create for intrusion detection, Alqahtani et al. [31]
them. Section 4 presents the evaluation results with a developed the GXGBoost model using genetic
variety of hyperparameters and discusses the algorithms and the XGBoost classifier, and Derhab et
implications of our findings. Finally, it is concluded al. [32] proposed a method to prevent command
in Section 5. forgery using blockchain and software-defined
networks (SDN). DL systems, which are widely used
95
K. Suresh Kumar et al.

in IDS [32], do not require manual feature extraction, are continuously proposed to address these evolving
increasing precision in threat detection. However, security concerns.
DL-based intrusion detection requires extensive
training data, which is challenging due to the scarcity [Link]
of attack traffic in real-world datasets like “KDD99”, 3.1Multimodal fuzzy-based convolutional neural
“NSL-KDD”, and “CICIDS2017” and the difficulty networks(mFCNN)
in capturing and simulating some attack traffic, This section analyzed a novel method for defending
making it hard to assess specific attacks in detail devices against various security-related attacks. The
using DL-based methods. While DL reduces overall mFCNN, an improved method for learning and
training time with large datasets, it does not require classifying traffic packets considering both the
prior feature selection for learning and testing temporal passage and physical location, was
classification rules. introduced. Consequently, class weight was utilized
to enhance the model's stability. The overall
The review indicates a focus on enhancing architecture of the proposed model was depicted in
cybersecurity in IoT using ML and DL, despite Figure 1.
challenges like zero-day attacks and the need for
extensive training data. Novel approaches and models

Figure 1 Architecture of proposed system

The classifier used here is the combination of sequences that correspond to the system's
mFCNN in association with Bi-LSTM and the ECEE membership in the various classes. The SoftMax
method. The mFCNN part consists of the input layer layer takes the most stringent classification as its
and the enhanced code element embedding layer. starting point and determines the final output from
Totally n number of convolution layer is used along there. A new estimate of the probability is produced
with the pooling layer and fully connected layer. as an output of the process.
After the content of the information that has been
preprocessed has been received, the mFCNN section The data packets, initially converted into two-
will begin to process it, and the processing will result dimensional matrices, undergo preprocessing to
in the production of a high-dimensional vector generate images. These images result from the
representing the BiLSTM section. The BiLSTM amalgamation of various bit values found in the
section consists of an output layer, a SoftMax layer, respective network traffic locations. The electronic
and three layers of the BiLSTM algorithm itself and data and other crucial information in network traffic
the output and SoftMax layers. A processing packets consist of bytes, the fundamental units of this
algorithm is applied to collect high-dimensional data. These bytes are valued on a scale ranging
vectors in the proposed model. The algorithm then between 0 and 255. This is just like the image bytes.
returns the results in the form of a probability of Consider the byte in the packet header ∝ and payload
96
 International Journal of Advanced Computer Research, Vol 13(65)

byte as β. These ∝ and β is combined with the traffic ( ) [∑ ( ) ∑ (

images for further processing. The mFCNN consists
) ] (4)
of input, convolution layers, pooling layer, fully
connected layer, SoftMax layer, and the output layer.
The combination of these layers allows the input 3.2Feature extraction using Bi LSTM
layer from various aspects. Totally n number of sets Both regular network connections and network
are there in the setup. Of these, the first combination attacks are conducted using a specific network
is used for extracting the local features of the traffic protocol. The packets that are affected by the attacks
images. The exact features and the stability-filled are encountered in the packets that are alongside the
results are obtained in the pooling layer. Then traffic. When considering the network protocol, it is
another set of combinations might yield a very large contained in the fixed regions as the normal
convolution kernel for analysing the similarities in connections are established by the key exchanges
between the two bytes that are not near which is same incorporating the connections and the other
as that of the information spread in the traffic disconnections. Normally, traffic attacks are used to
payload. After pre-processing, the network traffic determine the main cause of the attacks. This leads to
might establish the vector input in the input layer. data that is difficult to label, reveals an excessive
Then the input layer might set up with the data length amount of unwanted traffic, and ultimately produces
that is intercepted from the corresponding ith packet poor training results when using a single mFCNN to
i.e., ( ) which is followed by the train the properties of a single packet as the
combination C of m packets related information set foundation of the algorithm. This is because the data
( ). The convolution layer with side is difficult to label and reveals an excessive amount
length Sl is mathematically expressed as Equation 1. of unwanted traffic. This is due to the fact that the
algorithm relies on a single packet as its primary
( ) ∑ ∑ [ (
building block. Additionally, this exposes an
) ( )] ∑ ( ) (1) excessive amount of unwanted traffic. This could be
reduced by introducing the BiLSTM, which is
In addition to that the value of (a,b) is obtained as intended to take the information containing the single
Equation 2. connection or as a group that determines all the data
( ) * + (( ) packets in the mentioned groups and for determining
) (2) the basic nature of either the traffic details or the
packet information‟s. Natural language processing
Where Ck is the plotted channel count in the outperforms the processing scenario of the traffic or
characteristics graph, V denotes the output value in other network issues by using the same proposed
consideration with the activation function, the layer methodology.
that is used for convolution may consist of an
activation function that has an expression that has a The BiLSTM region contains the various layers used
lot of complex features due to the fact that V for classification purpose. It contains n number of
represents the output value in consideration with the convolutional layers in addition to the other fully
activation function. The vector representing the connected layers along with the SoftMax layers and
output is shown in the Equation 3 as, the output layers. Two BiLSTM layers carry out the
( ) ( ) (3) major functions. The BiLSTM is a unique neural
networks recurrent neural networks (RNN) that was
The convolutional layer is the one that is responsible created to overcome gradients disappearing in
for the extraction of the features, and after the addition with gradient explosion issues during
features have been extracted, the output image extended sequential training. Generally, the recurrent
undergoes a complete transformation so that it can be RNN might have the tanh layer but the other
sent to the max-pooling section, where a feature can BiLSTM networks will sequentially outperform the
be selected and filtered. situation by predicting the better performance time in
consideration of the memorable gates. Here the
The pooling section consists of an already set in BiLSTM node of a cell is represented as Nc, then the
pooling function which can replace the outcome in a input and output are expressed as x(t) and y(t). The
separate section corresponding to the adjacent region. initial step in the BiLSTM layer is to determine the
The pooling layer is computed using the model-related information that can discard the cell's
mathematical expression (Equation 4). state. The initial decision is made with the gate
forgetting. The gate terminal reads the input and the
97
K. Suresh Kumar et al.

outputs by considering the values between 0 and 1. 3.3Class weights

Each number in the Nc cell must contain a stable The data obtained after the pre-processing mode is
state by which „1‟ corresponds to the „completely shown in Table 1.
retained‟ and „0‟ corresponds to the „completely
discarded‟. Wt represents the weights and the bias in Table 1 Dataset related metrices
the neural network is represented by β. It is Label Types of attacks Number
mathematically expressed as Equation 5. 1 Physical Attacks 67432
( , - ) (5) 2 Encryption Attacks 45662
3 Denial of Service (DoS) 456
4 Firmware Hijacking 6532
The other step is for deciding the range of novel
5 Botnets 7654
information for adding the state of the cell. Initially 6 Man-in-the-Middle 9823
the sigmoid layer is used for determining the needed 7 Ransomware 4563
information that needs to be modified using Equation 8 Eavesdropping 784
6. 9 Privilege Escalation 8943
( , - ) (6) 10 Brute Force Password Attack 9045

The alternating vector for updating the security status Here the numbers of various types of data are
is generated by using the tanh layer. This is selected unevenly. When compared to the other
mathematically expressed using the Equation 7. types, the number of type '0' is the highest, while the
( , - ) (7) other types are ranked as the lowest. This may affect
the outcome of the classification section. The class
The two regions are then combined using the cell weight is calculated using the Equation 11.
state by updating the formula (Equation 8). ∑
̄ (11)
(8)

The gate at the output might determine the output [Link] and discussion
prediction of the cell (Equation 9). The dataset utilized for the experimentation was
( , - ) (9) sourced from the network intrusion detection section
on the
Then the cell state is processed through the tanh [Link]
function for obtaining the formula (Equation 10). rk-intrusion-detection website. It comprises specific
( ) (10) parameters relevant to the study. It normally specifies
the influence of the size of data packets in the process
The proposed model, which includes the feature of training, the impact of the number of packets per
maps, as each data packet containing a collection of flow, and the effect of the sample size chosen in
traffic images generated in conjunction with the addition to the total effect of the units used for
BiLSTM section. The relationships that exist between representing the BiLSTM by influencing the weight
the features are structured in the spaces that exist classes. The proposed model-related parameters are
between the 'd' data packets, which are then analysed optimized in such a way that the mFCNN is
using the BiLSTM layers. These initial few packets compared with the other neural network models
are also used. The following available packets could along with the BiLSTM. The ratio of the datasets
have very long payloads, which would include the used for training, testing, and validation is made in
data related to the attack. the ratio of 3:6:16. The parameters used for the
analysis are the accuracy, precision, recall,
The proposed system's BiLSTM layer may sensitivity, and the f score value. The overall
incorporate an activation function designed to reduce performance of the proposed model is specified in
training time. Typically, the rectified linear unit terms of accuracy, the precision represents the ratio
(ReLU) function triggers the BiLSTM's activation. of the positive samples in consideration with the total
The standard process involves a system used for samples and the negative sample consideration with
multi-classification, where the model is trained using the total samples is specified in terms of sensitivity.
multi-class cross-entropy. The recall values specify the number of true positive
samples divided in terms of true positives and false
negatives. The classifier‟s accuracy is defined with

98
 International Journal of Advanced Computer Research, Vol 13(65)

the harmonic weights which specifies the F score retraining with excessively long data packets
value. increases the proportion of smaller packets, thereby
raising the fraction of units with longer lengths and
4.1Experimentation reducing the model's accuracy. The median value of 1
The experiment was conducted on a Linux operating further lowers the calculation model's precision.
system using an Intel Core i5 processor, equipped
with a 256 GB SSD. The setup included Anaconda The study considers various attacks, including
4.5.10, Keras 2.2.20, and Python 3.9. Physical Attacks, Encryption Attacks, DoS,
Firmware Hijacking, Botnets, Man-in-the-Middle,
4.2Performance of the model based on BiLSTM Ransomware, Eavesdropping, Privilege Escalation,
In the framework of the BiLSTM, the value of the and Brute Force Password attacks. Table 2 presents
system's output parameter is modelled as a series of these types of attacks along with their corresponding
units. During the research, it was concluded that values, considering specified parameters like
increasing the number of BiLSTM units initially accuracy, precision, recall, and f-score.
improves the model's performance before it begins to
deteriorate. After careful consideration, the optimal It reflects the predicted changes in values relative to
number of BiLSTM units was determined to be 121. the increase in packet length during training. When
this length exceeds a maximum threshold, the
4.3Impact of Model performance in specifying the effectiveness of the data packets is ascertained by
packet length of the training packets evaluating their scientific credibility incorporated
The training results indicate an increase in the into the training sectors. Figure 2 demonstrates the
model's quality, which starts to decline when the impact on model performance as processed with
package length exceeds 70. This is likely because varying training packet lengths.

Table 2 Values for various types of attacks

Types of attacks Accuracy Precision Recall Sensitivity F Score
Physical Attacks 99.97 99.31 0.12 90.99 98.85
Encryption Attacks 99.98 99.32 0.14 91 98.86
DoS 99.96 99.3 0.31 90.98 98.84
Firmware Hijacking 100 99.34 0.22 91.02 98.88
Botnets 99.68 99.02 0.62 90.7 98.56
Man-in-the-Middle 99.78 99.12 0.01 90.8 98.66
Ransomware 99.98 99.32 0.05 91 98.86
Eavesdropping 99.91 99.25 0.08 90.93 98.79
Privilege Escalation 99.92 99.26 0.21 90.94 98.8
Brute Force Password Attack 99.93 99.27 0.36 90.95 98.81
Overall 99.911 99.251 0.212 90.931 98.791

Figure 2 Determining the Proposed model performance with respect to packet length

99
K. Suresh Kumar et al.

This also mitigates the effects of overfitting, thereby packets increases, consequently decreasing the
ensuring the most effective accuracy setting for the model's ability to extract information. The influence
classification and determining the accuracy in of the total number of packets in a flow on the results
identifying data packets along with their packet of a simulation, as demonstrated in Figure 3, shows
headers. that the model's performance significantly
deteriorates when the number of packets per flow
4.4Performance of the proposed model in the exceeds 12.
determination of the packet flow quantity Performance of the model in the effect of the
There exists a possibility that increasing the specified batch size
number of data packets in each flow involved in the Batch size is an important criterion to consider during
training process enhances the specificity of feature training models. Increasing the batch size while
extraction. However, setting this value too high may staying within a reasonable range can help improve
lead to a rise in the percentage of filled data packets. memory utilization and lead to a speedier processing
An increase in the specified number of data packets of data. However, if it is raised to an excessive range,
in each flow is thought to make the feature extraction it might severely slow down the process. The total
more specific. Nevertheless, if this value is size of the batch is proven to be the performance
excessively high, the percentage of already filled data effect in the proposed models, as shown in Figure 4.

Figure 3 Proposed model performance with respect to packet per-flow

Figure 4 Performance of the model in the effect of the batch size

4.5Performance model for determining the effects dataset's unbalance in the quantity of data from
of class weights various categories on simulation results. Figure 5
In Table 3, the results of two distinct groups of illustrates the performance model for determining the
experiments, one with class weights and one without, effects of class weights in consideration of the
are compared. The introduction of the class weight available dataset in determining the quantity of the
appears to lessen the influence of the existing data.
100
 International Journal of Advanced Computer Research, Vol 13(65)

Table 3 Performance model for determining the effects of Class Weights

Parameters With class weights (in%) Without class weights (in%)
Accuracy 99.25 97.45
Precision 96.35 96.25
Recall 0.87 0.45
Sensitivity 94.25 95.87
F Score 97.12 96.32

Figure 5 Performance model for determining the effects of class weights

The BiLSTM is used effectively for extracting the efficiency identification with more traffic attacks.
sequential connection between the packets. Table 4 The proposed mFCNN in combination with the
compares accuracy and other parameters obtained BiLSTM detects the intrusion more effectively with a
between the proposed models and other models in very less alarm rate and thus can be classified by
consideration of the types of attacks. The BiLSTM setting the parameters of network traffic more
appears to be improved effectively to prove the accurately when compared to the mFCNN or
efficiency of identification in consideration of the BiLSTM. The comparative results prove that the
parameters by specifying the time duration of the proposed model provides better performance when
attacks. When compared to BiLSTM model, the compared to the other existing methods. A complete
addition of the mFCNN will improve further for list of abbreviations is summarised in Appendix I.

Table 4 Comparison of the proposed model with other models

Methods Accuracy Precision Recall Sensitivity F Score
CNN 94.65 95.64 0.53 92.78 94.36
K-nearest neighbor algorithm (KNN) 93.65 94.64 0.62 91.78 93.36
LSTM 94.63 95.62 0.34 92.76 94.34
RNN 92.56 93.55 0.85 90.69 92.27
SVM 91.35 92.34 0.62 89.48 91.06
CNN+LSTM 96.62 97.61 0.12 94.75 96.33
RNN+LSTM 95.78 96.77 0.64 93.91 95.49
Proposed 99.911 99.251 0.212 90.931 98.791

[Link] designed using a hybrid model comprising mFCNN

DL algorithms address security threats in IoT in association with Bi-LSTM and the ECEE method.
environments. In this work, an intelligent Bi-LSTM is utilized to extract relevant spatial feature
methodology was proposed to defend devices against information, providing an effective intrusion
various security threats using modified DL detection mechanism. Meanwhile, mFCNN classifies
algorithms. The intrusion detection system is contextual inputs, including captured videos and
images, to extract essential temporal features. The
101
K. Suresh Kumar et al.

experimental results demonstrate that the proposed [11] Hao S, Shi C, Cao L, Niu Z, Guo P. Learning deep
hybrid model outperforms existing methodologies, relevance couplings for ad-hoc document retrieval.
achieving a benchmarked accuracy of 99.91% for Expert Systems with Applications. 2021; 183:115335.
protection, which is higher compared to other [12] Otoum Y, Liu D, Nayak A. DL‐IDS: a deep learning–
based intrusion detection framework for securing IoT.
baseline methods.
Transactions on Emerging Telecommunications
Technologies. 2022; 33(3):e3803.
Acknowledgment [13] Xu L, Zhou X, Tao Y, Liu L, Yu X, Kumar N.
None. Intelligent security performance prediction for IoT-
enabled healthcare networks using an improved CNN.
Conflicts of interest IEEE Transactions on Industrial Informatics. 2021;
The authors have no conflicts of interest to declare. 18(3):2063-74.
[14] Elsayed MS, Le-Khac NA, Dev S, Jurcut AD.
References Machine-learning techniques for detecting attacks in
[1] Kuzlu M, Fair C, Guler O. Role of artificial SDN. In 7th international conference on computer
intelligence in the internet of things (IoT) science and network technology 2019 (pp. 277-81).
cybersecurity. Discover Internet of Things. 2021; 1:1- IEEE.
4. [15] Binbusayyis A, Vaiyapuri T. Unsupervised deep
[2] Rahman SA, Tout H, Talhi C, Mourad A. Internet of learning approach for network intrusion detection
things intrusion detection: Centralized, on-device, or combining convolutional autoencoder and one-class
federated learning? IEEE Network. 2020; 34(6):310-7. SVM. Applied Intelligence. 2021; 51(10):7094-108.
[3] Dang QV. Studying machine learning techniques for [16] Kumar TA, Rajmohan R, Adithya M, Sunder R. A
intrusion detection systems. In future data and security novel security scheme using deep learning based low
engineering: 6th international conference, Nha Trang overhead localised flooding algorithm for wireless
City, Vietnam, proceedings,2019 (pp. 411-26). sensor networks. International Journal of Data
Springer International Publishing. Science. 2021; 6(1):19-32.
[4] Rajmohan R, Kumar TA, Julie EG, Robinson YH, [17] Sharma R, Vashisht V, Singh U. Performance analysis
Vimal S, Kadry S, et al. G-Sep:a deep learning of evolutionary technique based partitional clustering
algorithm for detection of long-term sepsis using algorithms for wireless sensor networks. In soft
bidirectional gated recurrent unit. International Journal computing: theories and applications: proceedings of
of Uncertainty, Fuzziness and Knowledge-Based SoCTA 2018 (pp. 171-80). Springer Singapore.
Systems. 2022; 30(Supp01):1-29. [18] Pontes CF, De Souza MM, Gondim JJ, Bishop M,
[5] Laghari AA, Wu K, Laghari RA, Ali M, Khan AA. A Marotta MA. A new method for flow-based network
review and state of art of Internet of Things (IoT). intrusion detection using the inverse Potts model.
Archives of Computational Methods in Engineering. IEEE Transactions on Network and Service
2021: 1-9. Management. 2021; 18(2):1125-36.
[6] Kumar KS, Mani AR, Sundaresan S, Kumar TA, [19] Martins I, Resende JS, Sousa PR, Silva S, Antunes L,
Robinson YH. Blockchain-based energy-efficient Gama J. Host-based IDS: a review and open issues of
smart green city in IoT environments. In blockchain an anomaly detection system in IoT. Future
for smart cities 2021 (pp. 81-103). Elsevier. Generation Computer Systems. 2022; 133:95-113.
[7] Magán-Carrión R, Urda D, Díaz-Cano I, Dorronsoro [20] He R, Schneider C, Ai B, Wang G, Zhong Z, Dupleich
B. Towards a reliable comparison and evaluation of DA, et al. Propagation channels of 5G millimeter-
network intrusion detection systems based on machine wave vehicle-to-vehicle communications: recent
learning approaches. Applied Sciences. 2020; advances and future challenges. IEEE Vehicular
10(5):1775. Technology Magazine. 2019; 15(1):16-26.
[8] Kumar A, Purohit V, Bharti V, Singh R, Singh SK. [21] Feng YH, Dai YJ, Wang RZ, Ge TS. Insights into
Medisecfed: Private and secure medical image desiccant-based internally-cooled dehumidification
classification in the presence of malicious clients. using porous sorbents: From a modeling viewpoint.
IEEE Transactions on Industrial Informatics. 2021; Applied Energy. 2022; 311:118732.
18(8):5648-57. [22] Zhou Z, Wang J, Ye Y. Exact BER analysis of
[9] Rajmohan R, Pavithra M, Kumar TA, Manjubala P. differential chaos shift keying communication system
Exploration of deep RNN architectures: LSTM and in fading channels. Wireless Personal
gru in medical diagnostics of cardiovascular and neuro Communications. 2010; 53:299-310.
diseases. In handbook of deep learning in biomedical [23] Angjelichinoski M, Trillingsgaard KF, Popovski P. A
engineering and health informatics 2021 (pp. 167- statistical learning approach to ultra-reliable low
202). Apple Academic Press. latency communication. IEEE Transactions on
[10] Sundaresan S, Kumar KS, Kumar TA, Ashok V, Julie Communications. 2019; 67(7):5153-66.
EG. Blockchain architecture for intelligent water [24] Tegos SA, Diamantoulakis PD, Karagiannidis GK. On
management system in smart cities. In blockchain for the performance of uplink rate-splitting multiple
smart cities 2021 (pp. 57-80). Elsevier.

102
 International Journal of Advanced Computer Research, Vol 13(65)

access. IEEE Communications Letters. 2022; T. Ananth Kumar received his

26(3):523-7. Bachelor's degree in Electronics and
[25] Arumugam D, Govindaraju K, Tamilarasan AK. Communication Engineering from
AIIoT-based smart framework for screening specific Anna University, Chennai, and his
learning disabilities. In machine learning for critical Master's degree in VLSI Design from
internet of medical things: applications and use cases Easwari Engineering College, affiliated
2022 (pp. 103-24). Cham: Springer International with Anna University. He earned his
Publishing. Ph.D. in VLSI Design from
[26] Kumar TA, John A, Kumar CR. 2. IoT technology and Manonmaniam Sundaranar University, Tirunelveli. He has
applications. Internet of Things. 2020:43-62. since become an Associate Professor at IFET College of
[27] Alhussein M, Muhammad G, Hossain MS, Amin SU. Engineering, India. He has presented papers at various
Cognitive IoT-cloud integration for smart healthcare: national and international conferences and journals. His
case study for epileptic seizure detection and fields of interest include Networks on Chips, Computer
monitoring. Mobile Networks and Applications. 2018; Architecture, and ASIC design. He is a recipient of the Best
23:1624-35. Paper Award at INCODS 2017 and holds several patents.
[28] Samuel TA, Pavithra M, Mohan RR. LIFI-based Email: tananthkumar@[Link]
radiation-free monitoring and transmission device for
hospitals/public places. In multimedia and sensory R. Nishanth serves as an Assistant
input for augmented, mixed, and virtual reality 2021 Professor in the Electronics and
(pp. 195-205). IGI Global. Communication Engineering
[29] Cao C, Liu X, Yang Y, Yu Y, Wang J, Wang Z, et al. Department at Cochin University
Look and think twice: Capturing top-down visual College of Engineering Kuttanad,
attention with feedback convolutional neural India. He obtained his Bachelor's
networks. In proceedings of the IEEE international degree in Electronics and
conference on computer vision 2015 (pp. 2956-64). Communication Engineering and his
[30] Khan FA, Gumaei A, Derhab A, Hussain A. A novel Master's degree in Applied Electronics from Anna
two-stage deep learning model for efficient network University, Chennai, Tamilnadu. Currently, he is pursuing
intrusion detection. IEEE Access. 2019; 7:30373-85. his Ph.D. in Information and Communication Engineering
[31] Alqahtani M, Gumaei A, Mathkour H, Maher Ben from Anna University, Chennai. With several years of
Ismail M. A genetic-based extreme gradient boosting experience in teaching and research, he has organized and
model for detecting intrusions in wireless sensor participated in numerous conferences, seminars,
networks. Sensors. 2019; 19(20):4383. workshops, and various other events. His research interests
[32] Derhab A, Guerroumi M, Gumaei A, Maglaras L, lie in VLSI Design and Image Processing.
Ferrag MA, Mukherjee M, et al. Blockchain and
random subspace learning-based IDS for SDN- Appendix I
enabled industrial IoT security. Sensors. 2019; S. No. Abbreviation Description
19(14):3119. 1 AF Amplify and Forward
2 AI Artificial Intelligence
K. Suresh Kumar is currently 3 Bi-LSTM Bidirectional Long-Short Term
pursuing his Ph.D. in Information and Memory
4 CNN Convolutional Neural Networks
Communication Engineering at Anna
5 CVE Common Vulnerabilities and
University, Chennai. He obtained his Exposures
Master's degree in Computer and 6 DL Deep Learning
Information Technology from 7 DoS Denial of Service
Manonmaniam Sundaranar University, 8 DRMM Deep Relevance Matching Model
Tirunelveli, during 2009-2011, and his 9 ECEE Enhanced Code Element
Bachelor's degree in Electronics and Communication Embedding
Engineering from Anna University, Chennai, during 2004- 10 IDS Intrusion Detection Systems
2008. Presently, he serves as an Assistant Professor at Sri 11 IoT Internet of Things
Krishna College of Technology, affiliated with Anna 12 KNN K-Nearest Neighbor Algorithm
University, Chennai. He has contributed papers to various 13 LSTM Long-Short Term Memory
national and international conferences and journals. His 14 mFCNN Multimodal Fuzzy-Based
areas of interest include Computer Networks, Data Convolutional Neural Networks
15 ML Machine Learning
Analytics, and Natural Language Processing. Additionally,
16 PDF Portable Document Format
he is a life member of ISTE and holds memberships in
17 ReLU Rectified Linear Unit
several other bodies. 18 RNN Recurrent Neural Networks
Email: sureshkumar.k@[Link] 19 SDN Software-Defined Networks
20 SVM Support Vector Machine

103