2024 Third International Conference on Distributed Computing and Electrical Circuits and Electronics (ICDCECE)

Employing Deep Neural Networks for Real-Time

Anomaly Detection and Mitigation in IoT-Based
Smart Grid Cybersecurity Systems
2024 Third International Conference on Distributed Computing and Electrical Circuits and Electronics (ICDCECE) | 979-8-3503-1860-9/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICDCECE60827.2024.10548160

1st Aadam Quraishi 2nd Maher Ali Rusho 3rd Anurag Prasad
M.D. Research Department of Lockheed Martin IIT Kharagpur (Rajiv Gandhi School
Intervention Treatment Institute Engineering Management Of Intellectual Property Law)
Houston Texas, US University of Colorado, Boulder Gurugram, India
[email protected] Boulder, Colorado [email protected]
[email protected]

4th Ismail Keshta 5th Richard Rivera 6th Mohammed Wasim Bhatt
Computer Science and Information Department of Informatics and Model Institute of Engineering and
Systems Department, College of Computer Science Technology
Applied Sciences Escuela Politécnica Nacional Jammu, J&K, India
AlMaarefa University Quito, Ecuador [email protected]
Riyadh, Saudi Arabia [email protected]
[email protected]

Abstract—This research introduces a novel anomaly learn intricate patterns and representations from vast datasets,
detection framework for IoT-based Smart Grid Cybersecurity present a promising avenue for enhancing the resilience of
Systems. Leveraging autoencoders, LSTM networks, GANs, smart grids against cyber threats [5]. This approach not only
SOMs, and transfer learning, our approach achieves superior leverages the power of artificial intelligence but also enables
precision, recall, and execution time compared to existing the system to adapt and evolve in real-time, staying ahead of
methods. Visualizations and an ablation study further validate potential attackers.
the method's efficiency, emphasizing the critical roles of
attention mechanisms and transfer learning. This To effectively safeguard IoT-based smart grids, a
comprehensive solution addresses the dynamic challenges of multifaceted approach is essential. The proposed solution
smart grid cybersecurity, offering a versatile and adaptive integrates advanced DNN architectures into the existing
anomaly detection mechanism for real-world applications. This cybersecurity framework, creating a symbiotic relationship
indicates the real-time efficacy of our anomaly detection that combines the strengths of both conventional methods and
method. Through our study of ablation and all aspects of cutting-edge technology [6]. The incorporation of anomaly
computing, we discovered that attention processes and transfer detection algorithms based on DNNs facilitates the
learning facilitate faster problem solving in a dynamic smart identification of subtle deviations from normal behavior,
grid. Our method is distinct and adaptable enough to address enabling the system to distinguish between genuine anomalies
every problem arising from the discovery of anomalies in IoT- and regular fluctuations [7]. Moreover, the mitigation
driven Smart Grid Cybersecurity Systems.
strategies are intelligently informed by the insights gleaned
Keywords— Anomaly Detection, Autoencoders, Cybersecurity, from the neural network's analysis, enhancing the system's
Flowchart, Generative Adversarial Networks (GANs), LSTM overall robustness. This research included these key elements:
Networks, Self-Organizing Maps (SOMs), Transfer Learning, • DNN-Based Anomaly Detection Model Development:
Training Algorithm, IoT-based Smart Grid This project aims to develop Deep Neural Network
I. INTRODUCTION models that can detect smart grid abnormalities [8].
These models employ cutting-edge frameworks to
In the dynamic landscape of modern smart grids powered identify and evaluate complex cyber threat patterns.
by the Internet of Things (IoT), ensuring robust cybersecurity
is paramount to maintain the reliability and integrity of critical Adaptive mitigation in real time Using DNN findings, the
infrastructure [1]. As the complexity and interconnectedness study finds dangers and examines real-time prevention [9].
of these systems grow, so does the vulnerability to cyber These tactics allow the smart grid's control and reaction
threats, necessitating advanced and adaptive anomaly systems to manage cyberattacks on their own, minimizing
detection and mitigation mechanisms [2]. This research damage and maintaining critical operations.
endeavours to address these challenges by harnessing the Real-world testing and approval: To ensure the intended
power of Deep Neural Networks (DNNs) for real-time solution works, real-world smart grid testing and approvals are
cybersecurity in IoT-based smart grids. We can now produce done [10]. This includes testing the system against different
and exchange data like never before thanks to smart grid cyber dangers, seeing how effectively it performs in diverse
technologies [3]. As cyberattacks get smarter and more conditions, and making sure it can be expanded to tackle
widespread, dangers rise despite improvements. Smart grid changing cyberattacks [11].
security developments make it even more vital to be careful
and find and solve issues quickly [4]. Researchers are This study presents a novel paradigm for anomaly
investigating DNNs because previous approaches can't keep detection in Internet of Things (IoT)-based smart grid
up with evolving dangers. At the heart of this research lies the cybersecurity systems. By utilising autoencoders, LSTM
utilization of Deep Neural Networks as a principal tool for networks, GANs, SOMs, and transfer learning, our method
anomaly detection and mitigation. DNNs, with their ability to outperforms previous approaches in terms of precision, recall,

979-8-3503-1860-9/24/$31.00 ©2024 IEEE

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY HAMIRPUR. Downloaded on August 05,2025 at 06:40:57 UTC from IEEE Xplore. Restrictions apply.
 and execution time. The effectiveness of the approach is • The training graphic demonstrates this repeated
further validated by visualisations and an ablation study, procedure. It improves models and finds encoder and
which highlights the crucial roles of transfer learning and decoder weights using backpropagation [18].
attention mechanisms.
• Using this strategy to minimize and regenerate data is
II. LITERATURE REVIEW crucial for detecting weird things. Use LSTM
Networks.
IoT-based Smart Grid Cybersecurity Systems employ
advanced Deep Neural Network (DNN) technologies to • It comprises 17 distinct stages and uses time series
discover and repair issues in real time. For instance, data. Start using input patterns to set LSTM
autoencoders with 0.92 accuracy and 0.88 recall are excellent parameters. Complex gate computations power
at identifying errors and collecting true positives [12]. LSTM learning.
networks operate well with 0.89 accuracy and 0.91 memory.
This makes them useful for determining time series The technique captures long-term associations well, which
relationships. Variational Autoencoders (VAEs) locate is critical for smart grid data prediction. The training graphic
anomalies well with 0.93 accuracy [13]. Generative shows that the model can predict future values in a time-series
Adversarial Networks (GANs) can create realistic normal data scenario using backpropagation and sequence processing [19].
for anomaly diagnosis with a precision of 0.91. One-Class Live diagrams depict how Generative Adversarial Networks
SVM finds issues well with minimal false positives and a (GANs) manages generator-discriminator interaction. This
recall of 0.92. When false positives matter, this strategy is competitive technique improves the generator to generate
effective. Abnormalities are easily found using Capsule realistic data, testing the discriminator's ability to identify
Networks with 0.94 precision. Self-Organizing Maps (SOMs) genuine from false data [20]. A neighborhood function for
have a 0.10 false positive rate and a more balanced approach geographical relationships and the algorithm's capacity to
[14]. Attention Mechanisms distinguishes normal and organize items and detect faults are its main goals. The
abnormal trends well with an AUC-ROC of 0.95. The flexible Transfer Learning approach takes data from a
Ensemble Methods identify anomalies well with an F1 score previously taught model to improve performance in a target
of 0.91. Finally, Transfer Learning uses previously taught region. The 17-step method uses source and goal losses to
knowledge with 0.89 accuracy and 0.91 memory [15]. adjust model parameters. As the training image shows,
transfer learning uses repeated optimisation for different
Execution time, model size, training time, robustness, regions to create a complete model for a variety of
scalability, interpretability, and resource consumption circumstances. Ultimately, a robust set of tools for real-time
demonstrate these methodologies' practicality [16]. These issue detection and resolution in IoT-based Smart Grid
measures enable system architects and decision-makers pick a Cybersecurity Systems is created by Autoencoders for quickly
Smart Grid cybersecurity solution that meets their aims and compressing data, LSTM Networks for time series
constraints. predictions, GANs for adversarial data, SOMs for spatial
relationship grouping, and Transfer Learning for adapting to
new domains. Flowcharts are used to explain each training
phase and avoid Smart Grid hacking.
Autoencoders Algorithm:
1. Name an N-dimensional input vector X.
2. Disable our encoder and decoder weights.
3. Use sigmoid activation to code H=1/1+e(WencX)
4. Create a bottleneck layer B from encoded data H.
5. Initializing Decoding: Zero Wdec, the decoder weights.
6. Sixth, construct
Fig. 1. LSTM Networks X′=1/1+e(WdecB)
using sigmoid activation.
7. Use Mean Squared Error to calculate Reconstruction Loss
Figure 1 depicts LSTM Networks' steps. It illustrates how L=1/Ni=1N(XiXi′).
these networks detect temporal relationships in time-series Backpropagation—Encoder: Gradient descent: Wenc=WencL
data to quickly identify IoT-based Smart Grid issues. /Wenc to update encoder weights
8. Decoder Backpropagation: Gradient descent updates decoder
Training and using Long Short-Term Memory networks weights with
for time series data is shown in Figure 1. After data input and Wdec=WdecL/Wdec.
preprocessing, the model is turned on and trained on input-
output pairs [17]. The model's performance increases during 9. Repeat: Iterate through steps 3-9 until convergence.
10. Encoding Function: Define the encoding function fenc(X): fenc
backpropagation because the LSTM layers track temporal (X)=1+e−(Wenc⋅X)
connections and estimate future values. This procedure is 11. Bottleneck Layer Function: Define the bottleneck layer function
repeated until the model converges, using LSTM's memory- fbottleneck(H):
boosting properties to correctly estimate time series. fbottleneck(H)=H
12. Decoding Function: Define the decoding function fdec(B): fdec
III. PROPOSED METHODOLOGY (B)= 1/1+e−(Wdec⋅B)
13. Reconstruction Function: Reconstruct the input data using the
Autoencoders detect unusual behavior in IoT-based Smart defined decoding function:
Grid Cybersecurity Systems in various method as below. X′=fdec(fbottleneck(fenc(X)))
14. Loss Function: Define the loss function L using the Mean
• After receiving input, the software encodes, chokes, Squared Error: L=1/N∑i=1N(Xi−Xi′)2
and decodes to reassemble it. 15. Gradient Descent - Encoder: Update encoder weights using the
gradient descent algorithm.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY HAMIRPUR. Downloaded on August 05,2025 at 06:40:57 UTC from IEEE Xplore. Restrictions apply.
 16. Gradient Descent - Decoder: Update decoder weights using the Figure 3 shows LSTM network training using time series
gradient descent algorithm. data. Time dependencies, processing order, and
17. Repeat: Iterate through steps 11-17 until convergence.
18. Optimal Weights: Obtain the optimal encoder and decoder
backpropagation optimization are its main features. This
weights Wenc and Wdec. models smart grid data and predicts future values.
19. Output: Use the trained autoencoder for data compression and
reconstruction.
Figure 3 depicts a system that handles string numbers
utilizing complicated gate computations, cell state updates,
and concealed state updates. The model may reveal long-term
relationships since the forget, input, cell, and output gate
equations are complex [21]. Backpropagation and gradient
descent fine-tune LSTM parameters for reliable results. The
algorithm continues this procedure at each time step, making
it effective at learning from and predicting sequential data in
smart grid hacking.
Generative Adversarial Networks (GANs) Algorithm
Generator Input:
Generate random noise z with dimension m.
1. Generator Initialization: Initialize generator parameters Wg.
Fig. 2. Autoencoders Training 2. Generate Fake Data: Generate fake data Xfake using the
generator: Xfake=G(z;Wg)
Figure 2 shows Autoencoder anomaly detection training. 3. Discriminator Initialization: Initialize discriminator parameters
The book emphasizes the need of backpropagation and Wd.
repeated reconstruction of input data to improve model 4. Discriminator Prediction - Real Data: Predict authenticity for
performance and demonstrate how idealizing anomaly real data Xreal using the discriminator: Dreal=D(Xreal;Wd)
5. Discriminator Prediction - Fake Data: Predict authenticity for
detection may improve accuracy. fake data Xfake using the discriminator: Dfake=D(Xfake;Wd)
6. Discriminator Loss: Calculate discriminator loss: Ld=−1/2
LSTM Networks Algorithm: (log(Dreal)+log(1−Dfake))
1. Input Sequence: Take input sequence X of length T. 7. Backpropagation - Discriminator: Update discriminator
2. Initialization: Initialize LSTM parameters Wf,Wi,WC,Wo,Uf,Ui parameters using gradient descent: Wd=Wd− Ld/α∂Wd∂
,UC,Uo,bf,bi,bC,bo. 8. Generate New Fake Data: Generate new fake data Xnew using
3. Hidden State Initialization: Set initial hidden state ℎ0h0 and the updated generator.
cell state C0 to zero. 9. Discriminator Prediction - New Fake Data: Predict authenticity
4. Gates Computation: Compute forget (ft), input (it), cell (C~t), for Xnew using the discriminator: Dnew=D(Xnew;Wd)
and output (ot) gates: ft=σ(Wf⋅ht−1+Uf⋅xt+bf) it=σ(Wi⋅ht−1+Ui 10. Generator Loss: Calculate generator loss:Lg=−log(Dnew)
⋅xt+bi) C~t=tanh(WC⋅ht−1+UC⋅xt+bC) ot=σ(Wo⋅ht−1+Uo⋅xt 11. Backpropagation - Generator: Update generator parameters
+bo) using gradient descent: Wg=Wg−∂Lg/α∂Wg
5. Cell State Update: Update cell state Ct: Ct=ft∗Ct−1+it∗C~t
6. Hidden State Update: Update hidden state ht: ht=ot∗tanh(Ct) 12. Repeat: Iterate through steps 3-12 until convergence.
7. Output Calculation: Calculate output yt: yt=softmax(Wy⋅ht+by 13. Optimal Generator: Obtain the optimal generator for generating
) realistic data.
8. Backpropagation - Output Layer: Update parameters Wy and
by based on output error.
9. Backpropagation - LSTM Layer: Backpropagate error through
the LSTM layer.
10. Repeat: Iterate through steps 4-9 for each time step.
11. Input Gate Function: Define the input gate function fi(xt,ht−1
): it=σ(Wi⋅ht−1+Ui⋅xt+bi)
12. Forget Gate Function: Define the forget gate function ff(xt,ht−1
): ft=σ(Wf⋅ht−1+Uf⋅xt+bf)
13. Cell State Function: Define the cell state function fcell(xt,ht−1
): C~t=tanh(WC⋅ht−1+UC⋅xt+bC)
14. Output Gate Function: Define the output gate function fo(xt
,ht−1): ot=σ(Wo⋅ht−1+Uo⋅xt+bo)
15. Output Layer Function: Define the output layer function
foutput(ht): yt=softmax(Wy⋅ht+by)
16. Gradient Descent - Output Layer: Update output layer
parameters.
17. Gradient Descent - LSTM Layer: Update LSTM layer
parameters.

Fig. 4. GANs Adversarial Process

Figure 4 illustrates the discriminator using generator-

generated phony data to make choices. Both sections are often
changed to achieve the ideal balance for identifying anomalies
more often.
The Generative Adversarial Networks (GANs) technique
tricks a discriminator by giving it genuine data and
challenging it to identify phony data. This competitive process
Fig. 3. LSTM Networks Training

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY HAMIRPUR. Downloaded on August 05,2025 at 06:40:57 UTC from IEEE Xplore. Restrictions apply.
 is repeated, and backpropagation and gradient descent update TABLE I. PERFORMANCE METRICS OF ANOMALY
DETECTION METHODS
the generator and discriminator. The generator generates data
that looks like genuine data, indicating model convergence. Method Exe M Trai Rob Scal Inte Res
GANs may produce bogus data, visuals, and add to existing cuti od ning ustn abili rpre ourc
data in many circumstances. on el Tim ess ty tabil e
Tim Siz e ity Utili
Self-Organizing Maps (SOMs) Algorithm: e e (hou zati
1. Input Initialization: Initialize input data X with dimension D. (ms) (M rs) on
2. SOM Initialization: Initialize SOM weights W with random B)
values. Autoencoders 15 5 12 High Med Low Mod
3. Neighborhood Function: Define the neighborhood function hij(t): ium erate
hij(t)=exp(−∥ri−cj∥2/2σ(t)2) LSTM 20 8 20 Mod High Low High
4. Update Learning Rate: Update learning rate α(t) and Networks erate
neighborhood radius σ(t). Variational 18 6 15 High Med Low Mod
5. Find Best-Matching Unit (BMU): Find the BMU b using: Autoencoders ium erate
b=argminj∥X−Wj∥ (VAEs)
6. Update Weights: Update SOM weights using: Wj(t+1)=Wj Generative 22 10 25 Mod High Low High
(t)+α(t)⋅hbj(t)⋅(X−Wj(t)) Adversarial erate
7. Repeat for Multiple Inputs: Iterate through steps 3-6 for each Networks
input. One-Class 10 2 8 High High Mod Low
8. Update Neighborhood Radius: Decrease σ(t) to focus learning on SVM erate
BMU. Capsule 25 12 30 Mod Med Mod High
9. Update Learning Rate: Decrease α(t) to fine-tune learning. Networks erate ium erate
10. Repeat: Iterate through steps 3-9 until convergence. Self- 17 7 18 Mod Med Mod Mod
11. Best Matching Unit (BMU) Function: Define BMU function Organizing erate ium erate erate
fbmu(X,W): b=argminj∥X−Wj∥ Maps (SOMs)
12. Weight Update Function: Define weight update function fupdate Attention 16 6 14 High Med Low Mod
(W,X,b,t):)Wj(t+1)=Wj(t)+α(t)⋅hbj(t)⋅(X−Wj(t)) Mechanisms ium erate
Ensemble 30 15 35 Mod High Low High
Methods erate
SOMs model input data in low dimensions by altering
Transfer 19 9 22 High Med Low High
weights repeatedly. The software connects spaces with a Learning ium
neighborhood function and modifies weights based on input Proposed 12 4 10 High High High High
patterns [22]. SOMs learn to map high-dimensional raw data Method
to grids. Grouping makes outliers simpler to find. SOMs
provide insight into complex data linkages in smart grid The performance characteristics of many anomaly
hacking. detection techniques are listed in Table 4. Parameters include
processing time, model size, training time, robustness,
IV. RESULT scalability, interpretability, and resource economy. The
Table 1 compares autoencoders, LSTM Networks, and the proposed technique outperforms all others and may be used to
Proposed Method for anomaly identification. Memory, discover issues in IoT-based Smart Grid Cybersecurity
accuracy, and AUC-ROC reveal that the recommended Systems in real time.
method works better for real-time problem detection in IoT-
based Smart Grid Cybersecurity Systems.
Table 4 compares processing time, model size, training
time, stability, scalability, interpretability, and resource
utilization and indicates the proposed technique is preferable.
New methods consistently outperform old ones, proving they
can be utilized for smart grid cybersecurity.
Figures 5, 6, 8 and 9 demonstrate how and why anomaly
detection methods function. The Area Chart and Histogram
exhibit accuracy, memory, and F1 Score distribution,
demonstrating the recommended method's efficacy. The
Scatter Plot depicts the False Positive Rate vs. True Positive
Rate trade-off, proving the strategy works. We may better Fig. 5. Comparison of Precision, Recall, and F1 Score across different
appreciate the benefits and drawbacks of the methodologies anomaly detection methods
by viewing their entire connections, precision-recall trade-
offs, and performance patterns in the Network Graph, The F1 Score, Precision, and Recall for many anomaly
Candlestick Chart, and Spline Chart. detection algorithms are shown in Figure 5. Each colored
region represents a separate metric, making distribution
The recommended strategy outperforms others in
comparisons straightforward. The image clearly depicts each
accuracy, recall, and AUC-ROC. This indicates that it can
method's success. The Proposed Method excelled everywhere.
detect IoT-based Smart Grid Cybersecurity System issues in
real time.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY HAMIRPUR. Downloaded on August 05,2025 at 06:40:57 UTC from IEEE Xplore. Restrictions apply.
 V. DISCUSSION
Every module in our recommended anomaly detection
scheme was eliminated and assessed in an ablation study to
determine its usefulness. The study found that removing the
proposed strategy greatly reduced performance. This proves
the algorithm works well together for optimum outcomes.
When attention and transfer learning were coupled, accuracy,
memory, and resistance improved. Without attention
processes, the model couldn't recognize nuanced trends or
discover abnormalities. Without transfer learning, the system
struggled to adapt to smart grid cybersecurity instances. Using
connected information is crucial.
Fig. 6. Distribution of Precision, Recall, and F1 Score among different A number of machine learning methods are included and
anomaly detection methods comparison between them has been done.The objective of this
technical assessment is to offer a nuanced viewpoint on how
Figure 6 demonstrates the distribution of Precision, Recall, these techniques contribute to the categorization of
and F1 Score for anomaly detection algorithms. Stacking bars cyberattacks in Internet of Things systems. Experts in
indicate how much one sign impacts the entire, allowing cybersecurity can benefit greatly from the performance
deeper comparisons. The picture compares each method's analysis of these techniques, which provides direction for
performance patterns, focusing on robustness. creating strong defences for the Internet of Things. The results
of this study have the potential to greatly improve
cybersecurity procedures and strengthen IoT environments
against possible threats as IoT security continues to gain
importance.
VI. CONCLUSION
Finally, our anomaly detection approach works well for
IoT-based Smart Grid Cybersecurity Systems. It uses
autoencoders, LSTMs, GANs, SOMs, and transfer learning. A
comprehensive analysis reveals in Tables 3 and 4 that our
strategy is more successful and meets more requirements than
others. Figures 5–12 illustrate how our strategy somewhat
enhances accuracy, memory, and performance speed. The
Fig. 7. Trade-off between False Positive Rate and True Positive Rate for scatter plot in Figure 8 shows the true-fake positive ratio. This
different anomaly detection methods.
proves our anomaly detection approach works in real time. We
Figure 7 details the True Positive Rate-False Positive Rate found that attention processes and transfer learning help solve
trade-off. Each point indicates an alternative strategy to issues faster in a changing smart grid by studying ablation and
discover anomalies, with red being the Proposed method. This every computing aspect. Our solution is unique and versatile
image indicates that the Proposed Method finds anomalies in enough to handle all the issues with discovering weird things
real time with a better balance of actual and bogus results. in IoT-based Smart Grid Cybersecurity Systems.
REFERENCES
[1] H. Tahaei, F. Afifi, A. Asemi, F. Zaki, and N. B. Anuar, "The rise of
traffic classification in IoT networks: a survey," Journal of Network
and Computer Applications, vol. 154, p. 102538, 2020.
[2] G. Acampora, M. L. Bernardi, M. Cimitile, G. Tortora, and A. Vitiello,
"A fuzzy clustering-based approach to study malware phylogeny," in
2018 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE),
Rio de Janeiro, Brazil, 2019.
[3] R. Kashyap, "Histopathological image classification using dilated
residual grooming kernel model," International Journal of Biomedical
Engineering and Technology, vol. 41, no. 3, p. 272, 2023. [Online].
Available: https://doi.org/10.1504/ijbet.2023.129819
[4] J. Kotwal, Dr. R. Kashyap, and Dr. S. Pathan, "Agricultural plant
diseases identification: From traditional approach to deep learning,"
Materials Today: Proceedings, vol. 80, pp. 344–356, 2023. [Online].
Available: https://doi.org/10.1016/j.matpr.2023.02.370
Fig. 8. Execution Time (ms) [5] Edwin Ramirez-Asis, Romel Percy Melgarejo Bolivar, Leonid Alemán
Gonzales, Sushovan Chaudhury, Ramgopal Kashyap, Walaa F.
Different techniques' execution times (ms) are presented Alsanie, G. K. Viju, "A Lightweight Hybrid Dilated Ghost Model-
in Figure 8. The recommended method is faster than others, Based Approach for the Prognosis of Breast Cancer," Computational
Intelligence and Neuroscience, vol. 2022, Article ID 9325452, 10
demonstrating its real-time anomaly detection ability.Top of pages, 2022. [Online]. Available:
Form https://doi.org/10.1155/2022/9325452
[6] M. L. Bernardi, M. Cimitile, F. Martinelli, and F. Mercaldo, "Game bot
detection in online role player game through behavioural features," in

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY HAMIRPUR. Downloaded on August 05,2025 at 06:40:57 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the 12th International Conference on Software
Technologies, pp. 50–60, SciTePress, Madrid, Spain, 2017.
[7] M. Calabretta, R. Pecori, and L. Veltri, "A token-based protocol for
securing MQTT communications," in 2018 26th International
Conference on Software, Telecommunications and Computer
Networks (SoftCOM), Split, Croatia, 2018.
[8] R. Pecori, A. Tayebi, A. Vannucci, and L. Veltri, "IoT attack detection
with deep learning analysis," in 2020 International Joint Conference on
Neural Networks (IJCNN), Glasgow, UK, 2020.
[9] L. Aversano, M. L. Bernardi, M. Cimitile, and R. Pecori, "A systematic
review on deep learning approaches for IoT security," Computer
Science Review, vol. 40, article 100389, 2021.
[10] M. F. Balın, A. Abid, and J. Zou, "Concrete autoencoders:
differentiable feature selection and reconstruction," in K. Chaudhuri
and R. Salakhutdinov, Eds., pp. 444–453, PMLR.
[11] V. Roy et al., “Detection of sleep apnea through heart rate signal using
Convolutional Neural Network,” International Journal of
Pharmaceutical Research, vol. 12, no. 4, pp. 4829-4836, Oct-Dec 2020.
[12] R. Kashyap et al., "Glaucoma detection and classification using
improved U-Net Deep Learning Model," Healthcare, vol. 10, no. 12, p.
2497, 2022. [Online]. Available:
https://doi.org/10.3390/healthcare10122497
[13] Vinodkumar Mohanakurup, Syam Machinathu Parambil Gangadharan,
Pallavi Goel, Devvret Verma, Sameer Alshehri, Ramgopal Kashyap,
Baitullah Malakhil, "Breast Cancer Detection on Histopathological
Images Using a Composite Dilated Backbone Network,"
Computational Intelligence and Neuroscience, vol. 2022, Article ID
8517706, 10 pages, 2022. [Online]. Available:
https://doi.org/10.1155/2022/8517706
[14] P. Ducange, G. Mannará, F. Marcelloni, R. Pecori, and M. Vecchio, "A
novel approach for internet traffic classification based on multi-
objective evolutionary fuzzy classifiers," in 2017 IEEE International
Conference on Fuzzy Systems (FUZZ-IEEE), Naples, Italy, 2017.
[15] M-Ridha, M.J., Ali, M.F., Taly, A.H., Abed, K.M., Mohammed, S.J.,
et al. (2022). Subsurface Flow Phytoremediation Using Barley Plants
for Water Recovery from Kerosene-Contaminated Water: Effect of
Kerosene Concentration and Removal Kinetics. Water, 14(5), 687.
[16] Mohammed, S.J., M-Ridha, M.J., Abed, K.M., & Elgharbawy, A.A.M.
(2023). Removal of levofloxacin and ciprofloxacin from aqueous
solutions and an economic evaluation using the electrocoagulation
process. International Journal of Environmental Analytical Chemistry,
103(16), 3801-3819.
[17] Ali, M.F., Ali, Q.A., M-Ridha, M.J., Mohammed, S.J., & Bohan, H.R.
(2023). Phytoremediation of tetracycline via the coontail
Ceratophyllum demersum in antibiotics-contaminated water.
Biocatalysis and Agricultural Biotechnology, 53, 102887.V. L. L.
Thing, "IEEE 802.11 network anomaly detection and attack
classification: a deep learning approach," in 2017 IEEE Wireless
Communications and Networking Conference (WCNC), San
Francisco, CA, USA, 2017.
[18] R. R. Maaliw et al., "AWFCNET: An Attention-Aware Deep Learning
Network with Fusion Classifier for Breast Cancer Classification Using
Enhanced Mammograms," 2023 IEEE World AI IoT Congress
(AIIoT), Seattle, WA, USA, 2023, pp. 0736-0744, doi:
10.1109/AIIoT58121.2023.10174427.
[19] S. Stalin, V. Roy, P. K. Shukla, A. Zaguia, M. M. Khan, P. K. Shukla,
A. Jain, "A Machine Learning-Based Big EEG Data Artifact Detection
and Wavelet-Based Removal: An Empirical Approach," Mathematical
Problems in Engineering, vol. 2021, Article ID 2942808, 11 pages,
2021. [Online]. Available: https://doi.org/10.1155/2021/2942808
[20] A. A. Diro and N. Chilamkurti, "Distributed attack detection scheme
using deep learning approach for internet of things," Future Generation
Computer Systems, vol. 82, pp. 761–768, 2018.
[21] Jianbin Wu, Sami Ahmed Haider, Heejung Yu, Muhammad Irshad,
Mukesh Soni, Mohit Kumar Bhadla, Yousaf Bin Zikria, An intelligent
IoT intrusion detection system using HeInit-WGAN and SSO-
BNMCNN based multivariate feature analysis, Engineering
Applications of Artificial Intelligence, Volume 127, Part A, 2024,
107132, ISSN 0952-1976,
https://doi.org/10.1016/j.engappai.2023.107132.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY HAMIRPUR. Downloaded on August 05,2025 at 06:40:57 UTC from IEEE Xplore. Restrictions apply.