Detection of DDoS Attacks on Urban IoT Devices Using Neural
Networks
Simon Onyebuchi Obetta and Arghir-Nicolae Moldovan
a
School of Computing, National College of Ireland, Dublin, Ireland
Keywords: IoT, DDoS Attacks, Intrusion Detection, Machine Learning, Neural Networks.
Abstract: As the Internet of Things (IoT) has grown in recent years, attackers are increasingly targeting IoT devices to
perform malicious attacks such as DDoS. Often, this is due to inadequate security implementation and
management of IoT devices. Sometimes, the infected IoT devices can be used as bots by attackers to launch
a DDoS attack on a target. Although various security methods have been introduced for IoT devices, effective
DDoS detection methods are still required. This paper compares the performance of four machine learning
algorithms for DDoS detection on a recent Urban IoT dataset: Feedforward Neural Network (FNN), Deep
Neural Network (DNN), Autoencoder (AEN) and Random Forest (RF). The results show that DNN achieved
the highest accuracy of 95.9% on train data and 88.6% on test data.
1 INTRODUCTION
The Internet of Things (IoT) is a technology that
connects smart electronic devices to the Internet for
data collection and transfer without human
intervention. Presently many IoT systems are
interconnected with several sensors and maintain
communication and exchange massive volumes of
data. For instance, in the context of smart-home
applications, large-scale IoT systems with numerous
sensor nodes are being used and proposed. Common
network architectures that utilize IoT services include
healthcare systems, institutions, organizations, and
home network systems. For communication between
the IoT devices and the controller, the majority of IoT
implementations in smart homes rely heavily on
home internet networks, either wireless or cable. IoT
devices enable smarter and more efficient homes by
allowing for automatic and remote control of
household equipment. For example, modern CCTV
cameras can now be monitored from afar using
smartphones.
IoT devices are also prone to security
vulnerabilities and increasingly targeted by attackers,
thus, IoT is becoming a major research area in the
realm of cybersecurity. The most prevalent IoT
security threats comprise code injection, middle-man
a
https://orcid.org/0000-0003-4151-1432
attack, sinkhole, Sybil attack, Denial of Service
(DoS), and Distributed Denial of Service (DDoS)
(Vashi et al., 2017). According to Cloudflare (2021),
DDoS attacks occur when an attacker floods the
target's network or application with fake requests
from a compromised botnet. It takes advantage of
vulnerabilities such as unsecured ports, use of default
passwords, unpatched software, to penetrate the
targets' system (Douligeris and Mitrokotsa, 2004).
The DDoS attackers aim to deny access to legitimate
users by taking down or slowing the network or
application.
Many recent anomaly-based detection studies
have compared the performance of different machine
learning algorithms (MLA) and showed they have
good potential in detecting malware and DDoS in
computer networks traffic. However, more research is
needed on detecting real-life or slow DDoS attacks,
especially in IoT devices.
This research aims to investigate DDoS attacks
detection in IoT devices using the recent Urban IoT
dataset (Hekmati et al., 2021). This will be performed
by doing a comprehensive performance comparison
of three Neural Network-based algorithms (i.e.,
Feedforward Neural Network, Deep Neural Network,
Autoencoder), and the traditional Random Forest
algorithm.
236
Obetta, S. and Moldovan, A.
Detection of DDoS Attacks on Urban IoT Devices Using Neural Networks.
DOI: 10.5220/0011998900003482
In Proceedings of the 8th International Conference on Internet of Things, Big Data and Security (IoTBDS 2023), pages 236-242
ISBN: 978-989-758-643-9; ISSN: 2184-4976
Copyright
c
2023 by SCITEPRESS Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
The rest of the paper is structured as follows.
Section 2 presents related works. Section 3 presents
the methodology. Section 4 presents the results, while
section 5 concludes the paper.
2 RELATED WORKS
This section provides an overview of several previous
research papers on DDoS detection using machine
learning methods.
Ashi et al. (2020) investigated DDoS attacks
detection with an emphasis on cloud computing
architecture. After collecting 256 Uniform Resource
Locators, the authors used four different systems to
simulate a DDoS attack simultaneously (URLs). A
dataset comprising the simulation's network traffic
flow was created, and Random Forest (RF) was
utilized for model testing.
Rahman et al. (2019) created an SDN framework
to identify and defend against DDoS attacks on the
controller and the switch. To predict DDoS attacks,
this framework requires training a machine learning
model with recorded data. The mitigation script then
uses the prediction to make decisions on the SDN
network. With an open-source DDoS dataset, they
tested and compared the results for Support Vector
Machine (SVM), K-Nearest neighbours (K-NN), J48,
and RF. The results of their experiment revealed that
J48 is the best classifier with accuracy, F-1, and recall
rate of 100%.
Reddy and Thilagam (2020) applied Naive Bayes
(NB) classifier to detect DDoS attack traffic by
considering the five most influential DDoS attack
network factors. Based on the probability of the
DDoS attack value, the proposed DDoS attack
classifier is applied on all monitor nodes to process
valid traffic and remove DDoS attack traffic.
According to simulation results, the proposed strategy
reduces the intensity of DDoS attacks and allows
network nodes to handle up to 80% of legal traffic.
Misbahuddin and Zaidi (2021) classified DDoS
attacks by using a semi-supervised machine learning
approach on the CICDS2017 dataset. They began
with unlabelled traffic information collected against
three aspects for victim-end defence, namely the
webserver. Two distinct clustering methods were
used to group unlabelled data, and a voting procedure
determines the final classification of traffic flows. To
detect DDoS attacks, the supervised learning
algorithms K-NN, SVM and RF are applied to
labelled data, with accuracy achieved of 95%, 92%,
and 96.66%, respectively.
Rios et al. (2021) tested and compared the Multi-
Layer Perceptron (MLP), K-NN, SVM, and
Multinomial Naive Bayes (MNB) machine learning
methods for detecting reduction of quality (RoQ)
attacks. They also suggested a method for detecting
RoQ attacks that combines three models: Fuzzy Logic
(FL), MLP, and Euclidean Distance (ED). They
tested these methods using both simulated and real-
world traffic patterns. They demonstrated that using
three parameters, namely the number of packets,
entropy, and average inter-arrival time, results in the
better categorization of the four machine learning
algorithms than using only entropy. MLP
outperformed the other four machine learning
algorithms when it comes to detecting RoQ attacks.
Doshi, et al. (2018) investigated multiple machine
learning algorithms K-NN, Linear SVM, Decision
Tree (DT), RF and Neural Network (NN) for DDoS
detection for consumer IoT. Their classification
algorithm was based on the idea that system traffic
conditions from these IoT nodes differ from those
from well-studied non-IoT network nodes. They used
data from a consumer IoT device that included both
normal and DoS attack traffic to test five different
machine learning classifiers. The results show
variations in accuracy, F1, recall, and precision across
the models. With K-NN, DT, RF, and NN having
99.9% accuracy while LSVM 99.1%.
Mishra et al (2021) investigated DDoS attacks
detection in cloud computing. The machine learning
algorithms adopted for classification were K-NN, NB
and RF. They generated a long feature vector by
merging all feature vectors of interest. Their focus
was more on supervised learning with the Random
Forest having the best accuracy of 99.58%.
Hekmati et al. (2021) proposed a simple Feed-
forward Neural Network for DDoS detection
employing 20 nodes out of 4060 in the original
dataset for the Urban IoT DDoS dataset. They also
provide a script for creating a benign dataset from the
original dataset to eliminate bias toward nodes with
higher activity. The authors used attack emulation to
generate an artificial DDoS attack for the attack ratio
of 1 on the 20 selected IoT nodes. The simple FNN
achieved a mean accuracy of 94% and 88% on the
train and test data, respectively.
Shaaban et al. (2019) proposed the use of a
Convolutional Neural Network (CNN) for DDoS
detection. For their research, the authors used two
datasets: a generated dataset and the NSL-KDD
dataset. The results showed that CNN achieved 99%
accuracy, and outperformed other algorithms like DT,
SVM, K-NN and NN.
Detection of DDoS Attacks on Urban IoT Devices Using Neural Networks
237
3 METHODOLOGY
Fig. 1 shows the steps of the research methodology.
Figure 1: Workflow of the research methodology.
3.1 Data Selection
The dataset selected for this research is the Urban IoT
that was captured from the activity status of genuine
event-driven IoT nodes installed in a city (Hekmati et
al., 2021). The data captured is the activity of 4060
urban IoT devices (nodes) for one month, making it a
very realistic IoT dataset. As it is a recent dataset
there are very few papers that have used it and there
are open research questions.
The original dataset contains the node ID and
location (latitude and longitude). The dataset also
contains each node's binary activity status at 30
seconds interval over a month in a benign (non-
attacked) environment. When a node's activity status
changes, a record is appended to the original dataset.
It was later supplemented with artificial attack
emulation to make it usable for training machine
learning models for DDoS detection. From the dataset
statistics, up to 65% of the nodes are active at midday,
but by midnight, only approximately 20% of the
nodes are active.
3.2 Data Cleaning
Data cleaning is part of the pre-processing activity
carried out on the original dataset. Data cleaning has
the potential to increase the efficiency and
effectiveness of the training process. To emphasize
the role of data cleaning in Machine Learning
processes, it has been discovered that even when
utilizing robust statistical techniques, the data
cleaning methodology chosen can have a
considerable impact on overall results (Krishnan et
al., 2016). It comprises work such as removing
extraneous data, dealing with missing values, label
conversion, categorization, and data standardization.
In this paper the data cleaning focused on generating
benign data for 20 randomly selected IoT nodes. The
benign dataset contains the occupancy status of each
node between the start and end dates available in the
original dataset, with a time step of 30 seconds. An
additional attribute, attacked, with the value set to 0,
was also added.
3.3 Attack Emulation
An artificial attack emulation was provided by the
dataset authors, to make it suitable for training
machine learning models for DDoS detection (ANRG
USC, 2021). Emulation overcomes the challenges
associated with performing real DDoS attacks on
active IoT nodes, which may not be allowed by the
IoT nodes owner, and it may not be feasible to
perform large scale DDoS attacks on many IoT nodes.
The attack emulation step created new attributes such
as begin_date, end_date, num_nodes, attack_ratio,
attack_duration.
3.4 Train Machine Learning Models
The ML algorithms employed in the experiment
include one traditional algorithm, Random Forest
(RF), and three neural network-based algorithms.
Feed-Forward Neural Network: This is the
simplest type of neural network that only has one
hidden layer (Jurafsky and Martin, 2021). The input
layer sends a multi-dimensional request to the hidden
layer and is processed using a weighted summation
and an activation function. It is trained using labelled
data and a learning algorithm that optimizes the
summation model's weights. The hidden layer is
linked to the input layer and the output layer link to
the hidden layer. The aim of using this model is to
reproduce the DDoS detection done by Hekmati et al.
(2021) using the same dataset. The FNN implemented
consists of a 12-neuron input layer, followed by a
single hidden layer with 8-neurons and ReLU
activation. At the end of the hidden layer, a 20%
dropout is employed, as well as batch normalization.
The output layer consists of a single neuron with the
Sigmoid activation function. To discover the attacked
IoTBDS 2023 - 8th International Conference on Internet of Things, Big Data and Security
238
time slots in the dataset, the neural network model is
trained for 500 epochs for each node.
Deep Neural Network: This model is made up of
feedforward neural networks that do not have any
feedback connections. DNN consists of the input and
output layers, but the main distinction from FNN is
that it has more than one hidden layer. Each layer
contains units with weights. The activation processes
of the units from the previous layer are carried out by
these units (Pande et al., 2021). Because the DNN
model's structure combines feature extraction and
classification operations, it benefits from both
supervised and unsupervised learning. In addition, the
multiple hidden layers architecture can automatically
uncover complex correlations and mappings from
input to output data that are not compatible with other
non-deep neural networks. This can lead to an
increase in the overall performance. The input layer
of the DNN implementation consists of 30 units with
a ReLU activation function, followed by two hidden
layers of 10 units each with a dropout of 0.4 and the
ReLU activation function. The sigmoid activation
function is used in the output layer.
Autoencoder Neural Network: This is a type of
neural network that shrinks multidimensional input
data within a hidden region before reconstructing the
data from the hidden region (Ozgur and Fatih, 2019).
In this research, the tanh activation function was used
for model training. The autoencoder is divided into
encoder, code, and decoder. The encoder is the region
that sits between the input layer and the hidden layer.
The encoding region enables the reduction of
multidimensional data to a lower size. The decoding
region is located between the hidden layer and the
output layer. The code region is between the encoder
and decoder. By increasing the size of the shrunk
hidden layers, the decoder attempts to reconstruct the
input. The reason why an Autoencoder model is used
is that like DNN, it is a multi-hidden layer neural
network that can uncover complex correlations and
mappings of data and increase the performance. For
example, Ozgur and Fatih, (2019) used this model to
propose DDoS attack detection in their study using
the kdd99 dataset because the model has an advantage
in terms of removing outliers and fixing complexes in
a dataset. In the implemented Autoencoder, encoder
is made up of three dense layers, which have 64, 32,
and 16 units respectively, and the "tanh" activation
function for each layer. The result of this encoder
generates code that the decoder subsequently uses to
reconstruct its input. The decoder on the other hand
comprises three dense layers with the same units and
tanh activation function. The output function
processes the result of the input function using a
single layer sigmoid activation.
Random Forest: This is a traditional machine
learning algorithms that is based on the construction
of numerous small trees in a decision tree (Dangwal
& Moldovan, 2021). By using a bagging method, the
results of each small tree are combined with a
weighted value to provide a final prediction outcome.
To reach the final predicted conclusion, this approach
employs the mean of the individual small trees.
According to Mishra et al., (2021), RF is
recommended for supervised learning since it
produces much better results than other machine
learning algorithms. This is because Random Forest
is less prone to overfitting than the alternative
Decision Tree since it employs an ensemble of
Decision Trees, with the values in the tree being a
random, independent sample. The implemented
random forest classifier uses ‘n_samples=1000’,
‘n_features=20’, 'random state=3', n split=10, n
repeats=3, and n jobs =-1.
3.5 Evaluation
The performance of the four models to detect DDoS
attacks is compared based on different metrics,
namely accuracy, recall, and precision. The data is
split into 80% for training and 20% for test. The
performance of the models is compared for 2 different
approaches, (1) implementing 20 models one for each
of the 20 IoT nodes, and (2) implement one model on
all the combined data from 20 IoT nodes. The 2
nd
approach was not used in previous papers on this
dataset. Moreover, two values were used for the
attack ratio of 1 and 0.8.
4 RESULTS
The results are presented in Tables 1 to 4. Tables 1
and 2 illustrate the results for the four ML algorithms
used for attack ratio of 1, which show that overall
DNN achieved the highest accuracy. For the 20
models on 20 IoT nodes, DNN achieved an accuracy
of 95.9% and 88.6% for train and test datasets,
respectively. These results are slightly better than the
results of prior work on the same dataset that used a
Feedforward Neural Network to identify DDoS, with
accuracy 94% (Hekmati et al. 2021). For the 1 model
built on the combined data of 20 IoT nodes, DNN
again achieved the highest accuracy of 87.4% and
85.4% on the train and test datasets, respectively.
Detection of DDoS Attacks on Urban IoT Devices Using Neural Networks
239
Table 1: Results for 20 models on 20 IoT nodes, attack ratio
1.
Model Mean
Accurac
y
Mean
Recall
Mean
Precision
Train
Data
FNN 0.943 0.936 0.794
DNN 0.959 0.942 0.842
Autoencode
r
0.957 0.936 0.839
RF 0.958 0.942 0.842
Test
Data
FNN 0.870 0.835 0.680
DNN 0.886 0.824 0.694
Autoencode
r
0.883 0.791 0.687
RF 0.886 0.819 0.694
Table 2: Results for 1 model on 20 IoT nodes, attack ratio
1.
Model Mean
Accuracy
Mean
Recall
Mean
Precision
Train
dataset
FNN 0.796 0.236 0.662
DNN 0.874 0.894 0.645
Autoencode
r
0.846 0.955 0.666
RF 0.846 0.955 0.666
Test
dataset
FNN 0.801 0.242 0.699
DNN 0.854 0.875 0.640
Autoencode
r
0.839 0.916 0.643
RF 0.839 0.916 0.643
Table 3: Results for 20 models on 20 IoT nodes, attack ratio
0.8.
Model Mean
Accuracy
Mean
Recall
Mean
Precision
Train
Data
FNN 0.933 0.942 0.726
DNN 0.956 0.945 0.799
Autoencode
r
0.958 0.938 0.803
RF 0.958 0.938 0.803
Test
Data
FNN 0.857 0.817 0.583
DNN 0.883 0.801 0.610
Autoencode
r
0.885 0.762 0.616
RF 0.885 0.762 0.616
Table 4: Results for 1 model on 20 IoT nodes, attack ratio
0.8.
Model Mean
Accuracy
Mean
Recall
Mean
Precision
Train
dataset
FNN 0.832 0.593 0.622
DNN 0.854 0.920 0.569
Autoencode
r
0.864 0.905 0.591
RF 0.854 0.920 0.569
Test
dataset
FNN 0.827 0.624 0.605
DNN 0.832 0.917 0.554
Autoencode
r
0.839 0.894 0.574
RF 0.832 0.917 0.554
Overall, the algorithms show higher performance
for the 1
st
approach of training 20 models for each of
the 20 IoT nodes than for the 2
nd
approach of training
one model on the combined data of 20 IoT nodes. For
example, the test accuracy for 20 models on 20 nodes
approach, ranges from 87.0% to 88.6%, whereas the
accuracy for one model on 20 nodes varies from
80.1% to 85.4%.
Tables 3 and 4 illustrate the results for the four
ML algorithms used for attack ratio of 0.8, which
show that overall Autoencoder achieved the highest
accuracy. For the 20 models on 20 IoT nodes,
Autoencoder achieved an accuracy of 95.8% and
88.5% for train and test datasets. For the 1 model built
on the combined data of 20 IoT nodes, Autoencoder
again achieved the highest accuracy of 86.4% and
83.9% on the train and test datasets, respectively.
These results are based on data from 20 IoT nodes
out of a total of 4060 IoT nodes, and only for two
attack ratios. As a result, future studies are still
required that could employ more IoT nodes and
different attack ratios.
Figures 2 to 3 show the true positive (TP) and
false positives (FP) for 20 models on 20 nodes for
attack ratio 1, on the training and test datasets,
respectively. The duration was set to 16 hours in these
experiments. The results indicate that there are fewer
FP for training data than for test data. Figures 4 and 5
show the DNN performance when using one model
on 20 nodes for arrack ratio 1. The explanation for
poorer performance of the one model on 20 nodes is
the lower true positive rates and during the attack
windows.
5 CONCLUSIONS
The research addresses the improvement in the
performance of machine learning models at detecting
DDoS attacks on Internet of Things devices using the
Urban IoT DDoS dataset. Four machine learning
algorithms were investigated: Feedforward Neural
Network (FNN), Deep Neural Network (DNN),
Autoencoder, and Random Forest (RF). Two
approaches were used for the comparison, building 20
models for 20 IoT nodes each, and building 1 model
on the combined data of the 20 nodes.
The results showed that DNN can classify DDoS
data with slightly better accuracy than the other three
algorithms. However, when compared to the other
algorithms, the DNN took a long time to train and
test. As a result, there is an opportunity for
improvement, and fine-tuning the model that may
enable it to train faster. Future work may investigate
other algorithms such as Convolutional Neural
Networks (CNN) with larger number of IoT nodes to
do further research on the same dataset.
IoTBDS 2023 - 8th International Conference on Internet of Things, Big Data and Security
240
Figure 2: DNN training dataset attack prediction vs. time,
for 20 models on 20 IoT nodes, attack ratio 1.
Figure 3: DNN test dataset attack prediction vs. time, for 20
models on 20 IoT nodes, attack ratio 1.
REFERENCES
Ashi, Z., Aburashed, L., Al-Fawa’reh, M., and Qasaimeh,
M. (2020) ‘Fast and Reliable DDoS Detection using
Dimensionality Reduction and Machine Learning' 2020
15th International Conference for Internet Technology
and Secured Transactions (ICITST), London, United
Kingdom, December 2020, pp. 1-10, doi: 10.23919/
ICITST51030.2020.9351347.
Cloudflare (2021) ‘What is a DDoS attack?’ Available at:
https://www.cloudflare.com/learning/ddos/what-is-a-
ddos-attack/.
Dangwal, S., & Moldovan, A.-N. (2021). Feature Selection
for Machine Learning-based Phishing Websites
Detection. 2021 International Conference on Cyber
Situational Awareness, Data Analytics and Assessment
(CyberSA), 1–6. doi: 10.1109/
CyberSA52016.2021.9478242
Doshi, R., Apthorpe, N. and Feamster, N. (2018) ‘Machine
Learning DDoS Detection for Consumer Internet of
Figure 4: DNN training dataset attack prediction vs. time,
for 1 model on 20 IoT nodes, attack ratio 1.
Figure 5: DNN test dataset attack prediction vs. time, for 1
model on 20 IoT nodes, attack ratio 1.
Things Devices’, 2018 IEEE Security and Privacy
Workshops (SPW), San Francisco, CA, USA, May
2018, pp. 29-35, doi: 10.1109/SPW.2018.00013.
Douligeris, C. and Mitrokotsa, A. (2004) ‘DDoS attacks
and defense mechanisms: classification and state-of-
the-art’, Proceedings of the 3rd IEEE International
Symposium on Signal Processing and Information
Technology (IEEE Cat. No.03EX795), 44(5), pp. 190-
193, doi: 10.1109/ISSPIT.2003.1341092.
ANRG USC, ‘Urban_IoT_DDoS_Data’. Available at:
https://github.com/ANRGUSC/Urban_IoT_DDoS_Dat
a/.
Hekmati, A., Grippo, E., and Krishnamachari, B., (2021)
'Dataset: Large-scale Urban IoT Activity Data for
DDoS Attack Emulation'. ACM Conference on
Embedded Networked Sensor Systems, Coimbra,
Portugal, November 2021, pp. 560–564 doi: org/
10.1145/3485730.3493695.
Jurafsky, D and Martin, J., (2021) 'Neural Networks and
Neural Language Models' Available at: https://
web.stanford.edu/~jurafsky/slp3/7.pdf.
Detection of DDoS Attacks on Urban IoT Devices Using Neural Networks
241
Krishnan, S., Franklin, J., Goldberg, K., Wang, J. and Wu,
E. (2016) 'ActiveClean: An Interactive Data Cleaning
Framework For Modern Machine Learning' In
Proceedings of the 2016 International Conference on
Management of Data (SIGMOD '16), New York, June
2016, pp. 2117–2120, doi: 10.1145/2882903.2899409.
Misbahuddin, M. and Zaidi, S. M., (2021) 'Clustering based
semi-supervised machine learning for DDoS attack
classification' Journal of King Saud University –
Computer and Information Sciences 33, pp. 436–446,
doi: 10.1016/j.jksuci.2019.02.003.
Mishra, A., Gupta, B., Perakovic, D., Penalvo, F. and Hsu,
C. (2021) 'Classification Based Machine Learning for
Detection of DDoS attack in Cloud Computing' 2021
IEEE International Conference on Consumer
Electronics (ICCE), Las Vegas, NV, USA, January
2021, pp. 1-4, doi: 10.1109/ICCE50685.2021.9427665.
Ozgur, C. F. and Fatih, M. A. (2019) ‘Distributed Denial of
Service Attack Detection Using Autoencoder and Deep
Neural Networks’ Journal of Intelligent & Fuzzy
Systems. 37. pp. 1-11, doi: 10.3233/JIFS-190159.
Pande S., Khamparia A., Gupta D. and Thanh D.N.H.
(2021) ‘DDOS Detection Using Machine Learning
Technique. In: Khanna A., Singh A.K., Swaroop A.
(eds) ‘Recent Studies on Computational Intelligence.
Studies in Computational Intelligence’, Springer,
Singapore. Studies in Computational Intelligence, 921,
pp. 59-68, doi: org/10.1007/978-981-15-8469-5_5
Rahman, O., Ali, M., Quraishi, G. and Lung, C. (2019)
'DDoS Attacks Detection and Mitigation in SDN using
Machine Learning' 2019 IEEE World Congress on
Services, Milan, Italy, July 2019, 2642-939X, pp. 184-
189, doi: 10.1109/SERVICES.2019.00051.
Reddy, K. and Thilagam, P. (2020) 'Naïve Bayes Classifier
to Mitigate the DDoS Attacks Severity in Ad-Hoc
Networks' International Journal of Communication
Networks and Information Security (IJCNIS) 12(2), pp.
221-226, doi: 10.54039/ijcnis.v12i2.4574 ,
Rios, V. M., Pedro, R.M., Magoni, D. and Freire, M. M.
(2021) 'Detection of reduction-of-quality DDoS attacks
using Fuzzy Logic and machine learning algorithms',
Computer Networks, 186, pp.107792, doi: 10.1016/
j.comnet.2020.107792
Shaaban, A. R., Abd-Elwanis, E. and Hussein, M. (2019)
'DDoS attack detection and classification via
Convolutional Neural Network (CNN)' 2019 Ninth
International Conference on Intelligent Computing and
Information Systems (ICICIS), Cairo, Egypt, December
2019, pp. 233-238, doi: 10.1109/
ICICIS46948.2019.9014826.
Vashi, S., Verma, S., Ram, J., Modi, J. and Prakash, C.
(2017) ‘Internet of Things (IoT): A vision, architectural
elements, and security issues' 2017 International
Conference on I-SMAC (IoT in Social, Mobile,
Analytics and Cloud) (I-SMAC). Palladam, India,
February 2017, pp. 492-496. doi: 10.1109/I-
SMAC.2017.8058399.
IoTBDS 2023 - 8th International Conference on Internet of Things, Big Data and Security
242