Content uploaded by Reena Rathee
Author content
All content in this area was uploaded by Reena Rathee on Mar 06, 2018
Content may be subject to copyright.
Scalable and Robust ANN Based Cooperative Spectrum
Sensing for Cognitive Radio Networks
Reena Rathee Jaglan
1
•Rashid Mustafa
1
•Sunil Agrawal
1
Springer Science+Business Media, LLC, part of Springer Nature 2018
Abstract Cognitive radio network (CRN) supports dynamic spectrum access addressing
spectrum scarcity issue experienced by today’s wireless communication network. Sensing
is an important task and cooperative spectrum sensing is used for improving detection
performance of spectrum. The sensing information from individual secondary users is sent
to fusion center to infer a common global decision regarding primary user’s presence.
Various fusion schemes for decision making are proposed in the literature but they lack
scalability and robustness. We have introduced artificial neural network (ANN) at fusion
center thereby achieving significant improvement in detection performance and reduction
in false alarm rate as compared to conventional schemes. The proposed ANN scheme is
found capable to deal with scalability of CRN with consistent performance. Further, SNR
of individual Secondary user is taken into consideration in decision making at fusion
center. Moreover the proposed scheme is tested against security attack (malicious users)
and inadvertent errors occurring at SUs are found to be robust.
Keywords Cooperative spectrum sensing Artificial neural network Fusion
center Malicious user Signal to noise ratio
1 Introduction
Advancement and progress in wireless communication field led to escalation in wireless
applications leading to problem of spectrum scarcity as spectrum is limited. Static spec-
trum sensing is used at present which makes inefficient usage of available radio spectrum.
Cognitive Radio (CR) is an efficient technology dealing with spectrum scarcity problem by
using the authorized idle radio resource at need [1–3]. In Cognitive Radio Network (CRN),
unlicensed user referred as Secondary User (SU) dynamically uses the idle spectrum
&Reena Rathee Jaglan
reenarathee5@gmail.com
1
Department of Electronics and Communication, U.I.E.T., Panjab University, Chandigarh, India
123
Wireless Pers Commun
https://doi.org/10.1007/s11277-017-5168-1
licensed to authorized users referred as Primary Users (PUs). Accurate determination of
vacant spectrum (holes) is very important thus a vital role is played by Spectrum Sensing
(SS) process.
In literature, plethora of SS techniques are proposed, existing schemes like Energy
Detection (ED), matched filter detection, cyclo-stationary feature extraction, entropy
detection are classified as single-user SS. Among them ED is widely used due to its
simplicity [4–8]. However, single-user SS techniques cannot tackle shadowing, multipath
fading, and noise uncertainty issues. Thus, Cooperative Spectrum Sensing (CSS) fits best
into picture dealing with the disadvantages of single-user SS techniques and providing
sensing reliability [9–11].
CSS is an efficient technique which improves the sensing results as different SUs share
their own information or decision to reach single combined decision known as global
decision. Parameters of each SU which effect spectrum sensing are diverse due to its
location and environmental conditions. Thus, decision of each SU plays an important part
at Fusion Center (FC) to reach a global decision [12–14].
There are M spatially distributed SUs following marginal probabilistic model as H
1
:
xm *p(x
m
|h
i
) where m =1, 2, 3…M and i 2f0;1gfor binary hypothesis. The general
concept of Fig. 1can be summarized as:
•Each SU independently senses the spectrum making a local decision regarding PU’s
status (presence/absence) in the channel.
•All SUs forward their own local decisions to a central body FC.
•FC fuses all independent local decisions to infer a decisive global verdict about H
0
or
H
1
.
Fusion schemes at FC has drawn substantial research attention over the last decade. FC
can employ hard combination schemes which collects one bit local decision or soft
combination schemes which collects multi-bit local information for decision making
[15,16]. Soft or data fusion rules [17,18] include Equal gain combining (ECG), weighted
gain combination (WGC) and Mean cumulative sum (MCS). In ECG, equal weight is
assigned to each SU’s sensed data while in WGC weight is assigned to each SU in
proportion to the level of detected energy by SU. The authors in [19] has reported work on
MCS data fusion rule. These fusion schemes yields to larger overhead, transmitting
accurate sensed energies to the fusion center as compared to decision fusion scheme. Hard
or decision fusion rules include AND, OR and Majority rule which can be generalized as
SU 1
SU 2
SU M
FUSION
CENTER
X1
X2
XM1
d1
d2
dM1
Globaldecision
Fig. 1 Basic model of cooperative spectrum sensing
R. R. Jaglan et al.
123
K-out-of-N rule. The authors in [20] considered OR fusion rule for optimization of user
allocation and sensing time. These techniques yield lower overhead [21] as compared to
soft fusion schemes. Soft-hard combination scheme or softened scheme is discussed in the
literature but at the cost of more reporting time and increased overhead [22]. The whole
range of witnessed energy is divided into four regions, allowing two bits in each region. In
[23], semi-soft fusion scheme is discussed where two additional thresholds are considered
generating one or two-bit local decision based on test statistics observation by each SU.
Though soft schemes provides better accuracy and improved performance but at the cost of
increased overhead and bandwidth requirement. Thus, while proposing our ANN based
scheme both the factors i.e. accuracy and overhead have been taken into consideration. The
ANN based scheme implemented at FC outperforms than conventional schemes as it has
brilliant capability to learn from previous dataset or examples. It exploits pattern recog-
nition and learning capabilities of ANN making it viable for making decision at FC. Scaled
conjugate gradient back propagation has been used for fusion, offering good generalization
performance. Moreover it is simple and includes less computational complexity making it
more prominent.
However, in the environment or network considered, there may be a user which conveys
wrong decision to FC [24]. Such users are referred as Malicious Users (MUs) and presence
of such users can affect the network performance adversely [25,26].
There are many researchers who investigated detection performance in CRNs
employing number of SUs with ED for local sensing of spectrum and a FC. Many methods
have been proposed for improving detection and reducing false alarm. Earlier works on
CSS considered hard and soft rules which lack learning. Moreover, consideration of
learning by the network along with notorious or MUs is missing in the existing literature.
Back propagation based cooperative spectrum sensing algorithm using machine learning is
proposed in [27], which can get an improved sensing performance. Machine learning is
basically studied as part of Artificial Intelligence including involvement of statistics and
mathematics. Artificial Neural Network (ANN) is a tool that can help in solving most
difficult problems or it can be referred as a computational model used in machine learning
for solving problems efficiently. It is attracting significant attention among researchers in
the field of CR for predicting status of channel, detection performance evaluation,
improvement in QoS parameters etc. [28–32].The network keeps on updating itself till
output error meets target value by continuously adjusting its weight and error between
genuine and target data.
Remaining paper is systematized in following manner. Section 2describes system
model along with its mathematical significance. Section 3gives the description of simu-
lation model and algorithm of proposed ANN based trained SS model. Results based on
performance of proposed scheme are discussed in Sect. 4. The paper concludes with
Sect. 5.
2 System Model
A CSS environment with one PU, ‘M’ SUs, and one FC is considered as shown in Fig. 1.
Each SU uses energy detection for making their own local decision about PU’s presence.
Optimal threshold (k) calculation based on SNR values is considered for good performance
of energy detection algorithm. These individual local decisions are transmitted to FC by
Scalable and Robust ANN Based Cooperative Spectrum Sensing…
123
each SU through AWGN channel to infer a decisive global verdict about H
0
or H
1
indi-
cating PU’s status.
2.1 Local Sensing (ED) Model
ED is simple and easy to use, less complex, no necessity of prior information of primary
signal being utmost used algorithm [33,34]. So, these are the main reasons for carrying out
our work using mentioned single-user algorithm.
The signal received by each SU is a binary hypothesis problem and can be represented
as:
YðnÞ¼WðnÞ:H0ð1Þ
YðnÞ¼hXðnÞþWðnÞ:H1ð2Þ
where signal samples expected at detector are represented by Y(n), W(n) is AWGN, h is
the channel gain, and signal samples of PU are represented by X(n).Detector needs to
choose one of the hypothesis depending on the test statistics. Energy of the sensed signal is
estimated, termed as test statistics and can be represented as:
TðYÞ¼ 1
NX
N
n¼1ðY½nÞ2ð3Þ
where N =f
s
is sample size, is sampling time and f
s
is sampling frequency. The test
statistics is function of received signal and is compared against predefined threshold kfor
an observational time period in order to conclude a decision. The threshold [6] and
decision statistics can be given as:
k¼
Q1PfðÞ
ffiffiffi
N
pþ1
hi
SNR ð4Þ
where Q (.) depicts standard Gaussian’s complementary distribution function
Z¼TðYÞk:H0
TðYÞ[k:H1
ð5Þ
Each SU will either detect presence of PU signal or will detect its absence. If decision
statistics i.e. Z is more than threshold, PU exists else it doesn’t exist.
The system performance can be measured using three important performance evaluation
parameters i.e. probability of detection (Pd), probability of missed detection (Pm) and
probability of false-alarm (Pf).Mathematically they can be given as:
Pd ¼PfTðYÞ[kjH1ð6Þ
Pm ¼PfTðYÞ\kjH1ð7Þ
Pf ¼PfTðYÞ[kjH0ð8Þ
Pe ¼Pf þPm ð9Þ
where Pe constitutes probability of error which is amalgamation of Pf and Pm.
R. R. Jaglan et al.
123
2.2 Cooperative Spectrum Sensing Model
Considered M SUs in considered cooperative network where k users participate in sensing
process. The received signal at FC by kth SU be given as:
yk¼hkmkþnk;k21;2...M
fg ð10Þ
where hkis the kth channel coefficient and nkis noise. The PSK message signal ðmkÞ
depends upon decision taken at kth SU i.e. mk¼1 1 indicates presence of PU while 0
points towards its absence. If yk[0, FC takes a decision (u
k
) favoring H1 otherwise favors
H0 in decision making as following:
uk¼1;H1
0;H0
ð11Þ
These individual local one-bit decisions are fused by FC where a fusion rule is
employed. Different variants k among M SUs specifies fusion rule (K) aiding in global
decision making. These widely used fusion rules are categorized and expressed as:
•OR Rule: (k =1) FC dictates PU presence i.e. favors H1 when at least one SU reports
existence of PU otherwise favors H0 dictating absence of PU. The general OR can be
mathematically expressed as:
K¼H1;if PM
k¼1uk1
H0;if PM
k¼1uk\1
ð12Þ
•AND Rule: (k =N) FC takes a decision favoring H1 when all SUs report PU existence
i.e. number of decisions =M otherwise it takes a decision favoring H0. The general
AND rule can be mathematically expressed as:
K¼H1;if PM
k¼1uk¼M
H0;if PM
k¼1uk\M
ð13Þ
•Majority (k-out-of-M or half voting) Rule: k ¼M
2
FC takes a decision favoring H1 i.e.
PU if the number of decisions favoring H1 are greater than or equal to number of
decisions favoring H0 or vice versa. The general MAJORITY rule can be mathemat-
ically expressed as:
K¼
H1;if PM
k¼1ukM
2
H0;if PM
k¼1uk\M
2
8
>
<
>
:ð14Þ
Likewise Eqs. (12)–(14), the performance evaluation parameters for cooperative envi-
ronment can be formulated as:
Global false alarm probability;Qf ¼Pu¼1jH0
ðÞ ð15Þ
Global detection probability;Qd ¼Pu¼1jH1
ðÞ ð16Þ
Global miss-detection probability;Qm ¼Pu¼0jH1
ðÞ¼1Qd ð17Þ
Scalable and Robust ANN Based Cooperative Spectrum Sensing…
123
Qe ¼Qf þQm ð18Þ
where ‘u’ points to combined sensing result; u =1 depicts PU’s ON (present) status while
u=0 depicts PU’s OFF (absent) status. Qe constitute global probability of error.
2.3 ANN Model
A neural network is a network of ‘neurons’ organized in layers as predictors (input layer),
intermediate layer (hidden layer) and forecasts (output layer) (Fig. 2).
In feed-forward network, each layer of nodes receives input from previous layer of
nodes [35,36]. In the input layer, either individual local decision of SUs or decision along
with corresponding SNR values is considered constituting neurons in input layer equivalent
to SUs considered in network while for second case neurons’ quantity is double providing
more input dataset for learning. Number of hidden layer neurons can be varied based on
learning. Neural pattern recognition [37,38] is chosen as per the requirement for a neural
network which can classify inputs into target categories set, helps in training of the net-
work and evaluating its performance using confusion matrices [39].The output layer
constitutes a single neuron as we need a final binary decision f0;1g. The input dataset or
samples are randomly divided for the following three process i.e.
•Training: samples offered to network during training and based on error network is
adjusted.
•Validation: samples for measuring network’s generalization or learning while halting if
improvement in generalization stops.
•Testing: these samples don’t have any consequence on training, thus providing an
autonomous measure for performance of network.
Training dataset is used to prepare a model and to train the model. Preparing a decision-
making program is called ‘training’, where collected samples are called training set and the
program is referred as ‘model ‘and this model of the problem classify spam (correct
decisions) and non-spam (incorrect decisions). It can be referred as trained artificial neural
network model which can be used to make predictions on new data. Scaled conjugate
gradient back propagation is one of the training algorithms (trainscg) used for training any
network until there exists derivative functions for its weight, transfer functions and net
Local
decision
and
SNR
values
Input layer with
15 neurons
Hidden layer
with 10 neurons
Output layer
with 1 neuron
Output
(decision)
Fig. 2 Artificial Neural network architecture
R. R. Jaglan et al.
123
input. Calculation of performance derivatives is done in reference to weight and ‘X’ bias
variables through back propagation. Trainscg is a supervised learning algorithm for feed-
forward neural networks.
Test dataset is new data where output values are withheld from the algorithm. Pre-
dictions are gathered from the trained model and compared with the withheld output values
of the test set. Thus, allowing to compute performance measure for the model on test
dataset. The performance evaluation parameters of the proposed scheme model can be
calculated and evaluated using confusion fusion as shown in Table 1.
•TP (True Positive) is the number of correct decisions or classifications of positive
dataset.
•FN (False Negative) is the number of incorrect decisions or classifications of positive
dataset.
•FP (False Positive) is the number of incorrect decisions or classifications of negative
dataset.
•TN (True Negative) is the number of correct decisions or classifications of negative
dataset.
Using confusion matrix, the performance evaluation parameters of the system can be
given as:
Accuracy ¼TP þTN
TP þFN þFP þTN ð19Þ
False alarm ¼FP
TP þFN þFP þTN ð20Þ
Miss detection ¼FN
TP þFN þFP þTN ð21Þ
3 Simulation Model and Algorithm
The Simulink model considering a cooperative environment is developed in MATLAB.
CSS facilitates FC to take decision in better way such that larger the number of SUs better
is the accuracy. Thus better performance of the network. The simulation is carried out
according to the following steps. AWGN channel is considered throughout the simulation
study. ANN is used to prepare a trained model for spectrum sensing [36].
Table 1 Confusion matrix retrieving performance information
Decision Detected positive Detected negative
Genuine positive TP (sensitivity) FN (miss detection)
Genuine negative FP (false alarm) TN (specificity)
Scalable and Robust ANN Based Cooperative Spectrum Sensing…
123
Our proposed algorithm includes both Algorithms 1 and 2 sequentially for creating a
trained neural network for spectrum sensing which can be given by Fig. 3. Here, 91 is the
input dataset and y1 is the output.
The first step of the customized ANN based CSS trained model is data acquisition or
generating data by processing signals received at different SUs. Neural pattern recognition
is used as per the requirement for a neural network which can classify inputs into target
categories set, helps in training of the network and evaluating its performance. The input
dataset or samples are randomly divided for training, validation and testing processes. The
output layer constitutes a single neuron as we need a final binary decision f0;1g. The
proposed customized model is further used for testing the network performance consid-
ering three cases i.e. increase in number of users, SNR consideration along with individual
decision in the input dataset and security attack.
Input (generation of
data)
ANN based CSS trained algorithm (neural pattern
recognition)
Output
x1 y1
Validation Testing
Training
Fig. 3 Customized ANN based CSS trained model
Algorithm 1: For generation of training and target dataset for considered system model
i. PU signal XðnÞwith equally likely hypothesis H2H0;H1
fg
is modeled for OFF/ON activity using a
random integer generator and Bernoulli generator.
ii. AWGN channel is considered (h =1) for simplicity.
iii. The received signal under hypothesis H1 is hXðnÞþWðnÞwhile under hypothesis H0 is WðnÞ.
iv. Detection threshold ðkÞcan be obtained from eq. (4) based on SNR received by a SU.
v. Energy of received signal or test statistics at each SU is evaluated using eq. (3) and is compared with
threshold ðk) value to reach decision statistic (Z).
vi. The decision statistic at each SU is evaluated using eq. (5).
vii. Steps ii–vi are repeated for M number of SUs.
viii. The received signal at FC from kth SU is yk¼hkmkþnkwhich is given by eq. (10) where
nkis the AWGN noise:Using eqs. (6), (7) and (8), average Pd, Pm and Pf are assessed using different
fusion rules or logics given by equations (15), (16) and (17), over a large number of simulation
samples.
ix. The input and output data are saved in workspace and from here we get our training dataset and the
target data.
Algorithm 2: For generation of a trained network model for spectrum sensing
i. Number of neurons in the hidden layer of pattern recognition neural network (pattern net) is selected.
ii. Scaled conjugate gradient back propagation training algorithm is used for training or preparing a
decision-making program.
iii. The input dataset or samples are randomly divided for training, validation and testing processes.
Network performance is evaluated after testing using confusion matrix given by Table 1.
R. R. Jaglan et al.
123
4 Results and Discussion
We propose the use of Artificial Neural Network (ANN) at fusion center to infer better
performance even if the number of secondary users (SUs) changes in the network. The
specifications of ANN used are summarized in the Table 2.
The proposed ANN spectrum sensing model considered bipolar datasets i.e. ‘1’
depicting PU’s presence while ‘-1’ depicting absence because learning is more in case of
bipolar datasets. The performance is estimated in terms of accuracy, false alarm rate and
miss detection. False alarm is one of the most important parameters as any wrong infor-
mation send regarding PU’s presence even when it is not present can cause interference
with primary users which is not tolerable. In any case hindrance to licensed user’s action is
not acceptable. Thus, there is need for reduced false alarm, ideally close to 0.
First of all, the performance parameters are investigated for proposed fusion scheme and
compared with those of traditional schemes considering a network where the number of
SUs is fixed say 15, to establish the superiority of proposed scheme.
We consider a CRN, with a single PU and M number of SUs situated at different
distances from the base station thereby having different SNR values ranging from -20 to
10 dB. Assuming a practical scenario where most of the SUs are having low SNR values
i.e. they are situated far from the base station.
In Fig. 4, performance of CSS based on different fusion schemes and proposed ANN
scheme is compared. The performance by means of accuracy, false alarm and miss
detection for CSS is evaluated considering AWGN channel.
It is observed from Fig. 4that the proposed ANN scheme is able to give an accuracy
increased by 17.99, 22.68, 17.78, 18 and 18.3% as compared to conventional AND, OR,
Majority, semi-soft [24] and Peirce’s algorithm [40] fusion schemes respectively. More-
over, the proposed scheme provides reduction in false alarm by 12.08, 21.37, 0.29, 0.3 and
0.1% while reduction in missed detection is by 12.98, 1.29, 7.59, 0.2 and 0.5% for AND,
OR, Majority, semi-soft [24] and Peirce’s algorithm [40] respectively. For a CRN,
achieving increased accuracy, reduced false alarm and missed detection are of great sig-
nificance and the most desirables. It can be clearly seen from Fig. 4that the proposed ANN
scheme performs far better than the existing schemes.
After making sure that the proposed scheme can outperform other approaches found in
the literature, we look at another issues of CRN like scalability, robustness and security
issues. In upcoming sections, we address these issues taking one at a time.
Table 2 Specifications of pro-
posed artificial neural network
(ANN) model
Parameters Values
Network Three-layer feed-forward
Hidden layer function Sigmoid
Learning rate 0.01
Output neuron function Softmax
Training algorithm Trainscg
Number of hidden neurons 10
Scalable and Robust ANN Based Cooperative Spectrum Sensing…
123
4.1 Scalability Issue
For any CRN, the number of SUs may vary with time and it is desirable that the perfor-
mance of the system should not be adversely affected. To explore this issue, with an
increase in number of SUs, the performance of different schemes along with the proposed
one have been evaluated and summarized in Table 3.
It is found that the proposed (ANN) scheme performs consistently superior to con-
ventional fusion schemes with an increase in number of SUs. It can be observed from the
table that percentage of miss detection drops from 0.3 to 0.1. This can be attributed to the
fact that in cooperative sensing, one can have more cooperation when number of SUs
increases, thereby leading to less chance of missing a hole in the spectrum. The perfor-
mance of the ANN scheme is separately depicted in Fig. 5. So it can be concluded that the
80.51
13.28
13.28
75.82
22.57
1.59
80.72
1.59
7.89
80.5
1.5
0.5
80.2
1.3
0.8
98.5
1.2
0.3
ACCURACY FALSE ALARM MISSED DETECTION
PERFORMANCE PARAMETERS (%)
FUSION SCHEMES
AND OR Majority Semi-so [24] Peirce's algo [40] Proposed ANN scheme
Fig. 4 Comparative performance analysis for different fusion schemes in CSS network
Table 3 Effect of increase in number of SUs on performance of fusion schemes
Number of secondary
users (M)
Performance evaluation
parameters (%)
Fusion schemes at fusion center
AND OR Majority Proposed
scheme (ANN)
15 Accuracy 80.51 75.82 80.72 98.5
False alarm 13.28 22.57 7.89 1.2
Miss detection 13.28 1.59 12.38 0.3
30 Accuracy 80.51 75.82 80.72 98.7
False alarm 13.28 22.57 7.89 1.2
Miss detection 13.28 1.59 12.38 0.1
45 Accuracy 80.51 75.82 80.72 98.7
False alarm 13.28 22.57 7.89 1.2
Miss detection 13.28 1.59 12.38 0.1
R. R. Jaglan et al.
123
performance of CRN can be maintained even with a variation in number of secondary
users, thereby allowing the network to be scalable.
Figure 5shows that the proposed ANN scheme achieves accuracy of 98.5% when
number of SUs involved in cooperation are 15. It gains a 0.2% increased accuracy when M
is increased to 30 and maintains the accuracy level with further increase in number of SUs.
However, the false alarm remains same while there is significant reduction of 0.2% in
missed detection with increase in number of SUs. This leads to conclusion that if a number
of SUs leave or join the CRN, the performance would not be affected adversely.
Although the proposed ANN scheme is performing well, even it is desirable to explore
ways to enhance its performance further. It can be said without any doubt that at an SU the
decision accuracy will be more at larger value of SNR. Therefor if the decision of each SU
is clubbed with their rrspective SNR in the data set for training of ANN, then we can
expect improved performance.
4.2 Role of SNR for Performance Improvement of Proposed Scheme
Whenever an SU explores the spectrum regarding absence/presence of PU and takes
decision, the SNR value at that time is also captured and stored along with decision in the
dataset. Then this dataset is used for training and testing of the proposed scheme. The SNR
values have been normalized in order to prevent the network from getting saturated. The
performance of the network now tested and results are shown in the Fig. 6.
Though the detection performance or accuracy remains almost same for both the cases
but a reduction in false alarm is clearly observed in Fig. 6when SNR values at the time of
decision making are considered in the input dataset along with individual decisions of the
users in the network. There is 8.33% reduction in false alarm as compared to the scenario
without considering SNR. Hence it can be concluded that consideration of SNR in input
dataset is an important factor which could be used to enhance the network performance
further.
After having established the positive impact of SNR inclusion in the dataset for training,
now it is required to explore the scalability of the network. For this we allow an increase in
the number of SUs in the network and evaluating the performance of the network
simultaneously. The comparative performance can be analyzed from the Fig. 7.
From Figs. 6and 7it can be observed that there is 8.33% reduction in false alarm as
compared to the scenario where SNR is not considered in input dataset of the network. So
98.5
98.7
98.7
1.2
1.2
1.2
0.3
0.1
0.1
15 30 45
PERFRORMANCE PARAMETERS (%)
NUMBER OF SECONDARY USERS (M)
Accuracy False alarm Missed detecon
Fig. 5 Effect of increase in number of SUs on performance of proposed ANN fusion scheme
Scalable and Robust ANN Based Cooperative Spectrum Sensing…
123
it can be concluded that the network is scalable and one can expect better performance with
increased numbers of users in the network.
In the CRN with large number of users, it cannot be ruled out that one or more SU may
malfunction or may send wrong decision inadvertently/intentionally. So it necessary to
have a fusion scheme which is robust against such issues. In the next section the capability
of the proposed fusion scheme will be explored against such kind of security attacks from
malicious users.
4.3 Security Attack and Robustness of Fusion Scheme
We considered a CSS environment for CRNs with one PU, M =15 SUs and number of
Malicious Users (MUs) =m. We assume that one SU is malicious i.e. provides false
information regarding PU’s presence. The performance of proposed scheme is investigated
for CRN consisting of one malicious user and the results are summarized in Table 4.
98.5
1.2
0.3
98.2
0.1
0.2
ACCURACY FALSE ALARM MISS DETECTION
INPUT DATASET WITH & WITHOUT
SNR
PERFORMANCE PARAMETERS (%)
without SNR with SNR
Fig. 6 Impact of SNR on Performance of proposed ANN scheme for M =15
98.2 98.2 98.2
0.1 0.1 0.1
0.2 0.2 0.2
15 30 45
PERFORMANCE PARAMETERS (%)
NUMBER OF SECONDARY USERS (M)
Accuracy False alarm Miss detecon
Fig. 7 Effect of increase in number of SUs on performance parameters for proposed ANN scheme con-
sidering SNR along with decision
R. R. Jaglan et al.
123
Table 4summarizes the effect of increase in number of MUs (m) on performance
parameters for proposed ANN scheme. It is found that the false alarm is condensed
drastically while maintaining detection accuracy even with a variation in number of MUs
in the CRN. This significant improvement in the False alarm can be attributed to the
concept that few errors in dataset help the ANN to generalize the learning thereby
improving its performance. This can be visualized more clearly in the Fig. 8.
Figure 8demonstrates comparative performance analysis by means of accuracy, false
alarm and miss detection for increase in malicious users while considering 15 SUs among
network. The proposed ANN scheme can withheld or maintain an average accuracy of
98.2%. Moreover, the false alarm is reduced by 0.9% and 1.0% in case of one and two
malicious users respectively. Hence the proposed scheme is robust enough to maintain the
detection performance with reduced false alarm even when malicious users’ scenario is
considered establishing significant importance of such schemes at fusion center of CRN.
5 Conclusion
In CRN it is always desirable not to interrupt the communication of PU. It is always
preferable to miss a detection of a hole in the spectrum rather than making a wrong
decision. This capability of the CRN is measured in terms of miss-detection and false
alarm rate. So it is advisable to reduce false alarm rate even at the cost of an increase in
miss-detection. Keeping this consideration in view, an ANN based CSS scheme for CRN
has been proposed which has been found to be giving significant improvement in detection
Table 4 Effect of increase in number of MUs (m) for proposed ANN scheme with 15 SUs
Performance evaluation parameters (%) Number of malicious users (m)
012
Accuracy 98.5 98.2 98
False alarm 1.2 0.3 0.2
Miss detection 0.3 8.9 8.8
98.5
98.2
98
1.2
0.3
0.2
0.3
0.8
0.7
0 1 2
PERFORMANCE PARAMETERS (%)
NUMBER OF MALICIOUS USERS
Accuracy False alarm Missed detecon
Fig. 8 Performance analysis with an increase in number of MUs (m) considering SUs =15
Scalable and Robust ANN Based Cooperative Spectrum Sensing…
123
performance and reduction in false alarm as compared to conventional schemes. It has
been evaluated that detection performance or accuracy is increased by 17.99, 22.68, 17.78,
18 and 18.3% as compared to conventional AND, OR, Majority, semi-soft [24] and
Peirce’s algorithm [40] fusion schemes respectively. Moreover, the proposed
scheme provides reduction in false alarm by 12.08, 21.37, 0.29, 0.3 and 0.1% while
reduction in missed detection is by 12.98, 1.29, 7.59, 0.2 and 0.5% for AND, OR, Majority,
semi-soft [24] and Peirce’s algorithm [40] respectively for 15 SUs in the network. The
trained ANN based spectrum sensing model is further used for testing the network con-
sidering following three cases. In case 1, the effect of increase in number of SUs (M) is
observed and it can be formulated that the proposed scheme is successful in maintaining
the detection performance and constant false alarm rate for further increase in M. Hence,
the proposed scheme is able to deal with scalability of CRN. In case 2, SNR of individual
Secondary User is taken into consideration in decision making at Fusion Center. There is
significant reduction in false alarm by 8.33% when SNR is included in input database of
the proposed neural network. Thus, SNR is one of the important factors that need to be
considered for efficient performance. In case 3, the proposed scheme is tested against
security attack (Malicious Users). There is significant reduction in false alarm by 0.9 and
1.0% for m =1 and m =2 respectively. In all the above discussed three cases, it is found
that the proposed ANN scheme is robust enough to withstand the detection performance as
compared to conventional fusion schemes.
References
1. Federal Communications Commission (FCC). (2003). Spectrum policy task force report. ET docket no.
02-135.
2. Haykin, S. (2005). Cognitive radio: Brain-empowered wireless communication. IEEE Journal on
Selected Areas Communication, 23(2), 201–220.
3. Wyglinski, A. H., Nekovee, M., & Hou, Y. T. (2010). Cognitive radio communications and networks.
New York: Academic Press.
4. Jaglan, R. R., Sarowa, S., Mustafa, R., Agrawal, S., & Kumar, N. (2015). Comparative study of single-
user spectrum sensing techniques in cognitive radio networks. Procedia Computer Science, 58(2015),
121–128.
5. Yucek, T., & Arslan, H. (2009). A survey of spectrum sensing algorithms for cognitive radio appli-
cations. IEEE Communications Surveys & Tutorials, 11(1), 116–130.
6. Bargavi, D., &Murthy, C. R. (2010). Performance comparison of energy, matched filter and cyclosta-
tionary based spectrum sensing. In 2014 Eleventh international workshop on signal processing
advances in wireless communication (SPAWC) (pp. 1–6).
7. Garhwal, A., & Bhattacharya, P. P. (2011). A survey on spectrum sensing techniques in cognitive radio.
International Journal of Computer Science & Communication, Networks, 1, 196–206.
8. Jaglan, R. R., Mustafa, R., Sarowa, S., & Agrawal, S. (2016). Performance evaluation of energy
detection based cooperative spectrum sensing in cognitive radio networks. In First international con-
ference on information & communication technology for intelligent systems (Vol. 2, No. 1,
pp. 585–593). Springer.
9. Chaudhri, S., Lunden, J., Koivunen, V., & Poor, H. V. (2012). Cooperative sensing with imperfect
reporting channels, hard decisions or soft decisions. IEEE Transactions on Signal Processing, 60(1),
18–28.
10. Letaief, K. B., & Zhang, W. (2009). Cooperative communications for cognitive radio networks. Pro-
ceedings of IEEE, 97(5), 878–893.
11. Yao, C., & Wu, Q. (2014). A hybrid combination scheme for cooperative spectrum sensing in cognitive
radio networks. Mathematical Problems in Engineering, 2014(1), 1–7.
12. Bouraoui, R., & Besbes, H. (2016). Cooperative spectrum sensing for cognitive radio networks: Fusion
rules performance analysis. In International wireless communication & mobile computing conference
(IWCMC) (pp. 136–143). IEEE.
R. R. Jaglan et al.
123
13. Chen, Y. (2010). Analytical performance of collaborative spectrum sensing using censored energy
detection. IEEE Transactions on Wireless Communications, 9(12), 3856–3865.
14. Duan, D., Yang, L., & Scharf. L. L. (2012). The optimal fusion rule for cooperative spectrum sensing
from a diversity perspective. In Forty sixth Asilomar conference on signals, systems & computers
(ASILOMAR) (pp. 1056–1062). IEEE.
15. Althunibat, S., Di Renzo, M., & Granelli, F. (2013). Optimizing the k-out-of-N rule for cooperative
spectrum sensing in cognitive radio networks. In Global communications conference (GLOBECOM)
(pp. 1–5). IEEE.
16. Farag, H. M., & Mohamed, E. M. (2017). Soft decision cooperative spectrum sensing with noise
uncertainty reduction. Pervasive & Mobile Computing, 35(1), 146–164.
17. Chawdhury, M., & Kader, M. F. (2013). Performance analysis of local and cooperative spectrum
sensing in cognitive radio networks. International Journal of Signal Processing, Image Processing and
Pattern Recognition, 6(6), 397–410.
18. Pudi, S. K., Sundara, T. S., & Padmaja, D. N. (2013). Performance analysis of cognitive radio based on
cooperative spectrum sensing. International Journal of Engineering Trends & Technology Innovation
(IJETI), 4(4), 821–827.
19. Sriharipriya, K. C., & Baskaran, K. (2014). Collaborative spectrum sensing of cognitive radio networks
with simple and effective fusion scheme. Circuits, Systems, and Signal Processing, 33(9), 2851–2865.
20. Liu, X., Zhong, Wei-Zhi, & Chen, Kun-qi. (2015). Optimization of sensing time and cooperative user
allocation for OR rule cooperative spectrum sensing in cognitive radio network. Journal of Central
South University, 22(7), 2646–2654.
21. Do, N. T., & An, B. (2015). A soft-hard combination-based cooperative spectrum sensing for cognitive
radio networks. Sensors, 15(2), 4388–4407.
22. Ma, J., Zhao, G., & Li, Y. (2008). Soft combination and detection for cooperative spectrum sensing in
cognitive radio networks. IEEE Transactions on Wireless Communications, 7(11), 4502–4507.
23. Du, J., Guo, D., Zhang, B., & Su, Y. (2015). A robust cooperative spectrum sensing-assisted multiuser
resource allocation scheme. Mathematical Problems in Engineering, 1, 1–12.
24. Verma, P., & Singh, B. (2016). On the decision fusion for cooperative spectrum sensing in cognitive
radio networks. Wireless Networks, 16(1), 1–10.
25. Chen, C., Cheng, H., & Yao, Y. D. (2011). Cooperative spectrum sensing in cognitive radio networks in
the presence of the PUEA. IEEE Transactions on Wireless Communications, 10(7), 2135–2141.
26. Li, H., Cheng, X., Li, K., Hu, C., Zhang, N., & Xue, W. (2014). Robust collaborative spectrum sensing
schemes for cognitive radio networks. IEEE Transactions on Parallel and Distributed Systems, 25(8),
2190–2200.
27. Chen, Y., Zhang, H., Hu, H., & Wang, Q. (2014). An efficient cooperative spectrum sensing algorithm
based on BP neural network. In International conference on wireless communication and sensor net-
work (pp. 297–301).
28. Pattanayak, S., Venkateswaran, P., & Nandi, R. (2013). Artificial intelligence based model for channel
status prediction: A new spectrum sensing technique for cognitive radio. International Journal on
Communication, Network and System Sciences, 6(3), 139–148.
29. Giribone, P., Revetria, R., Antonetti, M., & Tablacci, R. (2000). Use of artificial neural networks as
support for energy saving procedures in telecommunications. In Twenty-second international
telecommunications energy conference (INTELEC) (Vol. 2000, No. 1, pp. 159–162). IEEE.
30. He, A., Bae, K. K., Newman, T. R., Gaeddert, J., Kim, K., Menon, R., et al. (2010). A survey of artificial
intelligence for cognitive radios. IEEE Transactions on Vehicular Technology, 59(4), 1578–1592.
31. Zhu, X. L., Liu, Y. A., Wey, W. W., & Yuan, D. M. (2008). Channel sensing algorithm based on neural
networks for cognitive wireless mesh networks. In Fourth international conference on wireless com-
munication, networks & mobile computing (WiCOM) (pp. 1–4). IEEE.
32. Liang, Y. C., Zeng, Y., Peh, E. C. Y., & Hoang, A. T. (2008). Sensing-throughput tradeoff for cognitive
radio networks. IEEE Transactions on Wireless Communications, 7(4), 1326–1337.
33. Chatziantoniou, E., Allen, B., & Velisavljevic, V. (2015). Threshold optimization for energy detection
based spectrum sensing over hyper-rayleigh fading channels. IEEE Communications Letters, 19(6),
1077–1080.
34. Haykin, S. (1999). Neural networks: A comprehensive foundation. Upper Saddle River, NJ: Prentice-
Hall.
35. Karray, F. O., & De Silva, C. W. (2004). Soft computing and intelligent systems design: Theory, tools
and applications. London: Pearson.
36. Baldo, N., & Zorzi, M. (2008). Learning and adaptation in cognitive radios using neural networks. In
Fifth IEEE consumer communications and networking conference (CCNC) (pp. 998–1003).
Scalable and Robust ANN Based Cooperative Spectrum Sensing…
123
37. Orcay, O., & Ustundag, B. (2008). Pattern recognition in cognitive communication. In Twenty-third
International symposium on computer & information sciences (ISCIS) (pp. 1–6). IEEE.
38. www.dataschool.io/simple-guide-to-confusion-matrix-terminology. Accessed May 01, 2017.
39. https://www.mathworks.com/products/neural-network/features.html. Accessed Jan 25, 2017.
40. Lavanis, N., & Jalihal, D. (2017). Performance of p-norm detector in cognitive radio networks with
cooperative spectrum sensing in presence of malicious users. Wireless Communications and Mobile
Computing, 17(1), 1–8.
Reena Rathee Jaglan is a Ph.D. research scholar working in the
Department of Electronics and Communication in UIET, Panjab
University, Chandigarh, India. She obtained her master’s degree in
Electronics and Communication from ITM University, Gurgaon. She
received her bachelor’s degree in Electronics and Instrumentation from
Maharshi Dayanand University, Rohtak. Her current research interests
include cognitive radio, spectrum sensing, artificial neural networks.
Rashid Mustafa is a Ph.D. research scholar working in the Depart-
ment of Electronics and Communication in UIET, Panjab University,
Chandigarh, India. He received his master’s degree in Instrumentation
Engineering from Panjab University, Chandigarh. He obtained his
bachelor’s degree from Kanpur University. His research interests
include cognitive radio networks and wireless sensor networks.
R. R. Jaglan et al.
123
Dr. Sunil Agrawal is actively working in the field of Cognitive Radio
Networks, Digital Signal Processing, Wireless Communication and
Artificial Intelligence. He did his Ph.D. from Panjab University,
Chandigarh in 2013 and M.Tech. in Electronics and Communication
from Thapar University, Patiala. He received his B.Tech. degree from
Jodhpur University, Rajasthan. He has guided several M.Tech. stu-
dents. He is currently working as a Professor in the Department of
Electronics and Communication, UIET, Panjab University, Chandi-
garh, India.
Scalable and Robust ANN Based Cooperative Spectrum Sensing…
123
A preview of this full-text is provided by Springer Nature.
Content available from Wireless Personal Communications
This content is subject to copyright. Terms and conditions apply.