Content uploaded by Mohamed S. Abdalzaher
Author content
All content in this area was uploaded by Mohamed S. Abdalzaher on Apr 28, 2022
Content may be subject to copyright.
IEEE SYSTEMS JOURNAL 1
A Deep Autoencoder Trust Model for Mitigating
Jamming Attack in IoT Assisted by Cognitive Radio
Mohamed S. Abdalzaher , Member, IEEE, Mohamed Elwekeil , Taotao Wang , Member, IEEE,
and Shengli Zhang , Senior Member, IEEE
Abstract—This article proposes deep learning (DL) framework
constructed using deep autoencoder (DAE) to detect the malicious
nodes in an Internet of Things (IoT) network assisted by the cogni-
tive radio (CR) technology. In the IoT era, a plethora of nodes
are connected to the network for the purpose of collecting and
exchanging data. CR technology finds its role in IoT applications
because of its ability to efficiently exploit the available spectrum.
In this article, we consider IoT nodes as secondary users that
perform cooperative spectrum sensing (CSS). Specifically, these
IoT nodes sense the spectrum and send their reports to the fusion
center (FC) to determine whether the spectrum is occupied by the
primary user or not. In such a scenario, the existence of malicious
IoT nodes might mislead the FC. To determine which of these
IoT nodes are benevolent and which are malicious is crucial. We
adopt a DAE-based DL framework, called DAE-TRUST, to detect
malicious nodes in CR-assisted IoT. The proposed DAE-TRUST is
able to identify the malicious nodes, whose reports can be excluded
from the spectrum detection process. Simulation results, in six
different real-world environments, show that our DAE-TRUST
enhances the performance of CSS in IoT applications.
Index Terms—Autoencoder (AE), cognitive radio (CR),
convolutional neural network, cooperative spectrum sensing
(CSS), Internet of Things (IoT) security, jamming attacks,
machine learning (ML).
I. INTRODUCTION
THE INTERNET of Things (IoT) network is a fast boom-
ing technology that can be effectively utilized in many
applications. In an IoT network, the huge number of connected
nodes with the presence of day-to-day communication chal-
lenges may lead to open issues such as resource allocation,
Quality of Service (QoS) management, energy consumption, and
security. Accordingly, an adaptive and intellectual solution has
become desirable. Cognitive radio networks (CRNs) are among
Manuscript received January 3, 2021; revised June 8, 2021; accepted July
11, 2021. The work was supported in part by research grants from the Chi-
nese National Science Foundation Project under Grant 61771315 and in part
by Guangdong Basic and Applied Basic Research Foundation under Grant
2019B1515130003. (Mohamed S. Abdalzaher and Mohamed Elwekeil are co-
first authors.) (Corresponding author: Shengli Zhang.)
Mohamed S. Abdalzaher is with the Seismology Department, National
Research Institute of Astronomy, Geophysics, Cairo 11421, Egypt (e-mail:
msabdalzaher@nriag.sci.eg).
Mohamed Elwekeil is with the Department of Electronics, Electrical Commu-
nications Engineering, Faculty of Electronic Engineering, Menoufia University,
Menouf 32952, Egypt (e-mail: mohamed.elwekeil@el-eng.menofia.edu.eg).
Taotao Wang and Shengli Zhang are with the College of Electronics, Infor-
mation Engineering, Shenzhen University, Shenzhen 518060, China (e-mail:
ttwang@szu.edu.cn; zsl@szu.edu.cn).
Digital Object Identifier 10.1109/JSYST.2021.3099072
the most promising technologies that can contribute to solving
the aforementioned problems.
In the literature context, several CR techniques have been
studied concerning the QoS constraint relying on the role of
secondary users (SUs) [1]–[6]. In [7] and [8], other efforts
have been exerted to set up a primary user (PU) aware routing
scheme using compressive sensing to detect the PU sparsity
existence. Moreover, optimizing the packet size to prolong the
CRN lifetime has been considered in [9]. Many spectrum sensing
detectors have been proposed aiming at enhancing the spec-
trum utilization such as energy detection (ED), cyclostationary,
and matched filter [10]–[12]. Here, we focus on the ED as
the simplest detection mechanism employed in CR spectrum.
Nevertheless, the proposed approach is applicable to work with
different spectrum detection techniques. Due to the dramatic
growth of wireless sensor networks (WSNs) based IoT in sev-
eral applications, it can be efficiently utilized in CR spectrum
sensing. In other words, the fusion center (FC) can rely on these
allotted nodes to determine the CR spectrum. On the contrary,
WSNs suffer from some weaknesses such as security and power
management. In fact, the security front is the most vulnerable as-
pect of WSNs; and, thus, it was considered by many researchers,
e.g., [8], [13]–[18]. However, the security aspect is still an open
issue that needs further investigations. In this trajectory, modern
technologies such as machine learning (ML) can be an intrinsic
solution.
ML has evolved over the last few years as an effective
methodology for handling complex data. In wireless commu-
nication systems, traditional methods are based on mathemat-
ical modeling where approximations are inevitable due to the
complexity of the considered system. On the other hand, ML
can learn directly from past observations without the need for
any approximations. Accordingly, it can tackle many research
problems [19], [20]. With the emergence of deep learning (DL),
it becomes convenient to build very complex relational models
that classical approaches might not be able to capture, given their
models’ restrictions [21], [22]. The applications of ML tools
have witnessed great success in addressing many challenging
research problems, ranging from recommendation systems to
autonomous driving cars [23], [24]. This success is attributed
to the flexibility of ML in learning complex systems and due
to the fact that all you need is a suitable ML model and good
datasets to train this ML model. Therefore, ML facilitates the
requirements of traditional methods, which are mainly based
on complex/approximated mathematical modeling. ML algo-
rithms can be even used to extract important features from the
underlying data, i.e., they can “learn” the features. Indeed, the
2IEEE SYSTEMS JOURNAL
autoencoder (AE) is among the flourished ML techniques that
are merged into DL.
In the literature context, various approaches have been pro-
posed aiming to combat malicious attacks in CRN [25]–[29].
Game theory has taken a valuable role in mitigating security
threats. In [28], a Stackelberg game was used to treat the false in-
jected noise power in cognitive radio sensor networks (CRSNs)
due to the spectrum sensing data falsification (SSDF) attack. The
attack’s aim was to deceive the FC by corrupting the delivered
report. More particularly, the attack attempts to deviate reports’
signal-to-noise ratio from their original values. The same attack
type has also been mitigated with the presence of hardware
failure [29]. Game theory was also employed to enhance the
security and data trustworthiness in IoT applications [30].
ML has been widely employed to foster wireless commu-
nication systems. Besides, CR technology has been employed
to improve the spectral efficiency of wireless communication
systems. Specifically, the applications of ML in CRNs have been
highlighted in [31]. Generally speaking, the ML/DL approaches
can be classified into supervised and unsupervised learning.
The authors in [32] proposed ML-based cooperative spectrum
sensing (CSS) techniques in the frameworks of both supervised
learning and unsupervised learning. Specifically, they employed
both support vector machines (SVMs) and K-nearest neighbors
(KNN) as supervised learning tools to perform CSS in CR net-
works. Furthermore, they utilized unsupervised learning tools
such as the Gaussian mixture model and K-means clustering to
solve the same problem. The authors concluded that the SVM
with a linear kernel can achieve the best CSS performance.
However, the authors in [32] depended on small datasets that
have less than 1000 samples. Furthermore, they did not consider
the existence of malicious nodes. The authors in [33] adopted a
fuzzy c-means based clustering algorithm for training a classifier
to perform sepctrum sensing (SS) in CR. The work in [34]
employed a naive Bayes classifier and trained it for SS in
orthogonal frequency-division multiplexing systems. Moreover,
DL has been utilized to relieve the complexity of the energy
efficiency optimization in SS. In [35], the authors have developed
a DL model to optimize the energy utilization in cognitive SS.
In [36], convolutional long short-term deep neural networks are
utilized for improving the performance of the SS in CRNs.
In the context of ML applications in CRNs, research works
have been done not only to investigate the defense strategies
but also the attack strategies. For instance, the work in [37]
presented an online learning mechanism for implementing an
optimal primary user emulation attack strategy. Furthermore,
the authors in [38] investigated an adversarial ML scheme for
performing jamming attacks on CRNs. Then, they introduced
the corresponding defense mechanism. The defense mechanism
presented in [38] is based on exploiting the reactiveness of DL
to training errors. Specifically, this can be done by allowing
nodes to intentionally take wrong actions sometimes. Thus, the
attacker cannot predict the transmitter’s behavior. Along with
the exerted efforts in the security front, DL has exploited the
relay selection to maximize the end-to-end signal-to-noise ratio
in cognitive IoT networks [39].
Although the huge efforts exerted in the literature, more
research efforts are still required, especially, with the WSN
vulnerabilities such as low power consumption and limited pro-
cessing capabilities. To this end, ML/DL can effectively handle
these concerns as no extra hardware is needed to merge the
learned techniques into the existing infrastructure. To the best
of our knowledge, no similar work has been done before for
utilizing the AE-based trust model for ensuring security against
jamming attacks in CR-assisted IoT.
In this article, we focus on the automatic detection of ma-
licious attacks in CR-assisted IoT networks. Specifically, we
consider a WSN-based smart grid (SG), where the WSN nodes
are regarded as SUs which share the spectrum dedicated for
the PU, i.e., the licensed user. The SUs should cooperatively
sense the spectrum before using it, where each SU senses the
spectrum and sends a report to the FC, which determines whether
the spectrum is occupied by the PU or not based on the reports
received from the SUs. If the spectrum was used by the PU, the
FC informs the SUs to defer their transmissions. On the other
hand, if the PU was not transmitting on the spectrum, the SUs can
utilize the spectrum to do their transmissions. In such a scenario,
it is crucial to correctly identify whether the PU is utilizing the
spectrum or not. Furthermore, the existence of malicious nodes,
which falsify their reports sent to the FC, might mislead the FC
causing false decisions which would degrade the performance
of the CRNs. We employ a deep autoencoder (DAE) model
for malicious nodes’ detection, which we call “DAE-TRUST.”
We exploit the ability of DAE to identify abnormalities in its
input to detect the malicious nodes. Specifically, the DAE is
trained on the benevolent data, and, thus, when it is deployed in
real settings, it can identify anomalous nodes through the mean
square error (mse) between its input vector and the predicted
output vector. By identifying the malicious nodes, the FC can
exclude the reports of these malicious nodes; thus, the decision
taken will be more robust. In addition, the effectiveness of the
proposed DAE-TRUST model is verified in six different smart
grid environments. The main contributions of the article can
therefore be listed as follows.
1) The article proposes a DAE-based trust model, namely,
DAE-TRUST to detect jamming attacks in CR-assisted
IoT networks. The DAE is utilized to learn in an un-
supervised learning way, which can detect the disrupted
delivered reports at the FC. Based on the trained network,
the model is able to identify the malicious reports from
the expected benevolent ones.
2) The proposed DAE-TRUST approach is also used to
discard the malicious reports relying on the correlation
of encoder and decoder modules’ structure. Particularly,
the obtained results show the effectiveness of employing
the DAE-TRUST model in enhancing the CR spectrum
utilization and detecting the malicious reports received at
the FC. To the best of our knowledge, this work is a novel
approach that employs a DAE-based scheme to detect
jamming attacks in CR-assisted IoT networks. Besides,
the proposed DAE-TRUST model can be utilized to detect
other types of attacks without a change in the designed
network structure.
3) The proposed approach considers realistic conditions that
incorporate the shadow fading (log-normal distribution)
measurements on real experiments presented in [40] to
ABDALZAHER et al.: DAE TRUST MODEL FOR MITIGATING JAMMING ATTACK IN IOT ASSISTED BY COGNITIVE RADIO 3
TAB LE I
LIST OF ACRONYMS BPF : BAND PASS FILTER,SG:SMART GRID,SS:
SPECTRUM SENSING,ROC:RECEIVER OPERATING CHARACTERISTIC.
collect real datasets. Furthermore, the considered CR-
assisted IoT relies on the Tmote Sky nodes [41] with
the standard of IEEE 802.15.4. This node type is widely
employed in IoT applications because it guarantees both
scalability and accurateness of locations [42]. The DAE-
TRUST model effectiveness is verified over six different
environments: 1) outdoor line of sight (OL); 2) outdoor
nonline-of-sight (ON); 3) underground line of sight (UL);
4) underground nonline-of-sight (UN); 5) indoor line of
sight (IL); and 6) indoor nonline of sight (IN).
The remainder of the article is organized as follows. In
Section II, the system model is discussed. The proposed DAE-
TRUST approach is then presented in Section III. Section IV
shows the obtained results. Finally, the article is concluded in
Section V. Table I summarizes the list of acronym and Table II
presents the list of symbols used in the paper.
II. SYSTEM MODEL
Fig. 1 shows the proposed CR-assisted IoT, in which the FC
relies on a set of Tmote Sky nodes [41], N, to capture the CR
spectrum, where the number of nodes is given by the cardinality
of the set N, which is represented by |N |. As figured out by
Fig. 1, the sensor nodes (SNs), i.e., SUs, transmit their own
observed reports to the FC. Our aim is to allow the SUs to access
the spectrum without any collisions to protect the PU rights
and, at the same time, to provide the FC with immunity against
jamming attacks. We assume that the jammer attempts to deviate
the observed SNs’ energy reports from their original levels to
higher ones. In other words, the attacker aims at disrupting the
links between the PU and the SUs so that the individual decisions
of the SNs are interrupted. This might lead the FC to take a wrong
decision about the spectrum. We adopt a scenario of a disk-
shaped network with the maximum radio transmission radius
Rmax, where the PU is located at the center of the disk and
SNs are either uniformly distributed within the radius or allotted
at the circumference of the disk. Specifically, the latter case
TAB LE I I
NOTATIONS OF PARAMETERS AND VARIABLES
Fig. 1. CR-assisted IoT, where the malicious attackers may attack certain SUs
to mislead the FC.
4IEEE SYSTEMS JOURNAL
Fig. 2. Energy detection process in the CR-assisted IoT with the existence of
attack.
represents the worst case of transmission capability to prove the
effectiveness of the proposed model.
Fig. 2 shows the process of ED relying on the observed signals
by the IoT nodes (Tmote Sky). The sensing stage represents
the four steps that the detected signal moves through. First, the
signal x(t)is filtered by band pass filter (BPF). The filtered
signal is then squared in the second step. In the third step,
the signal samples are integrated. Here, the role of jamming
attacks comes by injecting their false energy level to disrupt
the received reports before the fourth step (the thresholding
level) in which the report will represent either the channel is
occupied by PU (H1)oridle(H0). It is worth mentioning that
the jammer continually tries to mislead the FC by injecting false
energy level so as to disrupt the reports received by the FC.
Therefore, if the jammer managed to do this before the fourth
step, the attack will maliciously affect the decision of the FC.
However, if the jammer’s energy level injection is performed
after the fourth step, this will not impact the decision of the
FC. Hereafter, the FC starts the decision-making relying on
the proposed AE-based trust model. The trust model depends
on the gathered delivered reports by the SNs, i.e., SUs. In
other words, by increasing the number of the received benev-
olent reports, the decision taken by the FC will be more
accurate.
III. PROPOSED DAE-TRUST APPROACH
This section illustrates the different aspects of the proposed
approach. On the one hand, we employ the ED mechanism to
determine the spectrum status. To build a realistic model, we
have taken into our consideration the shadow fading (log-normal
distribution) affecting the communications between the PU and
SUs throughout six different environments. On the other hand,
we have developed a DAE model to detect the jamming attacks
affecting the links between PU and SUs so that the delivered
reports at the FC are impacted; thus, wrong decisions can be
taken by the FC about the spectrum. Therefore, the proposed
model can identify the malicious reports received at the FC and
reach an adequate number of delivered trusted reports at the FC,
leading to accurate spectrum detections.
A. Energy Detection
In any spectrum sensing application, the ultimate decision can
be defined by a binary hypothesis testing problem. Assume that
the hypothesis (H0)represents the absence of the PU where the
noise only can be sensed by the SUs, and the hypothesis (H1)
represents the existence of the PU where the noise plus the PU’s
signal can be sensed by the SUs. These two hypotheses include
the impact of the interference power of the attack, where all SUs
are assumed to have the same noise floor. Thus, the spectrum
sensing problem for the ith SU can be formulated as follows:
H0:xi(t)=wi(t),PU absent
H1:xi(t)=γi.s(t)+wi(t),PU exist (1)
where xi(t)is the received signal at the ith SU, γis the channel
gain of the sensing channel from the PU to the ith SU, s(t)is
the PU transmitted signal, w(t)is the additive white Gaussian
noise with zero mean and variance σ2
i, and i∈{1,2,...,|N |}.
The output of the sampling and the integration stages of the
ED can be written as follows:
αi=1
L
L
n=1
|xi(n)|2(2)
where n∈{1,2,...,L}, Lis the number of samples in the
sensing interval.
From the Neyman–Pearson (NP) detector, in order to distin-
guish between the two hypothesis (H1and H0), a test statistics
(αi) with threshold ξiis employed as follows [43]:
αi
H1
>
<
H0
ξi,i=1,2,...,|N | .(3)
Accordingly, the probability of detection Pd,i and the proba-
bility of false alarm Pf,i at ith SU are already impacted by the
attack interference power and can be given by
Pd,i=Pr(αi>ξ
i|H1)=Qξi−E
Eσ2
i,(4)
Pf,i=Pr(αi>ξ
i|H0)=Qξi
Eσ2
i(5)
where Q(·)is a function that represents the tail probability of
the standard normal distribution at ith SU. The threshold level
of every ith report depends on its corresponding received power
level. Eis the PU signal energy. The sensing threshold (ξ)isa
function of the PU’s signal energy and noise variance. Thus, ξi
can be given by
ξi=Q−1(Pd,i)Eσ2
i.(6)
ABDALZAHER et al.: DAE TRUST MODEL FOR MITIGATING JAMMING ATTACK IN IOT ASSISTED BY COGNITIVE RADIO 5
TABLE III
PATH LOSS PARAMETERS [44]
Here, we give more attention to the majority rule, which is
widely used as a hard decision rule (HDR) for CSS in CRNs.
1) Hard Decision Rule: In this method, the FC has to decide
the availability of the spectrum from the received uncorrelated
hard decisions from rSUs out of the available |N | SUs. This
can be defined mathematically by
Pd=
|N |
m=r|N |
mPm
d,i(1 −Pd,i )|N|−m.(7)
Various schemes can be derived from this general formula in
(7) for decision fusion at the FC, e.g., AND,OR, and Majority
combining rules. In this article, we concentrate only on the
Majority rule for HDRs.
Majority Combining: This rule applies the majority voting
concept to identify availability of the spectrum by setting r=
|N |/2. Here, the probability of detection is given by
Pd,MAJ =
|N |
m=|N |/2|N |
mPm
d,i(1 −Pd,i )|N|−m.(8)
B. Path Loss Model
The employed path loss between the PU and the ith SU is
given by
PL
i[dB]=PL
0[dB]+10nlog10
di
d0
+σ[dB](9)
where PL
0is the free space path loss at the reference distance
d0of the antenna far field, nrepresents the path loss exponent,
diis the distance between the PU and ith SU, and σdenotes the
standard deviation in dB of the shadow fading. Table III shows
the path loss parameters for the considered six smart grid (SG)
measurements presented in [44].
The received antenna signal power at the ith SU from the
transmitting PU can be obtained by
PA
r,i [dBm]=PA
t[dBm]−PL
i[dB](10)
where PA
tis the transmission antenna signal power for the PU.
C. Autoencoder-Based Trust Model
Fig. 3 shows the framework of the proposed DAE-TRUST
approach. The proposed DAE-TRUST is employed to tackle
the problem of malicious nodes’ identification in a CR-assisted
IoT system. The DAE-TRUST is the binary individual SU
decisions, i.e., the input features are these binary decisions. The
proposed model adopts unsupervised learning strategy, where
there is no labeled output for every training sample. Instead, the
DAE-TRUST takes the input features and tries to replicate it
Fig. 3. AE process for malicious nodes detection.
on its output. The basic idea is that the DAE-TRUST model is
trained on clean data (i.e., the training samples where all SUs are
benevolent). Specifically, we can obtain the training data from
real field measurements or employing computer-aided simula-
tions. The training data incorporates the sensing decisions of a
multiplicity of benevolent SUs. Upon training the DAE-TRUST
on the benevolent SUs sensing decision, the DAE-TRUST can
be deployed on the FC to distinguish between benevolent and
malicious decision vectors based on a threshold on the mse
between the input vector and the predicted output vector. In other
words, based on this training step, we can determine a certain
mseth below which the input features can be considered clean
without any attacks. Later on, when the trained DAE-TRUST is
deployed for examining the receivedfeatures of new samples, the
actual mse value between the input of the DAE-TRUST model
and its predicted version in the output is determined. If this mse
is less than the mseth, there is no attack. On the other hand, if
this mse is larger than the mseth, attack is detected. Accordingly,
if the DAE-TRUST detected the existence of malicious SUs, the
received decisions of these SUs will be discarded and the FC will
take the spectrum availability decision based on the remaining
benevolent SUs’ decisions.
D. Proposed DAE Structure
The proposed DAE incorporates four convolutional layers,
two fully connected layers, two max-pooling layers, two flatten
layers, and two upsampling layers as depicted in Fig. 4. The
encoder part starts with the first convolutional layer that has
64 filters each with 3×1size. Then, a max-pooling layer with
pool size 2 is employed. After that, the second convolutional
layer that has 32 filters each with 3×1size is added. This is
followed by a max-pooling layer with pool size 2 and a flatten
layer, respectively. The last layer of the encoding part is a fully
connected layer with 32 neurons. Then, the decoder part begins
with the third convolutional layer that has 32 filters each with 3×
1size. This is followed by a 2:1upsampling layer. Afterward,
the fourth convolutional layer that has 64 filters each with 3×
1size is appended. Later on, a 2:1upsampling layer and a
flatten layer are added, respectively. Finally, the last layer of
the encoding part is a fully connected layer with 30 neurons
employed. The rectified linear unit activations are employed in
all convolutional layers and fully connected layers.
6IEEE SYSTEMS JOURNAL
Fig. 4. Structure of the proposed DAE-TRUST for malicious nodes detection.
E. Evaluation Process
The main loss function metric used in the evaluation process
is the mse, which is denoted by
mse =1
NYT−YP(11)
where YTis the set of true values, .represents the Euclidean
norm, and YPdenotes the set of predicted values which can be
given by
YP=f(X;θ)(12)
where θis the parameters including the sets of weights (W)
and bias values (b). The following section discusses the AE
formulations.
F. Autoencoder Formulations
Basically, the AE notion is to filter the added noise aiming
at matching the input to the output based on the concept of
unsupervised learning. The output of AE can be denoted by
YAE =FPθ(XAE),(13)
X=[X1,X
2,...,X
N](14)
where Xrepresents the input dataset of the AE, FP represents
the full-forward propagation function, and NFis the number of
features. When the target values match the input ones, the FP
satisfies the following:
FPθ(X)=X.(15)
Relying on the backpropagation base, the output is gradually
enhanced with respect to the learning rate (r) by updating the
parameters as follows:
θi+1 =θi±mi(16)
where miis the ith parameter update, which is computed by
mi=r∂LF
∂θi
(17)
TAB LE I V
SIMULATION PARAMETERS
where ∂LF
∂θiis the gradient of the ith parameter, and LF represents
the loss function represented by the mse, which given by
LF =1
NF
TS
S
(YS−XS)2(18)
where TS is the total number of samples representing the input
dataset matrix (X), Sis a vector sample, and XSand YS
represent the input and output of the AE’s sample S, respectively.
IV. SIMULATION RESULTS AND DISCUSSION
In this section, the obtained results for evaluating the proposed
DAE-TRUST model are presented. The utilized simulation pa-
rameters are summarized in Table IV. In our evaluations, we
employed the Tmote Sky nodes as SUs. The Tmote sky node,
which is usually used for IoT applications, is equipped by Texas
Instruments MSP430 microcontroller with Chipcon CC2420
radio [41] and follows the IEEE 802.15.4 standard [42]. We
assumed a 100% attack model and studied the impact of chang-
ing the number of malicious SUs. Specifically, we considered
the number of malicious SUs are 7, 5, or 3 SUs out of the total 30
SUs utilized in the spectrum detection. Furthermore, we studied
the impact of the attack model by adopting attack scenarios with
100%,70%, and 30% attack probability when the number of
malicious SUs is fixed to 5.
We have generated training dataset for every environment to
be employed for training the DAE-TRUST model to identify the
malicious SUs in the corresponding environment. Specifically,
we generated training samples for the Pfin the range from 0.01
to 1 with step 0.1. For every Pf, we generated 104samples.
The training features are the binary decisions taken by every
SU. It is noteworthy that in the training phase, we employ the
trustful samples to avoid training the DAE-TRUST on corrupted
samples. In fact, the mseth applied for a certain value of Pfis
calculated by
mseth =0.0001 ×1+10Pf
2.(19)
In this regard, we emphasize that the overhead in the pro-
posed scheme mostly lies in the training process which can be
done offline. After the training process, the DAE-TRUST model
will be deployed online where only simple operations such as
matrix multiplication, addition, and nonlinear activations are
performed. Then, the mse between the input and the output of
the DAE-TRUST model is calculated. This is the total overhead
that will be repeated for every decision vector received from
the SUs. It should be noted that this online repeated overhead
is more significant than the training overhead. Our experiments
ABDALZAHER et al.: DAE TRUST MODEL FOR MITIGATING JAMMING ATTACK IN IOT ASSISTED BY COGNITIVE RADIO 7
Fig. 5. Accuracy and loss of the training and validation datasets for the proposed DAE-TRUST model when random distribution of SUs is assumed. (a)
Environment 1 (OL) random. (b) Environment 2 (ON) random. (c) Environment 3 (UL) random. (d) Environment 4 (UN) random. (e) Environment 5 (IL) random.
(f) Environment 6 (IN) random.
Fig. 6. Accuracy and loss of the training and validation datasets for the proposed DAE-TRUST model when static distribution of SUs is assumed. (a) Environment
1 (OL) static. (b) Environment 2 (ON) static. (c) Environment 3 (UL) static. (d) Environment 4 (UN) static. (e) Environment 5 (IL) static. (f) Environment 6 (IN)
static.
8IEEE SYSTEMS JOURNAL
Fig. 7. Majority rule ROC curve for the proposed DAE-TRUST based malicious SUs detection compared with both cases of reputation-based malicious SUs
detection and the scenario without any protection against malicious SUs over six environments OL, ON, UL, UN, IL, and IN when attack attempts are 100%, 70%,
and 30%, respectively. (a) Environment 1 (OL). (b) Environment 2 (ON). (c) Environment 3 (UL). (d) Environment 4 (UN). (e) Environment 5 (IL). (f) Environment
6 (IN).
Fig. 8. Majority rule ROC curve for the proposed DAE-TRUST based malicious SUs detection compared with both cases of reputation-based malicious SUs
detection and the scenario without any protection against malicious SUs over six environments OL, ON, UL, UN, IL, and IN when number of malicious nodes is 7,
5, and 3, respectively. (a) Environment 1 (OL). (b) Environment 2 (ON). (c) Environment 3 (UL). (d) Environment 4 (UN). (e) Environment 5 (IL). (f) Environment
6 (IN).
ABDALZAHER et al.: DAE TRUST MODEL FOR MITIGATING JAMMING ATTACK IN IOT ASSISTED BY COGNITIVE RADIO 9
Fig. 9. DAE-TRUST model accuracy of detecting malicious SUs’ IDs over the six environments OL, ON, UL, UN, IL, and IN. (a) Environment 1 (OL). (b)
Environment 2 (ON). (c) Environment 3 (UL). (d) Environment 4 (UN). (e) Environment 5 (IL). (f) Environment 6 (IN).
show that the online repeated overhead is in the range of 48.3μs
per sample.
In the proposed approach, the SUs are assumed to be randomly
allotted around the PU within a circular disc of 100 m radius;
we refer to this scenario in the rest of the article as a random
distribution of SUs. By doing so, we can study the effect of
large-scale fading on the performance of the proposed DAE-
TRUST model. In this article, we specifically concentrate on
the jamming attack problem on the link between the PU and the
SUs. Thus, we assume that the links between the SUs and the
FC are error-free.
Besides, the received signal at the ith SU is affected by noise
power Pn(i.e., the variance of the noise signal wi(t)), path loss,
and the log-normal shadowing fading with standard deviation
σ. We considered six different environment scenarios, i.e., OL,
ON, UL, UN, IL, and IN, where their path loss parameters are
introduced in Table III [44]. Note that the Rayleigh fading has
not been considered in our model.
Generally speaking, the learning behavior of the proposed
DAE-TRUST model is depicted by the accuracy and loss on both
the training and the validation datasets as shown in Figs. 5 and 6.
Fig. 5 shows the learning behavior based on the random topology
between the SUs and the PU. On the other hand, Fig. 6 illustrates
the learning behavior using the static topology between the
SUs and the PU, where the SUs are equally allocated at the
maximum radio range from the PU. The accuracy and loss are
represented by the percentage and mse, respectively. It is clearly
denoted that the network structure of the proposed DAE-TRUST
model proves beneficial regardless of the environment type (i.e.,
OL, ON, UL, UN, IL, and IN) as the accuracy on the train-
ing/validation dataset and loss on the training/validation dataset
reach 100% and less than 0.02, respectively. Although, the ON
environment is the worst-case scenario and UL environment is
the best one due to the maximal and minimal path loss exponent,
respectively, as mentioned in Table III, the DAE-TRUST model
loss on both training and validation datasets are sufficiently
low. Accordingly, this reflects the effectiveness of the proposed
DAE-TRUST model structure.
Fig. 7 depicts the receiver operating characteristic (ROC)
curves based on the majority rule for the proposed DAE-TRUST
based malicious SUs detection compared with both cases of
reputation-based malicious SUs detection and the traditional
ED scenario without any protection against malicious SUs over
six environments OL, ON, UL, UN, IL, and IN when the SUs
are randomly distributed taking into consideration three attack
models (100%, 70%, and 30% attack attempts). Specifically,
the performance of the proposed DAE-TRUST is the best com-
pared to the other cases when the reputation-based malicious
SUs detection or when there is no protection against malicious
SUs in all the considered environments, where DAE-TRUST
can achieve the highest detection probability Pdat a certain
probability of false alarm Pf. It is clear that the ROC perfor-
mance of our proposed DAE-TRUST model is always better
than other competing algorithms for both cases of the randomly
distributed SU regardless of the change of the attack model. This
affirms the effectiveness of the proposed DAE-TRUST model in
different scenarios. It is noted from the obtained results that the
reputation-based SS scheme shows performance degradation as
compared to without protection case when the probability of
false alarm Pfis increased. This is because the reputatuion-
based SS scheme relies on accumulated reputation score calcu-
lated for each SU, as the reputation is strictly negatively affected
by the increase of Pf. More concretely, the increase of Pfis
represented by more false reports that are depicted by those
SUs, and, hence, their reputation is deteriorated at the FC to be
wrongly considered malicious SUs besides the real malicious
SUs. Consequently, the reputation-based gets worse with the
increase of Pf.
Moreover, to make a strong consensus about whether the pro-
posed model can achieve an enhanced performance compared
to the other benchmarks (reputation-based and traditional ED
methodology), we have examined the proposed DAE-TRUST
10 IEEE SYSTEMS JOURNAL
to the benchmarks with the change of the number of malicious
SUs over the six environments. We have adopted three numbers
of malicious SUs 7, 5, and 3 in each environment. Fig. 8 clearly
shows that the proposed DAE-TRUST outperforms both the
reputation-based and the traditional ED technique regardless of
the number of malicious SUs and the environment type. This
assures the proposed DAE-TRUST model efficiency.
After the end of the training phase, the trained DAE-TRUST
model can be deployed on the FC to detect malicious SUs. The
accuracy of the proposed DAE-TRUST model in detecting the
malicious SUs over the considered six environments is depicted
in Fig. 9. In this figure, we show the obtained results based
on the random topology of the SUs around the PU along with
the change of Pf. We denote that, in the cases of line-of-sight
environments, the obtained results show a better performance of
detecting the malicious nodes compared to the performance of
the nonline-of-sight environments. It is noted that the proposed
approach achieves feasible accuracy of detection in the six
adopted environments. In the OL environment, about 100%
accuracy of detecting the malicious SUs is achieved till 40%
Pf, while a slight decrease of the accuracy occurs when ON is
utilized. The UL shows a better performance till 60% Pf. Finally,
the best performance is achieved with the IN environment in
which the detection accuracy covers most of the Pf.
V. C ONCLUSION
This article proposes a novel DAE-based trust model for
CR-assisted IoT to achieve a confident decision about the spec-
trum status against the jamming attacks that affect the delivered
reports at the FC. The proposed model relied on the ED for
detecting the PU signal using the sensed reports of the distributed
SUs. To build a realistic model with the genuine dataset, the
FC relied on the Tmote Sky nodes, the most common SNs
with the standard of IEEE 802.15.4 for IoT applications, to
determine the spectrum status taking into account the shadow
fading effect. The model effectiveness has been verified over
six different environments (i.e., OL, ON, UL, UN, IL, and
IN). The simulation results demonstrate the robustness of the
proposed model regardless of the environment type and the SUs
distribution scenario attaining an efficient CR-assisted IoT.
REFERENCES
[1] Y. Wang, X. Tang, and T. Wang, “A unified QoS and security provi-
sioning framework for wiretap cognitive radio networks: A statistical
queueing analysis approach,” IEEE Trans. Wireless Commun., vol. 18,
no. 3, pp. 1548–1565, Mar. 2019.
[2] A. H. Anwar, K. G. Seddik, T.ElBatt, and A. H. Zahran, “Effective capacity
of delay constrained cognitive radio links exploiting primary feedback,”
IEEE Trans. Veh. Technol., vol. 65, no. 9, pp. 7334–7348, Sep. 2016.
[3] G. A. Shah, V. C. Gungor, and O. B. Akan, “A cross-layer QoS-aware
communication framework in cognitive radio sensor networks for smart
grid applications,” IEEE Trans. Ind. Inform., vol. 9, no. 3, pp. 1477–1485,
Aug. 2013.
[4] A. Alshamrani, X. S. Shen, and L.-L. Xie, “QoS provisioning for hetero-
geneous services in cooperative cognitive radio networks,” IEEE J. Sel.
Areas Commun., vol. 29, no. 4, pp. 819–830, Apr. 2011.
[5] A. M. Arafa, K. G. Seddik, A. K. Sultan, T. ElBatt, and A. A. El-Sherif,
“A feedback-soft sensing-based access scheme for cognitive radio net-
works,” IEEE Trans. Wireless Commun., vol. 12, no. 7, pp. 3226–3237,
Jul. 2013.
[6] A. Guirguis, M. Ibrahim, K. Seddik, K. Harras, F. Digham, and M. Youssef,
“Primary user aware k-hop routing for cognitive radio networks,” in Proc.
IEEE Global Commun. Conf., 2015, pp. 1–6.
[7] A. M. Bedewy, A. A. El-Sherif, K. G. Seddik, and T. ElBatt, “Coop-
erative MAC for cognitive radio network with energy harvesting and
randomized service policy,” in Proc. IEEE Globecom Workshops, 2015,
pp. 1–7.
[8] I. Kakalou, K. E. Psannis, P. Krawiec, and R. Badea, “Cognitive ra-
dio network and network service chaining toward 5G: Challenges and
requirements,” IEEE Commun. Mag., vol. 55, no. 11, pp. 145–151,
Nov. 2017.
[9] C. Majumdar, D. Lee, A. A. Patel, S. Merchant, and U. B. Desai, “Packet-
size optimization for multiple-input multiple-output cognitive radio sensor
networks-aided internet of things,” IEEE Access, vol. 5, pp. 14419–14440,
2017.
[10] T. Yucek and H. Arslan, “A survey of spectrum sensing algorithms for
cognitive radio applications,” IEEE Commun. Surv. Tut., vol. 11, no. 1,
pp. 116–130, Jan./Mar. 2009.
[11] H. Sun, A. Nallanathan, C.-X. Wang, and Y. Chen, “Wideband spectrum
sensing for cognitive radio networks: A survey,” IEEE Wireless Commun.,
vol. 20, no. 2, pp. 74–81, Apr. 2013.
[12] I. F. Akyildiz, W.-Y. Lee, M. C. Vuran, and S. Mohanty, “A survey on
spectrum management in cognitive radio networks,” IEEE Commun. Mag.,
vol. 46, no. 4, pp. 40–48, Apr. 2008.
[13] A. Ahmad, S. Ahmad, M. H. Rehmani, and N. U. Hassan, “A survey
on radio resource allocation in cognitive radio sensor networks,” IEEE
Commun. Surv. Tut., vol. 17, no. 2, pp. 888–917, Apr./Jun. 2015.
[14] S. Salim and S. Moh, “An energy-efficient game-theory-based spectrum
decision scheme for cognitive radio sensor networks,” Sensors, vol. 16,
no. 7, 2016, Art. no. 1009.
[15] S. H. R. Bukhari, M. H. Rehmani, and S. Siraj, “A survey of channel
bonding for wireless networks and guidelines of channel bonding for
futuristic cognitive radio sensor networks,” IEEE Commun. Surv. Tut.,
vol. 18, no. 2, pp. 924–948, Apr./Jun. 2016.
[16] A. A. Khan, M. H. Rehmani, and M. Reisslein, “Cognitive radio for smart
grids: Survey of architectures, spectrum sensing mechanisms, and net-
working protocols,”IEEE Commun. Surv. Tut., vol. 18, no. 1, pp. 860–898,
Jan. 2016.
[17] Z. Quan, D. Li, and Y. Gong, “Cooperative signal classification using
spectral correlation function in cognitive radio networks,” in Proc. IEEE
Int. Conf. Commun., 2016, pp. 1–6.
[18] H. Zhang, Y. Qi, H. Zhou, J. Zhang, and J. Sun, “Testing and defending
methods against dos attack in state estimation,” Asian J. Control, vol. 19,
no. 4, pp. 1295–1305, 2017.
[19] F.-L. Luo, Machine Learning for Future Wireless Communications. Hobo-
ken, NJ, USA: John Wiley & Sons, 2020.
[20] J. Wang, C. Jiang, H. Zhang, Y. Ren, K.-C. Chen, and L. Hanzo, “Thirty
years of machine learning: The road to pareto-optimal wireless networks,”
IEEE Commun. Surv. Tut., vol. 22, no. 3, pp. 1472–1514, Jul./Sep. 2020.
[21] M. Elwekeil, S. Jiang, T. Wang, and S. Zhang, “Deep convolutional neural
networks for link adaptations in MIMO-OFDM wireless systems,” IEEE
Wireless Commun. Lett., vol. 8, no. 3, pp. 665–668, Jun. 2019.
[22] I. Goodfellow, Y. Bengio, and A. Courville, “Deep learning,” MIT Press,
2016. [Online]. Available: http://www.deeplearningbook.org.
[23] B. Pradhan, “A comparative study on the predictive ability of the decision
tree, support vector machine and neuro-fuzzy models in landslide sus-
ceptibility mapping using GIS,” Comput. Geosci., vol. 51, pp. 350–365,
2013.
[24] M. Elwekeil, T. Wang, and S. Zhang, “Deep learning for environment
identification in vehicular networks,” IEEE WirelessCommun. Lett.,vol.9,
no. 5, pp. 576–580, May 2019.
[25] Z. Gao, H. Zhu, S. Li, S. Du, and X. Li, “Security and privacy of col-
laborative spectrum sensing in cognitive radio networks,” IEEE Wireless
Commun., vol. 19, no. 6, pp. 106–112, Dec. 2012.
[26] A. Attar, H. Tang, A. V. Vasilakos, F. R. Yu, and V. C. Leung, “A
survey of security challenges in cognitive radio networks: Solutions and
future research directions,” Proc. IEEE, vol. 100, no. 12, pp. 3172–3186,
Dec. 2012.
[27] A. G. Fragkiadakis, E. Z. Tragos, and I. G. Askoxylakis, “A survey on
security threats and detection techniques in cognitive radio networks,”
IEEE Commun. Surv. Tut., vol. 15, no. 1, pp. 428–445, Jan./Mar. 2012.
[28] M. S. Abdalzaher, K. Seddik, and O. Muta, “Using Stackelberg game to
enhance cognitive radio sensor networks security,” IET Commun., vol. 11,
no. 9, pp. 1503–1511, 2017.
[29] M. S. Abdalzaher and O. Muta, “Employing game theory and TDMA
protocol to enhance security and manage power consumption in WSNS-
based cognitive radio,” IEEE Access, vol. 7, pp. 132923–132936, 2019.
[30] M. S. Abdalzaher and O. Muta, “A game-theoretic approach for enhancing
security and data trustworthiness in IoT applications,” IEEE Internet
Things J., vol. 7, no. 11, pp. 11250–11261, Nov. 2020.
ABDALZAHER et al.: DAE TRUST MODEL FOR MITIGATING JAMMING ATTACK IN IOT ASSISTED BY COGNITIVE RADIO 11
[31] C. Clancy, J. Hecker, E. Stuntebeck, and T. O’Shea, “Applications of
machine learning to cognitive radio networks,” IEEE Wireless Commun.,
vol. 14, no. 4, pp. 47–52, Aug. 2007.
[32] K. M. Thilina, K. W. Choi, N. Saquib, and E. Hossain, “Machine learning
techniques for cooperative spectrum sensing in cognitive radio networks,”
IEEE J. Sel. Areas Commun., vol. 31, no. 11, pp. 2209–2221, Nov. 2013.
[33] S. Zhang, Y. Wang, J. Li, P. Wan,Y. Zhang, and N. Li, “A cooperativespec-
trum sensing method based on information geometry and fuzzy C-means
clustering algorithm,” EURASIP J. Wireless Commun. Netw., vol. 2019,
no. 1, pp. 1–12, 2019.
[34] J. Tian et al., “A machine learning-enabled spectrum sensing method
for OFDM systems,” IEEE Trans. Veh. Technol., vol. 68, no. 11,
pp. 11374–11378, Nov. 2019.
[35] H. He and H. Jiang, “Deep learning based energy efficiency optimization
for distributed cooperative spectrum sensing,” IEEE Wireless Commun.,
vol. 26, no. 3, pp. 32–39, Jun. 2019.
[36] J. Gao, X. Yi, C. Zhong, X. Chen, and Z. Zhang, “Deep learning for spec-
trum sensing,” IEEE Wireless Commun. Lett., vol. 8, no. 6, pp. 1727–1730,
Dec. 2019.
[37] M. Dabaghchian, A. Alipour-Fanid, K. Zeng, Q. Wang, and P. Auer,
“Online learning with randomized feedback graphs for optimal PUE
attacks in cognitive radio networks,” IEEE/ACM Trans. Netw., vol. 26,
no. 5, pp. 2268–2281, Oct. 2018.
[38] Y. Shi, Y. E. Sagduyu, T. Erpek, K. Davaslioglu, Z. Lu, and J. H. Li,
“Adversarial deep learning for cognitive radio security: Jamming attack
and defense strategies,” in Proc. IEEE Int. Conf. Commun. Workshops,
2018, pp. 1–6.
[39] C. D. Ho, T.-V. Nguyen, T. Huynh-The, T.-T. Nguyen, D. B. da Costa,
and B. An, “Short-packet communications in wireless-powered cognitive
IoT networks: Performance analysis and deep learning evaluation,” IEEE
Trans. Veh. Technol., vol. 70, no. 3, pp. 2894–2899, Mar. 2021.
[40] M. Elwekeil, M. S. Abdalzaher, and K. Seddik, “Prolonging smart grid
network lifetime through optimising number of sensor nodes and packet
length,” IET Commun., vol. 13, no. 16, pp. 2478–2484, 2019.
[41] Tmote Sky: Datasheet. [Online]. Available: https://insense.cs.st-andrews.
ac.uk/files/2013/ 04/tmote-sky-datasheet.pdf , 2013.
[42] S. Raza, S. Duquennoy, T. Chung, D. Yazar, T. Voigt, and U. Roedig,
“Securing communication in 6LoWPANwith compressed IPsec,” in Proc.
Int. Conf. Distrib. Comput. Sensor Syst. Workshops, 2011, pp. 1–8.
[43] F. Salahdine, H. El Ghazi, N. Kaabouch, and W. F. Fihri, “Matched filter
detection with dynamic threshold for cognitive radio networks,” in Proc.
Int. Conf. Wireless Netw. Mobile Commun., 2015, pp. 1–6.
[44] V. C. Gungor, B. Lu, and G. P. Hancke, “Opportunities and challenges
of wireless sensor networks in smart grid,” IEEE Trans. Ind. Electron.,
vol. 57, no. 10, pp. 3557–3564, Oct. 2010.
Mohamed S. Abdalzaher (Member, IEEE) received
the B.Sc. (Hons.) degree in electronics and commu-
nications engineering in 2008, from Obour High In-
stitute for Engineering and Technology, the M.Sc. de-
gree in electronics and communications engineering
from Ain Shams University, Cairo, Egypt, in 2012,
and the Ph.D. degree in electronics and communi-
cations engineering from the Electronics and Com-
munications Engineering Department, Egypt-Japan
University of Science and Technology, Madinet Borg
Al Arab, Egypt, in 2016.
He is currently an Assistant Professor with the National Research Institute
of Astronomy and Geophysics, Cairo, Egypt. He was a special research student
with Kyushu University, Fukuoka, Japan, from 2015 to 2016. In 2019, he joined
the Center for Japan-Egypt Cooperation in Science and Technology, Kyushu
University, where he was a Postdoctoral Researcher. His research interests
include earthquake engineering, data communication networks, wireless com-
munications, WSNs security, IoT, and deep learning.
Dr. Abdalzaher is a TPC Member with Vehicular Technology Conference
and International Japan-Africa Conference on Electronics, Communications
and Computers and a Reviewer for the IEEE Internet of Things Journal,IEEE
Systems Journal,IEEE Access,Transactions on Emerging Telecommunications
Technologies,Applied Soft Computing,Journal of Ambient Intelligence and
Humanized Computing,andIET Journals.
Mohamed Elwekeil received the B.Sc. degree in
electronics and electrical communications engineer-
ing from the Faculty of Electronic Engineering,
Menoufia University, Al Minufiyah, Egypt, in 2007,
and the M.Sc. and Ph.D. degrees in electronics and
communications engineering from Egypt-Japan Uni-
versity of Science and Technology(E-JUST), Alexan-
dria, Egypt, in 2013 and 2016, respectively.
He was a Teaching Assistant with the Department
of Electronics and Electrical Communications Engi-
neering, Faculty of Electronic Engineering, Menoufia
University in the period from 2007 to 2016. In 2016, he was promoted to a
Lecturer (Assistant Professor) position in the same department. In the period
from 2014 till 2015, he was a Research Intern with Alcatel-Lucent Bell N.V.
(now NOKIA), Antwerp, Belgium, where he was working on WiFi optimization
project. In 2015, he joined Kyushu University, Fukuoka, Japan, as a special
research student for a period of nine months. In the period from 2018 to 2020,
he was with the College of Information Engineering, Shenzhen University,
Shenzhen, China, where he was working as a Postdoctoral Researcher. His
research interests include radio resource management for wireless networks,
spatial modulation, signal processing for communications, and earthquake
engineering.
Taotao Wang received the B.S. degree in electrical
engineering from the University of Electronic Science
and Technology of China, Chengdu, China, in 2008,
the M.S. degree in information and signal processing
from Beijing University of Posts and Telecommu-
nications, Beijing, China, in 2011, and the Ph.D.
degree in information engineering from The Chinese
University of Hong Kong, Hong Kong, in 2015.
From 2015 to 2016, he was a Postdoctoral Research
Fellow with the Institute of Network Coding, The
Chinese University of Hong Kong. After that, he
joined the College of Information Engineering, Shenzhen University, Shenzhen,
China, as an Assistant Professor, and he was promoted as a tenured Associate
Professor in 2021. His main research interests include blockchain technology,
wireless communications and networking, statistical signal, and data processing.
Dr. Wang was a recipient of the Hong Kong Ph.D. Fellowship.
Shengli Zhang (Senior Member, IEEE) received
the B.Eng. degree in electronic engineering and the
M.Eng. degree in communication and information
engineering from the University of Science and Tech-
nology of China (USTC), Hefei, China, in 2002 and
2005, respectively. He received the Ph.D. degree in
information engineering from the Department of In-
formation Engineering, The Chinese University of
Hong Kong (CUHK), Hong Kong, in 2008.
After studies, he joined the Communication Engi-
neering Department, Shenzhen University, Shenzhen,
China, where he is currently a Full Professor. From 2014 to 2015, he was a Visit-
ing Associate Professor with Stanford University, Stanford, CA, USA. He is the
pioneer of physical-layer network coding (PNC). He has authored or coauthored
more than 20 IEEE top journal papers and ACM top conference papers, including
IEEE Journal on Selected Areas in Communications, IEEE TRANSACTIONS ON
WIRELESS COMMUNICATIONS, IEEE TRANSACTIONS ON MOBILE COMPUTING,
IEEE TRANSACTIONS ON COMMUNICATIONS,andACM Mobicom. His research
interests include blockchain, PNC, and wireless networks.
Dr. Zhang servedas an Editor for IEEE TRANSACTIONS ON VEHICULAR TECH-
NOLOGY,IEEE WirelessCommunications Letters,andIET Communications.He
has also served as a TPC member in several IEEE conferences.
