Received 1 July 2023, accepted 23 July 2023, date of publication 15 August 2023, date of current version 25 August 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3305388

Deep Neural Networks for Spectrum

Sensing: A Review
SADAF NAZNEEN SYED 1 , (Graduate Student Member, IEEE), PAVLOS I. LAZARIDIS 1, (Senior Member, IEEE),
FAHEEM A. KHAN 1 , (Member, IEEE), QASIM ZEESHAN AHMED 1 , (Member, IEEE),
MARYAM HAFEEZ1 , (Member, IEEE), ANTONI IVANOV 2,3 , (Member, IEEE),
VLADIMIR POULKOV 2,3 , (Senior Member, IEEE),
AND ZAHARIAS D. ZAHARIS 4 , (Senior Member, IEEE)
1 School of Computing and Engineering, University of Huddersfield, HD1 3DH Huddersfield, U.K.
2 Facultyof Telecommunications, Technical University of Sofia, 1000 Sofia, Bulgaria
3 Research and Development and Innovation Consortium, Sofia Tech Park, 1784 Sofia, Bulgaria
4 School of Electrical and Computer Engineering, Aristotle University of Thessaloniki, 541 24 Thessaloniki, Greece

Corresponding author: Pavlos I. Lazaridis ([email protected])

This work was supported by the European Union through the Horizon 2020 Marie Skłodowska-Curie Innovative Training Networks
Programme ‘‘Mobility and Training for beyond 5G Ecosystems (MOTOR5G)’’ under grant agreement no. 861219. The authors would like
to acknowledge the support of the Intelligent Communication Infrastructures Laboratory, Research and Development and Innovation
Consortium, Sofia Tech Park, Bulgaria.

ABSTRACT As we advance towards 6G communication systems, the number of network devices continues
to increase resulting in spectrum scarcity. With the help of Spectrum Sensing (SS), Cognitive Radio (CR)
exploits the frequency spectrum dynamically by detecting and transmitting in underutilized bands. The
performance of 6G networks can be enhanced by utilizing Deep Neural Networks (DNNs) to perform SS.
This paper provides a detailed survey of several Deep Learning (DL) algorithms used for SS by classifying
them as Multilayer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), Long Short-Term Memory
(LSTM) networks, combined CNN-LSTM architectures and Autoencoders (AEs). The works are discussed
in terms of the input provided to the DL algorithm, data acquisition technique used, data pre-processing
technique used, architecture of each algorithm, evaluation metrics used, results obtained, and comparison
with standard SS detectors. This survey further provides an overview of traditional Machine Learning (ML)
algorithms and simple Artificial Neural Networks (ANNs) while highlighting the drawbacks of conventional
SS approaches for completeness. A description of some publicly available Radio Frequency (RF) datasets is
included and the need for comprehensive RF datasets and Transfer Learning (TL) is discussed. Furthermore,
the research challenges related to the use of DL for SS are highlighted along with potential solutions.

INDEX TERMS Autoencoders, cognitive radio, convolutional neural networks (CNNs), deep learning (DL),
deep neural networks (DNNs), machine learning (ML), multilayer perceptrons (MLPs), long short-term
memory (LSTM) networks, spectrum sensing, 6G.

I. INTRODUCTION Internet of Everything, Holographic Telepresence, Extended

The global network traffic continues to rise exponentially Reality, Industry 5.0, and Intelligent Healthcare have moti-
due to the growing popularity and ease of accessing wireless vated the development of 6G technology [6]. The possibility
devices [1], [2], [3], [4]. As per a forecast by Ericsson [5], of finding a new vacant spectrum to accommodate these
the number of mobile subscriptions will rise from about service requirements has become an increasingly difficult
8.4 billion in 2022 to around 9.2 billion by the end of 2028. and expensive process [7]. The problem of spectrum scarcity
The heterogeneous demands of the emerging applications like can be attributed to the traditional fixed spectrum allocation
policies which allocated a large part of the radio spectrum to
The associate editor coordinating the review of this manuscript and licensed Primary Users (PUs) and left a smaller portion of
approving it for publication was Usama Mir . the spectrum for unlicensed Secondary Users (SUs) [8], [9].

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 11, 2023 89591
 S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

These allocation techniques permitted only the PUs to utilize TABLE 1. List of acronyms in alphabetical order.
the spectrum despite the resources being idle [10]. Since a
considerable portion of the expensive frequency resources
provided by wireless systems are not always utilized in many
regions, the static spectrum assignment of most current and
legacy networks is quite inefficient [11]. These techniques
inherently protect the PUs from interference but can deny
critical radio requirements of SUs even when multiple bands
of PU are unused. With the massive rise in the number
of network devices, static allocation techniques will rapidly
exhaust the radio resources.
Dynamic Spectrum Management (DSM) is a flexible
spectrum allocation technique that allows the SUs to access
the PU spectrum if it is idle, or to even share the PU spectrum
if the transmission of PU is protected from interference [12].
Cognitive Radio (CR) is the principal enabler of DSM [13]
and ensures that the spectrum is utilized in an efficient
manner by allowing the SUs to opportunistically access the
idle frequency bands of the PUs [14]. The term ‘CR’ was
first proposed by Joseph Mitola [15] and Simon Haykin
introduced the basic cognitive cycle in [16]. Wireless regional
area network (IEEE 802.22) is the first international standard
based on CR, and the CR technology is now being used
in several standards, including Zigbee (IEEE 802.15.4) and
Wi-Fi (IEEE 802.11) [17]. Spectrum Sensing (SS) is one of
the most crucial techniques of CR which provides real-time
occupancy information of frequency bands that are available
for SUs without interfering with PUs’ operations [18].
An accurate detection of the occupancy of spectrum bands
is critical in CR operation, since all secondary transmission
strategies are based on it [19].
The PU spectrum has been traditionally sensed with
popular SS techniques like Energy Detection (ED), Matched
Filter Detection (MFD) and Cyclostationary-based Detection
(CBD). These conventional techniques result in inefficient
use of radio resources due to missed detection of the PU and
false alarms [20]. Due to the drawbacks of conventional SS
techniques and increasing popularity of Artificial Intelligence
(AI), many recent works have used traditional Machine
Learning (ML) algorithms like Support Vector Machines
(SVMs) and K-Nearest Neighbor (KNN) for SS. However,
the manual feature extraction process involved in traditional
ML algorithms requires expert knowledge and is a time-
consuming process. Deep Learning (DL) is a data-driven
approach and also a subset of ML that can automatically
capture complex patterns and features from input data [21].
Deep Neural Network (DNN) techniques being quickly
adaptable are robust to uncertain radio environment. The
main objective of this article is to survey the latest research
efforts towards the application of DL for the important task
of sensing the PU spectrum.

A. INTRODUCTION TO SPECTRUM SENSING In CR-based DSM, SS is a crucial step to learn the radio
The radio spectrum is inefficiently exploited because of the environment [12] and help increase spectrum utilization.
fixed allocation policies and the growth in user needs [22]. SS is used to continuously sense the licensed user spectrum

89592 VOLUME 11, 2023

S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

TABLE 1. (Continued.) List of acronyms in alphabetical order. detector receives and utilizes a priori information if required
and generates a decision about PU presence or absence.
SS techniques can be grouped based on the number of users
as multi-user or Cooperative Spectrum Sensing (CSS) and
single user or local or non-CSS [25]. In CSS, a group of SUs
share their sensed data (soft fusion) or decisions (hard fusion)
with fusion center to improve the sensing accuracy. Local
sensing results are different for SUs due to the differences
in sensing capabilities [26]. The SS performance of a CR
system is determined by the global decision combination rule
of fusion center along with other factors like the number
of SUs, the environment for SS, and the capabilities of
SUs [27]. The combination of soft decisions results in an
optimal detection performance but needs an infinite amount
of bandwidth in theory [28]. Hard fusion on the other
hand produces inferior results while saving bandwidth [29].
Performance and bandwidth efficiency can be balanced by
using a combination of hard and soft decisions from the
SUs [27].
Non-CSS techniques suffer from the hidden PU prob-
lem [30]. By using CSS, the performance gain of a CR
system can be increased by the cooperation of multiple SUs to
detect spectrum holes [27]. CSS can overcome problems like
multipath fading and shadowing, ensuring that PU constraints
are met for SS [31], [32].

B. RELATED SURVEYS
There are several survey papers that have reviewed the
various types of SS algorithms and have discussed the
research challenges and future directions associated with it.
For instance, [33] emphasizes the need of exploring the code
and angle dimensions along with frequency, time, and space
dimensions for obtaining complete spectrum awareness.
SS approaches of ED, waveform-based sensing, CBD, radio
identification-based sensing and MFD are presented. The
work further discusses the challenges involved in sensing
the PU spectrum and the concept of CSS and its types.
The survey [34] reviews SS techniques by categorizing
them into three classes based on whether they need both
source signal and noise power information, only noise power
information (semi-blind detection) or no prior information
(totally blind detection), with a particular focus on semi-
blind and blind techniques. An analysis on detection
threshold and test statistics distribution is provided and the
by the SUs to detect PU activity and spectrum holes challenges in developing a practical SS device are discussed.
in terms of duration, frequency, and location [23]. SS is Axell et al. [35] provide the fundamentals of signal detection
modelled as a binary hypothesis testing problem with the two and conventional narrowband and wideband SS detectors.
hypotheses being the presence or absence of the PU [24]. The work describes CSS in detail and discusses energy
The functions of spectrum sharing, decision making, and efficiency in CSS.
resource allocation are implemented if the spectrum is Sharma et al. [36] discuss the enablers of CR along with
available for transmission [17]. Fig. 1 depicts a block diagram the practical imperfections in a CR system. In addition,
representing sensing of the PU spectrum by an SU. The the work provides a classification of popular SS techniques
received radio signals at the SU are first passed through any based on the signal processing techniques used, signal
signal processing, feature extraction and data pre-processing bandwidth, coordination between SUs and number of RF
steps before being provided as input to an SS detector. The SS chains. Another survey [37] provides a detailed review of

VOLUME 11, 2023 89593

 S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

FIGURE 1. Block diagram representing sensing of the PU spectrum by an SU.

traditional SS techniques by categorizing them on the basis A study in [11] reviews probabilistic SS approaches by
of bandwidth as narrowband and wideband SS approaches. grouping them on the basis of what features they extract
Besides describing narrowband sensing techniques like ED, from the samples of received signals. ED, CBD, EBD,
MFD and Eigenvalue-based Detection (EBD) and wideband MFD and blind detection techniques and their sub-categories
sensing techniques like multiband sensing, wavelet-based are described in detail along with a discussion on the
sensing and compressive sensing, the paper discusses the implementation challenges. Based on the applications and
practical implementation aspects for various sensing tech- technologies already envisioned for 6G, the role of SS is
niques. Work [19] first describes the different access modes conceptualized for use in future networks.
for CR and then summarizes four common SS methods: The survey in [23] is based on the applications of ML
ED, MFD, Covariance Absolute Value (CAV) detector and for CSS and dynamic spectrum sharing by focusing on the
Hadamard ratio-based detector. The concept of Signal-to- feature vector extracted from the received signal, the type of
Noise Ratio (SNR) wall is explained and the performance of ML algorithm and evaluation metrics. The ML-based sensing
the four SS detectors is analyzed and compared under various techniques are categorized as supervised, unsupervised and
conditions. reinforcement learning techniques, and are analyzed based
The survey in [25] groups the SS techniques on the basis on the features, type of SUs, and performance metrics used.
of bandwidth as narrowband and wideband, on the basis of The work further classifies the spectrum sharing techniques
number of users as non- cooperative and cooperative, on the and summarizes the use of ML algorithms for spectrum
basis of detection as transmitter and receiver detection, and sharing. Although numerous survey papers exist in the field
finally on the basis of need of prior knowledge as blind, of SS, there is a lack of literature on the latest DL-based
semi-blind and non-blind detection. In addition, a detailed developments for sensing the spectrum. ML-based SS has
comparison of popular SS techniques: ED, MFD, feature been covered in some surveys, but the papers mainly focus on
detection, Waveform Detection (WD) and EBD is provided. traditional ML algorithms and not on the emerging research
The work also discusses the system modelling methods efforts using DL. Table 2 summarizes the key focus of the
for SS, the challenges associated with sensing the PU surveys discussed in this sub-section.
spectrum and CR standards. The work in [17] summarizes the
fundamentals of CR and SS. SS techniques are described, and C. CONTRIBUTION
mathematical models are provided by classifying the schemes In contrast to the existing research, this paper provides a
into conventional methods, such as ED and MFD and recent detailed survey of the recent works that have used DNN
advanced sensing schemes such as wideband compressive algorithms for SS. The key contributions of this paper are as
and adaptive compressive techniques. The survey further follows:
discusses the challenges involved in sensing the PU spectrum • After introducing the concept of SS, this work sum-
and various applications of Cognitive Radio Networks marizes the contribution of several fundamental review
(CRNs). papers in the field of SS. The survey then discusses
A review in [38] provides a classification of SS techniques about conventional SS algorithms and their drawbacks.
based on bandwidth. The narrowband sensing techniques: This will provide a comprehensive guide for readers to
ED, CBD, MFD, covariance-based detection and traditional develop an understanding of the SS research.
ML-based SS are discussed. Works based on traditional ML • We then provide an overview of some works that
techniques like K-means clustering, SVM and KNN are are based on traditional ML algorithms and simple
surveyed. Wideband sensing techniques are further grouped Artificial Neural Networks (ANNs) before discussing
into Nyquist-based and compressive sensing techniques. DNNs for completeness. Summarizing the early works

89594 VOLUME 11, 2023

S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

FIGURE 2. Summary of ML-based SS algorithms discussed in this paper.

that used ML for SS helps create a timeline and • We describe some publicly available Radio Frequency
understanding of research efforts towards adopting AI (RF) datasets and discuss the need for comprehensive
for SS. RF datasets and the concept of Transfer Learning (TL).
• The major portion of this paper is dedicated to • The work further highlights the research challenges
surveying the latest works using DL algorithms for related to the use of DL for SS along with potential
SS by categorizing them into five types of DL solutions.
algorithms: Multilayer Perceptrons (MLPs), Convo-
lutional Neural Networks (CNNs), Long Short-Term D. PAPER ORGANIZATION
Memory (LSTM) networks, combined CNN-LSTM Fig. 2 provides a summary of the ML-based SS algorithms
architectures and Autoencoders (AEs). A detailed discussed in this paper along with their years of publication.
analysis of various DL approaches in terms of the The figure also highlights if a particular work is based
input provided to the DL algorithm, method of on CSS or non-CSS scenario. Fig. 3 summarizes the
data acquisition, data pre-processing technique used, paper organization. We introduced the concept of DSM
architecture of each algorithm, evaluation metrics used, and the problem of spectrum scarcity in Section I. This
results obtained and comparison with standard SS section further covers the related surveys and highlights our
detectors is presented. This discussion will create contribution. Section II describes conventional and ML-based
awareness about using DL techniques for SS. For SS approaches. Section III reviews various DL-based SS
instance, which pre-processing technique and DNN techniques. Section IV provides information about various
architecture should be selected for a high detection RF signal datasets, the concept of TL and software used.
performance. Section V includes research challenges of applying DL for

VOLUME 11, 2023 89595

 S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

FIGURE 3. Survey paper organization.

TABLE 2. Related surveys. II. SPECTRUM SENSING APPROACHES

A. CONVENTIONAL SPECTRUM SENSING
The PU spectrum has traditionally been sensed using popular
sensing techniques like ED, MFD, EBD, CBD, and WD.
These conventional SS algorithms have been extensively
reviewed in several research works, as summarized in
section I. However, these techniques result in an inefficient
utilization of radio resources due to the problem of missed
detection of the licensed user and false alarms [20]. For
instance, there is no need for prior information about the
PU signal before using ED, but it has a high false alarm
rate and is inefficient in environments with low SNRs [38].
The MFD technique becomes impractical to implement
because it requires prior knowledge of the PU signals [25].
EBD technique has high computational complexity [37].
SS using CBD involves high power consumption, processing
complexity and sensing time [17]. There is a possibility of
synchronization errors with WD [25]. SS techniques must
ensure a low Probability of False Alarm (PFA) and a high
Probability of Detection (PoD) to lower the impact of harmful
interference [38]. Due to the drawbacks of conventional SS
techniques and growing popularity of the field of AI, a lot
of recent works have adopted ML algorithms for sensing the
spectrum accurately.

B. MACHINE LEARNING (ML)-BASED SPECTRUM

SENSING
ML-based SS algorithms are based on extracting feature
vectors from patterns and classifying them into either
null hypothesis (absence of PU) or alternative hypothesis
(presence of PU) [27]. These techniques are more adaptive
than conventional sensing techniques due to their learning
ability and when adopted for CSS can achieve a better
detection performance due to their capacity to describe more
SS along with potential solutions while section VI concludes optimized decision region on the feature space [39]. ML can
this paper. help address the problem of spectrum scarcity by increasing

89596 VOLUME 11, 2023

S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

spectrum utilization [40]. ML techniques treat SS as a binary Work [39] proposes unsupervised ML techniques:
classification problem and use energy or probability vectors K-means clustering and Gaussian Mixture Model (GMM)
to predict the status of RF channel [38]. The ML algorithms along with supervised ML techniques: SVM and weighted
can be broadly classified into three classes [41]: KNN. The energy levels received at SUs are considered
• Supervised Learning: where model is trained using input as feature vectors and fed into ML models to predict the
samples and the corresponding labels. channel availability. K-means clustering works by dividing
• Unsupervised Learning: where input samples are distin- the features into ‘K’ clusters and mapping the clusters
guished by the model without any output labels. to the status of PU based on the centroids of clusters.
• Reinforcement Learning: where an agent learns to map GMM is a probabilistic approach which models feature
input to actions by communicating with an environment. vectors as a gaussian mixture distribution so that each
A review of the three types of ML algorithms used for gaussian distribution corresponds to a cluster. For using
CSS is provided in a recent survey [23]. Out of the three SVM, the training energy vectors are made linearly separable
categories, supervised and unsupervised ML techniques are by mapping them to a higher dimensional feature space
popular among researchers in the context of SS. In this work by means of a non-linear mapping function. The SVM
ML techniques for SS are analyzed by dividing them into two algorithm then finds a hyperplane which is at the maximum
classes: traditional ML and ANNs. The literature on ANN- distance from data points of the two classes corresponding
based sensing is scarce, and this survey aims to provide a to availability of PU. In weighted KNN technique the nearest
detailed review of various ANN algorithms with a particular ‘K’ points are assigned weights inversely proportional to their
focus on DL techniques by describing their architectures, distances and classes are predicted based on the majority
input provided to algorithms, outputs produced, evaluation voting of neighbors.
metrics and comparison with standard models. The system model consists of multiple PUs and the
channel is considered available only if all PUs are inactive.
1) TRADITIONAL MACHINE LEARNING-BASED SPECTRUM Multiple SUs estimate the energy levels of received baseband
SENSING complex signal samples and report it to the fusion center
In this work, traditional ML refers to ML algorithms that which generates the energy vector. The ML models are
are not based on neural networks. The use of traditional ML evaluated by comparing their training times, classification
for sensing of radio spectrum has been widely adopted in delays and ROC curves by modelling a system with 25 SUs
several works. For example, to utilize ML algorithms for and 2 PUs. KNN classifier takes the least training time while
CSS, [42] explored the use of Fisher Linear Discriminant SVM needs the highest training duration. On comparing
Analysis (FLDA) for fusing the sensing results from SUs. the average classification delay for various classifiers, it is
FLDA can be considered as a supervised ML technique observed that the classification time remains constant even
and is used to separate two or more classes by determining on increasing the number of training samples for Fisher linear
a linear combination of features [43]. The PU network is discriminant [42], K-means clustering, and GMM techniques.
modelled as a random geometric network and the SUs sense On comparing the ROC curves of various CSS schemes
the spectrum using ED to determine spectrum availability. by varying the number of SUs it is found that a greater
The sensing performance is made accurate by incorporating number of SUs results in better performance for the proposed
location information and reliability of the decision of each classifiers. SVM with linear kernel outperforms SVM with
SU with the help of a linear fusion rule whose coefficients are polynomial kernel, K-means clustering, KNN with Euclidean
determined by FLDA. The Receiver Operating Characteristic and Cityblock distances, GMM as well as the other CSS
(ROC) plots of the proposed scheme are compared with schemes like Fisher linear discriminant, OR rule, and AND
equal coefficient model, AND rule, OR rule and Maximum rule.
Likelihood Detector (MLD)-based rule by considering two Following a similar approach, Lu et al. [44] proposed
circular detection areas. the use of two-dimensional probability vectors in place of
In equal coefficient model, the sensing results are com- high-dimensional energy vectors for K-means clustering and
bined in a linear manner like the proposed model but with SVM-based CSS. The high dimensional feature vector is
same linear coefficients for all SUs without considering the transformed into a low dimensional probability vector to
SU network topology. The OR and AND rules are hard fusion achieve smaller training time and classification delay. The
rules. The OR rule determines the presence of PU if at least performance of K-means clustering, SVM with linear kernel
one SU reports the presence of PU whereas in AND rule the and SVM with polynomial kernel techniques with energy
presence of PU is confirmed when all SUs detect the PU. vector and probability vector is compared to determine their
The MLD following Neyman-Pearson criterion is the optimal probabilities of detection, training durations and classifica-
detector for the problem of random PU network detection. tion delays. When considering a CRN of one PU and 2 SUs,
The detection performance for all models improves with the ML techniques report higher detection probabilities than
the increase in radius of detection and the proposed model OR and AND fusion rules while SVM with probability vector
outperforms the equal coefficient, AND and OR rules-based outperforms all other sensing techniques. In a CRN with
models for both circular detection areas. one PU and 9 SUs, SVM with linear kernel and probability

VOLUME 11, 2023 89597

 S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

vector has the highest detection accuracy. While the training Table 3 summarizes the traditional ML-based research
duration of K-means clustering is the longest, it has the works discussed in this subsection by providing a descrip-
least classification delay. Due to the low dimension of the tion of the input to the traditional ML algorithm, any
probability vector-based ML algorithms, they report lower pre-processing technique used, the ML algorithm used,
training duration and classification delay than energy vector- key evaluation metrics used and performance of the SS
based ML algorithms. model. Several studies have used both supervised and
In [27], the ED of received signals is viewed as an analogy unsupervised traditional ML techniques to sense the PU
to feature vectors for training a KNN classifier [45] for spectrum. Unsupervised learning is practically easier to
CSS. The KNN algorithm considers the sensing classes from implement than supervised learning as it does not require
training phase and current sensing report as neighbors and information about the PU availability, but supervised learning
the Smith-Waterman algorithm [46] calculates the distance techniques demonstrate better sensing performance as they
between the neighbors. Posterior probability is used to have additional information about the status of PU [39].
determine the nearest neighbor and KNN determines the Despite their popularity, the manual feature extraction in
prior and conditional probabilities. Multiple SUs are used traditional ML techniques requires the knowledge of field
to provide spatial diversity and the sensing slot is divided experts and is a time-consuming process. Additionally, the
into mini slots to add temporal diversity. At each mini accuracy of the selection of input features influences the
slot, the energy signal is quantized into discrete zones and detection results [23]. Moreover, most of the works based
multiple bits corresponding to the zones are transmitted on traditional ML are limited to CSS scenario and rely
instead of transmitting soft or hard decisions. The local on features extracted from conventional SS techniques to
sensing reports from SUs are combined by the fusion center provide input to the sensing algorithms. These disadvantages
with the help of a weight-based decision combination rule in of traditional ML algorithms have accelerated the research
which SUs are assigned weights based on their effectiveness. towards adopting neural networks for SS.
The performance of the scheme with training sizes of
330 and 100 is evaluated in Additive White Gaussian Noise 2) ARTIFICIAL NEURAL NETWORK (ANN)-BASED
(AWGN) as well as fading channels and compared with the SPECTRUM SENSING
conventional OR rule. ANNs are inspired by the structure and functioning of a
In AWGN channel, the proposed technique with 330 train- human brain and are widely used to model complex real-
ing size reports the highest PoD while the scheme with world problems in many disciplines [51]. These networks
100 training size has comparable performance with the primarily consist of an input layer, hidden layers, and an
conventional OR method for low SNR values. When output layer. Neural networks are efficient in learning non-
comparing the Probability of Error (PoE), the proposed linear functions and adapting the non-linear features of PU
model with higher training size reports low PoE for low signals [52]. In this paper the ANN-based SS is grouped into
SNR values. For high SNR ranges, both proposed schemes simple ANN techniques which include neural networks with
have similar performance, confirming that the model gives a single hidden layer and DL which is covered in detail in the
more reliable performance than conventional SS even with a next section.
smaller training phase. It was concluded that for low SNR In an early attempt to utilize simple ANN for SS, [53]
values, a higher number of training samples is needed to proposed a joint detection method which combined ED and
accurately predict the status of the PU. When considering a Cyclostationary Feature Detection (CFD) with ANN. The
fading environment, the proposed method with larger training PU signal is assumed to be an amplitude modulated signal
size reports a better detection performance and a lower PoE in AWGN. Four features are extracted to provide input to
than the OR scheme. the ANN, one of which is energy and the other three are
In a recent work, Tian et al. [47] formulated the SS cyclostationary feature values. These features form the input
problem as an SNR-based multi-class classification problem of a neural network having one hidden layer and determines
to adapt to SNR variations. A Naïve Bayes Classifier (NBC) the status of the PU.
is trained for the SS of Orthogonal Frequency Division Vyas et al. [52] proposed a hybrid SS scheme by utilizing
Multiplexing (OFDM) signals. NBC is a supervised ML energy from traditional ED and Zhang test statistics from like-
algorithm which performs classification based on Bayesian lihood ratio test statistic [54] as features for training a simple
decision theorem [48]. The work also proposed a class- ANN with a single hidden layer. The model is evaluated with
reduction assisted prediction method to reduce the time the help of real-world PU signals obtained with an experi-
needed for SS. On comparing the PoD versus SNR curves and mental test setup inspired from [55]. The hardware part of the
the ROC curves, the NBC-based detection outperforms ED, platform consists of a Universal Software Radio Peripheral
Cyclic Prefix (CP)-based detector [49], asymptotic simple (USRP-N210), WBX daughter board, RF-Explorer and
hypothesis test-based detector [50] and a neural network with D3000N Super Discone antenna. GNU Radio and MATLAB
two hidden layers. Furthermore, the performance bounds of form the software part of the measurement setup.
the SS error rates are calculated, and the performance of The combined features and labels from four radio
NBC-based detector is found within the bounds. technologies are utilized for training four different ANN

89598 VOLUME 11, 2023

S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

TABLE 3. Summary of traditional ML-based approaches for spectrum sensing.

architectures and the best ANN model is identified with radio environments. DL enhances model performance by
the help of a cross-validation set. The model is evaluated utilizing the non-linear relationship in training data opti-
for the four radio technologies by considering different sets mally [58]. The multiple hidden layers enable the DNN to
of features: only energy of current sample, only Zhang learn patterns from datasets layer by layer. Low-level data
statistics of current sample, energy and Zhang statistics of features are transformed into high-level abstract features as
current sample, and energy and Zhang statistics of current the output from a lower layer serves as input for a higher
and previous samples. For all radio technologies, using only layer [59].
the current sensing event gives the worst sensing performance Fig. 4 depicts a generalized representation of DNN-based
while using energy and Zhang statistics of current and SS in a non-CSS scenario. With DL, the SS problem is treated
previous sample achieves the best accuracy. The proposed as a binary classification or hypothesis testing problem
technique achieves a better detection performance than the with the two classes representing absence of PU or null
Improved Energy Detection (IED) [56] and Classical Energy hypothesis (H0 ) and presence of PU or alternative hypothesis
Detection (CED) [55] approaches for all radio technologies. (H1 ). Firstly, either real-world data is acquired, or samples
Table 4 summarizes the simple ANN-based research works are synthetically generated to represent spectrum data. This
discussed in this subsection. Even though simple ANN data can be collected in the format of In-phase/Quadrature
methods improve sensing performance, their detection results (I/Q) samples, spectrograms, Covariance Matrices (CMs),
depend directly on the accuracy of the input features obtained etc., or various features like energy and cyclostationary
from the received signals [57]. The extraction of specific features can be derived from this data. Secondly, the acquired
features from the original received signal is limited to spectrum data can be pre-processed by techniques like
obtaining partial information, which inevitably leaves out data standardization, data normalization, filtering, matrix
implicitly hidden but helpful features. As simple ANNs manipulation, etc., to be in an appropriate form and increase
do not have multiple layers, they cannot perfectly capture the detection performance of the DNN. The pre-processed
complex data features and non-linearity. To overcome these data is split into training, validation, and test sets. The model
shortcomings of simple ANN algorithms, DL algorithms is trained by the training set and then tuned by the validation
have been widely adopted for the task of SS. data in an offline process. Next, the well-trained model with
its hyperparameters optimized is used to classify the test data
III. DEEP LEARNING (DL)-BASED SPECTRUM SENSING into either H0 or H1 in an online detection process.
DL is a subset of ML that can automatically capture Fig. 5 represents the generalized use of a DNN for decision
complex patterns and features from input data [21]. As these fusion in a CSS scenario involving multiple SUs. For CSS, the
algorithms are quickly adaptable, they are robust to uncertain DNN is used to combine the sensing results obtained from

VOLUME 11, 2023 89599

 S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

TABLE 4. Summary of simple ANN-based approaches for spectrum sensing.

FIGURE 4. Generalized representation of DNN-based spectrum sensing in a non-CSS scenario.

individual SUs. The dataset formed of the sensing results can A. MULTILAYER PERCEPTRONS (MLPs)
be pre-processed and is then divided into training, validation, An MLP is a feedforward ANN having one or more hidden
and test sets. The DL model acts as a fusion center and layers. When the number of hidden layers is higher than one,
generates prediction about the status of the PU by combining an MLP is considered as a DNN. The number of neurons in
sensing information from multiple SUs. the input layer is determined by the dimensions of dataset.
DL algorithms are data-driven and can efficiently capture The neurons in output layer are equal to the number of output
the non-linearities of input data. With data-driven methods, labels or classes. The number of hidden layers and the count
the model has the advantage of being trained by extracting of neurons in each hidden layer are determined with the aim
inherent patterns from data and not relying on signal and of optimizing the MLP accuracy. The architecture of a simple
noise assumptions, thus ensuring reliable performance when MLP having multiple hidden layers is depicted in Fig. 7.
used in practical with real signals [60]. In this paper, Du et al. [61] proposed an MLP with 3 hidden layers
the DNN architectures used for SS have been classified for the centralized CSS of the PU spectrum by combining
as MLPs, CNNs, LSTM networks, combined CNN-LSTM information geometry with DL. Input to this method called
architectures, and AEs, as shown in Fig. 6. As a result of ‘IG-DNN’ is a dataset consisting of geodesic distances
their simple architecture, MLPs are easy to implement and derived from covariance matrices of sensing signals and
are preferred for the task of CSS for fusion of the sensing noise. The energy values of noise and signal mixed with noise
data from multiple SUs. CNNs are also preferred for fusion of in a range of SNR values are sensed by multiple SUs to form
sensing results in CSS as the sensing outcomes are correlated. the input for the MLP. The experiments were performed using
To efficiently capture the spatial and temporal details of multiple signals and by fixing the PFA it was concluded that
sensing data, CNNs and LSTMs are popularly used with a greater number of SUs and a higher SNR result in better
diverse and complex radio data in non-CSS scenarios. AEs sensing performance. IG-DNN outperformed IG-FCM [62]
being an unsupervised ML approach are suitable when only and MME-K-means [63] algorithms when their performances
a limited amount of labeled data is available. were compared under various simulation settings.
89600 VOLUME 11, 2023
S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

FIGURE 5. Representation of decision fusion using a DNN in CSS scenario.

TABLE 5. Summary of MLP-based approaches for spectrum sensing.

FIGURE 6. Deep Learning-based architectures used for SS. FIGURE 7. An MLP having multiple hidden layers.

In another work [64], an MLP with two hidden layers

is designed after optimizing the number of hidden layers, input to the MLP are four features: the energy values
the neurons in each hidden layer, optimization algorithm, of the current and previous sensing event and the Zhang
activation function and learning rate for SS. The PU data statistics of current and previous sensing event while the
is formed by four different radio technologies captured output is the status of PU channel. A total of four ANN
using an empirical setup similar to that in work [52]. The architectures are utilized, each assigned to a particular radio
signals were filtered, the transient peaks were removed technology. The proposed approach is compared with a neural
and AWGN was added to obtain desired SNR levels. The network without hyperparameter tuning, CED, IED and with

VOLUME 11, 2023 89601

 S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

NBC-based sensing [47]. On comparing the PoD versus SNR of overfitting. MobileNetV2 [75] was proposed for mobile
curves, it is found that the proposed scheme has similar devices and has 19 residual bottleneck layers following the
performance to NBC at low SNRs. CED and IED techniques initial fully convolutional layer of 32 filters. Its architecture
are computationally simple, but the MLP reports a better is inverted residual with bottleneck layers having residual
PoD. On averaging the detection performance of the model connections between them.
over four radio technologies, it reports a 63% performance VGG-16 and VGG-19 [76] are classical CNN architectures
improvement over CED and IED. designed by exploring the impact of architecture depth on
Nasser et al. [65] proposed a hybrid SS scheme wherein model performance. VGG-16 has 16 layers out of which
the test statistics of six detectors were combined to train an 13 layers are convolutional layers and 3 layers are fc layers.
MLP having 2 hidden layers. By considering 16- Quadrature VGG-19 has 19 layers among which 16 are convolutional
Amplitude Modulation (16QAM) signals and AWGN, the layers and 3 layers are dense layers. ResNet-18 and ResNet-
data pertaining to the test statistics of ED, autocorrelation 50 [77] were proposed to facilitate the training of deep
detector [66], maximum eigenvalue detector [67], cumulative networks with the help of a deep residual learning framework.
power spectral density detector [68], maximum–minimum In deeper networks, as the network depth increases, accuracy
eigenvalue detector [67], and goodness-of-fit detector [69] becomes saturated and then rapidly degrades due to a higher
and the SNR values were used as features to train the MLP. training error. In ResNet-18 there are 17 convolutional layers
The performance of the DNN is evaluated with metrics PoD while ResNet-50 has 49 convolutional layers [78]. As the
and False Alarm Rate (FAR). It is observed that increasing network becomes deeper, it suffers from the problem of
the number of detectors used for training the MLP increases vanishing gradients. To address this, in DenseNet-121 [79]
PoD, while decreasing the FAR. More than three detectors all layers are interconnected and the feature maps of all
result in an average PoD of 0.93 and an almost zero FAR. preceding layers form inputs for each layer. DenseNet-121
Table 5 summarizes the research works discussed in this has a total of 120 convolutional layers.
subsection by providing details about the dataset used, key The similarity between images and signal covariance
pre-processing technique used, key features of the MLP matrices makes CNNs widely suitable for SS problems [21].
architecture, key evaluation metrics used and performance Another reason for their popularity is that the operation of
notes on the SS model. a CNN filter or kernel is similar to the filtering operations
at communications receivers [80]. In an early attempt to use
CNN architecture for SS, [81] proposed a CNN-based SS
B. CONVOLUTIONAL NEURAL NETWORKS (CNNs) algorithm having a single convolutional layer to sense the
CNNs are DNNs popularly deployed in the areas of computer presence of PU signal in environments having low SNRs. The
vision and natural language processing. Besides input and energy and cyclostationary features of signals are extracted,
output layers, a simple CNN network consists of convolution, standardized, and provided as input to the CNN model.
pooling and fully connected (fc) layers [70]. Features are Based on [82], it is considered that the PU signals are
extracted automatically from input samples with the help of cyclostationary signals while noise is stationary and hence,
kernels or filters in convolutional layers and are called feature they can be distinguished on the basis of cyclostationary
maps. Pooling layers perform down-sampling to decrease the features. The proposed model is evaluated by using Binary
complexity for further layers in the network and to avoid Phase-Shift Keying (BPSK) signals in an AWGN channel.
overfitting [71]. Fully connected layers are further used to The CNN with standardized input (S-CNN) is compared with
classify the input data with the help of features extracted from CNN model trained without standardizing the input (N-CNN)
previous layers. Fig. 8 shows the architecture of a basic CNN and CFD models by varying the SNR between -20 dB and
model. -5 dB. While S-CNN has a better detection performance
There are some standard CNN architectures most of than N-CNN, both S-CNN and N-CNN outperform the
which have been trained on a large image database called conventional CFD algorithm at all SNRs.
ImageNet [72]. By utilizing these models, a different task can In [60], a CNN-based algorithm called Activity Pattern
be recognized without the need to train from scratch, making Aware SS (APASS) is proposed which learns the PU activity
these models particularly useful in cases where limited pattern to perform SS. The input to the algorithm includes
training data is available [73]. This is the process of TL, and CM of current frame and a matrix formed by stacking CMs
the pre-trained weights of these models can be accessed with from past frames to enable the CNN model to exploit the PU
the help of various DL libraries like Keras and PyTorch. For activity pattern and improve detection accuracy. The model
instance, the GoogLeNet/Inception-V1 architecture proposed architecture is inspired from the standard CNN architecture
in [74] consists of 22 layers having a total of 9 inception called LeNet [83] and consists of a total of seven layers,
blocks. The work defines inception modules as blocks that including 2 convolutional layers and 2 dense layers. Both
facilitate better computation and deeper networks by reducing correlated and uncorrelated signal models are adopted, and
the dimensionality through stacked 1 × 1 convolutions. PU signal vector is considered Gaussian with zero mean. The
By using these modules, the dimensions are reduced thus work analyzes the convergence behavior of loss function at
lowering the computational costs and addressing the issue different SNR levels, and it is observed that at high SNR

89602 VOLUME 11, 2023

S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

FIGURE 8. Architecture of a basic CNN model.

levels the loss function quickly converges to zero due to DenseNet-121, along with CNN and LSTM models proposed
a large difference in CMs of the two hypotheses. APASS in this work is evaluated. Two test sets, i.e., Test A and
detector is compared with optimal Estimator-Correlator (EC) Test B, were defined. Test A had an equal proportion of
detector and Hidden Markov Model (HMM) with the help of spectrograms in categories: SPN-43, Radar 3 Out-Of-Band
PoD versus SNR curves. When considering the uncorrelated Emissions (OOBE), both SPN-43 and Radar 3 OOBE and
signal model, the APASS detector with both deterministic neither. Test B was a subset of Test A without Radar 3 OOBE.
and random PU activity model outperforms the EC and For the proposed LSTM architecture, the 10 MHz channel
HMM models. The detection performance of all algorithms is divided into sequential slices along the time axis. The
improves with the increase in the number of receiving outputs of LSTM cells are passed to dropout cells with a
antennae. On adopting a correlated signal model, similar 50% probability. An fc layer with 50 neurons receives the
trends are observed with an improvement in performance of output of the last cell after all the time slices have been
all three detectors. provided to the LSTM. Next a layer with a single neuron and
Liu et al. [84] designed DNN-based Likelihood Ratio sigmoid activation generates a prediction between 0 and 1.
Test (DNN-LRT) and proposed the CM-aware CNN (CM- The proposed CNN architecture is called CNN-3 and consists
CNN) model for SS. Sample CMs form the input of the of a convolutional layer, a dense layer with 150 neurons and
DNN which is based on LeNet-5 [83] architecture and a layer with a single neuron which generates output between
consists of 2 convolutional layers. The PU signals are 0 and 1. After the convolutional layer, the work uses a novel
generated by simulation using independent and identically averaging step for activation maps which is equivalent to
distributed (i.i.d.) and exponential correlation models. Noise single filter 1∗1 convolutional layer.
samples either follow a Gaussian distribution or are real For narrowband detection, all 13 models are evaluated and
sea clutter samples [85]. The performance of CM-CNN is compared based on their ROC curves. Out of all the models,
compared with EC, ED [86], Maximum Eigenvalue Detector CNN-3 has the best performance for Test A set while it
(MED) [87], blindly combined energy detector [88] and closely follows the performance of the Inception-V1 model
CAV-based detector [89] with the help of ROC and PoD for Test B set. For both sets A and B, the best performing
versus SNR curves. Under Gaussian noise, CM-CNN obtains model amongst standard CNN models is Inception-V1,
a comparable performance with the optimal EC detector traditional ML models is SVM with linear kernel and full
for both i.i.d. and exponential correlation models. When input, conventional models is SI-ED and proposed models
considering sea clutter and exponential correlation model, is CNN-3. The best performing models from the single
CM-CNN has a satisfactory detection performance while EC channel evaluation for each model category are compared
detector cannot be implemented as sea clutter lacks statistical for wideband detection of SPN-43 across multiple channels
model. CM-CNN is robust in conditions of low SNR and observed simultaneously with a single receiver by using
outperforms EC detector when noise uncertainty is 1 dB. Free-response ROC (FROC) curves. For set A, CNN-3
Lees et al. [78] compared the narrowband and wideband outperforms the other 3 models while for set B Inception-V1
detection performance of 13 different models for detection of reports the best area under the curve value followed by SVM
SPN-43 air traffic control radar. With the help of 3.5 GHz and then CNN-3. It is further observed that the CNN-3 has the
band low-resolution spectrograms [90], [91], the perfor- fastest detection time among ML models. CNN-3 is further
mance of conventional algorithms: ED and Sweep-Integrated used for the classification of the complete set of spectrograms
Energy Detection (SI-ED) [92], traditional ML algorithms: following which a spectrum occupancy estimate for SPN-
SVM, KNN and GMM, standard CNN architectures: 43 is provided, and the power of non-SPN-43 emissions is
ResNet-18, ResNet-50, Inception-V1, VGG-16, VGG-19 and characterized.

VOLUME 11, 2023 89603

 S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

By presenting SS as a binary classification problem, generate complex signal frames using eight different mod-
Zheng et al. [24] proposed a CNN architecture which in ulation techniques: 64-QAM, 16-QAM, Continuous Phase
addition to 2 basic convolution layers consists of 6 resid- Frequency Shift-Keying (CPFSK), Gaussian Frequency-Shift
ual blocks in cascade. The model was made robust to Keying (GFSK), BPSK, 8-PSK, QPSK and PAM4 and each
noise power uncertainty by normalizing the received signal of these frames is separately impacted by an independent
power. Signal data for eight modulation techniques is Rayleigh multipath fading channel, clock offset, and AWGN.
simulated, including 32QAM, 16QAM, 8- Pulse Ampli- The SNR range is kept between −20 dB to +25 dB,
tude Modulation (8PAM), 4PAM, Quadrature Phase-Shift in increments of 5 dB. Equal number of AWGN samples
Keying (QPSK), 4-Frequency-Shift Keying (4FSK), 2FSK are generated by changing the noise power in a range
and BPSK. An equal number of AWGN or colored noise of −100 dBm to −5 dBm in steps of 5 dB.
samples are simulated and used for training the DL model. The model achieves an accuracy of 99.74% in differentiat-
An accuracy of 90.55% was achieved on the test data. ing between the signal and noise frames. The performance of
The model outperformed two conventional SS models, Deep-CRNet for over the air signals is assessed by using sig-
frequency domain entropy-based method and maximum- nal frames from RadioML [93], [94], [95], [96], [97] dataset
minimum eigenvalue ratio-based method by reporting a having the modulation techniques of 64-QAM, 16-QAM,
higher PoD. 8-PSK, BPSK and QPSK. The model outperforms other state-
To test the performance of the model with untrained data, of-the-art pre-trained DNN architectures of GoogLeNet [74]
the work also simulated test samples having modulation and MobileNetV2 [75]. Deep-CRNet demonstrates superior
types 8- Phase-Shift Keying (8PSK), 8FSK and 64QAM. detection performance when compared with other benchmark
After fixing the PFA at 0.01 it was observed that the signals traditional and DL-based SS schemes.
were sensed with high probability. To observe the model’s In [22], 2000 FSK and Amplitude-Shift Keying (ASK)
performance with real-world signals it was tested against modulated signals are synthetically generated by using
Aircraft Communications Addressing and Reporting System an Arduino Uno microcontroller board and a 433 MHz
(ACARS) signals. The ACARS samples when used to fine transmitter. The signals are received by an RTL-SDR receiver
tune the model by adopting a TL-based approach outperforms which is connected to MATLAB. The received signals are
the two conventional techniques. The model is robust to transformed into time-frequency representations and are
pink noise in contrast to the conventional approaches whose classified as PU signals or noise by a CNN classifier having
performance degrades with pink noise proving that DL can two convolutional layers. If an SU is positioned closer to
automatically extract noise characteristics from data. PU it can sense the spectrum more reliably therefore the
The ‘Deep Sensing’ CNN introduced in work [80] authors generate their data by varying the distances between
comprises of 2 convolutional and 2 dense layers. The received the sender and receiver. CNN surpasses the performance of
radio signals are filtered to limit noise using a rectangular ED, ANN, and SVM models.
band limited filter and sampled in MATLAB to produce CNNs are also suitable for CSS as just like the adjacent
a discrete time sequence. The work detects narrowband pixels in an image are correlated, the sensing outcomes from
Gaussian-distributed signal in AWGN to compare the perfor- nearby SUs and adjacent bands have spectral and spatial
mance of Deep Sensing with an optimal sensing algorithm correlations [98]. Lee et al. [98] used a CNN structure having
whose analytical expression is available in accordance with three convolutional and two dense layers, ‘Deep Cooperative
log-likelihood ratio. Deep Sensing outperforms ED and Sensing’ (DCS) in a CSS scenario to combine the sensing
has a performance close to the optimal sensing algorithm decisions from multiple mobile SUs. It was assumed that
when compared by using PoD and PFA. The authors the PU can simultaneously occupy multiple bands and that
further examined the robustness of Deep Sensing using SUs do not transmit when sensing is performed. Each
narrowband Gaussian signals with zero mean in AWGN SU senses the spectrum using ED and DCS combines
and QPSK signals. Since Deep Sensing was not effective the decisions regardless of if they are Hard Decisions
on communication scenarios that differed from the training (HDs) or Soft Decisions (SDs). DCS is evaluated with a
set, experiments with TL without labels and fine tuning parameter ‘sensing error’ which is determined by averaging
were conducted. When a limited amount of labelled data is the probabilities of missed detection and false alarm. The
available, the model has proven to be robust across various performance of DCS is compared with the conventional
domains. sensing methods: K-out-of-N and SVM with linear kernel
Ahmed et al. [45] proposed a CNN-based approach and it is found that DCS with SD shows the lowest sensing
called ‘Deep-CRNet’ which consists of 85 layers in total error followed by DCS with HD. DCS is robust in the
and has 5 convolution blocks coupled with 2 intermediate conditions of high noise power densities and lower count of
residual- inception blocks. In a communication network training samples or SUs but takes the highest computation
comprising of Internet of Things (IoT) and Unmanned time.
Aerial Vehicles (UAVs), the model performs opportunistic In another example of CNN-based CSS, [32] used a CNN
spectrum access-based SS for SUs. Complex waveforms of with one convolutional layer for performing data fusion with
PU and noise signals are generated artificially. The authors five SUs and a mobile PU. BPSK modulated random bits with

89604 VOLUME 11, 2023

S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

FIGURE 9. Architecture of a simple LSTM model with an LSTM cell [102].

Rayleigh and Nakagami-m fading are transmitted and raw I/Q and in a CR network, this statistical information will be
samples for SUs are generated and split into training and test helpful in predicting spectrum occupancy trends, planning
sets. The performance of the proposed model is analyzed with SS, selecting the right spectrum band and channel for CR
the help of classification accuracy and is compared with CSS system, maximizing system performance, and improving
with ED-based HD fusion rules: AND majority rules and OR. spectral efficiency [29]. PAS-SS consists of an LSTM model
CNN-based CSS scheme outperforms ED-based AND and with 3 hidden layers for prediction and an ANN with a
majority rules. It is observed that for all models, performance single hidden layer for classification. For acquiring data for
under Rayleigh fading is superior to Nakagami-m fading. experiments with LSTM-SS and PAS-SS, two empirical bed
Table 6 provides a comprehensive summary of the CNN- setups are used with a USRP and a digital spectrum analyzer
based SS algorithms surveyed in this work. respectively. When LSTM-SS is compared with CNN and
ANN with the help of PoD versus SNR curves, LSTM-
SS achieves the best detection performance. On comparing
C. LONG SHORT-TERM MEMORY (LSTM) NETWORKS
the classification accuracy versus SNR curves, training and
Recurrent Neural Networks (RNNs) are a class of ANNs used execution times of LSTM-SS with the ML techniques ANN,
with time series data in which output of a layer is provided as Gaussian Naïve Bayes, and Random Forest it is observed
feedback to the input to determine the output of that layer. that LSTM-SS achieves the best classification accuracy but
LSTMs are a type of RNNs which can capture long term reports longer training and execution times.
time dependencies and can exploit correlation from the time- A summary of research works utilizing LSTM networks
series spectrum data. Fig. 9 depicts the architecture of a for SS has been provided in Table 7.
simple LSTM model and an LSTM cell. The LSTM algorithm
proposed in [78] was summarized in the previous subsection D. COMBINED CNN-LSTM ARCHITECTURES
on CNNs. A CNN can extract spatial features from the input data while
In another work, Soni et al. [99] analyzed the temporal an LSTM network captures temporal variations. The CNN
correlation within the spectrum data captured through an and LSTM techniques can be combined to extract complex
empirical setup with the help of an LSTM network. To make features from data. A simple CNN-LSTM model is shown
the LSTM unbiased, data at very low SNR values is in Fig. 10. In the presence of noise uncertainty, when the
included. An LSTM with a single hidden unit has the best SNR value falls below a threshold called the SNR-wall, an SS
validation accuracy. LSTM-based SS (LSTM-SS) which can detector will fail to perform. The problem of SNR-wall can
capture the temporal correlation from the input data and PU be solved by utilizing the structure of the PU signal, adding
Activity Statistics-based SS (PAS-SS) models are introduced diversity, and reducing the noise uncertainty [100].
in this work. PU activity statistics and occupancy patterns As an alternative to conventional energy detectors that
can be estimated using the sequence of sensing decisions suffer from the SNR-wall problem, work [101] uses DL

VOLUME 11, 2023 89605

 S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

TABLE 6. Summary of CNN-based approaches for spectrum sensing.

TABLE 7. Summary of LSTM-based approaches for spectrum sensing.

to extract the hidden structures of the PU signals and RadioML2016.10a dataset [93], [94], [95], [96], [97] in eight
propose ‘DetectNet’ which consists of convolutional, LSTM, modulation types: 8PSK, BPSK, QPSK, GFSK, CPFSK,
and fc layers. The signal samples are generated using the QAM64, QAM16 and PAM4. The negative samples are

89606 VOLUME 11, 2023

S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

FIGURE 10. Architecture of a simple CNN-LSTM model.

additive noises having zero mean Circularly Symmetric Com- sensing data. The PU signals are QPSK modulated signals
plex Gaussian (CSCG) distribution. Energy normalization is having unit energy while the noise signals are simulated by
performed on the training samples as the simulation results following Gaussian and Laplace distributions. For the PU
are minimally impacted by energy. Furthermore, the signal activity pattern, the work considers lognormal state sojourn
modulation structure can be better exploited, and detector time model where state transitions are represented by a
model can have better generalization ability. The paper uses semi-Markov process [104], [105] and the real state sojourn
a customized two-stage training strategy based on constant time model for which real sensing data is collected using
false alarm rate detector to control the performance of the DL the USRP-2922. The CNN-LSTM detector outperforms the
detector. The first stage of the model involves early stopping detectors MED [87], Signal Subspace Eigenvalues (SSE)
at 6 epochs while in the second step the model observes detector [106], Arithmetic to Geometric Mean (AGM)
metrics trade-off characteristic where the validation loss and detector [106] and the DL-based APASS detector [60] in
accuracy are both kept stable while PFA and PoD at different scenarios with and without noise uncertainty.
SNRs vary with each epoch. In [20] ‘DLSenseNet’ which is based on combined
The authors further propose and optimize other DNN CNN-LSTM architecture is used to capture both spatial
models which are an MLP with four fc layers, a CNN with and temporal details of data and for sensing the spec-
two convolutional and one fc layer and an LSTM with two trum. DLSenseNet is made of a modified inception block,
LSTM layers. The detection performance of these models LSTM layers and fc layers. Samples of eight different
is compared with that of DetectNet by considering QAM16 types of digital modulation from the RadioML2016.10b
modulated signals. It is found that DetectNet and CNN have dataset [93], [94], [95], [96], [97] are used to represent PU
superior performance than MLP and LSTM and DetectNet signals and the absence of PU is represented by CSCG
outperforms CNN by achieving a low value of PFA for similar noise vector with zero mean. I/Q components of these
PoD. For a fixed sample length, DetectNet gives the best signals form input to the model which then predicts the
performance for FSK modulated signals. While evaluating status of the PU. The signals are energy normalized to
the generalization ability of DetectNet it is found that make DLSenseNet independent of energy and have greater
when the training and testing data have similar modulation generalization capacity in environments where background
schemes, the model has good generalization power, and the noise changes. The work analyzed the effect of various
performance deteriorates when the training and test sets have modulation schemes and length of data samples and com-
different modulation types. For a higher sample length of pared the performance of DLSenseNet with other DNN
data, the performance of DetectNet improves. The work models: CNN, residual network (ResNet), LeNet, inception
further proposes ‘SoftCombinationNet’ for a CSS scenario module, LSTM, Convolutional Long short-term Deep Neural
which provides a global decision by combining the soft Network (CLDNN) and previously reported SS techniques of
information from sensing nodes. DetectNet is deployed at DetectNet [101] and CNN-LSTM [103]. DLSenseNet out-
each sensing node and the probability vectors from these performs other models because it integrates the advantages
nodes are processed by a neural network having three fc of CNN, LSTM, and inception models. The performance of
layers to make the final decision about the status of the PU. DLSenseNet is compared by using eight different modulation
Xie et al. [103] proposed a CNN-LSTM-based model schemes and there is very little difference in detection
which first extracts energy correlation features from covari- performance between various modulated signals, suggesting
ance matrices generated by sensing data with the help of that DLSenseNet is insensitive to the order of modulation.
CNN layers and then inputs the series of energy correlation Xing et al. [21] proposed a BiLSTM-based DNN by
features corresponding to multiple sensing periods into an combining convolutional, concatenated, BiLSTM, Self-
LSTM network to learn PU activity patterns. This DNN Attention (SA) layers and fc layers. The model utilizes the
architecture consists of two convolutional layers followed by hidden states by simultaneously scanning data in opposite
an LSTM layer and a dense layer. The work mentions that directions. The PU signals in eight modulation schemes:
DL-based SS architectures are not susceptible to the signal- BPSK, QPSK, 8PSK, CPFSK, GFSK, QAM16, QAM64
noise model assumptions since they learn directly from the and PAM4 are generated using GNU radio while complex

VOLUME 11, 2023 89607

 S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

AWGN signals are used as noise signals. In an ablation

study, the impact of BiLSTM, SA, and concatenated layers is
observed on sensing error by comparing five models: CNN,
CNN-LSTM, CNN-BiLSTM, CNN-BiLSTM-SA, and CNN-
BiLSTM-SA-CONCAT by using QAM 16 signals. The CNN
model comprises of two convolutional blocks each having
a one-dimensional convolutional layer which helps extract
sufficient local features, especially in environments with low
SNR [80]. The features lost during transmission by the one-
dimensional CNN are compensated by the concatenated layer
which is denoted by CONCAT and combines the input and
output from the CNN. BiLSTM layers help capture the long-
and short-term time dependencies from the input data in
opposite directions and with the help of the SA layer, the
model highlights the most important features obtained by the
BiLSTM layers. It is found that for SNR values of less than
−5 dB, there is a significant difference in the performance of
the five models. CNN-BiLSTM-SA-CONCAT achieves the
best sensing performance followed by CNN-BiLSTM-SA, FIGURE 11. Structure of a conventional autoencoder.
CNN-BiLSTM, CNN-LSTM and CNN. The models achieve
comparable sensing results in an SNR range above −5 dB.
The CNN-BiLSTM-SA-CONCAT model is compared the original data [107]. An AE has three layers: input, hidden,
against the CNN architecture used in [80], the ResNet model and output or reconstruction layers. Each AE undergoes an
in [24], the LSTM network in [99] and the CLDNN proposed encoding-decoding process during training. The encoding
in [101] with their default hyperparameter settings using process maps the input data into a hidden representation,
QAM 16 signals. It is observed that the models using LSTM and the decoding process reconstructs input data from the
take more time for both training and detection as unlike hidden representation [108]. Fig. 11 shows the structure of
CNNs, LSTM operations cannot be parallelized by a GPU. a conventional AE. AE architectures having two or more
The CNN-BiLSTM-SA-CONCAT model takes the longest hidden layers in both encoder and decoder have been utilized
training and detection times but reports the least sensing error for SS.
under all SNR conditions. The authors further observed the Cheng et al. [57] proposed two novel Stacked Autoencoder
results of the model under eight modulation schemes and (SAE) frameworks having two hidden layers for sensing
different SNR values by keeping a fixed sample length of 128. of OFDM signals. To create the trained SAE model, the
The model has comparable performance for all modulation input and hidden layers of all the trained AEs are stacked
schemes, but the GFSK modulated signals have the lowest together layer by layer. The first framework is termed
sensing error in the SNR range of −12 dB to −5 dB. Stacked Autoencoder-based SS (SAE-SS). SAE-SS is first
The robustness of CNN-BiLSTM-SA-CONCAT is studied pre-trained wherein the features of the PU are extracted
by training it with QAM16 modulated signals and testing in an unsupervised manner following which a Logistic
it individually with signals having a different modulation Regression classifier is used to fine tune the model. As the
scheme (BPSK, QAM64, PAM4 and GFSK). The classifier sensing accuracy of SAE-SS degrades for low SNR values,
achieves good detection results with minimal deterioration another model, Stacked Autoencoder-based SS using Time-
in performance. When the model is trained individually with Frequency domain signals (SAE-TF) is proposed. With SAE-
QAM16 signals having sample lengths of 64, 128, 256 and TF, both time and frequency domain signals are used as input
512, it is established that the sample length of 512 reports the which allows the model to extract more hidden features and
least sensing error as longer signal sample consists of more perform better in conditions of low SNR. The received signals
temporal information. in time domain are transformed to frequency domain with
Table 8 provides a summary of CNN-LSTM-based the help of Fast Fourier Transform and then manipulated
approaches discussed for SS by describing the input to the to provide input to the SAE-TF model along with the time
CNN-LSTM algorithm, the data pre-processing technique domain input.
used, key features of the CNN-LSTM architecture, main SAE-TF has a better sensing performance than SAE-SS
evaluation metrics and performance of the SS model. but has twice as many input units as SAE-SS resulting in a
higher training complexity. The OFDM system is generated
with BPSK modulation. For training of SAE-SS and SAE-
E. AUTOENCODERS (AEs) TF models, the training data is divided according to the
An AE neural network is an unsupervised learning algorithm SNR values for training different SAE architectures for a
that reduces the dimensionality of input data and reconstructs higher detection performance. The performance of SAE-SS

89608 VOLUME 11, 2023

S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

TABLE 8. Summary of CNN-LSTM-based approaches for spectrum sensing.

and SAE-TF models are compared with conventional OFDM LSTM AE consists of three hidden layers both in encoder and
signal sensing techniques: ED, CP [109] and CM [110] -based decoder. USRP B210 and GNU Radio were used to capture
SS and with neural network-based techniques: ANN model the LTE signals while MATLAB was used to generate Wi-
in [52] and CNN model in [81] under various conditions. Fi signals. The signal strength of LTE signals was matched to
When comparing the Probability of Miss detection (PM) that of Wi-Fi signals by addition of 20 dB. Four features were
while varying the SNR values, it is observed that the provided as inputs for all AEs: I/Q samples and the phase
proposed SAE models have the least PM values, with SAE- and amplitude values derived from the I/Q samples. The AEs
TF outperforming SAE-SS. were evaluated by taking different combinations of signals
In work [111], a Variational Autoencoder (VAE) and from 802.11ax and 802.11ac Wi-Fi protocols. With the main
unsupervised DL-based detector termed Unsupervised Deep evaluation metrics being precision and recall, the Deep AE
SS (UDSS) is designed to limit the amount of labelled using exponential linear unit activation was found to be most
data required for training the SS model. The VAE proposed efficient for the classification task.
in [112] has a probabilistic model parameter layer after The AE-based SS algorithms discussed in this subsection
the hidden layer as opposed to a conventional AE [113]. are summarized in Table 9.
A VAE- GMM-based approach having three hidden layers
each in encoder and decoder is used to separate data into IV. RADIO FREQUENCY (RF) SIGNAL DATASETS,
two clusters and a small amount of labeled noise data is TRANSFER LEARNING (TL) AND SOFTWARE USED
then used to identify the clusters representing PU signals The application of ML algorithms, especially DL algo-
and noise. PU signals are represented by unit energy QPSK rithms is becoming increasingly popular in the field of
modulated symbols while noise samples are either Gaussian SS. The training, validation, and testing of these models
or Laplacian. CMs are calculated and vectorized before being requires huge amounts of data. There are a few pub-
input to the neural network. The sensing performance of licly available RF datasets having a collection of var-
UDSS is compared to MED, SSE detector, AGM detector, ious modulations with ‘RadioML’ released by DeepSig
CNN [84] and Kernel K-Means [114] with the help of ROC Inc. [93], [94], [95], [96], [97] being the most popular.
curves. Under all simulation settings, the performance of RadioML datasets have been widely used by researchers
UDSS is close to that of CNN even with limited labelled data for addressing modulation classification and SS prob-
and it outperforms all other detectors. lems [20], [45], [101] with the help of ML. The latest
Subray et al. [107] classified LTE and Wi-Fi signals for SS version is ‘RadioML 2018.01A’ which comprises of synthetic
with three types of AE neural networks: Deep, Variational simulated channel effects and over-the-air recordings of
and LSTM AEs. Both encoders and decoders of the deep AE 24 digital and analog modulation techniques. This data is
and VAE consist of two hidden layers. In LSTM AE, the available in hdf5 format as complex floating-point values,
input data is encoded and decoded using LSTM cells. The with 2 million examples, each having a sample length

VOLUME 11, 2023 89609

 S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

TABLE 9. Summary of AE-based approaches for spectrum sensing.

TABLE 10. Description of some publicly available RF datasets. In addition to these published RF datasets, many works
have generated data specific to their experiments using
MATLAB and Python. Some works have captured real-world
data for their experiments with the help of a USRP and GNU
Radio Software. However, the wireless communications and
RF signals domain lacks robust and comprehensive datasets
comparable to those in domains like speech, handwriting, and
object recognition [40]. The use of TL can be an effective
strategy when a limited amount of training data is available.
ImageNet [72] is a large database of images that has been
used to train many standard CNN architectures such as
VGG-16 [76] and ResNet-50 [77]. With the help of TL,
the knowledge gained by these standard models to classify
the ImageNet dataset can be transferred to RF datasets
of 1024. A previous version of this dataset, ‘RadioML for SS. Works [45] and [78] have explored some standard
2016.10A’ consists of 8 digital modulated and 3 analog CNN models and have evaluated their performances for SS.
modulated signals in varying SNRs and is available as a Table 11 provides a description of the architectures of the
pickle file. standard CNN models used in these works.
Apart from RadioML datasets, there are some other RF
datasets that can be explored by researchers for testing the V. RESEARCH CHALLENGES RELATED TO THE USE OF
robustness of their SS models. For example, an RF dataset DL FOR SS
by Panoradio SDR [115] was generated synthetically by A. REQUIREMENT OF LARGE AMOUNTS OF DATA AND
the application of Gaussian noise, Watterson fading and LACK OF COMPREHENSIVE RF DATASETS
with random frequency and phase offsets to speech, music, Although DL-based approaches are becoming increasingly
and text signals. It is designed for signal and modulation popular in SS, the lack of comprehensive RF datasets poses
classification tasks using novel ML algorithms and has a major challenge to the deployment of DL algorithms [40].
signals from 18 different transmission modes. It consists of DL techniques are data-hungry and require huge amounts of
172,800 signal vectors with each vector having 2048 I/Q data for training, validating, and testing their models. The
samples. Another dataset, MIGOU-MOD [116], [117] is concept of TL can be explored to utilize the model weights
a dataset acquired from a low-power IoT platform called of pre-trained DL architectures to enhance the sensing
‘MIGOU’ and has over-the-air measurements of real radio performance. In addition, data augmentation could be used
signals modulated with 11 modulation types. The signals to increase the amount and quality of data and add diversity
were generated with the help of a USRP and GNU Radio to the dataset [118]. By augmenting the dataset with relevant
software and recorded using the MIGOU platform in an data, the trained model can be made more robust, improving
office environment. The main properties of these datasets are its overall performance significantly [119]. Furthermore, the
summarized in Table 10. use of unsupervised DNN architectures which require only

89610 VOLUME 11, 2023

S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

TABLE 11. Description of the architectures of standard CNN models. different SAE architectures with data in different SNR
conditions. Along with it, the SAE-TF model used both
time and frequency domain samples as input to ensure
robust performance in regions of low SNRs. The focus of
future DL-based SS research should be to ensure robust
performance at low SNRs as a model that performs better
at low SNRs also performs well in conditions of high
SNR [20]. It is also important for researchers to explore the
interpretability of radio signal data for achieving a higher
detection performance.

VI. CONCLUSION
As a principal enabler of DSM, CR technology helps
SUs access the unutilized bands of the PUs and improves
spectrum utilization. The SS capability of CR determines
the availability of radio resources of PUs in order for the
SUs to utilize the vacant frequency bands. The use of
DL algorithms for SS can enhance the performance of 6G
networks. This paper surveyed various DNNs used for the
task of SS by classifying them as MLPs, CNNs, LSTM
networks, combined CNN-LSTM architectures, and AEs.
a limited amount of labelled data like in work [111] is a
The DL algorithms are compared based on the dataset
potential solution for this research problem.
used, data acquisition technique, data pre-processing method,
algorithm architecture, evaluation metrics, obtained results
B. LACK OF VALIDATION OF RESULTS ON
and comparison of results with standard SS detectors. DNNs
REAL-WORLD DATA
are increasingly being used to sense the PU spectrum due
Just as in the case of modulation classification, most of
to their automated feature extraction capabilities and their
the literature on ML-based SS is based on simulations
ability to adapt to changing radio environments. This work
or theoretical results, so it is unclear how well a model
also highlighted the shortcomings of conventional SS and
performs in real-life settings [120]. To add to it, the
traditional ML approaches while presenting an overview of
algorithms are often trained for specific modulation and noise
traditional ML algorithms and simple ANNs. Traditional
considerations and their performance cannot be generalized.
ML techniques involve manual feature extraction, while
To ensure the robustness of DL algorithms in practical
conventional sensing techniques suffer from missed detection
environments, real-world datasets should be used to validate
of PUs and generate false alarms. In addition, this paper
their performance [23].
summarized some publicly available RF signal datasets, the
concept of TL and the need to have diverse RF signal datasets.
C. HIGHER OFFLINE TRAINING TIME
Finally, the research challenges associated with the use of DL
Even though DNN techniques result in enhanced detec- techniques to identify vacant frequency bands were discussed
tion performance, these models have high computational along with potential solutions.
complexity. DNN-based models require huge amounts of
data for training which results in higher offline training
times. Consequently, more research should be conducted on REFERENCES
enhancing detection performance through limited training [1] S. N. Syed, P. I. Lazaridis, F. A. Khan, Q. Z. Ahmed, M. Hafeez,
V. Holmes, I. P. Chochliouros, and Z. D. Zaharis, ‘‘Deep learning
data to decrease training time. This issue can also be mitigated approaches for spectrum sensing in cognitive radio networks,’’ in Proc.
by the use of specialized hardware with high processing 25th Int. Symp. Wireless Pers. Multimedia Commun. (WPMC), Oct. 2022,
capabilities like GPUs which can significantly cut down the pp. 480–485.
training time [57]. Moreover, the training is usually required [2] G. K. Papageorgiou, K. Voulgaris, K. Ntougias, D. K. Ntaikos,
M. M. Butt, C. Galiotto, N. Marchetti, V. Frascolla, H. Annouar,
only occasionally when the sensing conditions change, and A. Gomes, A. J. Morgado, M. Pesavento, T. Ratnarajah, K. Gopala,
the well-trained model can then generate fast predictions. F. Kaltenberger, D. T. M. Slock, F. A. Khan, and C. B. Papadias,
‘‘Advanced dynamic spectrum 5G mobile networks employing licensed
shared access,’’ IEEE Commun. Mag., vol. 58, no. 7, pp. 21–27, Jul. 2020.
D. POOR PERFORMANCE AT LOW SNR
[3] A. Ahmed, Z. Elsaraf, F. A. Khan, and Q. Z. Ahmed, ‘‘Cooperative
DNN-based SS algorithms generally perform well in con- non-orthogonal multiple access for beyond 5G networks,’’ IEEE Open
ditions of high SNR, but their performance drastically falls J. Commun. Soc., vol. 2, pp. 990–999, 2021.
at low SNRs. The SNR for SS is usually low, and the [4] F. C. Okogbaa, Q. Z. Ahmed, F. A. Khan, W. B. Abbas, F. Che,
S. A. R. Zaidi, and T. Alade, ‘‘Design and application of intelligent
received signal energy and noise levels at the SU fluctuate reflecting surface (IRS) for beyond 5G wireless networks: A review,’’
over time [111]. Work [57] addressed this issue by training Sensors, vol. 22, no. 7, p. 2436, Mar. 2022.

VOLUME 11, 2023 89611

 S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

[5] Ericsson. Mobile Subscriptions Outlook. Accessed: Feb. 28, 2023. [27] H. A. Shah and I. Koo, ‘‘Reliable machine learning based spectrum
[Online]. Available: https://www.ericsson.com/en/reports-and- sensing in cognitive radio networks,’’ Wireless Commun. Mobile Comput.,
papers/mobility-report/dataforecasts/mobile-subscriptions-outlook vol. 2018, pp. 1–17, Sep. 2018.
[6] C. D. Alwis, A. Kalla, Q.-V. Pham, P. Kumar, K. Dev, W.-J. Hwang, [28] J. Ma, G. Zhao, and Y. Li, ‘‘Soft combination and detection for
and M. Liyanage, ‘‘Survey on 6G frontiers: Trends, applications, cooperative spectrum sensing in cognitive radio networks,’’ IEEE Trans.
requirements, technologies and future research,’’ IEEE Open J. Commun. Wireless Commun., vol. 7, no. 11, pp. 4502–4507, Nov. 2008.
Soc., vol. 2, pp. 836–886, 2021. [29] S. Kyperountas, N. S. Correal, and Q. Shi, ‘‘A comparison of
[7] Y.-C. Liang, J. Tan, H. Jia, J. Zhang, and L. Zhao, ‘‘Realizing intelligent fusion rules for cooperative spectrum sensing in fading channels,’’
spectrum management for integrated satellite and terrestrial networks,’’ EMS Res., Motorola, Tech. Rep., 2010. [Online]. Available:
J. Commun. Inf. Netw., vol. 6, no. 1, pp. 32–43, Mar. 2021. https://api.semanticscholar.org/CorpusID:16716527
[8] K. A. Yau, J. Qadir, C. Wu, M. A. Imran, and M. H. Ling, ‘‘Cognition- [30] I. F. Akyildiz, B. F. Lo, and R. Balakrishnan, ‘‘Cooperative spectrum
inspired 5G cellular networks: A review and the road ahead,’’ IEEE sensing in cognitive radio networks: A survey,’’ Phys. Commun., vol. 4,
Access, vol. 6, pp. 35072–35090, 2018. no. 1, pp. 40–62, Mar. 2011.
[9] S. T. Muntaha, P. I. Lazaridis, M. Hafeez, Q. Z. Ahmed, F. A. Khan, and [31] R. Fan and H. Jiang, ‘‘Optimal multi-channel cooperative sensing in
Z. D. Zaharis, ‘‘Blockchain for dynamic spectrum access and network cognitive radio networks,’’ IEEE Trans. Wireless Commun., vol. 9, no. 3,
slicing: A review,’’ IEEE Access, vol. 11, pp. 17922–17944, 2023. pp. 1128–1138, Mar. 2010.
[10] Q. Z. Ahmed, M. Hafeez, F. A. Khan, and P. Lazaridis, ‘‘Towards beyond [32] S. P, K. R. Sudhindra, and S. M. N, ‘‘Deep learning for cooperative
5G future wireless networks with focus towards indoor localization,’’ spectrum sensing,’’ in Proc. 2nd PhD Colloq. Ethically Driven Innov.
in Proc. IEEE 8th Int. Conf. Commun. Netw. (ComNet), Hammamet, Technol. Soc. (PhD EDITS), Nov. 2020, pp. 1–2.
Tunisia, Oct. 2020, pp. 1–5. [33] T. Yucek and H. Arslan, ‘‘A survey of spectrum sensing algorithms for
[11] A. Ivanov, K. Tonchev, V. Poulkov, and A. Manolova, ‘‘Probabilistic spec- cognitive radio applications,’’ IEEE Commun. Surveys Tuts., vol. 11,
trum sensing based on feature detection for 6G cognitive radio: A survey,’’ no. 1, pp. 116–130, 1st Quart., 2009.
IEEE Access, vol. 9, pp. 116994–117026, 2021. [34] Y. Zeng, Y.-C. Liang, A. T. Hoang, and R. Zhang, ‘‘A review on spectrum
[12] Y.-C. Liang, Dynamic Spectrum Management. Singapore: Springer, 2019, sensing for cognitive radio: Challenges and solutions,’’ EURASIP J. Adv.
doi: 10.1007/978-981-15-0776-2. Signal Process., vol. 2010, no. 1, pp. 1–16, Dec. 2010.
[13] L. Li, F. Khan, M. Pesavento, T. Ratnarajah, and S. Prakriya, ‘‘Sequential [35] E. Axell, G. Leus, E. G. Larsson, and H. V. Poor, ‘‘Spectrum sensing
search based power allocation and beamforming design in overlay for cognitive radio: State-of-the-art and recent advances,’’ IEEE Signal
cognitive radio networks,’’ Signal Process., vol. 97, pp. 221–231, Process. Mag., vol. 29, no. 3, pp. 101–116, May 2012.
Apr. 2014. [36] S. K. Sharma, T. E. Bogale, S. Chatzinotas, B. Ottersten, L. B. Le,
[14] G. Ding, J. Wang, Q. Wu, Y.-d. Yao, R. Li, H. Zhang, and Y. Zou, ‘‘On the and X. Wang, ‘‘Cognitive radio techniques under practical imper-
limits of predictability in real-world radio spectrum state dynamics: From fections: A survey,’’ IEEE Commun. Surveys Tuts., vol. 17, no. 4,
entropy theory to 5G spectrum sharing,’’ IEEE Commun. Mag., vol. 53, pp. 1858–1884, 4th Quart., 2015.
no. 7, pp. 178–183, Jul. 2015. [37] A. Ali and W. Hamouda, ‘‘Advances on spectrum sensing for cognitive
[15] J. Mitola and G. Q. Maguire, ‘‘Cognitive radio: Making software radio networks: Theory and applications,’’ IEEE Commun. Surveys Tuts.,
radios more personal,’’ IEEE Pers. Commun., vol. 6, no. 4, pp. 13–18, vol. 19, no. 2, pp. 1277–1304, 2nd Quart., 2017.
Aug. 1999. [38] Y. Arjoune and N. Kaabouch, ‘‘A comprehensive survey on spectrum
[16] S. Haykin, ‘‘Cognitive radio: Brain-empowered wireless communica- sensing in cognitive radio networks: Recent advances, new challenges,
tions,’’ IEEE J. Sel. Areas Commun., vol. 23, no. 2, pp. 201–220, and future research directions,’’ Sensors, vol. 19, no. 1, p. 126, Jan. 2019.
Feb. 2005. [39] K. M. Thilina, K. W. Choi, N. Saquib, and E. Hossain, ‘‘Machine
[17] M. S. Gupta and K. Kumar, ‘‘Progression on spectrum sensing for learning techniques for cooperative spectrum sensing in cognitive radio
cognitive radio networks: A survey, classification, challenges and future networks,’’ IEEE J. Sel. Areas Commun., vol. 31, no. 11, pp. 2209–2221,
research issues,’’ J. Netw. Comput. Appl., vol. 143, pp. 47–76, Oct. 2019. Nov. 2013.
[18] Z. Zhang, W. Zhang, S. Zeadally, Y. Wang, and Y. Liu, ‘‘Cognitive radio [40] T. A. Hall, R. Caromi, M. Souryal, and A. Wunderlich, ‘‘Reference
spectrum sensing framework based on multi-agent arc hitecture for 5G datasets for training and evaluating RF signal detection and classification
networks,’’ IEEE Wireless Commun., vol. 22, no. 6, pp. 34–39, Dec. 2015. models,’’ in Proc. IEEE Globecom Workshops (GC Wkshps), Waikoloa,
[19] L. Claudino and T. Abrão, ‘‘Spectrum sensing methods for cognitive HI, USA, Dec. 2019, pp. 1–5.
radio networks: A review,’’ in Wireless Pers. Commun., vol. 95, no. 4, [41] J. Kaur, M. A. Khan, M. Iftikhar, M. Imran, and Q. E. U. Haq,
pp. 5003–5037, Apr. 2017. ‘‘Machine learning techniques for 5G and beyond,’’ IEEE Access, vol. 9,
[20] S. Solanki, V. Dehalwar, and J. Choudhary, ‘‘Deep learning for spectrum pp. 23472–23488, 2021.
sensing in cognitive radio,’’ Symmetry, vol. 13, no. 1, p. 147, Jan. 2021. [42] K. W. Choi, E. Hossain, and D. I. Kim, ‘‘Cooperative spectrum sensing
[21] H. Xing, H. Qin, S. Luo, P. Dai, L. Xu, and X. Cheng, ‘‘Spectrum under a random geometric primary user network model,’’ IEEE Trans.
sensing in cognitive radio: A deep learning based model,’’ Trans. Emerg. Wireless Commun., vol. 10, no. 6, pp. 1932–1944, Jun. 2011.
Telecommun. Technol., vol. 33, no. 1, p. e4388, Jan. 2022. [43] T. Li, S. Zhu, and M. Ogihara, ‘‘Using discriminant analysis for multi-
[22] M. Saber, A. Chehri, A. E. Rharras, R. Saadane, and M. Wahbi, ‘‘A cog- class classification: An experimental investigation,’’ Knowl. Inf. Syst.,
nitive radio spectrum sensing implementation based on deep learning vol. 10, no. 4, pp. 453–472, Mar. 2006.
and real signals,’’ in Innovations in Smart Cities Applications, vol. 4. [44] Y. Lu, P. Zhu, D. Wang, and M. Fattouche, ‘‘Machine learning techniques
Cham, Switzerland: Springer, 2021, pp. 930–941, doi: 10.1007/978-3- with probability vector for cooperative spectrum sensing in cognitive
030-66840-2_70. radio networks,’’ in Proc. IEEE Wireless Commun. Netw. Conf., Doha,
[23] D. Janu, K. Singh, and S. Kumar, ‘‘Machine learning for cooperative Qatar, Apr. 2016, pp. 1–6.
spectrum sensing and sharing: A survey,’’ Trans. Emerg. Telecommun. [45] R. Ahmed, Y. Chen, and B. Hassan, ‘‘Deep learning-driven opportunistic
Technol., vol. 33, no. 1, p. e4352, Sep. 2021. spectrum access (OSA) framework for cognitive 5G and beyond 5G
[24] S. Zheng, S. Chen, P. Qi, H. Zhou, and X. Yang, ‘‘Spectrum sensing based (B5G) networks,’’ Ad Hoc Netw., vol. 123, Dec. 2021, Art. no. 102632.
on deep learning classification for cognitive radios,’’ China Commun., [46] T. F. Smith and M. S. Waterman, ‘‘Comparison of biosequences,’’ Adv.
vol. 17, no. 2, pp. 138–148, Feb. 2020. Appl. Math., vol. 2, no. 4, pp. 482–489, Dec. 1981.
[25] V. Sharma and S. Joshi, ‘‘A literature review on spectrum sensing in [47] J. Tian, P. Cheng, Z. Chen, M. Li, H. Hu, Y. Li, and B. Vucetic,
cognitive radio applications,’’ in Proc. 2nd Int. Conf. Intell. Comput. ‘‘A machine learning-enabled spectrum sensing method for OFDM
Control Syst. (ICICCS), Madurai, India, Jun. 2018, pp. 883–893. systems,’’ IEEE Trans. Veh. Technol., vol. 68, no. 11, pp. 11374–11378,
[26] M. Usman and K. Insoo, ‘‘Secure cooperative spectrum sensing Nov. 2019.
via a novel user-classification scheme in cognitive radios for future [48] T. Kim and J.-S. Lee, ‘‘Exponential loss minimization for
communication technologies,’’ Symmetry, vol. 7, no. 2, pp. 675–688, learning weighted naive Bayes classifiers,’’ IEEE Access, vol. 10,
May 2015. pp. 22724–22736, 2022.

89612 VOLUME 11, 2023

S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

[49] Z. Lei and F. P. S. Chin, ‘‘Sensing OFDM systems under frequency- [71] S. Albawi, T. A. Mohammed, and S. Al-Zawi, ‘‘Understanding of a
selective fading channels,’’ IEEE Trans. Veh. Technol., vol. 59, no. 4, convolutional neural network,’’ in Proc. Int. Conf. Eng. Technol. (ICET),
pp. 1960–1968, May 2010. Antalya, Turkey, Aug. 2017, pp. 1–6.
[50] M. Jin, Q. Guo, J. Xi, Y. Li, and Y. Li, ‘‘On spectrum sensing of OFDM [72] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei,
signals at low SNR: New detectors and asymptotic performance,’’ IEEE ‘‘ImageNet: A large-scale hierarchical image database,’’ in Proc. IEEE
Trans. Signal Process., vol. 65, no. 12, pp. 3218–3233, Jun. 2017. Conf. Comput. Vis. Pattern Recognit., Miami, FL, USA, Jun. 2009,
[51] I. A. Basheer and M. Hajmeer, ‘‘Artificial neural networks: Fundamentals, pp. 248–255.
computing, design, and application,’’ J. Microbiol. Methods, vol. 43, [73] L. Alzubaidi, J. Zhang, A. J. Humaidi, A. Al-Dujaili, Y. Duan,
no. 1, pp. 3–31, Dec. 2000. O. Al-Shamma, J. Santamaría, M. A. Fadhel, M. Al-Amidie, and
[52] M. R. Vyas, D. K. Patel, and M. Lopez-Benitez, ‘‘Artificial neural L. Farhan, ‘‘Review of deep learning: Concepts, CNN architectures,
network based hybrid spectrum sensing scheme for cognitive radio,’’ in challenges, applications, future directions,’’ J. Big Data, vol. 8, no. 1,
Proc. IEEE 28th Annu. Int. Symp. Pers., Indoor, Mobile Radio Commun. Mar. 2021, Art. no. 53.
(PIMRC), Montreal, QC, Canada, Oct. 2017, pp. 1–7. [74] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, and A. Rabinovich, ‘‘Going deeper with convolutions,’’ in
[53] Y.-J. Tang, Q.-Y. Zhang, and W. Lin, ‘‘Artificial neural network based
Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Boston, MA,
spectrum sensing method for cognitive radio,’’ in Proc. 6th Int. Conf.
USA, Jun. 2015, pp. 1–9.
Wireless Commun. Netw. Mobile Comput. (WiCOM), Chengdu, China,
Sep. 2010, pp. 1–4. [75] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen,
‘‘MobileNetV2: Inverted residuals and linear bottlenecks,’’ in Proc.
[54] D. K. Patel and Y. N. Trivedi, ‘‘LRS-G2 based non-parametric spectrum
IEEE/CVF Conf. Comput. Vis. Pattern Recognit., Salt Lake City, UT,
sensing for cognitive radio,’’ in Proc. CROWNCOM, 2016, pp. 330–341.
USA, Jun. 2018, pp. 4510–4520.
[55] M. López-Benitez, F. Casadevall, and C. Martella, ‘‘Performance of [76] K. Simonyan and A. Zisserman, ‘‘Very deep convolutional networks for
spectrum sensing for cognitive radio based on field measurements of large-scale image recognition,’’ 2014, arXiv:1409.1556.
various radio technologies,’’ in Proc. Eur. Wireless Conf. (EW), Lucca, [77] K. He, X. Zhang, S. Ren, and J. Sun, ‘‘Deep residual learning for
Italy, 2010, pp. 969–977. image recognition,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recognit.
[56] M. López-Benítez and F. Casadevall, ‘‘Improved energy detection (CVPR), Las Vegas, NV, USA, Jun. 2016, pp. 770–778.
spectrum sensing for cognitive radio,’’ IET Commun., vol. 6, no. 8, p. 785, [78] W. M. Lees, A. Wunderlich, P. J. Jeavons, P. D. Hale, and M. R.
2012. Souryal, ‘‘Deep learning classification of 3.5-GHz band spectrograms
[57] Q. Cheng, Z. Shi, D. N. Nguyen, and E. Dutkiewicz, ‘‘Sensing OFDM with applications to spectrum sensing,’’ IEEE Trans. Cognit. Commun.
signal: A deep learning approach,’’ IEEE Trans. Commun., vol. 67, no. 11, Netw., vol. 5, no. 2, pp. 224–236, Jun. 2019.
pp. 7785–7798, Nov. 2019. [79] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, ‘‘Densely
[58] Z. Elsaraf, F. A. Khan, and Q. Z. Ahmed, ‘‘Deep learning based power connected convolutional networks,’’ in Proc. IEEE Conf. Comput. Vis.
allocation schemes in NOMA systems: A review,’’ in Proc. 26th Int. Conf. Pattern Recognit. (CVPR), Honolulu, HI, USA, Jul. 2017, pp. 2261–2269.
Autom. Comput. (ICAC), Portsmouth, U.K., Sep. 2021, pp. 1–6. [80] Q. Peng, A. Gilman, N. Vasconcelos, P. C. Cosman, and L. B. Milstein,
[59] X. Du, Y. Cai, S. Wang, and L. Zhang, ‘‘Overview of deep learning,’’ ‘‘Robust deep sensing through transfer learning in cognitive radio,’’ IEEE
in Proc. 31st Youth Academic Annu. Conf. Chin. Assoc. Autom. (YAC), Wireless Commun. Lett., vol. 9, no. 1, pp. 38–41, Jan. 2020.
Wuhan, China, Nov. 2016, pp. 159–164. [81] D. Han, G. C. Sobabe, C. Zhang, X. Bai, Z. Wang, S. Liu, and B. Guo,
[60] J. Xie, C. Liu, Y.-C. Liang, and J. Fang, ‘‘Activity pattern aware spectrum ‘‘Spectrum sensing for cognitive radio based on convolution neural
sensing: A CNN-based deep learning approach,’’ IEEE Commun. Lett., network,’’ in Proc. 10th Int. Congr. Image Signal Process., Biomed. Eng.
vol. 23, no. 6, pp. 1025–1028, Jun. 2019. Informat. (CISP-BMEI), Shanghai, China, Oct. 2017, pp. 1–6.
[61] K. Du, P. Wan, Y. Wang, X. Ai, and H. Chen, ‘‘Spectrum sensing method [82] V. Prithiviraj, B. Sarankumar, A. Kalaiyarasan, P. P. Chandru, and
based on information geometry and deep neural network,’’ Entropy, N. N. Singh, ‘‘Cyclostationary analysis method of spectrum sensing for
vol. 22, no. 1, p. 94, Jan. 2020. cognitive radio,’’ in Proc. 2nd Int. Conf. Wireless Commun., Veh. Technol.,
[62] S. Zhang, Y. Wang, J. Li, P. Wan, Y. Zhang, and N. Li, ‘‘A cooperative Inf. Theory Aerosp. Electron. Syst. Technol. (Wireless VITAE), Chennai,
spectrum sensing method based on information geometry and fuzzy India, Feb. 2011, pp. 1–5.
c-means clustering algorithm,’’ EURASIP J. Wireless Commun. Netw., [83] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner, ‘‘Gradient-based
vol. 2019, no. 1, pp. 1–12, Jan. 2019. learning applied to document recognition,’’ Proc. IEEE, vol. 86, no. 11,
pp. 2278–2324, Nov. 1998.
[63] Y. Zhang, P. Wan, S. Zhang, Y. Wang, and N. Li, ‘‘A spectrum sensing
method based on signal feature and clustering algorithm in cognitive [84] C. Liu, J. Wang, X. Liu, and Y.-C. Liang, ‘‘Deep CM-CNN for spectrum
wireless multimedia sensor networks,’’ Adv. Multimedia, vol. 2017, sensing in cognitive radio,’’ IEEE J. Sel. Areas Commun., vol. 37, no. 10,
pp. 1–10, Oct. 2017. pp. 2306–2321, Oct. 2019.
[85] S. Haykin, C. Krasnor, T. J. Nohara, B. W. Currie, and D. Hamburger,
[64] D. K. Patel, M. López-Benítez, B. Soni, and Á. F. García-Fernández,
‘‘A coherent dual-polarized radar for studying the ocean environment,’’
‘‘Artificial neural network design for improved spectrum sensing in
IEEE Trans. Geosci. Remote Sens., vol. 29, no. 1, pp. 189–191, Jan. 1991.
cognitive radio,’’ Wireless Netw., vol. 26, no. 8, pp. 6155–6174, Jul. 2020.
[86] F. F. Digham, M.-S. Alouini, and M. K. Simon, ‘‘On the energy detection
[65] A. Nasser, M. Chaitou, A. Mansour, K. C. Yao, and H. Charara, ‘‘A deep
of unknown signals over fading channels,’’ IEEE Trans. Commun.,
neural network model for hybrid spectrum sensing in cognitive radio,’’
vol. 55, no. 1, pp. 21–24, Jan. 2007.
Wireless Pers. Commun., vol. 118, no. 1, pp. 281–299, Jan. 2021.
[87] Y. Zeng, C. L. Koh, and Y.-C. Liang, ‘‘Maximum eigenvalue detection:
[66] M. Naraghi-Pour and T. Ikuma, ‘‘Autocorrelation-based spectrum sensing Theory and application,’’ in Proc. IEEE Int. Conf. Commun., Beijing,
for cognitive radios,’’ IEEE Trans. Veh. Technol., vol. 59, no. 2, China, 2008, pp. 4160–4164.
pp. 718–733, Feb. 2010.
[88] Y. Zeng, Y. Chang Liang, and R. Zhang, ‘‘Blindly combined energy
[67] Y. Zeng and Y.-c. Liang, ‘‘Eigenvalue-based spectrum sensing algo- detection for spectrum sensing in cognitive radio,’’ IEEE Signal Process.
rithms for cognitive radio,’’ IEEE Trans. Commun., vol. 57, no. 6, Lett., vol. 15, pp. 649–652, 2008.
pp. 1784–1793, Jun. 2009. [89] Y. Zeng and Y.-C. Liang, ‘‘Spectrum-sensing algorithms for cognitive
[68] A. Nasser, A. Mansour, K. C. Yao, H. Abdallah, and H. Charara, radio based on statistical covariances,’’ IEEE Trans. Veh. Technol., vol. 58,
‘‘Spectrum sensing based on cumulative power spectral density,’’ no. 4, pp. 1804–1815, May 2009.
EURASIP J. Adv. Signal Process., vol. 2017, no. 1, pp. 1–19, May 2017. [90] P. Hale, J. Jargon, P. Jeavons, M. Lofquist, M. Souryal, and A.
[69] D. Teguig, V. Le Nir, and B. Scheers, ‘‘Spectrum sensing method based Wunderlich, ‘‘3.5 GHz radar waveform capture at Point Loma,’’ Nat. Inst.
on likelihood ratio goodness-of-fit test,’’ Electron. Lett., vol. 51, no. 3, Standards Technol., Gaithersburg, MD, USA, Tech. Rep. NIST TN 1954,
pp. 253–255, Feb. 2015. 2017. [Online]. Available: https://doi.org/10.6028/NIST.TN.1954
[70] MathWorks. What is a Convolutional Neural Network? 3 Things [91] P. Hale, J. Jargon, P. Jeavons, M. Lofquist, M. Souryal, and A.
You Need to Know. Accessed: Mar. 15, 2022. [Online]. Available: Wunderlich, ‘‘3.5 GHz radar waveform capture at fort story,’’ Nat. Inst.
https://uk.mathworks.com/discovery/convolutional-neural-network- Standards Technol., Gaithersburg, MD, USA, Tech. Note 1967, 2017.
MATLAB.html [Online]. Available: https://doi.org/10.6028/NIST.TN.1967

VOLUME 11, 2023 89613

 S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

[92] M. A. Richards, Fundamentals of Radar Signal Processing. Chicago, IL, [117] R. Utrilla, E. Fonseca, A. Araujo, and L. A. Dasilva, ‘‘Gated recurrent unit
USA: McGraw-Hill, 2005. neural networks for automatic modulation classification with resource-
[93] X. Liu, D. Yang, and A. E. Gamal, ‘‘Deep neural network architectures for constrained end-devices,’’ IEEE Access, vol. 8, pp. 112783–112794,
modulation classification,’’ in Proc. 51st Asilomar Conf. Signals, Syst., 2020.
Comput., Pacific Grove, CA, USA, Oct. 2017, pp. 915–919. [118] A. Mumuni and F. Mumuni, ‘‘Data augmentation: A comprehensive sur-
[94] T. J. O’Shea, J. Corgan, and T. C. Clancy, ‘‘Convolutional radio vey of modern approaches,’’ Array, vol. 16, Dec. 2022, Art. no. 100258.
modulation recognition networks,’’ in Proc. Int. Conf. Eng. Appl. Neural [119] L. Li, Z. Zhang, and L. Yang, ‘‘Influence of autoencoder-based data
Netw. Cham, Switzerland: Springer, 2016, pp. 213–226. augmentation on deep learning-based wireless communication,’’ IEEE
[95] Deepsig Inc. RF Datasets For Machine Learning. Accessed: Jun. 8, 2022. Wireless Commun. Lett., vol. 10, no. 9, pp. 2090–2093, Sep. 2021.
[Online]. Available: https://www.deepsig.ai/datasets [120] C. d. Vrieze, L. Simic, and P. Mahonen, ‘‘The importance of being
[96] T. J. O’Shea, T. Roy, and T. C. Clancy, ‘‘Over-the-air deep learning based earnest: Performance of modulation classification for real RF signals,’’
radio signal classification,’’ IEEE J. Sel. Topics Signal Process., vol. 12, in Proc. IEEE Int. Symp. Dyn. Spectr. Access Netw. (DySPAN), Seoul,
no. 1, pp. 168–179, Feb. 2018. South Korea, Oct. 2018, pp. 1–5.
[97] T. J. Shea and N. West, ‘‘Radio machine learning dataset gen-
eration with GNU radio,’’ in Proc. GNU Radio Conf., vol. 1,
no. 1, Sep. 2016. Accessed: Jul. 2, 2022. [Online]. Available:
https://pubs.gnuradio.org/index.php/grcon/article/view/11
[98] W. Lee, M. Kim, and D.-H. Cho, ‘‘Deep cooperative sensing: Cooperative
spectrum sensing based on convolutional neural networks,’’ IEEE Trans.
Veh. Technol., vol. 68, no. 3, pp. 3005–3009, Mar. 2019.
[99] B. Soni, D. K. Patel, and M. López-Benítez, ‘‘Long short-term memory
based spectrum sensing scheme for cognitive radio using primary activity
statistics,’’ IEEE Access, vol. 8, pp. 97437–97451, 2020. SADAF NAZNEEN SYED (Graduate Student
[100] R. Tandra and A. Sahai, ‘‘SNR walls for signal detection,’’ IEEE J. Sel. Member, IEEE) received the B.Tech. degree in
Topics Signal Process., vol. 2, no. 1, pp. 4–17, Feb. 2008.
electronics and communication engineering and
[101] J. Gao, X. Yi, C. Zhong, X. Chen, and Z. Zhang, ‘‘Deep learning
for spectrum sensing,’’ IEEE Wireless Commun. Lett., vol. 8, no. 6,
the M.Tech. degree in digital communication from
pp. 1727–1730, Dec. 2019. the Maulana Azad National Institute of Tech-
[102] G. Loye. Long Short-Term Memory: From Zero to Hero nology (MANIT), India, in June 2018 and June
With PyTorch. Accessed: Mar. 22, 2022. [Online]. Available: 2021, respectively. She is currently pursuing the
https://blog.floydhub.com/long-short-term-memory-from-zero-tohero- Ph.D. degree with the University of Huddersfield.
with-pytorch/ Her M.Tech. thesis was based on breast cancer
[103] J. Xie, J. Fang, C. Liu, and X. Li, ‘‘Deep learning-based spectrum sensing classification from histopathology images using
in cognitive radio: A CNN-LSTM approach,’’ IEEE Commun. Lett., transfer learning and deep neural networks. She was a System on Chip
vol. 24, no. 10, pp. 2196–2200, Oct. 2020. Verification Intern with Intel Corporation, Bengaluru, from August 2020 to
[104] Y. Saleem and M. H. Rehmani, ‘‘Primary radio user activity models for July 2021, where she performed data analysis using machine learning
cognitive radio networks: A survey,’’ J. Netw. Comput. Appl., vol. 43, techniques and got familiar with formal verification. She is an MSCA Early
pp. 1–16, Aug. 2014. Stage Researcher for the EU H2020 MOTOR5G Consortium. Her research
[105] T. Nguyen, B. L. Mark, and Y. Ephraim, ‘‘Spectrum sensing using a
interests include the application of deep learning techniques in wireless
hidden bivariate Markov model,’’ IEEE Trans. Wireless Commun., vol. 12,
no. 9, pp. 4582–4591, Sep. 2013. communication and healthcare.
[106] R. Zhang, T. Lim, Y.-C. Liang, and Y. Zeng, ‘‘Multi-antenna based
spectrum sensing for cognitive radios: A GLRT approach,’’ IEEE Trans.
Commun., vol. 58, no. 1, pp. 84–88, Jan. 2010.
[107] S. Subray, S. Tschimben, and K. Gifford, ‘‘Towards enhancing spectrum
sensing: Signal classification using autoencoders,’’ IEEE Access, vol. 9,
pp. 82288–82299, 2021.
[108] G. Liu, H. Bao, and B. Han, ‘‘A stacked autoencoder-based deep neural
network for achieving gearbox fault diagnosis,’’ Math. Problems Eng.,
vol. 2018, pp. 1–10, Jul. 2018. PAVLOS I. LAZARIDIS (Senior Member, IEEE)
[109] E. Axell and E. G. Larsson, ‘‘Optimal and sub-optimal spectrum sensing received the Diploma degree in electrical engineer-
of OFDM signals in known and unknown noise variance,’’ IEEE J. Sel. ing from the Aristotle University of Thessaloniki,
Areas Commun., vol. 29, no. 2, pp. 290–304, Feb. 2011. Thessaloniki, Greece, in 1990, the M.Sc. degree
[110] W. Xu, W. Xiang, M. Elkashlan, and H. Mehrpouyan, ‘‘Spectrum sensing
in electronics from Université Pierre and Marie
of OFDM signals in the presence of carrier frequency offset,’’ IEEE Trans.
Curie (Paris 6), Paris, France, in 1992, and the
Veh. Technol., vol. 65, no. 8, pp. 6798–6803, Aug. 2016.
[111] J. Xie, J. Fang, C. Liu, and L. Yang, ‘‘Unsupervised deep spectrum Ph.D. degree from École Nationale Supérieure des
sensing: A variational auto-encoder based approach,’’ IEEE Trans. Veh. Télécommunications (ENST) Paris and Université
Technol., vol. 69, no. 5, pp. 5307–5319, May 2020. Paris 6, in 1996. From 1991 to 1996, he was
[112] D. P Kingma and M. Welling, ‘‘Auto-encoding variational Bayes,’’ 2013, involved in research with France Télécom and
arXiv:1312.6114. teaching with ENST Paris. In 1997, he became the Head of the Antennas
[113] Q. Xu, Z. Wu, Y. Yang, and L. Zhang, ‘‘The difference learning of and Propagation Laboratory, Télédiffusion de France/the France Télécom
hidden layer between autoencoder and variational autoencoder,’’ in Proc. Research Center (TDF–C2R Metz). From 1998 to 2002, he was a
29th Chin. Control Decis. Conf. (CCDC), Chongqing, China, May 2017, Senior Examiner with the European Patent Office (EPO), The Hague,
pp. 4801–4804. The Netherlands. From 2002 to 2014, he was involved in teaching
[114] I. S. Dhillon, Y. Guan, and B. Kulis, ‘‘Kernel k-means: Spectral clustering and research with the ATEI of Thessaloniki, Thessaloniki, and Brunel
and normalized cuts,’’ in Proc. 10th ACM SIGKDD Int. Conf. Knowl. University, London, U.K. He is currently a Professor of electronics and
Discovery Data Mining, Aug. 2004, pp. 551–556.
telecommunications with the University of Huddersfield, U.K. He has been
[115] Machine Learning Dataset for Radio Signal Classification.
Accessed: Jun. 12, 2022. [Online]. Available: https://panoradio- involved as a Principal Investigator in several international research projects,
sdr.de/radio-signal-classification-dataset/ such as EU Horizon 2020 MOTOR5G and RECOMBINE and NATO-SfP
[116] R. Utrilla. MIGOU-MOD: A Dataset of Modulated Radio ORCA. He has published over 200 research articles and several national and
Signals Acquired With MIGOU, a Low-Power IoT Experimental European patents. He is a member of IET (MIET), a Senior Member of URSI,
Platform. Accessed: Jun. 8, 2022. [Online]. Available: and a fellow of the Higher Education Academy (FHEA). He is serving as an
https://data.mendeley.com/datasets/fkwr8mzndr/1 Associate Editor for IEEE ACCESS.

89614 VOLUME 11, 2023

S. N. Syed et al.: Deep Neural Networks for Spectrum Sensing: A Review

FAHEEM A. KHAN (Member, IEEE) received the ANTONI IVANOV (Member, IEEE) received
bachelor’s degree in electronics engineering from the master’s degree in innovative communica-
Aligarh Muslim University, India, the master’s tion technologies and entrepreneurship from the
degree in communication and radar engineering Technical University of Sofia (TUS), Bulgaria,
from the Indian Institute of Technology Delhi, and Aalborg University, Denmark, in 2016, and
India, and the Ph.D. degree in electrical and the Ph.D. degree in communication networks and
electronic engineering from Queen’s University systems from TUS, in 2020. He is currently a
Belfast, in 2012. He is currently a Senior Lec- Postdoctoral Researcher with the Teleinfrastruc-
turer in electronic engineering with the School ture Laboratory, Faculty of Telecommunications,
of Computing and Engineering, University of TUS. His research interests include cognitive radio
Huddersfield. He has more than 15 years of teaching and research experience networks, adaptive algorithms for dynamic spectrum access, deep learning-
at academic and research institutions in the U.K., Sultanate of Oman, based solutions for cognitive radio applications, radio frequency data
United Arab Emirates, and India. Before joining as a Lecturer with the analytics, volumetric spectrum occupancy assessment, and graph signal
University of Huddersfield, in 2016, he was a Research Associate in wireless processing for resource allocation in current and future wireless networks.
communications and signal processing with the Institute for Digital Commu-
nications, The University of Edinburgh, where he contributed to research and
project management in several EU/EPSRC research projects. His research
is funded by the following research grants: Proof of Commercial Concept
(PoCC) research grant on ‘‘non-orthogonal multiple access techniques in
beyond 5G networks,’’ EU H2020 ETN Research Project MOTOR5G, and
EU H2020 RISE Research Project. His current research interests include VLADIMIR POULKOV (Senior Member, IEEE)
spectrum sharing/spectrum management for beyond 5G wireless networks, received the M.Sc. and Ph.D. degrees from the
licensed shared access and cognitive radio, non-orthogonal multiple access, Technical University of Sofia (TUS), Sofia, Bul-
machine learning and deep learning in wireless communications, full-duplex garia. He has been the Dean of the Faculty of
communications, MIMO and millimeter wave communications, and the Telecommunications, TUS, and the Vice Chair-
Internet of Things (IoT) networks. He is a fellow of HEA. man of the General Assembly of the European
Telecommunications Standards Institute (ETSI).
He is currently a TUS Professor with more than
30 years of teaching, research, and industrial expe-
QASIM ZEESHAN AHMED (Member, IEEE)
rience, managing numerous educational, research
received the Ph.D. degree from the University
and development, and engineering projects in the field of telecommu-
of Southampton, Southampton, U.K., in 2009.
nications. He is also the Head of the Teleinfrastructure Research and
He was an Assistant Professor with the National
Development Laboratory, TUS; the Head of the Intelligent Communication
University of Computer and Emerging Sci-
Infrastructures Research and Development Laboratory, Sofia Tech Park; and
ences (NUCES-FAST), Islamabad, Pakistan,
the Chairperson of the Cluster for Digital Transformation and Innovation,
from November 2009 to June 2011. He was a
Bulgaria. He is the author of many scientific publications, tutoring B.Sc.,
Postdoctoral Fellow with the Computer, Electrical
M.Sc., and Ph.D. courses in the field of information transmission theory
and Mathematical Sciences and Engineering
and wireless access networks. He is a fellow of the European Alliance for
Division, King Abdullah University of Science and
Innovation (EAI).
Technology, Thuwal, Saudi Arabia, from June 2011 to June 2014. He joined
the University of Kent, U.K., as a Lecturer, from January 2015 to January
2017. He was a Lecturer, then a Senior Lecturer, and currently a Reader
of electronic engineering with the School of Computing and Engineering,
University of Huddersfield, U.K., in 2017, 2018, and 2020, respectively.
He is also a Principal Investigator of the U.K. for Erasmus + DigiHealth-
Asia Project and the MSCA Staff Exchanges EVOLVE Project. He is a
Co-Investigator of the EU H2020 ETN Research MOTOR5G Project and ZAHARIAS D. ZAHARIS (Senior Member, IEEE)
the EU H2020 RISE Research RECOMBINE Project. His research interests received the B.Sc. degree in physics, the M.Sc.
include ultrawide bandwidth systems, millimeter waves, device-to-device, degree in electronics, the Ph.D. degree in antennas
digital health, and cooperative communications. He is a fellow of HEA. and propagation modeling for mobile commu-
nications, and the Diploma degree in electrical
and computer engineering from the Aristotle
University of Thessaloniki, Thessaloniki, Greece,
MARYAM HAFEEZ (Member, IEEE) received in 1987, 1994, 2000, and 2011, respectively.
the Ph.D. degree in electrical engineering From 2002 to 2013, he was with the Admin-
from the University of Leeds, U.K., in 2015. istration of the Telecommunications Network,
From 2015 to 2018, she was a Research Fellow Aristotle University of Thessaloniki. Since 2013, he has been with the
with the Institute of Robotics, Autonomous Sys- School of Electrical and Computer Engineering, Aristotle University of
tems and Sensing (IRASS), University of Leeds. Thessaloniki. He has been involved in several international research projects,
Since 2018, she has been a Senior Lecturer with such as EU Horizon 2020 MOTOR5G and RECOMBINE. He is the author
the Department of Engineering and Technology, of 78 scientific journal articles, 62 international conference papers, one
University of Huddersfield. Her current research national patent, five book chapters, and one textbook. His current research
is funded by the EU Horizon 2020 Program. Her interests include the design and optimization of antennas and microwave
research interests include the design and analysis of protocols for next- circuits, signal processing on smart antennas, development of evolutionary
generation green intelligent wireless networks by employing tools from game optimization algorithms, and neural networks. He is a member of the
theory and stochastic geometry along with the Internet of Things (IoT) and Technical Chamber of Greece. Recently, he was elected as the Chair
Industry 4.0 related research. She is also serving as a member of the Editorial of the Electron Devices/Microwave Theory and Techniques/Antennas and
Board for Frontiers in Communications and Networks. She has worked in the Propagation Joint Chapter of the IEEE Greece Section. He is also serving as
area of dynamic spectrum access had received the Best Paper Award from an Associate Editor for IEEE ACCESS.
the IEEE International Conference on Communications (ICC), in 2013.

VOLUME 11, 2023 89615