Identifying Malicious Secondary User Presence Within Primary User Range in Cognitive Radio Networks

Cognitive Radio (CR) Network is a wireless communication technology, in which a detection device smartly detects occupied and unoccupied channels. During traffic, unused primary user spectrum space is allotted to a secondary user, without causing any intervention with the primary user. Though this has reduced network traffic to a great extent, many issues related to security has became an alarming problem, in which the primary user’s space is misused by some malicious secondary users without the knowledge of primary user. To address this issue, in this paper we have proposed a Boundary detection method that uses the estimated location of each SU, which is obtained using the Recurrent Neural Network algorithm, to determine the boundary of PU coverage. Then Malicious User Detection by Ordering (MUDO) methodology is proposed, in which all secondary users are weighed using Basic Probability Analysis (BPA), and based on the orders the SUs are paired with corresponding PUs. The SUs with the least orders are discarded as they might be malicious users. The proposed methodology possesses higher detection speed and precise detection thereby enhancing the performance of CR.


Introduction
After technological advancements, various wireless network technologies like Wi-Fi, Bluetooth, LTE, and ZigBee, and other satellite technologies came into the telecommunication field.But as the number of users and usage increases, there were severe issues with the emergence of spectrum scarcity.There were also reports showing a wide range of spectrum remains unused at a particular frequency and in a particular region [1].It was stated by the Federal Communication Commission [FCC] that, nearly 15% to 85% of the spectrum frequency remains unutilized [2].
To overcome this problem, Cognitive Radio networks were proposed by Joseph Mitola [3,4].The Cognitive Radio (CR) Networks make the efficient usage of available spectrum, by allocating unused spectrum space of Primary or Legitimate User (PU or LU) to the Secondary Users (SU).Whenever the LU is found to appear, the SU has to suspend from using the channel [5].
To improve the detection mechanism, Cooperative Spectrum Sensing (CSS) came to the spectrum field, where all the SUs send their detection to a fusion center where a decision will be made whether SUs can be granted to use the Spectrum space of a particular PU.There are two types of CSS.One is Centralized CSS, in which the SUs send the presence of LU information to a central Data Center Unit (DCU), where the decision is made about the presence of LU.The second is decentralized CSS, where the information gathered by SUs is shared and they make the decision about the presence of LU among themselves.
But the major problem with Centralized Cooperative Spectrum Sensing is the Data Centre Unit (DCU), which will not know whether the LU information from SUs is legitimate or not.So there is a possibility of giving LU spectrum space access to some anonymous or illegal users.To resolve this problem, this paper has proposed a two-step detection process.In step-1, a boundary detection method using Recurrent Neural Network is used to predict the estimated location of SU within the PU boundary location.In the next step, the Malicious User Detection by Ordering (MUDO) methodology is proposed, where all the SUs who had sent the information to the DCU have been allotted number orders.Based on the ordering, the LUs and SUs are paired for accessing the unused LU frequency spectrum space.SUs with the highest ordering are only allowed to pair with LU.SUs with the least order are discarded as they might be considered a malicious user.
The remainder of the paper is organized as follows: Sect. 2 discusses the Literature review of papers related to the security of Cognitive Radio networks, Sect. 3 discusses the proposed model, Sect. 4 discusses results analysis and Sect. 5 concludes the paper.

Literature Survey
In [6], the author proposed a PU boundary detection mechanism using a Support Vector Machine (SVM).In this method, there will be sensing nodes that detect the signal strength of the PU and compares that to a Threshold value.The fusion center will receive signals from SU. From this, it is possible to detect whether the SUs are within the boundary of PU.The performance of the algorithm is improved by the kernel-SVM [7], where the data are applied to high dimensional space and transforming data into Euclidean space, for applying varying data classification methods.But both [6] and [7] proposed methods were unable to detect the location of SU in the Fusion Centre.
To address this issue, range-based and range-free localization algorithms were proposed [8].These types of techniques identify the devices that fall within the sensor range.Several detection techniques like matched-filter techniques, wave-form-based detection techniques, and cyclostationary based detection techniques, have been sort out as energy detectors in [9].In [10], it was discussed that though there are more advancement in Cognitive Radio networks, there are many security issues and different threats which was discussed in this literature.In [11], the author has discussed that Primary User Emulation (PUEA) and Spectrum Sensing Data Falsification (SSDF) are the two most dangerous attacks in the CR networks.It was stated that in PUEA, the malicious users act as the primary user and make the SUs free the available spectrum.In SSDF, the malicious user sends wrong information about whether PU exists or not.
In [12], the author has proposed a methodology named Cooperative scheme based on adaptive spectrum sensing, in which the energy detector is used as a first stage detector and the matched filter detector is used as the second stage detector, in the regions between the signal and noise.But this methodology has a drawback of inefficient detection when there is a low Signal-noise ratio.This problem was resolved in [13], by applying bi-level threshold holding, for energy detection.In [14], SUs reputation degree was utilized for allocating weights to the secondary users by proposing a soft-fusion-based algorithm, to detect the malicious user.
In [15], the Monte Carlo Localization method is proposed, where the localization accuracy is improved utilizing Connectivity, node range, and mobility information.But this method suffered from high complexities, in calculating, Angle of Arrival (AOA), Time Difference Of Arrival (TDOA), Received Signal Strength Indicator (RSSI), and Time Of Arrival (TOA).In [16], to address this problem, proposed multiple anchor nodes, to sense for each sensor in the entire network.
In [17], to detect the malicious user, K Nearest Neighbor K-NN, has been used, in which detection can be made without any knowledge of the malicious user.In [18] & [19], advanced K-NN and Dempster-Shafer (D-S) were implemented to detect malicious SUs in the PU boundary coverage.In [20], a discussion on Cognitive Radio Sensor Networks (CRNS) was done, where some general implementation issues with CRNS has been signified.
In [21], a novel temporal learning algorithm has been implemented, for achieving improved results through synchronization between input and output patterns.This algorithm eliminates backward integration as this methodology is based on forward instantaneous weight updation principle.Artificial Neural Network is implemented as Switched-Capacitors (SC) in [22], to analyze the effects of real biological neurons.This approach also depicts how SC structures to input data vectors namely binary and bipolar coding.This approach has enhanced pattern and characters recognition based applications in more effective manner.
In [23], the author has proposed Canonical Correlation Analysis methodology to detect and extract overlying and numerous nuclei scraps from the images of blood cells.This method employs CNN-LSTM network framework, which showed better performance with increased accuracy.In [24], for analyzing the liquid saturated steam heat process, the author has proposed Focused Time Lagged Recurrent Neural Network algorithm.This methodology along with dynamic NN approach outperforms the static NN in terms of Mean square Error (MSE), normalized MSE and correlation coefficients.
For modeling and identification of artificial Neural Networks in Binary Distillation Column, topologies of Feed forward network and recurrent neural network was implanted in [25].Hyperbolic tangent sigmoid functions and pure liner functions were used as activation functions in this observation.From experimental results it was proved that RNN performed better in modeling and identification that FFNN.
In [26], critical analysis of optimization techniques was proposed in order to fine tune the wavelet parameters of MRPID thermal controller system.For enhanced results, the optimization is performed using Elephant Herding Optimization (EHO) with RNN.With this implementation, the controller performed better in terms system stability and ephemeral response.In [27], for intelligent information retrieval based on text and image, Deep Convolution Neural Network is proposed.This methodology employs VGGNet-19 for retrieving images and Bi-directional-long short-term memory technique is used for retrieving texts.When compared to existing techniques, this intelligent model performed better in terms of precision, recall and F-Score.
In [28], the author has proposed fault diagnosis protocol using neural networks, for identification falsified sensor node effects in a wireless sensor network.The simulation results showed that, the proposed protocol performed better than existing methodologies in terms of false alarm rate, detection accuracy, false positive, false negative and detection latency.

Proposed Methodology
Consider there are one Primary User (PU) and many secondary users (SU) who are waiting to utilize the unused spectrum of primary users.Each SU sends its request to the DCU to know the status of the Primary User spectrum.The received primary user signal can be detected for the ith SU as, Probability of detection, P (di) as, di is the distance between the ith SU and PU, di is the reference distance and Po is the signal strength received at d0 and k is the path-loss exponent.

Boundary Detection by Recurrent Neural Network
Secondary Users confirm the availability of Primary User spectrum by contacting the Fusion Center.Once the availability is confirmed, the Secondary User tries to access the PU spectrum space.At this stage, it is necessary to check whether the SU lies within the boundary of PU coverage.This can be done by employing the Recurrent Neural Network algorithm.As Recurrent Neural Network is Long term memory, they are used to remember the previous states effectively, and hence they will be able to identify the presence of SU is within the PU coverage.The equation for calculating the current state is: where h t is the current state, h t-1 is the previous state and X t is the input state.The output can be calculated as 0' where W hy is the weight at the output layer.
Lets us consider that the PU coverage is dth.So, if ht for the ith SU lies within dth, SU is bounded to PU coverage else, the particular SU can be discarded.

Malicious User Detection by Ordering(MUDO)
To assess whether the SU is legitimate or malicious, weight of every SU has to be calculated.This is calculated using Basic Probability Assignment (BPA), in which the secondary users requesting access to primary user spectrum, is assigned weights.Upon ( 1) assigning weights, SU with least ordering is considered as malicious user and discarded.The SU's location-sensing is assumed such that, the Primary User is present under Hypothesis H0 and not present under Hypothesis H1.
where a(n) denotes additive white Gaussian noise, t(n) is the signal transmitted from the Primary User.The received signal estimation of every Secondary User is given by the following equation as follows, where F = 2DB, where D is the sensing duration, and B is the bandwidth.|l j | denotes the received sample of j th signal.As per Central Limit Theorem, when D > 200, then the received signal can be calculated with means µ 0 & µ 1 and variances 0 & 1, under H 0 and H 1 .
Under H0, Under H1, where ∅ is the signal-to-noise ratio of the Primary User at the Secondary User location.The Basic Probability Assignment determines whether the Primary User is present or absent in the following equations, where p i ( b Gj | H 0 ) denotes the probability of absence of primary User and p i ( b Gj | H 1 ) denotes the presence of the Primary User.
The weight can be calculated as follows: where SNR (i) is the i-th SU's Signal-to-noise ratio.After obtaining the mass of every SU calculate the Ordering of every SU as follows, where o'H0 (i) is the updated ordering by ith SU when PU is absent and o H1 (i) is the updated ordering by ith SU when PU is present.Now if the order of ith SU is less than the predefined threshold (th), the corresponding SU is not legitimate and ordering is assigned zero.If the order of ith SU is greater than the Threshold (th) then, the SU is legitimate and orderings are updated accordingly.

Identifying Legitimate Secondary User Algorithm
The proposed algorithm first takes the SUs request for accessing the unused spectrum of Primary Users by sending the request to DCU.As a first step, it is necessary to ascertain whether the Secondary Users lie within the PU coverage.For this, boundary detection is done using a Recurrent Neural Network, to identify whether the SU lies within PU coverage.
If it is false, the corresponding SUs are discarded and if it is true the mass of every SU is calculated, and based on the mass, the ordering is calculated to distinguish whether the SU is legitimate or not.If the ordering of SU is less than th, the SU is discarded and orders are assigned zero.If the reordering of SU is greater than the threshold, then they are legitimate SU and the orderings are updated.
The algorithm has been depicted in Table 1 and the flow chart of the proposed methodology has been depicted in Fig. 1.

Result Analysis
In this section, the simulation results of the proposed Boundary detection algorithm and MUDR methodology has been discussed.The simulations were performed using Mat-lab2018r to analyze the results of Pd, Pf, and SNR.The simulation parameters to carry out this simulation are depicted in Table 2.
In Fig. 2, a graph is plotted to compare the performance analysis of the Probability of detection and the probability of false detection, by taking the warning limit of false detection at 0.01, with SNR 10 dB and the broadcasting factor at d = 1.It is found from the result that when the false warning limit is 0.01, the probability of SNR is 10, which run through 10,000 iterations to distinguish between the malicious and genuine secondary users.
In Fig. 3, the Secondary Users that falls within the range of Primary User is detected using a recurrent neural network The secondary users, who are within the range of primary users, are allowed to access the unused spectrum of primary users.For 30 dB threshold, with modeling at 5, acts as the boundary between primary user and the secondary user.The secondary user modeling range that falls between 20Db threshold and ( 8) 45 dB threshold is takes as the valid region, Hence the secondary users who whose threshold fall within the region, is termed as genuine users.
In Fig. 4, a ROC curve was plotted which considers the extended probability for Pf from 0.01 to 0.05, and 0.1 is depicted.When the SNR probability is 1.7 times, there was a 5% expansion in the probability of false detection with a false alarm rate at SNR = -10 dB.From the figure, the extended probability for Pf from 0.01 to 0.05 and 0.1 is depicted.When the SNR probability is 1.7 times, there was a 5% expansion in the probability of false detection.
In Fig. 5, the performance evaluation of the ROC curve for Pd and SNR has been depicted.The probability of false detection is 0.02, -15 SNR, d = 1 and -11db.In the graph, you will be noticing that there is an increasing pace in the probability of detection.This clearly shows that when there is no malicious user addition, the probability of detection of this proposed scheme is 0.92 and when the malicious user is added, the detection probability falls to 0.8.  3 compares the overall probability detection for K = 10 Users for different existing methodologies and the proposed methodology.From the comparison, it is found that the proposed methodology shows the better result at all working conditions.Figure 6 depicts the graphical comparison of various existing methods and proposed methods, plotted against SNR VS detection probability.From the graph, it is found that at SNR -20 DB the proposed has reached the maximum detection probability of 1.This proves our proposed methodology performs better than the existing methodologies.

Fig. 1
Fig. 1 Flow chart of the proposed methodology

Table 1
Identifying legitimate secondary user algorithmTable