436 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 12, NO.

3, MARCH 2023

Robust Channel Invariant Deep Noncooperative Spectrum Sensing

Zhengyang Su , Kah Chan Teh , Senior Member, IEEE, Sirajudeen Gulam Razul ,
and Alex C. Kot , Life Fellow, IEEE

Abstract—Deep learning (DL) has been introduced to cogni- Recently, deep learning (DL) technology has been widely
tive radio network to solve the problem of spectrum scarcity adopted in communications and signal processing as it can
and further enhance the spectrum utilization. However, many extract higher levels of representations of the received signals.
DL-based spectrum sensing methods are sensitive to the environ-
ment, which means the sensing model needs to be re-trained with
For example, in [7], a two-step DL-based spectrum sensing
a large number of labelled samples in a new environment. In this framework was introduced, which is applicable to several types
letter, we propose a novel DL-based channel environment-robust of modulations.
spectrum sensing network named ER-SNet, which contains the Generally, most of DL-based spectrum sensing methods out-
encoder part extracting channel invariant features and the clas- perform traditional feature-based methods in scenarios where
sifier part for true hypothesis prediction. Extensive simulations they are trained. However, the environment of electromag-
have been conducted to show the performance improvement and
robustness of the proposed algorithm in sensing weak signals
netic wave propagation is considerably diverse, which means
over different channel conditions. the received signal distribution is highly impacted. Hence, the
sensing model trained well under one environment may work
Index Terms—Deep learning, spectrum sensing, channel
poorly in another one. One approach to address the challenge
robustness, cognitive radio network.
of the environment variations is to gather sufficient labeled
samples in various contexts, which is difficult to realize in
I. I NTRODUCTION practice. In [8], the authors exploited the cross-correlation
OWADAYS, the enormous growth of mobile terminal property of the cyclic prefix of orthogonal frequency-division
N device connectivity has led to increasing demand for
spectrum resources and a severe shortage of spectrum. The
multiplexing (OFDM) signals to sense the full-duplex OFDM
system via matching network (MN). However, the proposed
cognitive radio network (CRN) has shown tremendous poten- scheme is only applicable to a certain type of signal. In [9],
tial in tackling new requirements and challenges brought by the authors introduced adversarial learning strategy for deal-
the fifth generation (5G) mobile communication system [1]. ing with performance degradation with untrained testing SNR
Spectrum sensing performs critical task in cognitive radio sets, but it requires training SNR labels to extract the SNR-
(CR) technology, allowing a secondary user (SU) to detect invariant features. In [10], the authors proposed a deep transfer
and access idle spectrum opportunistically [2]. cooperative sensing (DTCS) approach to enhance the adap-
Energy detection (ED) is a traditional approach with the tation ability, but it could only obtain good performance on
lowest computational complexity, but it requires the estimation source and target domains rather than unknown domains.
of the noise power, and performs poorly at low signal-to- In [11], the authors investigated the phase difference (PD)
noise ratio (SNR) levels [3]. In [4], the matched-filter detection distribution of the received samples and proposed a deep
method has been adopted to achieve better performance, but it neural network (DNN)-based spectrum sensing approach that
requires prior information of primary user (PU) which is usu- is robust to noise uncertainty. However, the computation of
ally not available. In [5], a cyclo-stationary feature detection PD distribution would be highly affected by the sampling
method was proposed by leveraging the intrinsic periodicity of length.
the modulated signal. However, it has high computational com- Motivated by the aforementioned obstacles, we propose
plexity. As compared to the traditional approaches, machine an end-to-end spectrum sensing network named ER-SNet
learning (ML)-based spectrum sensing methods could obtain to improve the model generalization ability for untrained
higher detection accuracy [6], but it requires huge dataset, channel conditions. The contributions are summarized as
and the detection performance is affected by feature selection. follows:
• We propose a non-cooperative spectrum sensing scheme
Manuscript received 8 November 2022; accepted 9 December 2022. Date of ER-SNet, which is applicable for detecting the presence
publication 15 December 2022; date of current version 9 March 2023. This of unknown PU signals in new propagation environments.
work was supported in part by the Temasek Laboratories and Rapid-Rich
Object Search (ROSE) Lab, NTU, Singapore. The associate editor coordinat- • ER-SNet operates on complexed received signals, with
ing the review of this article and approving it for publication was R. Wang. advantage of extracting underlying features than manual
(Corresponding author: Kah Chan Teh.) extraction. The core idea is to extract the channel-
Zhengyang Su, Kah Chan Teh, and Alex C. Kot are with the School
of Electrical and Electronic Engineering, Nanyang Technological University,
insensitive features that are discriminative for PU state
Singapore 639798 (e-mail: [email protected]; [email protected]; but robust to channel variation to some extent.
[email protected]). • The effectiveness of ER-SNet is demonstrated via sim-
Sirajudeen Gulam Razul is with the Temasek Laboratories, Nanyang ulations. It can achieve better sensing performance as
Technological University, Singapore 637553 (e-mail: esirajudeen@
ntu.edu.sg). compared to other baseline schemes including DTCS,
Digital Object Identifier 10.1109/LWC.2022.3229491 DetectNet, TF-ResNet, etc.
2162-2345 
c 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
thorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on November 26,2023 at 22:20:06 UTC from IEEE Xplore. Restrictions app
 SU et al.: ROBUST CHANNEL INVARIANT DEEP NONCOOPERATIVE SPECTRUM SENSING 437

module’s output is labeled as (Channel, Length). Both streams

of the neural network (NN) E encode the different input raw
signals into high-dimensional features z = Eθe (x ) separately,
where θe denotes the training parameters of all the layers of
NN E. Since the high-dimensional features obtained by most
of the existing DL-based spectrum sensing networks are highly
sensitive to channel variations, it is difficult to infer the spec-
trum status precisely when the testing channel condition space
Ω2 is different from the training space Ω1 , i.e., Ω2 ∩ Ω1 = ∅.
Adopting the contrastive loss function, a domain augmen-
tation is conducted, which reduces the channel-dependency
of the encoding feature z . In addition, the dependency of
Fig. 1. The proposed ER-SNet network architecture.
z is partly related to the size of training space, i.e., |Ω1 |.
With the increased number of source domains, the detection
II. S YSTEM M ODEL performance could be enhanced accordingly.
In this letter, we consider a non-cooperative spectrum sens-
ing scenario where only one antenna is deployed to receive B. Training and Prediction of ER-SNet
the signal samples. A general spectrum sensing system can be Generally, the predicting process f of the proposed ER-SNet
formulated as a binary hypothesis test problem: H0 and H1 can be modelled by the composition of two functions, i.e.,
represent the absence or presence of PU signals in the sensing f = E ◦ H, where E : x → z aims to map the input I/Q
frequency band, respectively. Let x n ∈ C1×L denote the n th vectors x into a high-dimensional embedding feature z , and
complex observed vector of SU in one sensing period, where H : z → p is to produce the classification probabilities of each
L denotes the sampling length. The binary hypothesis system hypothesis of the encoding features, i.e., [p 0 , p 1 ]T . Given the
can be formulated as: constraints of the available source datasets, the purpose of the
 network is to find out the best approximation of functions E
H0 : x n = u n ,
H1 : x n = s n + u n ,
(1) and H with the designed loss.
1) Training Dataset Generation: To achieve the channel
where u n ∈ C1×L represents the identically and indepen- invariant spectrum sensing, we collect the labelled samples
dently distributed (i.i.d.) additive noise following circularly from different channel conditions. The dataset Ds can be
symmetric complex Gaussian (CSCG) distribution with zero expressed as follows:
mean and variance σw 2 , namely, u 2
n ∼ CN (0, σw ), s n ∈
C1×L is the n th modulated signal through the fading channel {X , Ψ, D} = {(x 1 , y 1 , d1 ), . . . , (x K , y K , dK )}, (3)
transmitted by the PU. The solution of the above problem can where X = {x k }K k =1 denotes the K received samples with
be expressed as:
random class labels (0 or 1). Additionally, Ψ = {y k }K k =1
H1 shuffles and randomizes the orders of samples in X, where x k
fθ|H1 (x ) ≷ η, (2) and y k are sampled under |Ω1 | distinct channel conditions.
H0
D = {dk }K k =1 denotes whether the class labels of x k and y k
where x ∈ C1×L is the received I/Q vectors and fθ|H1 (·) are the same (dk = 0) or not (dk = 1). During the process of
denotes the output probability of hypothesis H1 with trained training, one-hot encoding will be conducted for label dk .
parameters θ after the transformation formula f (·) of the 2) Training Process: During the training of ER-SNet, the
proposed ER-SNet. We can determine the state of PU sig- encoder E aims to maximize the difference among those sam-
nals by comparing fθ|H1 (x ) with the detection threshold η. ples with different labels and minimize the Euclidean distance
Typically, two performance metrics are used to evaluate the between the samples of the same label, which encourages the
algorithm, namely, the probability of detection Pd , and the received signals from various channel environments to map to
probability of false alarm Pfa . For these two probabilities, we nearby area in the embedding feature space. The objective of
have Pd = Pr {fθ|H1 (x ) > η | H1 }, or true positive, and the classifier H is to correctly decide the spectrum status by
Pfa = Pr {fθ|H1 (x ) > η | H0 }, or false positive, where Pr {·} minimizing the cross-entropy (CE) loss function.
denotes the probability of the event. Specifically, the encoder E is separated into two streams
of the same CNN structure with shared training weights
III. E NVIRONMENT-ROBUST S PECTRUM S ENSING S CHEME as the Siamese network does. The encoder E computes the
In this section, we will first present the structure of the contrastive loss of the last encoding CNN layer, as given by
proposed ER-SNet. Then, we will introduce the training strat-
1 
B
 2
egy and procedures of the network. Lastly, online deployment Jcon (θe ) = (1 − dk )Eθe (x k ) − Eθe (y k ) +
will be presented for the detection of PU signals. B
k =1
  2 
dk (max 0, m − Eθe (x k ) − Eθe (y k ) ) , (4)
A. Network Architecture of ER-SNet
As depicted in Fig. 1, the proposed network contains two where · denotes the Frobenius norm, and Eθe (·) is the func-
parts, namely, the encoder E and the classifier H, and each tion that maps the k th input x k into less channel-dependent

thorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on November 26,2023 at 22:20:06 UTC from IEEE Xplore. Restrictions app
 438 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 12, NO. 3, MARCH 2023

feature space. In addition, B stands for the batch size, and m Algorithm 1 Environment-Robust Spectrum Sensing Scheme
denotes the margin value that specifies the separability of each (ER-SNet)
layer’s embedding features. Pre-processing: Collect training samples and constitute the dataset
Ds = {X , Ψ, D}
The classifier H is constituted with transposed CNN layers Initialize: Set θe , θh with random initial weights
and a dense layer. Represented by Hθh , where θh denotes 1: Repeat
the parameters of layers in the classifier H, the output of the 2: Compute the total loss J (θe , θh ) with (8);
3: Compute the backward gradients;
network can be expressed as 4: Update the network’s weights θe , θh by employing
⎡  ⎤ gradient descent method Adam;
  Hθh Eθe |H0 (x k ) 5: Until accuracy with no improvement;
Hθh Eθe (x k ) = ⎣  ⎦ 6: Return the probability of each hypothesis, i.e., fθ|H0 (x ), fθ|H1 (x ),
Hθh Eθe |H1 (x k ) and the well-trained optimal parameters θe∗ , θh∗ .
Prediction: Evaluation of samples under untrained channel conditions
fθ|H0 (x k ) utilizing the decision threshold η with (12).
= , (5)
fθ|H1 (x k )

with
Thus, the reordered sequence with V corresponding classi-
fθ|H0 (x k ) + fθ|H1 (x k ) = 1, (6) fication probabilities of hypothesis H1 is given by:
 
where fθ|Hi (x k )(i = 0, 1) is the output probability of each fθ∗ |H1 (x̂ 1 ), fθ∗ |H1 (x̂ 2 ), . . . , fθ∗ |H1 (x̂ V ) . (11)
hypothesis with hyper-parameters θ of the whole network. We Lastly, the decision threshold can be obtained by
adopt the CE loss function for classification according to the  
maximum likelihood criterion, as given by η = fθ∗ |H1 x̂ floor (V ∗(1−Pfa )) , (12)

1
B    where the function floor (·) rounds down to the nearest integer.
Jcross (θe , θh ) = − bk · log fθ|H1 (x k ) + Thus, we can regulate Pfa to be our desired value, and the PU
B
k =1
  state will be determined by equations (2) and (12).
(1 − bk ) · log 1 − fθ|H1 (x k ) , (7) Hence, the proposed scheme with training and testing
procedures is summarized in Algorithm 1.
where bk represents the class label of the k th sample x k . To
balance the portion of the contrastive loss and classification IV. S IMULATION R ESULTS AND A NALYSIS
loss, equations (4) and (7) are normalized and weighted by γ In this section, we will present the parameters of the dataset
and 1 − γ, respectively. Hence, the total loss of the ER-SNet in detail, and analyze the complexity of our proposed scheme.
can be expressed as: Following that, simulation results and analysis including
robustness to channel variations will be presented.
J (θe , θh ) = γJcon (θe ) + (1 − γ)Jcross (θe , θh ). (8)

As binary classification problem is our main focus, the CE A. Dataset Setting

loss function should dominate the total loss function, thus We establish the training dataset by generating quadrature
leading to a relatively small γ value. When the CE loss func- phase-shift keying (QPSK) signals over different Rayleigh
tion is solely adopted, the proposed ER-SNet is degraded channels with the roll-off factor of 0.35, which is referred as
into a conventional deep learning-based architecture with- a commonly used baseline dataset RadioML2016.10a [13] for
out any domain augmentation. When the CE loss is ignored, modulation classification problems. It considers typical issues
the network cannot train its classifier H, resulting in poor that impact the wireless communication quality such as carrier
detection performance. In our experiment, γ is set to 0.3 frequency offset (CFO) and sample rate offset (SRO).
through extensive simulations. The optimal weights (θe∗ , θh∗ ) For the ease of training, we will view the complexed-value
of ER-SNet can be obtained by samples as concatenations of real and imaginary parts for
input. The sample length and sampling frequency of signals
(θe∗ , θh∗ ) = arg minJ (θe , θh ). (9) are set to 512 and 200 KHz, respectively. Furthermore, in
θe ,θh
our training dataset, the SNR range varies between −20dB
3) Online Prediction: With the well-trained ER-SNet, the and −2dB with a 2dB increment and there are 4,000 sam-
prediction dataset Dt will be established from the testing chan- ples with random labels under one SNR level across one
nel space Ω2 . To better compare the detection performance, type of communication channel. The channel models for
we need to maintain a constant probability of false alarm Pfa . training and prediction are specified in Table I [14], includ-
As shown in (2), we shall derive a decision threshold η to ing maximum doppler shift, multi-path delays and gains. To
identify the presence or absence of PU signals. V random show the advantages of the proposed ER-SNet in different
samples {x 1 , x 2 , . . . , x V } with hypothesis H0 are picked and channel environments, we choose Rayleigh channel models
rearranged in an ascending order of fθ|H1 (x v ), as given by {Ds_Rayleigh1, Ds_Rayleigh2, Ds_Rayleigh3} as the training
{x̂ 1 , x̂ 2 , . . . , x̂ V }, which satisfy space Ω1 , where |Ω1 | = 3. Characterized by Ω2 , the other two
  different channel models {Dt_Rayleigh, Dt_Rician (K = 4)}
∀1 < i < j < V , fθ∗ |H1 (x̂ i ) < fθ∗ |H1 x̂ j . (10) are used for prediction, which infers Ω2 ∩ Ω1 = ∅.

thorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on November 26,2023 at 22:20:06 UTC from IEEE Xplore. Restrictions app
 SU et al.: ROBUST CHANNEL INVARIANT DEEP NONCOOPERATIVE SPECTRUM SENSING 439

TABLE I
C HANNEL S PECIFICATIONS

Fig. 2. Sensitivity to different channel models at SNR=-12dB.

B. Complexity Analysis
In this section, we use a measurement known as multiply-
accumulate operations (MACC) [15] to analyze the complexity
of the proposed approach. Specifically, the ER-SNet mainly
consists of multiple convolutional, deconvolutional, and dense
layers that may significantly increase the system complexity.
The complexity of 1D-CNN layer is O(Cin · Cout · ks · Lout ),
where Cin and Cout represent the number of input chan-
nel and output channel, respectively. ks stands for the kernel
size, and Lout denotes the output length. Besides, deconvolu-
tional layers contribute similar complexity expression as that
of 1D-CNN layers, namely, O(Cin · Cout · ks · Lin ), where Fig. 3. Comparisons of detection performance between training and
Lin is the input length of the layer. Lastly, fully connected prediction at Pfa = 0.1.
layers incur the computational complexity of O(Min · Md ),
where Min denotes the input dimension of the dense layer,
and Md represents the number of neurons. Overall, the MACC
those of other methods since the encoder E in ER-SNet can
of the proposed ER-SNet in one iteration is approximately
be viewed as a domain augmentation step, which has the
6.58 million for one sample, which is about 17% more than
ability to deal with the channel variations and help reduc-
that of DetectNet [16]. More intuitively, it takes 0.056 ms for
ing the channel-dependency of the extracted features. Hence,
ER-SNet to make a decision for one sample. As compared
for prediction purposes, it will have a relatively small accu-
to DTCS, ER-SNet can obtain the Pd improvement and the
racy drop for Rician testing channels. Although DTCS could
robustness to the communication environments at the expense
perform relatively better than TF-ResNet and DetectNet, it
of around 30% more prediction time per sample.
requires re-training the model to set the specified target domain
in order to obtain good adaptation performance to unknown
C. Performance Evaluation channel conditions. Similarly, utilizing the strategy of transfer
In the simulations, we compare the performance of learning, TF-ResNet needs a small portion of labeled samples
our proposed ER-SNet to that of deep learning-based or under the new channel condition for re-training to obtain sim-
other conventional spectrum sensing approaches, includ- ilar level of probability of detection Pd as training stage. By
ing DetectNet [16], transfer learning-based ResNet (TF- contrast, the channel variations will cause around 25% drop
ResNet) [17], DTCS [10], and ED [18]. To ensure a fair in detection performance of DetectNet over Rician channel
performance evaluation of each scheme, the training dataset model. Besides, even within the same Rayleigh channel model
is maintained consistent. It should be noted that the optimal as that of training dataset, the minor changes such as multi-
hyper-parameters of each network have been obtained through path delays and gains in the channel profile will also cause an
extensive numerical simulations. accuracy drop of around 10%. In general, the proposed ER-
As illustrated in Fig. 2, the accuracy drop at SNR=−12dB SNet can perform better in untrained environments without
caused by channel model variations can be clearly observed. re-training the model.
The ED scheme will be excluded for comparisons in Fig. 2 In Fig. 3, the probabilities of detection of various spec-
since the conventional approach does not require offline train- trum sensing approaches under different channel conditions are
ing phase. Note that the DetectNet and ResNet are typical plotted against different SNR levels at Pfa = 0.1. Assuming
DL-based methods, and it is difficult to derive a general {Dt_Rayleigh} is the target domain for transfer learning-based
function that can extract the common feature of training sam- methods, and {Dt_Rician (K = 4)} is the unknown domain.
ples with different channel models. The training accuracy of It is shown that the proposed ER-SNet performs the best for
the proposed ER-SNet is around 7% to 14% higher than target Rayleigh channel regardless of the SNR, indicating the

thorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on November 26,2023 at 22:20:06 UTC from IEEE Xplore. Restrictions app
 440 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 12, NO. 3, MARCH 2023

performance and predict the state of PU over different chan-

nel conditions instead of re-training the model with numerous
labelled samples. By combining the contrastive loss with clas-
sification loss, the optimized structure can effectively extract
the channel invariant features for further accurate prediction
of spectrum status. Numerical results have shown that the
proposed channel-robust spectrum sensing scheme can achieve
a greater probability of detection at the cost of a slightly higher
computational complexity and is adaptable to other untrained
channel environments. For the future work, a potential exten-
sion is to scale our approach to multiple secondary users
scenario, where their observations or decision metrics could
be shared over a distributed ad-hoc network.

R EFERENCES
Fig. 4. ROC curves at SNR = −12dB for prediction over Rician channel: [1] J. Lunden, V. Koivunen, and H. V. Poor, “Spectrum exploration and
{Dt_Rician (K = 4)}. exploitation for cognitive radio: Recent advances,” IEEE Signal Process.
Mag., vol. 32, no. 3, pp. 123–140, May 2015.
[2] Y. Chen and H.-S. Oh, “A survey of measurement-based spectrum occu-
pancy modeling for cognitive radios,” IEEE Commun. Surveys Tuts.,
efficiency of the proposed approach. Under low SNR levels, vol. 18, no. 1, pp. 848–859, 1st Quart., 2016.
Pd of ER-SNet is around 8% to 13% higher than that of other [3] H. Urkowitz, “Energy detection of unknown deterministic signals,” Proc.
IEEE, vol. 55, no. 4, pp. 523–531, Apr. 1967.
DL techniques. Besides, without re-training the experiment [4] F. Salahdine, H. E. Ghazi, N. Kaabouch, and W. F. Fihri, “Matched
using large amounts of labelled samples, DTCS and DetectNet filter detection with dynamic threshold for cognitive radio networks,” in
will suffer a significant degradation of Pd when predicting Proc. Int. Conf. Wireless Netw. Mobile Commun. (WINCOM), Oct. 2015,
under the Rician channel model. If a small portion of labelled pp. 286–291.
[5] K. Sherbin and V. Sindhu, “Cyclostationary feature detection for spec-
samples in the unknown domain are considered, TF-ResNet trum sensing in cognitive radio network,” in Proc. Int. Conf. Intell.
could maintain comparable performance. Although the prob- Comput. Control Syst. (ICCS), May 2019, pp. 1250–1254.
ability of detection of ED is not affected by prediction, it [6] K. M. Thilina, K. W. Choi, N. Saquib, and E. Hossain, “Machine
learning techniques for cooperative spectrum sensing in cognitive radio
performs the poorest in general over both Rayleigh and Rician networks,” IEEE J. Sel. Areas Commun., vol. 31, no. 11, pp. 2209–2221,
channel. Nov. 2013.
Lastly, Fig. 4 shows the receiver operating characteristic [7] Z. Su, K. C. Teh, S. G. Razul, and A. C. Kot, “Deep non-cooperative
(ROC) curve, which is a graph that depicts the performance spectrum sensing over Rayleigh fading channel,” IEEE Trans. Veh.
Technol., vol. 71, no. 4, pp. 4460–4464, Apr. 2022.
of the binary classification model for unknown Rician channel [8] Q. Cheng, Z. Shi, and J. Yuan, “Spectrum sensing in full-duplex OFDM
model at a specific SNR level with two metrics, Pd and Pfa . systems using one-shot learning,” in Proc. Int. Conf. Commn. (ICC),
By adjusting Pfa along x-axis in log scale, we can visualize Jun. 2021, pp. 1–6.
Pd along y-axis at SNR of −12dB. It is observed that the [9] C. Wang, Y. Xu, Z. Chen, J. Tian, P. Cheng, and M. Li, “Adversarial
learning-based spectrum sensing in cognitive radio,” IEEE Wireless
proposed ER-SNet outperforms other schemes at any level of Commun. Lett., vol. 11, no. 3, pp. 498–502, Mar. 2022.
Pfa . When Pfa = 0.1, the Pd of ER-SNet can reach over [10] L. Li, H. Jiang, and H. He, “Deep transfer cooperative sensing
70%, whereas the Pd of other schemes range between 22% in cognitive radio,” IEEE Wireless Commun. Lett., vol. 10, no. 6,
pp. 1354–1358, Jun. 2021.
to 60%. In addition, we consider the imperfect situation with [11] Y. Wang et al., “Deep neural network-based robust spectrum sensing:
noise uncertainty (NU), which may be inevitable in a real- Exploiting phase difference distribution,” in Proc. Int. Conf. Commn.
world scenario. In the case of NU, the noise power varies over (ICC), Jun. 2021, pp. 1–7.
time, and the estimated noise power σ̂n2 follows a uniform dis- [12] X. Fu, Z. Feng, Y. Zhang, Q. Zhang, and W. Xu, “Phase difference
tribution over [ 1ε σn2 , εσn2 ], where ε denotes the NU factor and
variance based low complexity spectrum sensing scheme,” in Proc. IEEE
Global Commun. Conf. (GLOBECOM), Dec. 2015, pp. 1–6.
σn2 represents the actual noise power. In our simulations, ε is [13] T. J. O’Shea and N. West, “Radio machine learning dataset generation
set to 2dB, namely, ε = 1.57, which is denoted as “NU-2dB” with GNU radio,” in Proc. GNU Radio Conf., Jan. 2016, pp. 1–6.
in Fig. 4. It is observed that the ED-based scheme is sensitive [14] E. Perenda, S. Rajendran, G. Bovet, S. Pollin, and M. Zheleva, “Learning
the unknown: Improving modulation classification performance in
to NU because it implements the blind detection through the unseen scenarios,” in Proc. IEEE INFOCOM Conf. Comput. Commun.,
estimation of noise power. Contrarily, other schemes do not May 2021, pp. 1–10.
require prior knowledge about noise power for blind sensing. [15] V. Sze, Y.-H. Chen, T.-J. Yang, and J. S. Emer, “Efficient processing
of deep neural networks: A tutorial and survey,” Proc. IEEE, vol. 105,
These results have validated the detection reliability and effi- no. 12, pp. 2295–2329, Dec. 2017.
ciency of our proposed ER-SNet regardless of the presences [16] J. Gao, X. Yi, C. Zhong, X. Chen, and Z. Zhang, “Deep learning
of NU. Additionally, the proposed ER-SNet could be extended for spectrum sensing,” IEEE Wireless Commun. Lett., vol. 8, no. 6,
to the generalization of SNR and signal types. pp. 1727–1730, Dec. 2019.
[17] S. Zheng, S. Chen, P. Qi, H. Zhou, and X. Yang, “Spectrum sens-
ing based on deep learning classification for cognitive radios,” China
V. C ONCLUSION Commun., vol. 17, no. 2, pp. 138–148, Feb. 2020.
[18] Y.-C. Liang, Y. Zeng, E. C. Y. Peh, and A. T. Hoang, “Sensing-
In this letter, a novel network architecture and the train- throughput tradeoff for cognitive radio networks,” IEEE Trans. Wireless
ing strategy have been designed to improve the detection Commun., vol. 7, no. 4, pp. 1326–1337, Apr. 2008.

thorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on November 26,2023 at 22:20:06 UTC from IEEE Xplore. Restrictions app