2022 International Conference on Futuristic Technologies (INCOFT)

Karnataka, India. Nov 25-27, 2022

Implementation of Deep Learning Techniques for

Secure IoT Networks
Rohini Mahajan, Dushyant Singh
Assistant Professor, Assistant Professor,
2022 International Conference on Futuristic Technologies (INCOFT) | 978-1-6654-5046-1/22/$31.00 ©2022 IEEE | DOI: 10.1109/INCOFT55651.2022.10094474

Department of Computer Science & Engineering, Department of Computer Science Engineering,

Chandigarh Engineering College, Vivekananda Global University,
Jhanjeri, Jaipur, India.
[email protected] [email protected]

Abstract - The medical industry, the military, and the security and privacy might be jeopardised if IoT devices
power grid are just a few examples of businesses that may aren't adequately safeguarded. A formidable security
benefit greatly from using IoT solutions. The safety of an solution, TI refers to artificially intelligent technologies that
Internet of Things network is less important than the safety of can automatically safeguard SAGS networks. On this
the underlying computer and communications systems. research, we provide a novel method for threat identification
However, IoT networks are vulnerable to a variety of threats (TI) that is grounded in deep learning techniques and is
because of the scarcity of available resources and capable of detecting cyber threats emanating from SAGS
computational power. Protecting the Internet of Things (IoT) infrastructure.
network from malicious assaults necessitates the creation and
use of effective security measures, such as an intrusion In view of the ever-increasing demands on top-notch
detection system. In this paper, we provide a unique method healthcare and the ever-increasing costs of treatment, a
for categorising the data flowing through an Internet of Things growing number of people are examining the possibility of
(IoT) network by using Deep Neural Networks (DNNs) adopting ubiquitous medicine as a technology solution to the
educated on deep learning data to spot outliers. We create world's healthcare concerns. Most notably, the Internet of
standard features from field-collected packet-level data using a Medical Things has emerged as a result of cutting-edge
recently released IoT dataset. To classify potential threats to innovation in IoT devices (IoMT). Security and privacy
Internet of Things (IoT) devices, such as denial-of-service
issues are typically overlooked, despite the fact that cheap
(DoS), distributed denial-of-service (DDoS), reconnaissance,
and widely distributed sensors have the ability to change the
and information theft, we create a feed-forward neural
networks model. On the sanitised dataset, the tested approach
present reactive treatment into preventive care. Connections
exhibits high classification accuracy. between medical equipment and amongst themselves must
be very secure because of the sensitive nature of the health
Keywords – IoT, Security, Deep Learning, DDoS, Attackers. information they gather and handle. Because of their limited
computing power, however, only a subset of security
I. INTRODUCTION protocols may be used in miniaturised IoT devices. Further,
The Mining Industry's Blockchain Strategy for the widespread use of IoMT devices makes it difficult to
Increasing Online Safety, Availability, and Data Integrity in monitor and guarantee the security of these networks. That is
the IoT: Problems and Prospects. The fast evolution of a significant hurdle for IoMT-based medicinal research.
Internet-enabled gadgets and related communication IoT, an abbreviation for "Internet of Things," is a system
technologies has led to the emergence of the "Internet of of interconnected gadgets that can share and gather
Things" (IoT). Although the IoT's numerous features and information in real time over the internet. This idea is also
benefits inspire innovative approaches to problems, they also often referred to as a "virtual of objects." Internet of
present certain dangers to users. Knowing the weaknesses of Everything technologies such as wireless sensor networks
these technologies is crucial for avoiding their misuse. There (WSN) and radio frequency identification (RFID) have the
are several known and undisclosed security vulnerabilities potential to improve many areas of society (RFID).
with the protocol-level approach of Message Queuing Miniature, high-tech devices are crucial to the functioning of
Telemetry Transfer (MQTT). A more efficient agronomic the IoT. There is a vast variety of devices in this category,
big data ant colony optimization algorithm (BigDataACO) is each with its own set of pros and downsides. Given that the
presented to round out the multispectral information found in great majority of currently available web technologies and
agricultural big data for use in feature sets and decision- protocols were not built to enable IoT, security is a
making. Our goal is to increase the efficiency with which significant issue for the Internet of Things. On the other
agricultural land is put to use by analysing and using hand, this was framed as a cornucopia of untapped
multisource, heterogeneous big data. Additionally, the opportunities. Numerous studies have examined the IoT's
problem of combining information from many sources was unanswered questions and security holes, but the full scope
tackled and handled. of the environment's susceptibility has yet to be investigated.
Connecting Space, Air, Grounds, and Sea (SAGS) In this paper, we use a three-layer neural network of self-
networks is becoming more important to reliably offer encoder to construct a hierarchical encroachment security
automated services to consumers and companies as the detection model and investigate the practical use of
Internet of Things (IoT) evolves. However, networks' computational intelligence denoisingautoencoder (DAE) in

978-1-6654-5046-1/22/$31.00 ©2022 IEEE 1

Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on April 18,2023 at 22:56:45 UTC from IEEE Xplore. Restrictions apply.
the context of IoT fusion security. A stacked DAE can now increase security. Once the IoT is working, these
supporting vector machine is the name for this concept cluster pieces may be deployed throughout the hosts. In this
(SDAE-SVM). In this research, we want to learn more about setup, a computer is known as a host if it is used to break
potential applications of deep learning denoisingautoencoder down a larger work into smaller chunks. Each packet in this
(DAE) in the field. Dimensional reductions were obtained by example consists of data and an associated address. The head
supervised training, and then iterative adjustments were of the packet, also called as the addresses part of the packet,
made to each level, and now the resultant data have been contains the addresses of the other cluster along the drain
gathered. In this study, common deep learning methods are machine's route. As the internet of things relies on a
presented and compared to a reference prototype via decentralised design, the anonymity of package information
computation. These methods include stacked noise will be jeopardised if we use the conventional approach to
autoencoder (SNAE), stacked autoencoder (SAE), stacked network security. To put this concept into action, we need to
contractive autoencoder (SCAE), stacked sparse autoencoder implement a decentralised system made up of many
(SSAE), and convolutional (DBN). The system's setup and individual devices that are all connected to the internet. Each
the use of IoT devices have allowed it to grow beyond its of these gadgets should operate as a sink machine, and a host
original scope, which has increased its ability to machine is required for sink management. Our planned
communicate, analyse, and be controlled. To survive and organisational structure is based on a hierarchical chain of
thrive in today's ultracompetitive economic climate, almost command, and all of our information is stored on the cloud.
every industry and company that operations anywhere in the This framework is also used to coordinate bitcoin
world will soon be wired with a system of Internet of Things transactions and device-to-device data transfers for increased
(IoT) devices. While technological advancements have trustworthiness. To assure a suburban-like topology, our
reduced certain threats, they have not eliminated them architecture takes use of a feature called distributed trust.
entirely. The same is true for IoT-enabled systems, where [13] B. Mishra and N. Padhy (2020). Using Blockchain
several security concerns have emerged and pose threats to Technology in Mining: Challenges and Solutions The
the reliability of the system's ongoing functioning.Related purpose of [14] is to fix the problems with the IoT's security,
Work dependability, and data integrity. In order to highlight the
shortcomings of the current ideas, this review article will
According to the results, a four-layer network generator examine and classify the many security approaches that are
produces the best accuracy rate (Ac) for the model (97.83%) now in use. The goal is to analyse and classify the various
and the lowest untrue rate (Fn) (1.27%) and false-positive forms of protection currently in use. Summary of what we'll
rate (Fp) (3.21%). The model's performance is quantified by talk about IoT MQTT protocol concepts, IIoT MQTT
its accuracy rate (Ac). The model's quality rises to roughly protocol common security solutions, and IIIoT MQTT
98% when there are around 110 vertices in the first hidden protocol security levels. In conclusion, investigators may
layer. With values of 98.12%, 3.21%, and 1.27%, utilise the values provided in this article to formulate an
respectively, for Ac, Fn, and Fp, this model outperforms the effective security approach for a variety of IoT-based use
competition when large amounts of data are reduced. This cases [15]. Swarm intelligence is a software that mimics the
was developed specifically for this study, and it served as the complex problem of crowds in nature by having its users
basis for the model. Compared to results obtained using other work together. The procedure is resilient, can be parallelized,
deep learning methods, the recognition rate, Ac, error rate, and does not depend on any one specific problem.
and separation factor all show promising results. Across Implementation strategies for the agricultural big data fusion
several datasets, the female's detection performance, Ac, problem are studied, along with its description, idea, and
margin of error, and amount of defects are consistently methodology. The limitations of the large data fusion
higher than those attained by traditional deep learning modelling approach are then discussed. This study's last
algorithms. In conclusion, the detect load can be reduced, the section delves into the massive data fusion approach used by
sensing effect can be improved, and the operation can be insect colonies, as well as the code editor and its key
made safer and more reliable when deep learning SDAE is components. Results show that the obtained data may be
used to IoT alignment intrusion risk detection [11]. This used to verify the improved BigDataACO approach. The
article presents a comprehensive analysis of the many upgraded technique [16] established in this work greatly
security issues plaguing IoT-enabled devices and the decreases the ambiguity associated with data fusion when
difficulties these issues create. Also, both technological and compared to K-means, D-S evidence theory, and the Bayes
administrative suggestions for countermeasures are included method. The novel technique consists of three parts: a deep
in the article. Finally, the paper concludes with a discussion patterns extraction stage, a TI-driven recognition stage, and a
of potential future research directions for the practitioner of TI-attack type identification stage. While the unit is designed
this discipline [12]. The present and future eras are going to to uncover dormant features contained in IoT, its output is
undergo an unexpected shift transformation from defunct used as input by the TI-driven identification. TI-attack type
systems into a pervasive p2p empowered entity that can identification is used to classify the damaging actions that
communicate with one other and generate data from which make up the different forms of attacks; this information is
useful insights may be drawn. The term "Internet of then utilised to help in problem solving. TON-IoT and N-
everything" describes a hypothetical future in which all BAIOT, two distinct sets of data, are used to evaluate the
networks are interconnected globally (IOT). The Internet of proposed approach. In terms of detection and false alarm
Things is crucial in improving governmental efficiency, rates, the experimental findings show that the approach
easing communication, and many other spheres of life. excels in comparison to other, similar methods [17]. The
Therefore, a large number of people will be relying on it in Internet of Things, or IoT, is a vast, adaptable global network
the near future, therefore its security and information architecture that enables virtualized items to communicate
integrity must be rock solid. Using the bitcoin strategy of with one another and with any object, service, or person
distributing the programme among several cluster nodes, we across the World Wide Web at any time. The article provides

Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on April 18,2023 at 22:56:45 UTC from IEEE Xplore. Restrictions apply.
a review, analysis, and high-level summary of an IoT We combine, drop, and encode the column values in
ecosystem. Definitions, components, features, architectures, feature preparation once the obtained fields data has been
technologies, application areas, and an examination of preprocessed as well as stored in arrays. Training traffic
security problems and challenges related to the Internet of (used during the training stage) and testing traffic (used
Things (IoT) make up this ecosystem [18]. The purpose of during the testing phase) were both generated using the
this research was to investigate the current state of IoMT processed data (used in the classification phase). Each
privacy and security, as well as to identify the areas where malicious and benign packet in both training and testing
more work is needed and the most promising avenues for traffic is labelled, and each damaging packet is labelled with
future investigation. This gives you a quick rundown of the the kind of attack it is part of. All of the labels in the test data
methods that are currently state-of-the-art [19]. This paper are utilised for is evaluating the accuracy of the predictions.
provides a brief overview of the motivation behind IoT, a
description of the protected IoT nicely textured structure, IoT Third, we feed the network the processed data (arrays)
Noted, a discussion of the security problems with multiple from the training traffic in order to educate it. In order to
attacks within each layer of the IoT, and a look at the categorise new incoming traffic, a neural network model has
limitations of the traditional approaches to having security been developed. The classifier has the potential to divide
remedies [20]. data into both binary and many categories.
Part 4: The Sorting Process The processed data from the
II. PROPOSED WORK testing traffic is sent into the classifier, which then assigns
In this section, we give a high-level overview of the normal/malicious labels to each packet. The classifier places
framework for IoT networks' intrusion detection. Figure 1 the malicious packet into the attack category that best
depicts the framework, which primarily consists of the four matches its features for the purpose of multi-classification.
phases of feature extraction, feature preprocessing, training,
Here is an example of how the suggested DNN may be
and classification.
used.
In a nutshell, the input layer is made from : The network
Prepare Dataset receives information from these neurons. Here, we lay out all
of the possible paths we may take, =r 1,r 2,...,r u.
In order to extract features from the input flow, a deep
neural network with several hidden layers is used (up to L).
Apply a crossover operator to the primary input pathways.
Preprocessing Through a mechanism called crossover, geneticists are able
to create previously unthinkable permutations of ancestry-
based pathways. The information acquired through feature
extraction is then used to detect infiltration at the topmost
layer.
Feature Extraction
III. EXPERIMENTAL RESULTS & DISCUSSION
A. Simulation setup
A simulation of the intended network has been
implemented in the network simulator version 3.26. (ns-
Deep Classification 3.26). Various network simulations are possible with the help
of the event-based ns-3.26 simulation tool. To begin, ns-3.26
is present on the machine running Ubuntu 14.04. The
configuration of the network is then adjusted such that it best
accommodates our studies. Initially, the algorithms are
written in C++, but once they're ready to be executed, they're
Normal Malicious transferred to Python. Before beginning the experiment, we
used ns3.26 to determine the node positions, initial energy,
and n value. With everything set up, we can run the whole
Fig. 1. Deep Learning based IDS in IoT simulation and see what happens.
An in-depth analysis of each stage is provided below. TABLE I. DATA GENERATED BY SIMULATION
First, we get raw traffic data from the network analyzer.
Parameter Values
Thereafter, we extract the necessary information from each
packet. Each extracted field is the name of a feature. Here, Number of flows 10000
we utilise the field data from each individual packet rather Normal 350
than relying on aggregated packets to infer features. DDoS 150
Therefore, we aim to capture general characteristics of traffic
rather than designing attack-oriented features that are tailored Man in the Middle 200
to identify specific attack behaviours. In particular, we focus Intrusion 300
on the frame, IPv4/IPv6, and TCP/UDP fields in IP packets'
headers.

Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on April 18,2023 at 22:56:45 UTC from IEEE Xplore. Restrictions apply.
B. Comparative analysis
Naïve Bayes Random Forest ANN DNN
We can see that the accuracy, precision, recall, and F1
120
score for classifying genuine traffic and malicious traffic for
each DDoS/DoS subcategory is all above 0.999, indicating
minimal FP and FN predictions. In particular, the extracted 100
dataset at hand detects Distributed Denial of Service (DDoS)
assaults using HTTP/UDP with a perfect 100% success rate.
80
Based on the prediction of DoS through UDP assaults, 222
out of 194480 (0.1%) malicious packets were incorrectly

Precision (%)
classified as benign. It's possible that this is due to the fact 60
that certain ICMP packets disguise a UDP packet within the
ICMP payload; this UDP packet might be malicious or
40
benign, depending on the context, and this could throw off
the classifier's general characteristics.
20

Accuracy 0
1 2 3 4 5 6
100 Iteration
Fig. 4. Analysis on Precision
95

90 Naïve Bayes Random Forest ANN DNN

120
85
100
80
80
75
Recall (%)

Naïve Random ANN DNN 60

Bayes Forest
40
Fig. 2. Analysis on Accuracy

20

Naïve Bayes Random Forest ANN DNN 0

1 2 3 4 5 6
Iteration
120
Fig. 5. Analysis on Recall

100
Naïve Bayes Random Forest ANN DNN
80
Accuracy (%)

120

60 100

40 80
F-score (%)

20 60

0 40
1 2 3 4 5 6
Iteration 20

Fig. 3. Analysis on Accuracy 0

1 2 3 4 5 6
Iteration
Fig. 6. Analysis on F-score

Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on April 18,2023 at 22:56:45 UTC from IEEE Xplore. Restrictions apply.
 REFERENCES
Naïve Bayes Random Forest ANN DNN [1] Lv, Z., Qiao, L., Li, J., & Song, H.H. (2021). Deep-Learning-Enabled
120 Security Issues in the Internet of Things. IEEE Internet of Things
Journal, 8, 9531-9538.
[2] Poonia, A.S., Banerjee, C., Banerjee, A., & Sharma, S.K. (2020).
100 Security Issues in Internet of Things (IoT)-Enabled Systems: Problem
and Prospects.
80 [3] Mishra, B., & Padhy, N. (2020). Enhancing the Security, Reliability,
and Data Integrity Issues in the Internet of Things by Implementing
AUC (%)

Blockchain Strategy in Mining: Challenges and Solutions.

60
[4] Tang, C., Wei, X., Liu, C., Jiang, H., Wu, H., & Li, Q. (2020). UAV-
Enabled Social Internet of Vehicles: Roles, Security Issues and Use
40 Cases. SocialSec.
[5] Hintaw, A.J., Manickam, S., Karuppayah, S., & Aboalmaaly, M.F.
(2019). A Brief Review on MQTT's Security Issues within the
20 Internet of Things (IoT). J. Commun., 14, 463-469.
[6] Zhang, F., & Zhang, Y. (2020). A Big Data Mining and Blockchain-
0 Enabled Security Approach for Agricultural Based on Internet of
1 2 3 4 5 6 Things. Wirel. Commun. Mob. Comput., 2020, 6612972:1-6612972:8.
Iteration [7] Al-Hawawreh, M.S., Moustafa, N., Garg, S., & Hossain, M.S. (2021).
Deep Learning-Enabled Threat Intelligence Scheme in the Internet of
Fig. 7. Analysis on AUC
Things Networks. IEEE Transactions on Network Science and
Engineering, 8, 2968-2981.
The categorization for reconnaissance assaults likewise [8] Pukkasenung, P. (2020). Internet of Things (IoT): A Basic Concept
achieves good precision, recall, and accuracy values, as well and Analysis Security Issues.
as a high F1 score (all above 0.99). 1201 over 203278 (0.6%) [9] Sun, Y., Lo, F.P., & Lo, B.P. (2019). Security and Privacy for the
normal packets were categorised as malicious packets in the Internet of Medical Things Enabled Healthcare Systems: A
prediction on service scan assaults, while 694 over 199700 Survey. IEEE Access, 7, 183339-183355.
(0.3%) malicious packets were classified as normal packets. [10] Krishna, B.V., & Gnanasekaran, T. (2017). A systematic study of
The ICMP payload may contain embedded UDP packets, security issues in Internet-of-Things (IoT). 2017 International
Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud)
which could be benign or malicious and cause confusion for (I-SMAC), 107-111.
the classifier based on existing features for assaults.
[11] Nayak, P., Kavitha, K., & Mallikarjuna Rao, C. (2019). IoT-Enabled
Agricultural System Applications, Challenges and Security
IV. CONCLUSION Issues. Studies in Big Data.
Deep learning is an intelligent method that may be used [12] Spathoulas, G.P., Negka, L., Pandey, P., & Katsikas, S.K. (2020).
to address the intrusion detection problem in IoT networks. Can Blockchain Technology Enhance Security and Privacy in the
Internet of Things?
After proposing a smart binary and multiclass [13] Dr.MushthaqAhammed, K. (2017). Online Transaction Security
classification scheme using a feed-forward neural networks Issues In A Digitally Enabled Business Environment.
model, extracting and preprocessing the field information in [14] Bangare, P.S., & Patil, K.P. (2022). Security Issues and Challenges in
individual packets as generic features, and testing its efficacy Internet of Things (IOT) System. 2022 2nd International Conference
for the recently published IoT dataset with realistic network on Advance Computing and Innovative Technologies in Engineering
(ICACITE), 91-94.
traffic, we have been able to demonstrate the performance of
[15] Kumar, R., Sharad, A., Jaiswal, S.B., Singh, V.K., & Kumari, S.
the proposed scheme. (2022). Privacy and Security Issues and Challenges in 5g Enabled
Results showing almost 0.99 on all assessment Vehicular Networks.
parameters, including accuracy, precision, recall, and F1 [16] Abishu, H.N., Seid, A.M., Yacob, Y.H., Ayall, T., Sun, G., & Liu, G.
(2022). Consensus Mechanism for Blockchain-Enabled Vehicle-to-
score, demonstrate the classifier's ability to categorise binary Vehicle Energy Trading in the Internet of Electric Vehicles. IEEE
data for DDoS/DoS and reconnaissance assaults. The Transactions on Vehicular Technology, 71, 946-960.
classifier claimed an accuracy of 0.99 or higher for detecting [17] Alqarni, H.M., Alnahari, W., & Quasim, M.T. (2021). Internet of
DDoS/DoS assaults in multiclass classification, while Things (IoT) Security Requirements: Issues Related to Sensors. 2021
achieving an accuracy of 0.98 when classifying normal National Computing Colleges Conference (NCCC), 1-6.
traffic. [18] Singh, O., Anand, A., Agrawal, A.K., & Singh, A.K. (2022).
Electronic Health Data Security in the Internet of Things through
Those results are significant, and they warrant more Watermarking: An Introduction. IEEE Internet of Things Magazine,
research into the topic of Internet of Things (IoT) network 5, 55-58.
security. [19] Liu, L., & Li, Z. (2022). Permissioned blockchain and deep
reinforcement learning enabled security and energy efficient
Healthcare Internet of Things. IEEE Access, PP, 1-1.
[20] Geetha, R., Suntheya, A.K., & Srikanth, G.U. (2020). Cloud
Integrated IoT Enabled Sensor Network Security: Research Issues and
Solutions. Wireless Personal Communications, 113, 747-771.

Authorized licensed use limited to: Consortium - Algeria (CERIST). Downloaded on April 18,2023 at 22:56:45 UTC from IEEE Xplore. Restrictions apply.