Using deep learning to detecting abnormal behavior in internet of things

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 12, No. 2, April 2022, pp. 2108~2120
ISSN: 2088-8708, DOI: 10.11591/ijece.v12i2.pp2108-2120  2108
Journal homepage: http://ijece.iaescore.com
Using deep learning to detecting abnormal behavior in internet
of things
Mohammed Al-Shabi, Anmar Abuhamdah
Department of Management Information System, College of Business, Taibah University, Tayba, Medina, Saudi Arabia
Article Info ABSTRACT
Article history:
Received Jun 11, 2021
Revised Jul 19, 2021
Accepted Aug 14, 2021
The development of the internet of things (IoT) has increased exponentially,
creating a rapid pace of changes and enabling it to become more and more
embedded in daily life. This is often achieved through integration: IoT is
being integrated into billions of intelligent objects, commonly labeled
“things,” from which the service collects various forms of data regarding
both these “things” themselves as well as their environment. While IoT and
IoT-powered devices can provide invaluable services in various fields,
unauthorized access and inadvertent modification are potential issues of
tremendous concern. In this paper, we present a process for resolving such
IoT issues using adapted long short-term memory (LSTM) recurrent neural
networks (RNN). With this method, we utilize specialized deep learning
(DL) methods to detect abnormal and/or suspect behavior in IoT systems.
LSTM RNNs are adopted in order to construct a high-accuracy model
capable of detecting suspicious behavior based on a dataset of IoT sensors
readings. The model is evaluated using the Intel Labs dataset as a test
domain, performing four different tests, and using three criteria: F1,
Accuracy, and time. The results obtained here demonstrate that the LSTM
RNN model we create is capable of detecting abnormal behavior in IoT
systems with high accuracy.
Keywords:
Deep learning
Detection of abnormal behavior
Internet of things
Long short-term memory
neural networks
Machine learning
This is an open access article under the CC BY-SA license.
Corresponding Author:
Anmar Abuhamdah
Department of Management Information System, Faculty of Business Administration, Taibah University
Janadah Bin Umayyah Road، Tayba, Medina 42353, Saudi Arabia
Email: aabuhamdah@taibahu.edu.sa; anmar_81@hotmail.com
1. INTRODUCTION
Internet of Things (IoT) is a technical term that has spread to both the digital and physical worlds
following its introduction by British scientist Kevin Ashton, a technology pioneer and the founder of the first
AutoID lab at Massachusetts Institute of Technology [1]. The driving concept of IoT was to connect common
digital devices (often everyday or household electrical appliances) with one another in ways that enable users
to monitor their status and take in accurate information from the network of devices without needing to be in
close proximity to any of these devices. Put differently, IoT is a system comprised of various
components (that may include machines, computing devices, people, or animals) that each possess unique
identifiers (UID) and a network connection that enables them to transfer data without human interaction. In
other words, IoT is describing the network's physical objects as the “things” of its title, which have been
enhanced using technologies that enable them to link to and exchange data with other systems or devices
using the internet [1]–[3].
Moreover, the IoT extends the internet of behavior (IoB), which is the interconnection of devices
that results in a massive diversity of new data sources. While the behavior of an IoT network and its devices
may sometimes be abnormal and could indicate a security threat, identifying this kind of behavior and

Int J Elec & Comp Eng ISSN: 2088-8708 
Using deep learning to detecting abnormal behavior in internet of things (Mohammed Al-Shabi)
2109
distinguishing valid concerns from actual security threats is considered a major challenge in IoT [4]–[6].
There are several methods that can be applied here, but machine learning is considered one of the most
promising areas capable of solving various different problems. Researchers have applied various techniques
to improve the accuracy of machine learning (ML) algorithms, which have done much work to extend and
improve the capacities of machines. ML processes work from the fact that learning, a natural part of human
behavior, can be made into a vital part of machines as well. Deep learning (or deep neural network [DNN]) is
a subcategory of ML that has been used in numerous applications to notable result, demonstrating its
applicability for many real-world applications [5], [7]. Also called hierarchical learning, DL has been
growing in popularity since 2006 as a new field in ML research, and deep learning (DL) techniques have
evolved over the past years and affected many areas. DL can be defined as a subfield within ML that depends
on the representations of several learning levels that correspond to a hierarchical structure of features, factors,
or concepts, and high-level features, where concepts are defined by relying on lower-level concepts, and
concepts at various levels may help the lower level itself define many top-level concepts [8].
The goal of the present study is to investigate and evaluate a DL model capable of detecting
abnormal behavior in IoT networks. The model we propose here depends on DL, taking advantage of
structures such as deep neural networks (DNN), deep certainty networks (DCN), and recurrent neural
networks (RNN). In doing so, it seeks results that may realize the vision behind the current trend of
researchers using DL in several fields of application, such as healthcare, banking services, industry, and
many more.
In this work, we adopt DL techniques in order to propose a model for detecting abnormal suspicious
behavior in IoT systems. Our model utilizes long short-term memory (LSTM) recurring neural
networks (RNNs), a subset of DNNs. In order to assess the model's performance, we perform four different
tests on Intel Labs dataset and compare the test results based on two criteria, F1 and accuracy. This remainder
of this work is organized as follows. Section 2 presents a literature review, while Section 3 describes the
proposed methodology for achieving our own model or algorithm. From there, Section 4 presents the results
of these experiments and a discussion of our findings, and finally Section 5 serves as a conclusion,
summarizing our research findings and providing recommendations for future work in this field.
2. LITERATURE REVIEW
There are several methods applied in different fields. However, ML is considered one of the current
areas that may solve many different problems. Researchers have applied various techniques to enhance the
performance and precision of ML algorithms in order to improve their results. Learning is a natural part of
human behavior that can be turned into a dynamic part of machines as well. deep learning (DL) and deep
neural networks (DNNs) constitute a subcategory of ML that has been analyzed and applied in numerous
applications, where it achieves notable results that hold promise for various real-time applications [5], [7].
DL also sometimes called hierarchical learning, has been in use since 2006 as a new subfield of ML research,
though over the past years it has evolved further. DL depends on learning several levels of representations
(which correspond to a certain hierarchical structure of features, factors, or concepts) and high-level features.
Here, important concepts are often defined by relying on lower-level concepts, while elsewhere this
organization may help the lower level itself define many top-level concepts [8]. The following sub-sections
describes various techniques of both ML and DL algorithms and then providers an overview of several
previous studies.
2.1. Machine learning
Machine learning (ML) describes a field of computer science intended to enable computers to learn
without programming directly from humans [5], [9]. ML was initially defined in 1959 by IBM engineer
Arthur Samuel, who described the nascent area as follows: “A field of study that gives computers the ability
to learn without being explicitly concluded” [10]. ML algorithms can be divided into supervised learning
algorithms (including classification, regression, standardization, support vector machines (SVM), decision
trees, Bayesian, and others) and unsupervised learning algorithms (such as clustering, neural networks, and
others) [11]–[13]. Beyond these two types, several ML algorithms perform multiple functions [5], [7].
2.2. Deep learning techniques
Deep learning (DL) is one facet of a larger family of representational ML methods. DL builds from
the fact that concepts and images can be represented in several ways, but at the same time there are certain
representations of these images that make learning tasks faster and more efficient [8]. DL is the field of
research that is interested in finding this definition of representations and how to teach machines using it.
Table 1 summarizes the most important differences between ML and DL approaches and methods [14].

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 12, No. 2, April 2022: 2108-2120
2110
DL also includes its own methods, such as neural networks (NN) and recurrent neural
networks (RNN) [15]. Different criteria are typically utilized in order to evaluate the performances of these
methods such as the accuracy and F1-score (F1). Accuracy, which is constituted of the ratio of correctly
projected views to total views, constitutes the most widely-used and accepted measure of performance.
Meanwhile, F1 results take both positive and false positives into account, providing results that are only more
useful than accuracy if researchers have a non-distribution equal for class.
Table 1. Comparison between machine learning and deep learning
Machine Learning Deep Learning
Data dependencies Excellent tool for small/medium data set An excellent tool for a Big data set
Devices dependencies It works on standard devices (low-end) Requires powerful devices
Engineering features Request to understand the features that represent data Not requested to better understand the features
that represent data
Execution time A few minutes to hours Up to weeks, because of the calculation of weights
Expandability
Some algorithms are easy to interpret, others are
nearly impossible
Difficult to understand or impossible
Neural networks (NN) are a set of algorithms inspired by the performance of the human brain,
which takes a large set of data and processes it to extract patterns from data. NN are classified according to
the model of the connection between neurons to single-layer networks and multi-layer networks and this is
called a structure. According to the method of assigning the weights accompanying the links to learning
(which may be done in the presence of a supervisor or learning unsupervised), this is called “training” and is
the process of the machine actually learning [15].
The main activity of NN is the classification and coding process. NN are approaches as general
functions, which means that neural networks can learn to approximate any function (i.e. fx()) of the type of
general approximation functions [16]. Recurrent neural networks (RNN) describe a NN model first proposed
during the 1,980 by researchers attempting to model time series [17]. RNN is considered to be the first
algorithm that uses a memory to save or remember the previous entries from a large set of serial data. The
structure of this network is like a standard multi-layer receiver but differs in allowing the connections
between hidden layers that associated with a specific time delay. The model will preserve knowledge about
the past through these connections, enabling it to discover the temporal relations in the data between events
far apart from one another. Traditional NNs assume that every input and every output is independent of every
other input and output, which in many cases does not provide optimal efficiency. For example, when
attempting to predict the next word or phrase in a full sentence, better and faster results will be possible if the
previous word(s) are known. RNNs rely on a similarly iterative chain of information, and then perform an
identical task at each segment of the chain, wherein the output relies on prior calculations while also
preserving memories that keep the information calculation). Figure 1 illustrate a pattern of RNNs [18], [19].
Figure 1. Pattern of recurrent neural networks (RNN)
In Figure 1, xT represent the input data (i.e. x1 , x2 , x3, …, xT), while, or represents the next expected
(predicate) data (i.e. o1 , o2 , o3, …, or), and hr represents the hidden layers that contains information about the
previous income data. The following equation is used in the network training process:
ℎ𝑟 = 𝑓(𝑊𝑊(ℎℎ)
ℎ𝑟 + 𝑊(ℎ𝑥)
) (1)

2111
where hT holds the information about the previous data in the chain and calculated by using the previous hr-1
beam and the xr input current beam. A non-linear activation function f (usually tanh, sigmoid, or ReLU) is
also applied upon the final assembly. Here, it is reasonable to take it that h0 is a zero beam:
𝑦𝑟 = 𝑠𝑜𝑓𝑡𝑚𝑎𝑥(𝑊(𝑠)
ℎ𝑟 ) that calculates the expected beam of data at a given time step, or t. In this instance,
we use the softmax function to get a beam (1, V) whose sum of all elements is equal to one. This probability
distribution provides us with the following (and most probable) income indicator from that income data:
𝐽(𝑟)
(𝜃) = ∑ (𝑦𝑦𝑟𝑖
|𝑉|
𝑖=1 log 𝑦𝑟𝑖 ) (2)
Here we use what is known as loss function to account for the continued loss of cross entropy at
each t in order to calculate the level of error between the expected statement and the actual result, which is
usually referred to as the average squared error or mean squared error (MSE). Regrettably, if we implement
the steps above, then the results will not be satisfactory. This is because the simple model of RNNs comes
with a major weakness, often termed the vanishing hierarchy problem that can prevent these networks from
achieving accuracy. In short, this issue stems from the way in which each step during the RNNs training will
utilize the exact same weights as calculation yr [20]. This result is also calculated during the back
propagation; here, whenever we move backwards, an error signal develops, which may be greater or smaller
depending on the particular circumstances. This error’s presence, though, means that the RNN itself is having
difficulty storing remote data in the chain and because of this, is making predictions based on simply its most
recent inputs. As a result, many recent models have had extreme difficulty in either (or both) training RNNs
on the one hand and the inability to handle long strings using tanh or ReLU as activation functions on the
other.
These limitations of RNNs can be handled with more powerful models, including LSTM and gated
recurrent unit (GRU) that can resolve the recurring issues of fading gradients in ways that enable researchers
to still implement RNNs usefully [20]. LSTMs in particular are able to learn short-term dependencies for a
longer period, as introduced by Olah [19] and later revised and circulated by many more researchers [19],
[21]. LSTM RNNs are designed specifically to try and avoid these long-term dependency problems that
cause the fading gradient issue in other RNNs. Its advantage in this area is that its default performance
method is memorizing information inputs for longer periods of time than many other RNNs do [21]. Because
of these advantages, our proposed model will adopt this type of RNN. All RNNs utilize a series of recurring
units, and in more standard types of RNN, this recurrent unit usually a very simple structure, for example one
tanh layer. With LSTM RNNs, though, the recurrent unit has a dissimilar structure, utilizing four neural
network layers that interact in particular ways, rather than a single layer that follows a simpler pattern.
Figure 2 depicts the LSTM RNN symbols, while Figure 3 shows the recurrence units in an LSTM RNN with
its gates and Figure 4 demonstrates the LSTM RNN gates [14], [21], [22].
Figure 2. LSTM RNN symbols [13]
Figure 3. Recurrence units in LTSM

 ISSN: 2088-8708
2112
Figure 4. LTSM RNN gates RNN with its gates
As shown in Figure 4, the recurring units in the LSTM have three gates (forget gate, input gate,
output gate):
1) Forgetting gate: decides how long you should remember and forget what is left as in the following
equation. This portal defines information that must be removed from the cell in a specific time step (t) as
decided using the 𝑠𝑖𝑔𝑚𝑜𝑖𝑑 (𝜎) function. The function considers the previous state (ℎ𝑟−1) and the
current content of both the current input (𝑥𝑟) and outputs, which will be a number whose value is either
0 (to ignore) or 1 (to keep) for each number in the case of cell (𝑐𝑟−1 ).
𝑓𝑟 = 𝜎(𝑊𝑓 . [ℎ𝑟−1, 𝑥𝑟] + 𝑏𝑓] (3)
2) Input gateway update: decides the amount added from this unit to the current state as in the following
equations and then defines any new information that will be stored in that cell case. There are two parts
of this particular step. First, the 𝑠𝑖𝑔𝑚𝑜𝑖𝑑(𝜎) layer determines which values require updating, and next,
the 𝑡𝑎𝑛ℎ layer creates a new candidate filter beam to be added on here.
𝑖𝑟 = 𝜎(𝑊𝑖 . [ℎ𝑟−1, 𝑥𝑟] + 𝑏𝑖) (4)
𝐶
̃ = 𝑡𝑎𝑛ℎ(𝑊
𝑐 . [ℎ𝑟−1, 𝑥𝑟] + 𝑏𝑐) (5)
The second step at this gate combines those two actions in order to create and then update the status.
Thus, we will update the state of the old cell 𝐶𝑟−1to the status of the new cell 𝐶𝑟 as demonstrated in the
following (6). The previous gate and the steps it has set in motion have already determined the actions
to be precipitated here, so all that is required now is there implementation. Here, we can multiply the
old case by 𝑓𝑟 in order to forget any things that we had earlier determined to let go of at this stage. From
here we then add 𝑖𝑟 ∗ 𝐶
̃. These are the new candidate values, which were measured by how much we
decided to update in each individual case.
𝐶𝑟 = 𝑓𝑟 ∗ 𝐶𝑟−1 + 𝑖𝑟 ∗ 𝐶
̃𝑟 (6)
3) Output gate: Decides which segment of the current cell will be made into the output as in the following
equations. The 𝑆𝑖𝑔𝑚𝑜𝑖𝑑 function determines the values that should be allowed by 0 and 1. The tanh
function gives weight to those values passed and determines their level of significance as calculated
from -1 to 1 and then multiplies it by the 𝑠𝑖𝑔𝑚𝑜𝑖𝑑(𝜎) output, thereby leaving only the parts that we have
already decided to output
𝑂𝑟 = σ(𝑊
𝑜 . [ℎ𝑟−1, 𝑥𝑟 ] + 𝑏𝑎) (7)
ℎ𝑟 = 𝑂𝑟 ∗ 𝑡𝑎𝑛 ℎ(𝐶𝑟) (8)
2.3. Previous studies
The literature includes several approaches that have been applied to solve different problems, A
selection of these includes the following. Argyros et al. [22] present the deep packet analysis mechanism

2113
used the REST call of websites. They attempt to learn correct communication patterns using the black box
analysis methodology. Their approach is based on symbolic automation and periodic exploration in order to
identify patterns.
DeMarinis and Fonseca [23] compare IoT data traffic with Internet data passage. They concluded
that IoT data traffic has more regular patterns than others, which shows the benefit of using periodic
exploration. Javaid et al. [24] proposed the use of DL applications to detect suspicious behavior in networks,
which they did using the NSL-KDD dataset. In their work, they followed a self-learning DL scheme in which
the advantage of obtaining unsupervised advantages from training data was employed using a sparse
autoencoder. The two attributes of “attack” or “natural” were learned and then the DL was applied to a new
test dataset to evaluate the speed and accuracy of its classification. They used the K-fold cross-validation
technique to evaluate the model's performance, obtaining reasonable but not outstanding results. Another
work related to the present study regards the application of automatic coding to detect abnormalities, which
was introduced by Sakurada and Yairi [25]. Here the natural state of the network was learned by means of
automatic coding through the mechanism of reducing non-linear features. Sakurada and Yairi were able to
demonstrate that normal records from their test dataset had a small reconfiguration plan but this error was
significant for revealing abnormal records in the dataset itself.
DL methods were also applied by Li et al. [26] to detect malicious code by using automatic code
extraction and deep belief networks (DBN) to constitute a detection class. Their article showed that a hybrid
method, such as the extraction technique they utilized, could be more efficient and yield more accurate
results than could be achieved by utilizing DBN alone. This work also verified that “deep” networks
outperform their shallow counterparts in detecting cyber attacks. Wang et al. [27] also investigated the
possibilities of DL systems for the detection of malicious code. In this work, de-noising autoencoders were
implemented to achieve deeper feature learning and thus distinguish between malicious and natural code,
particularly JavaScript. Results demonstrated a promising degree of accuracy and seemed effective when
applied to web applications, but it would be difficult to apply these findings to distributed loT/Fog systems.
Hasan et al. [28] applied several traditional machine learning techniques to predict attacks and
detect abnormalities within IoT systems and eventually reached the following values of accuracy: 98.3%
achieved when using the “logistic regression” mechanism, 98.2% achieved from the automated support beam
mechanism, and 99.4% resulting from use of the decision tree mechanism, the “random purpose” mechanism,
and the industrial neural networks mechanism all at once. Where Pahl and Aubet [29] used k-means
algorithm and the BIRCH algorithm for classification and obtained 96.3% accuracy, Diro and Chilamkurti
[30] developed an anomaly detector based on machine learning methodology using neural network
mechanisms and an open source data set NSL-KDD with an accuracy of 98.27%. Alghuried [31] also used
the same dataset on a model for detecting anomalies regarding IoT systems using traditional ML techniques
and a combination of two particular ML algorithms: inverse weight cluster (IWC) and decision tree algorithm
(C4.5). From this, their model achieved 97.0% accuracy. Meanwhile, D’angelo et al. [32] proposed an ML
model for classification doubling that achieved 94.1% accuracy. Finally, Bru et al. [33] used DNNs to detect
attacks within the home environments connected to the internet of things similar to [34], [35]. Table 2
summarizes the relevant research that has been conducted on the subject of detecting suspicious behavior in
IoT through the use of ML techniques.
Table 2. Several methods to detecting suspicious behavior in IOT using machine learning techniques
Publication
date
Publishing group Research methods Dataset Used
Accuracy
ratio
Error ratio
MSE
2021 Agarwal et al. [35] - Support vector machines (SVM)
- K-nearest neighbor (kNN)
- Naïve Bayes (NB)
UNSW-NB15 - 97.7777
- 93.3333
- 95.5555
-
2020 Zerkouk and Chikhaoui [34] - Long short term memory (LSTM)
- Convolutional neural network (CNN)
- CNN-LSTM
- Autoencoder-CNN-LSTM
SIMADL - 94%
- 93%
- 98%
- 84%
-
2019 Hasan et al. [28] - Logistic regression.
- support vector machines
- Decision trees
- Random forest
- Neural networks
Especially [36] - 98.3%
- 98.2%
- 99.4%
- 99.4%
- 99.4%
- 1.7%
- 1.8%
- 0.6%
- 0.6%
- 0.6%
2018 Pahl and Aubet [29] K-means and BIRCH Especially [36] 96.3% 3.7%
2018 Diro and Chilamkurti [30] Deep neuron networks NSL-KDD [37] 98.27% 1.73%
2017 Alghuried [31] IWC and C4.5 Intel Lab Data [38] 97.0% 3.0%
2015 D’Angelo et al. [32] U-BRAIN NSL-KDD [37] 94.1% 5.9%
2018 Brun et al. [33] Deep neuron networks Especially [36] unavailable unavailable

 ISSN: 2088-8708
2114
3. METHOD
This work is intended to develop an algorithm or model with its process to detect abnormal
suspicious behavior in IoT systems. As discussed earlier, powerful DL techniques such as RNNs with LSTM
as introduced by Olah [19], have been used to great effect in many fields. LSTM in particular has proven able
to learn short-term dependencies for a longer period, and LSTM RNNs are designed specifically to try and
avoid the common long-term dependency problem that causes the fading gradient of a neuron network. An
LSTM RNN can memorize the needed information for longer times, which enables it to avoid this problem
more easily by default [21]. LSTM RNNs also employ four layers that interact in particular ways as
described above, adding a further degree of protection. Because of these capacities, we adopted LSTM RNNs
to fulfill our proposed model for finding a solution that can identify malicious and unauthorized uses of IoT.
The followed subsection details both the datasets used to train our model and the algorithm's own design.
3.1. Dataset
Input data is a data set that we must process before beginning to utilize it in developing our
proposed model. This is necessary because some datasets may not include an attribute label, since they were
not initially intended or prepared for research purposes on the detection of suspicious behavior. Also, such
datasets may include a large number of features that such research will not require, so we should ignore these
and keep only the features that we need. Finally, we will generate anomalies (or “abnormal” results) and add
them to the input data manually.
The benefits of adopting such an approach to a training dataset [39] are i) the input data has been
addressed and defined in some way to the ML model; ii) the input data set has been configured into an
eligible data set that will help teach the ML model to detect anomalies, and thus is suitable for research
purposes; iii) the input data has been modified to meet research needs in ways that overcome the usual
difficulties of obtaining an IoT income dataset at prohibitively high cost.
The dataset used in this work is from Intel Labs and it includes over 2.3 million reads from
54 sensors distributed across the Intel Berkeley research laboratory [38]. As noted by Intel Labs, each of
these sensors here can collect particular types of data (light, voltage, temperature, and humidity) and will
transfer all of this data back to the gateway on a schedule of once every 31 seconds. In addition, each reading
in the dataset is marked by the date, the time, and the ID of the sending sensor. This Intel Labs dataset is
freely and readily available, which means that it has been used as a dataset for many different problems and it
was not gathered or presented specifically for research into uncovering data on suspicious behavior in IoT
devices. To overcome this potential shortcoming, we processed the dataset as discussed next.
Processing over 2.3 million pieces of data can be computationally expensive. To resolve this, we
took a random sample from the total number of readings and saved it in a separate file. We also omitted
unnecessary features (columns) such as the identifying information (time, date, and sensor ID) as well as both
light, and voltage: these were omitted due to our research focus on a proposed anomaly detection. Finally,
anomalies were generated and added to that processed file of input data manually [39].
3.2. The algorithm
As mentioned earlier, our proposed model or algorithm is an adapted LSTM RNNs [19] intended to
detect certain issues in IoT. The model consists of several steps as in Figure 5. As shown in Figure 5, there
are several steps represents the mechanism or process that we follow the build the model. These steps
(i.e., read dataset, feature definition, encoding of variables, divide the dataset [leaning and training], reducing
MSE error [actual output], obtained model trained, realization of proposed model, convert to TensorFlow
data, and predict the test data) are used to build the model in five stages (i.e., initialization, pre-processing,
realization, training, and results stages).
In the initialization stage, all the test intermediaries are first defined and then configured with their
fixed values. Also, both the training and the testing datasets are examined and stored in the specified
variables. Record defaults are also configured and variables that must be reconfigured are made in order to
prevent math errors. In the stage of data pre-processing, the feature and records names for each dataset are
examined and extracted from the variables specified in the initialization stage and then placed in their
respective elements. During the realization stage, a five-layer LSTM RNN model is utilized. At this point, the
random weights, biases, and nodes that are randomly generated for the models are defined as well as the
computing function (ReLU). The training stage of the algorithm architecture is part of the code that carries
out the model test run. To accomplish this, we utilized three datasets (one apiece for training, testing, and
frequency or epochs data) and recorded our model's results on each one. These results were then used in
order to calculate metrics for success such as the cost and accuracy function. The results of these success
metrics are used in order to improve the computed learning rate and various training steps.

2115
The TensorFlow platform is a second-generation ML system that was created to succeed the earlier
system DistBelief. TensorFlow is an open-source library collecting various models and algorithms for
numerical computing, ML, and DL. Our adopted model was programmed by using Python to provide a
suitable API for the front end of building applications which were then saved in high-performance C++
language. Figure 6 provides an overview of sample TensorFlow code for a basic LSTM cell.
Figure 5. the process of the proposed model or algorithm
Figure 6. Sample TensorFlow code for a basic LSTM cell
4. EXPERIMENTAL RESULTS
4.1. Experimental settings
This research conducted preliminary experiments to determine the suitable configuration of the
model for a proper data analysis. At first, it is implemented with its default parameter settings. Then, its
parameter values are carefully tuned to obtain the desired results. Using TensorFlow (Python), experiments
are conducted on a Windows 10 machine with CoreTM i7-7700HQ processor, 3.80 GHz CPU and 8 of
RAM.
4.2. Experimental results
This work conducted over four tests to evaluate the performance of the model. These tests were
performed on the basis of seven mediators (i.e., weights, biases, number of layers, learning rate, number of
nodes, size of the dataset, and iterative periods). In the seven mediators, the weights and biases were
randomly generated, while the values of layers, nodes, and learning rate remained constant and the set of
repetitive data and periods were variable in all tests.

 ISSN: 2088-8708
2116
To evaluate our model’s performance, we utilized two criteria: time spent (i.e., MSE) and accuracy.
These two metrics enabled us to evaluate the results that our model provided for the training data. Table 3
shows the mediators used in the first, second, third and fourth tests. Where in the first test our model
performed upon small datasets (e.g., 5,000 data points apiece) in 100 periods. In the second test our goal was
to determine whether our model could maintain a certain level of accuracy in detection: we were seeking a
high rate with reasonable time similar to the one achieved in the first test while also increasing the data sets'
volume and periods by 100%. Our third test and fourth test proceeded in the same way: for each one, we
increased the dataset volume and dataset period by 100% beyond the previous test and observed whether our
model could maintain a high degree of accuracy and a reasonable timeframe.
Following Table 3, Table 4 shows the first, second, third, and fourth test’s training results in terms
of scores for accuracy, time, and F1. Likewise, Figure 7 illustrates the graph of loss obtained by our model
from TensorFlow for the first test, Figure 8 for the second test, Figure 9 for the third test, and Figure 10 for
the fourth test.
Table 3. Tests mediators
Test mediators data sets volume Periods Learning rate nodes Layers Weights Biases
First test 5000 100 0.1 40 5 Random Random
Second test 10000 200 0.1 40 5 Random Random
Third test 20000 400 0.1 40 5 Random Random
Fourth test 40000 800 0.1 40 5 Random Random
Table 4. Tests training results
First test Second test Third test Fourth test
Epoch Loss Epoch Loss Epoch Loss Epoch Loss
0 0.65207316 100 3.89E-15 - - - -
10 1.35772E-13 110 3.89E-15 - - - -
20 1.35771E-13 120 3.89E-15 - - - -
30 1.35771E-13 130 3.89E-15 - - - -
40 1.35771E-13 140 3.89E-15 340 2.96E-08 740 6.20644E-08
50 1.35766E-13 150 3.88485E-15 350 3.01E-08 750 5.29224E-08
60 1.35755E-13 160 3.88253E-15 360 3.3E-08 760 6.13865E-08
70 1.35727E-13 170 3.88021E-15 370 3.25E-08 770 5.21E-08
80 1.35672E-13 180 3.87789E-15 380 3.19E-08 780 6.29E-08
90 1.35572E-13 190 3.87558E-15 390 3.08E-08 790 5.13865E-08
F1 Score: 0.499670616 0.499670616 0.519473631 0.494739875
Accuracy Score: 0.998683332 0.998683332 0.996681998 0.998568974
Time spent 507 2436 8931 23459
The results in Table 4 for the first test and Figure 7 demonstrate that the model we have developed
offers a remarkably high rate of detection and a correspondingly low rate of false alarms. Moreover, the
value of the loss continues to decrease with the continuation of the model’s training, but even the first test
already achieves a 99.87% accuracy rate in detection. These results demonstrate that LSTM RNN technology
is a highly valuable means of detecting suspicious behavior in IoT accurately and efficiently.
Results in Figure 8 shows that the loss decreased when the training model data was increased.
Meanwhile, the results in Table 4 for the second test show that the accuracy is 99.87%, which is comparable
with the first test despite the increased. Results also show that the training time of the second test is increase
by 4 times, or 480.473%. This increased time is due to increasing both the frequency and the dataset size.
These results also support our assertion that LSTM RNN technology can detect suspicious behavior in IoT
data efficiently and accurately.
Results in Table 4 for the third test and Figure 9 shows that the model has a very high detection rate
and a low false alarm rate, with 400 iterations the model completed the detection with accuracy of 99%.67.
Results also shows that the training time is decreased to 366.63 % (which is decreased 3.6 times) while
comparing with the second test. Meanwhile, the accuracy of the final detection decreased to 99.67%, but this
is still considered a very high accuracy, which supports the fact that LSTM RNN is able to detect suspicious
behavior in IoT accurately and efficiently.
Results in Figure 10 show the model’s different training stages, demonstrating how the entire
process took 6 hours, 30 minutes and 59 seconds to complete from beginning to end. Results in both
Figure 10 and Table 4 for the fourth test demonstrate how the model begins to decrease its loss rates as it
continues with training until it reaches a stage where the loss values become very close to each other. Results
also reveal that the developed model has a remarkably high rate of detection and a very low rate of false

2117
alarms. Over 800 iterations were conducted by the algorithm, and the model achieved a detection accuracy
of 99.86%. While comparing the results between the third and fourth tests, we notice that the accuracy is
increased by 0.19% in the model final detection, which indicates that the accuracy of detection in the fourth
test is higher than the third test. Furthermore, the training time is decreased to 262.67%, where in the third
test was 366.63%. This supports the fact that LSTM RNN technology is able to detect suspicious behavior in
IoT accurately and efficiently.
Figure 7. The graph of a loss in the first test Figure 8. The graph of a loss in the second test
Figure 9. The graph of a loss in the third test Figure 10. The graph of a loss in the fourth test
4.3. Results discussion
Figure 11 shows the relationship between 4 tests based on datasets size, final accuracy and
frequency, while Figure 12 shows the data set size and accuracy analysis (where the horizontal axis represent
the four tests results and the vertical axis represent the accuracy), Figure 13 illustrates the relationships
between frequency and accuracy of model detection (where the horizontal axis represent the four tests results,
the blue line respresent the accuracy results achieved and the dark red dash line represent the frequency
results achieved) and Table 5 summarizes the four tests’ results based on the following mediator:
a) The final accuracy of the fully-trained LSTM RNN model: represents ultimate results achieved by the
model.
b) Average mean square error (MSE): represents the MSE values achieved with training outputs.
c) Dataset: represents the volume of data points and records used during the model's training process.
d) Frequency: represents the overall number of times that the model was trained in behavior features.
e) Time: represents the time needed for the model to complete each test.
f) Percentage increase in time: represents time as a percentage increase across all four tests.
Figure 11 shows a comparison between the four tests based on the datasets used and the results
achieved for time, accuracy, and frequency. As this figure demonstrates, the first and second tests each
achieved 99.87% accuracy, where the model shows the data with distinctive features. However, as the size of
the dataset is increased by 100% in the third test, there was also a 0.2% decrease in accuracy, resulting in a
final accuracy of 99.67%, which may have resulted from the increased dataset size or the model’s exposure
to data with different characteristics. However, when we increased the size of the dataset and maintained

 ISSN: 2088-8708
2118
similar characteristics, the rate of accuracy detection increased again, rising back up to 99.86%. This result
supports our hypothesis that the model's accuracy depends at least in part on the characteristics of the data
that it is trained on.
Frequency describes the ultimate number of times that the algorithm has been exposed to different
data attributes over the course of its training. From the various forms of testing done on our proposed model,
it was determined that use of a larger dataset would mean that the model would require additional iterations
in order to learn differences between different data attributes and increase its own ability to distinguish
among them. Therefore, we did not use a uniform iteration of times across all stages of training, but instead
increased times as the datasets’ own sizes increased. Figure 12 shows a comparison between the four tests
based on the model accuracy and Figure 13 shows the relationships between frequency and accuracy of
model detection.
Table 5. The results of the model for the four tests
Standard First test Second test Third Test Fourth Test
Final accuracy 0.9987 0.9987 0.9967 0.9986
Average (MSE) 0.0013 0.0013 0.0033 0.0014
Date set 5000 10000 20000 40000
Frequency 100 200 400 800
Time (Second) 507 2436 8931 23459
Increase % in time 0 380.473 266.63 162.67
Figure 11. Relationship between 4 tests
Figure 12. Data set size and accuracy analysis

2119
Figure 13. Accuracy and frequency analysis
Figure 13 shows a comparison among the four tests (as in horizontal axis) based on the accuracy and
frequency the model achieved in each one, while also comparing the results of its training with the results
achieved by models trained using traditional ML algorithms, which the researchers suggested in [30]. Here
we noticed significant differences between the final detection of accuracy rates achieved by each model and
the weighting of the cuff of the model we developed with an average detection accuracy of 99.8% (across the
four tests) versus 97% for the traditional model. This difference indicates a profound learning proficiency in
detecting suspicious behavior within IoT networks.
5. CONCLUSION
This work has investigated the detection of abnormal behaviors in IoT networks, which constitute
one of the most important and widespread networks in use today. An adapted LSTM RNN was programmed
using Python and used to produce our model. The second-generation Google ML framework called
TensorFlow was used to determine the efficacy and efficiency of our proposed model, along with the
LSTM’s own features and the following mathematical formulas. To best evaluate our model’s performance,
four tests were performed based on seven mediators and two criteria, wherein the mediators of data size and
period were changed with each test in order to evaluate the effectiveness of the model. Where all of the tests
performed were assigned different values of volume and iterative periods, randomly-generated values are
used for weights and biases as well as the various layers, nodes, and learning rate until the model was able to
achieve the best rate of accuracy detection. The results obtained demonstrate that our model performs better
when the dataset increases, and its accuracy does not depend on the platform used to train it. The results also
shows that the model developed has an average detection accuracy of 99.8%, meaning that LSTM RNN is
able to detect suspicious behavior for different types of IoT sensor data. However, the model may produce
even better results in terms of accuracy when using a higher-quality processor. Thus, we recommend that
future work trains its model on computing platforms other than the CPU, perhaps such as graphics processors
(GPU) for example, and then compare the results achieved with each of these approaches.
REFERNCES
[1] K. Ashton, “Kevin Ashton May Change the World,” RFID Journal. pp. 25–36, 2002.
[2] D. R. Kiran, “Internet of Things,” Production Planning and Control, pp. 495–513, Jan. 2019, doi: 10.1016/B978-0-12-818364-
9.00035-4.
[3] A. Khanna and S. Kaur, “Internet of things (IoT), applications and challenges: a comprehensive review,” Wireless Personal
Communications, vol. 114, no. 2, pp. 1687–1762, May 2020, doi: 10.1007/s11277-020-07446-4.
[4] S. Y. Lee, S. R. Wi, E. Seo, J. K. Jung, and T. M. Chung, “ProFiOt: abnormal behavior profiling (ABP) of IoT devices based on a
machine learning approach,” in 2017 27th International Telecommunication Networks and Applications Conference, ITNAC
2017, Nov. 2017, vol. 2017-Janua, pp. 1–6, doi: 10.1109/ATNAC.2017.8215434.
[5] P. A. Savenkov and A. N. Ivutin, “Methods of machine learning in system abnormal behavior detection,” in In: Tan Y., Shi Y.,
Tuba M. (eds) Advances in Swarm Intelligence, ICSI 2020, Springer International Publishing, 2020, pp. 495–505.
[6] M. Koutli, N. Theologou, A. Tryferidis, and D. Tzovaras, “Abnormal behavior detection for elderly people living alone
leveraging IoT sensors,” in Proceedings - 2019 IEEE 19th International Conference on Bioinformatics and Bioengineering, BIBE
2019, Oct. 2019, pp. 922–926, doi: 10.1109/BIBE.2019.00173.
[7] M. Bertolini, D. Mezzogori, M. Neroni, and F. Zammori, “Machine learning for industrial applications: A comprehensive
literature review,” Expert Systems with Applications, vol. 175, p. 114820, Aug. 2021, doi: 10.1016/j.eswa.2021.114820.
[8] L. Deng and D. Yu, “Deep learning: Methods and applications,” Foundations and Trends in Signal Processing, vol. 7, no. 3–4.
Now Publishers, pp. 197–387, 2013, doi: 10.1561/2000000039.
[9] J. Bell, Machine learning: hands-on for developers and technical professionals. Wiley, 2020.
[10] A. L. Samuel, “Some studies in machine learning using the game of checkers,” in IBM Journal of Research and Development,
vol. 3, no. 3, pp. 210–229, Jul. 1959, doi: 10.1147/rd.33.0210.
[11] M. Al-kmali, H. Mugahed, W. Boulila, M. Al-Sarem, and A. Abuhamdah, “A machine-learning based approach to support
academic decision-making at higher educational institutions,” Oct. 2020, doi: 10.1109/isncc49221.2020.9297177.

 ISSN: 2088-8708
2120
[12] J. H.Alkhateeb, A. A.Turani, A. Abuhamdah, M. J. A. Sara, and M. F. J. Klaib, “An effective deep learning approach for
improving off-line Arabic handwritten character recognition,” International Journal of Software Engineering and Computer
Systems, vol. 6, no. 2, pp. 104–112, 2020, doi: 10.15282/ijsecs.6.2.2020.7.0076.
[13] J. Brownlee, “A tour of machine learning algorithms,” Machine Learning Algorithms, 2019.
[14] S. A. Bini, “Artificial intelligence, machine learning, deep learning, and cognitive computing: what do these terms mean and how
will they impact health care?,” International Journal of Software Engineering and Computer Systems, vol. 33, no. 8, pp. 2358–
2361, Aug. 2018, doi: 10.1016/j.arth.2018.02.067.
[15] L. K. Hansen and P. Salamon, “Neural network ensembles,” IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 12, no. 10, pp. 993–1001, 1990, doi: 10.1109/34.58871.
[16] J. Zurada, “Introduction to artificial neural systems,” West Publishing Co., 1992.
[17] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagating errors,” Nature, vol. 323, no.
6088, pp. 533–536, Oct. 1986, doi: 10.1038/323533a0.
[18] B. C. Csaji, “Approximation with artificial neural networks,” Etvs Lornd University, Hungary, 2001.
[19] C. Olah, “Understanding LSTM networks,” github. 2015, Accessed: Mar. 02, 2021. [Online]. Available:
http://colah.github.io/posts/2015-08-Understanding-LSTMs.
[20] S. Hochreiter, “The vanishing gradient problem during learning recurrent neural nets and problem solutions,” International
Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, vol. 06, no. 02, pp. 107–116, Apr. 1998, doi:
10.1142/s0218488598000094.
[21] R. C. Staudemeyer and E. R. Morris, “Understanding LSTM-a tutorial into long short-term memory recurrent neural networks,”
arXiv preprint arXiv:1909.0958, Sep. 2019.
[22] G. Argyros, I. Stais, A. Kiayias, and A. D. Keromytis, “Back in black: towards formal, black box analysis of sanitizers and
filters,” May 2016, doi: 10.1109/sp.2016.14.
[23] N. DeMarinis and R. Fonseca, “Toward usable network traffic policies for IoT devices in consumer networks,” in Proceedings of
the 2017 Workshop on Internet of Things Security and Privacy, Nov. 2017, pp. 43–48, doi: 10.1145/3139937.3139949.
[24] A. Javaid, Q. Niyaz, W. Sun, and M. Alam, “A deep learning approach for network intrusion detection system,” Proceedings of
the 9th EAI International Conference on Bio-inspired Information and Communications Technologies (formerly BIONETICS).
ACM, 2016, doi: 10.4108/eai.3-12-2015.2262516.
[25] M. Sakurada and T. Yairi, “Anomaly detection using autoencoders with nonlinear dimensionality reduction,” in Proceedings of
the MLSDA 2014 2nd Workshop on Machine Learning for Sensory Data Analysis-MLSDA’14, 2014, pp. 4–11, doi:
10.1145/2689746.2689747.
[26] Y. Li, R. Ma, and R. Jiao, “A hybrid malicious code detection method based on deep learning,” International Journal of Software
Engineering and its Applications, vol. 9, no. 5, pp. 205–216, May 2015, doi: 10.14257/ijsia.2015.9.5.21.
[27] Y. Wang, W. Cai, and P. Wei, “A deep learning approach for detecting malicious JavaScript code,” Security and Communication
Networks, vol. 9, no. 11, pp. 1520–1534, Jul. 2016, doi: 10.1002/sec.1441.
[28] M. Hasan, M. M. Islam, M. I. I. Zarif, and M. M. A. Hashem, “Attack and anomaly detection in IoT sensors in IoT sites using
machine learning approaches,” Internet of Things, vol. 7, p. 100059, Sep. 2019, doi: 10.1016/j.iot.2019.100059.
[29] M. O. Pahl and F. X. Aubet, “All eyes on you: distributed multi-dimensional IoT microservice anomaly detection,” in 14th
International Conference on Network and Service Management, CNSM 2018 and Workshops, 1st International Workshop on
High-Precision Networks Operations and Control, HiPNet 2018 and 1st Workshop on Segment Routing and Service Function
Chaining, SR+SFC 2, 2018, pp. 72–80.
[30] 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, May 2018, doi: 10.1016/j.future.2017.08.043.
[31] A. Alghuried, “A model for anomalies detection in internet of things (IoT) using inverse weight clustering and decision tree,”
2017.
[32] G. D’angelo, F. Palmieri, M. Ficco, and S. Rampone, “An uncertainty-managing batch relevance-based approach to network
anomaly detection,” Applied Soft Computing, vol. 36, pp. 408–418, Nov. 2015, doi: 10.1016/j.asoc.2015.07.029.
[33] O. Brun, Y. Yin, and E. Gelenbe, “Deep learning with dense random neural network for detecting attacks against IoT-connected
home environments,” Procedia Computer Science, vol. 134, pp. 458–463, 2018, doi: 10.1016/j.procs.2018.07.183.
[34] M. Zerkouk and B. Chikhaoui, “Spatio-temporal abnormal behavior prediction in elderly persons using deep learning models,”
Sensors, vol. 20, no. 8, p. 2359, Apr. 2020, doi: 10.3390/s20082359.
[35] A. Agarwal, P. Sharma, M. Alshehri, A. A. Mohamed, and O. Alfarraj, “Classification model for accuracy and intrusion detection
using machine learning approach,” PeerJ Computer Science, vol. 7, p. e437, Apr. 2021, doi: 10.7717/peerj-cs.437.
[36] F. X. Aubet, “DS2OS traffic traces,” Kaggle. 2018, Accessed: Mar. 06, 2021. [Online]. Available:
https://www.kaggle.com/francoisxa/ds2ostraffictraces.
[37] M. Tavallaee, E. Bagheri, W. Lu, and A. A. Ghorbani, “A detailed analysis of the KDD CUP 99 data set,” Jul. 2009, doi:
10.1109/CISDA.2009.5356528.
[38] P. Bodik, W. Hong, C. Guestrin, S. Madden, M. Paskin, and R. Thibaux, “Intel lab data,” Intel Berkeley Research lab. [Online].
Available: http://db.csail.mit.edu/labdata/labdata.html.
[39] R. Fu, K. Zheng, D. Zhang, and Y. Yang, “An intrusion detection scheme based on anomaly mining in Internet of Things,” in 4th
IET International Conference on Wireless, Mobile & Multimedia Networks (ICWMMN 2011), 2011, pp. 315–320, doi:
10.1049/cp.2011.1014.

Using deep learning to detecting abnormal behavior in internet of things

More Related Content

Similar to Using deep learning to detecting abnormal behavior in internet of things (20)

More from IJECEIAES (20)

Recently uploaded (20)

Using deep learning to detecting abnormal behavior in internet of things