Optimised Levenshtein centroid cross-layer defence for multi-hop cognitive radio networks

Cognitive radio networks (CRN) make use of dynamic spectrum access to communicate opportunistically. Unlicensed users severely affect the spectrum sensing outcomes in CRN. Primary user emulation attack (PUEA) and spectrum sensing data falsiﬁcation (SSDF) have become a paramount concern in CRNs. It is especially challenging when both masquerading (in the physical layer) and falsiﬁcation (in the data link layer) occur by providing false spectrum reports. Existing methods to detect such attacks cannot be utilised in scenarios with multi-hop CRN. In this study, to mitigate attack against PUEA and SSDF, a method called optimised sensing and Levenshtein nearest centroid classiﬁ-cation (OS-LNCC) for multi-hop CRN is presented. First, a network model for multi-hop CRN is designed. Next, a probable density optimal logical sensing model is designed to alleviate the problems related to falsiﬁcation of spectrum reports. Here, the falsiﬁcation of spectrum reports is overcome by exploiting dual factors, that is, probability for false alarm and probability for detection according to the departure rate of primary user (PU). With these dual factors, optimal logical sensing is made, therefore improving the throughput with minimum delay. Finally, each cognitive radio (CR) user evaluates its current sensing information to existing sensing classes through the Levenshtein distance function. Based on quantitative variables, the prediction function of each sensing class is measured using nearest centroid (NC) classiﬁer and the sensing report is classiﬁed into either presence or absence of PU. These predictive classes are then integrated at the fusion centre so that robust mitigation against PUEA and SSDF is made. Computer simulation outputs show that OS-LNCC method performs better than the conventional methods using metrics such as sensing delay by 47%, percentage of error in prediction by 46% and throughput by 45%.


INTRODUCTION
Cognitive Radio (CR) is one of the methods employed to address the issues related to insufficiency radio spectrum. In CR technology, there is a possibility of malicious users (MUs) accessing the radio frequencies across a spectrum hole. As MUs are not primary group of users for using CR technology, it poses a challenge to the exploitation of spectrum prediction. Thus, the users are classified into two types in CRN. Primary users (PUs)that take the specific part of the radio spectrum by holding an authority, and secondary users (SUs) that use the specific portions of the radio spectrum. With this, the licensed users only access resources and share information freely, anytime designed: control model, target model, and a control and a prediction possibility. By designing these models, immunity against DoS attack was generated. With this, two different attacks, namely, PUEA and SSDF were detected with better back-off and sensing delay (SD). However, high spectrum utility was not attained due to single hop CRN. With the design of multi-hop CRN, more spectrum resources are said to be supplied to the CRN user and falsification of spectrum is also reduced by means of STATLOG function, an integration of statistical and logical functions, therefore contributing to high spectrum utility. A routing and spectrum allocation algorithm based on the deadline was designed in [2].
The objective of developed algorithm was to maximise the resource utilisation and offer higher packet delivery. The method combined the interface of opportunistic routing, spectrum allocation to progress the efficient data delivery.
Here, a virtual queue (VQ) with weights was exploited for forming network utility function. Based on this, session optimality, transmit power, and frequency were measured by means of the distributed method to provide effective exploitation of existing resources. In this manner, maximisation of packet delivery was ensured within the specified deadline, therefore improving throughput and reliability.
Though spectrum distribution were made, with the existence of PUE, an assailant by posing as a PU starts broadcasting the data packets (DPs) in a licensed frequency, therefore preventing the SC from using that frequency. To address this issue, a novel machine learning classifier called nearest centroid (NC) classifier along with the Levenshtein distance function is used in this study as a means to classify between normal and MU, therefore contributing to spectrum detection accuracy at minimum time.

Motivation and contribution
In CRNs, PUEA and SSDF are a major concern. Recently, several research works have been developed for multi-hop CRN by using machine learning models. But, an inter-domain handoff communication and other emerging attack were not considered. In conventional methods, the SD and the percentage of error in prediction to avoid complexity are not solved. In this study, probable density logical-OR rule is used to detect the attack by using probability function; and the Levenshtein distance function among the test and training sample sensing reports is measured. If the distance between the test and training sample reports is minimum, this result is a robust classifier result.
In order to overcome the issues, a novel machine learning model called optimised sensing and Levenshtein nearest centroid classification (OS-LNCC) method for multi-hop CRN is introduced. In this study, combining probable density optimality for sensing and LNCC is proposed where the spectrum falsification and PU emulation are controlled by the fusion centre. The purpose is to adapt the probability for false alarm and detection of the CRN user to properly decide the spectrums exploited for communication. Next, by efficient means of classification between the malicious and normal users, the PU emu-lation is said to be reduced and therefore providing channel for SU. Accordingly, the OS-LNCC is planned to progress both throughput and percentage of error detection. The cross-layer defence using machine learning and queuing mechanism is carried out; the integration of decision-making functions was not considered.
The main contributions of OS-LNCC method are as follows: 1. At first, probable density optimal logical sensing model solve the spectrum sensing falsification difficulty in CRN by using STATLOG. An optimal decision about the spectrum sensing is made at the fusion centre with less delay. 2. In CRN node classification, node classification is performed according to NC based on Levenshtein distance which improves the throughput and reduce the error in prediction. By applying the Levenshtein distance function, distance between the test and training sample sensing reports is calculated. In addition, per-class centroids and prediction function are employed to categorise the normal user and MU. 3. The simulation experiment indicates that the cross-layer attack defence architecture based on novel machine learning can effectively reduce the SD and error in prediction to improve the throughput and obtain good network performance.
The rest of the study is given as follows. Section 2 offers a related work on CRN and mechanisms for defence against cross-layer attack. Section 3 gives the process of OS-LNCC method. Section 4 explains the simulation characteristics, results for the proposed OS-LNCC for multi-hop CRN. Finally, a conclusion is made in Section 5.

RELATED WORK
In recent years, CRN has been brought to the cutting edge owing to its views to handle the problem between constrained spectrum supply and demand. A new penalty mechanism based on the cognitive trust concentrating on security aspects was designed in [3]. However, with higher rate of resources being allocated and amount of interference exceeding, the threshold increase the cross-layer interference. To address this issue, a joint channel and power allocations were presented in [4] to reduce the interference. Despite minimising the interference, malicious activities are not said to be controlled. In [5], a crosslayer verification method was designed for cloud environment. An intelligent gateway model was designed in [6] to use licensed PU channels via CR for making reliable transmission to the healthcare monitoring. The main objective of wireless network remains in the efficient utilisation of the spectrum band. A survey of in-band full-duplex technology for CRN was presented in [7]. In [8], a comprehensive survey of CR research activities is studied. The major concerns in the CR communication networks were analysed and the related work in the literature was studied. Through the inception of several promising wireless devices, the exploitation of limited spectrum has risen in a significant manner in the recent few years. In [9], an in-depth discussion of channel assignment mechanisms including different functionalities for CRNs is presented. At first, an overview was given to both CRNs and wireless networks. Then, the several aspects and limitations in the channel assignment were also discussed. In this context, neural network and multi-resolution analysis [10] were integrated to improve spectral efficiency and data rate of CRs. An overview of security threats and challenges concerning CRN was proposed in [11]. The impact of SU node transmission on channel capacity scaling factor based on maximum concurrent multi commodity flow (MCMCF) was designed in [12] to ensure fairness. Yet another Bayesian learning approach for both detection and mitigation was presented in [13] to achieve desired trade-off between security and performance for taming cross-layer attack. Another spectrum mobility management model was proposed in [14] with the objective of improving spectral sensing accuracy and channel holding time. Detecting and defeating attacks in cross layer to increase network efficiency were presented in [15]. A double adaptive thresholding scheme to distinguish the legal and MUs using Dempster-Shafer (DS) evidence theory was proposed in [16].
Here, analysis between probability and false alarm detection was also made. An integrated scheme [17] using adaptive modulation with coding (AMC) was carried out in the physical layer and rate adaptation (RA) in the MAC layer. This resulted in offering optimised throughput. A secure routing based on the trust value was employed in [18] to lessen the attack in CRN. This was said to be achieved by monitoring the nodes' forwarding behaviours as malicious nodes were said to be identified based on the nodes' trust. With this design consideration, routing selection-based model was integrated with spectrum allocation. Finally, trusts were utilised to build obtained path trusts and delay for deciding suitable routing decisions. Attack in TCP and MAC layers was presented in [19] along with the defence mechanism based on round trip time (RTT) to improve throughput. A collaborative approach was designed in [20] utilising cluster analysis to achieve high spectrum accuracy. Channel hoppingbased defence method was discussed in [21] to lessen the PUEA in CRN. Different attacks that occurred in the physical layer were analysed and studied in [22]. The detection method and encounter measures were exploited to defend the attacked CRN. A novel scheme to defend against the PUE attacker was presented in [23] by using adaptive Bayesian learning automaton algorithm. A distributed method based on the uncoordinated frequency hopping technique was developed in [24] to mitigate the PUE attack. Trust-based multi-hop cooperative spectrum sensing (CSS) method was discussed in [25] to defend against SSDF attack. A novel hybrid optimisation was designed in [26] to detect the spectrum in CRN. In [27], new CSS method was employed with less delay. Also, soft CSS was described in [28] to perceive the occurrence of PUEA. It effectively lessens the error probability in CRN. In [29], CRN with CSS is applied to handle the issues of sensing, grouping and decision making. A generic Byzantine attack model was introduced in [30] to obtain the condition that constructs the fusion centre (FC) blind. A non-cooperative game-theory approach was presented in [31] for overcoming the Byzantine attack problem to understand an  [39] to efficiently mitigate the attack, but the designed system failed to consider the other emerging attack.
As explained above, many methods were addressing the issue of providing defence against attack in CRN. Critical issue in CRN is to plan a proper defence mechanism in the existence of attack, and capturing the physical and data link layer characteristics with single hop network. The SD ignores the percentage of error in prediction to avoid complexity and are not solved in conventional methods. Considering this issue converts the difficulty to multi-hop problem that needs a non-traditional customised solution method. In this study, a novel machine learning model called OS-LNCC method for multi-hop CRN is presented to solve this problem.

OS-LNCC FOR MULTI-HOP CRN
The open characteristics of cognitive radio networks (CRNs), a security vulnerability, is exploited with dissimilar types of attacks which is launched in the process of CSS, namely, PUEA and Byzantine attack called SSDF. PUEA is one of the most challenging attack types in the part of PU. In this section, cross-layer defence architecture against two different attacks, namely, PUEA and SSDF under multi-hop CRN using novel machine learning method is proposed. Figure 1 shows the block   Figure 1, a defence mechanism for cross-layer attack using novel machine learning model is presented. In this study, two layers, namely, physical layer (i.e. PUEA attack) and data link layer (i.e. SSDF attack) and the corresponding attacks in this layer are analysed and a defence architecture is presented accordingly. First, probable density optimal logical sensing model is designed to provide mechanism to mitigate against SSDF with higher throughput and minimum SD. Next, a robust algorithm called LNCC is designed for classification or emulation between normal and malicious CRN users (i.e. CRN nodes) made at the fusion centre, therefore contributing to robustness against PUEA attack. Thus, the normal users are identified and information shared across the network. To start with, the network model for multi-hop CRN is designed followed by the presentation of the elaborate description of the OS-LNCC method.

Network model
In this study, a multi-hop CRN consisting of m PUs (i.e. m Primary Users) and n SUs (i.e. Secondary Users) is taken into consideration and specific spectrum bands are assigned to PUs. On the other hand, the SUs do not possess any licensed or authorised channels and transmit DP in an opportunistic manner when identifying that the spectrum bands are not possessed by the PUs or the spectrum band is free. Each SU node n i possess a set of available channels which includes both data transmission channel (DTC ) and common control channel (CCC ). From the name itself, DTC refers to data transmission and is denoted as given below: From the above Equation (1), c 1 , c 2 , … , c m corresponds to the set of available channels. On the other hand, CCC is utilised for negotiation between the PUs and SUs. Therefore, at any instance, a link is said to be established between SU n i and n j , if the following condition prevails: With the above said assumptions as given in Equations (1) and (2), a network model with multi-hop CRN is designed and shown as follows. In the networking scenarios, with multi-hop CRN as basics, Figure 2 shows two centralised PU networks, where n ST transmitters SU-T with one FC are employed in the presence of two centralised PU transmitters PU-TP networks. On one hand,SU-T carry out spectrum sensing to discover PU-T condition. Then broadcast their corresponding sensing results to fusion centre FC .

3.2
Probable density optimal logical sensing SSDF attack [1] sends false observation causing DoS in a specific spectrum, therefore reducing the throughput. However, with falsified reports sent through the users, the performance of spectrum allocation is degraded. To address this issue, in this study, a probable density optimal logical sensing model is designed to reduce the falsification of spectrum reports by means of STATLOG. A mechanism is designed where a joint Let us further consider that the obtained signal is sampled at frequency f s . If PU-T is active, the obtained signal at the PU-T is represented as given below: From the above Equation (3), s(n) and (n) refer to the PUs signal and the noise factor, respectively, and is termed as H 1 . On the other hand, if PU-T is inactive, the received signal at PU-T is represented as given below: From the above Equation (4), (n) refers to the noise factor and is termed as H 0 . Here, mean and variance are represented as 0 and 0 for H 0 , and mean and variance are represented as 1 and 1 for H 1 . Then, the probability for false alarm Prob FA and probability for detection Prob D are represented as given below.
From the above Equations (5) and (6), Q and represent the Q-function for multi-hop CRN with mean and variance for both H 0 and H 1 , respectively. Finally, an optimal decision rule based on the probable density logical-OR rule is made: From the above Equations (7) and (8), Prob FA denotes the probability for false alarm and Prob D represents the probability for detection. Prob FA,i represents the users probability for Measure first hypothesis H 1 when the PU is active using Equation (3) 5: Measure second hypothesis H 2 when the PU is inactive using Equation (4) 6: Evaluate probability for false alarm using Equation (5) 7: Evaluate probability for detection using Equation (6) 8: Obtain optimal decision rule with probable density logical-OR rule using Equations (7) and (8)   9: Return (sensed spectrums)

10:
End for

11: End
false alarm, Prob D,i denotes the users probability for detection. Probability of Prob FA,i and Prob D,i are computed based on the departure rate dr to improve the attack detection rate.
In the probable density logical-OR rule, when one of the decisions decides that there is a PU, then the final decision decides that there is a PU along with the departure rate dr of With this, an optimal decision is made by the fusion centre thereby mitigating spectrum falsification. The pseudo code representation of probable density optimal logical sensing is given below.
With the design consideration of multi-hop CRN using probable density optimal logical sensing algorithm, two different objectives, that is, minimising the SD and maximising the throughput are said to be attained. To start with, in this algorithm, to design cross-layer defence architecture with the objective of minimising the falsification of spectrum in multi-hop CRN, two hypotheses with PU active and inactive model are measured. At first, departure rate of PU − − T is initialised. Next, PU and SUs with available channel, the two hypotheses are measured for each CRN node. When the PU is in active state, the first hypothesis is measured. The second hypothesis is measured when the PU is in an inactive state. In the next step, the probability of false alarm rate is calculated, then probability of detection is computed. The probability of false alarm rate and detection is computed based on the departure rate of PU. In the last step, the probable density logical-OR rule is employed to acquire the optimal decision rule. Here, an optimal decision regarding the spectrum sensing is made at the fusion centre with minimum delay, therefore reducing the SSDF attack with an improved throughput with less complexity. With this, the information is freely transmitted anywhere in the network without any data loss.

LNCC model
In a PUEA, an attacker [1] poses as a PU for executing the broadcasting in licensed spectrum bands that evades other SUs to exploit services. With proper classification mechanism, normal and MU can be differentiated so that only normal user makes transmission in a licensed frequency, making room for SU also. The proposed study employs Levenshtein distance calculation to represent the features of user signal which is then fed into the NC classifier for classification and is called as the LNCC model. In this study, the LNCC model is initiated through determining the energy for predicting the users on the spectrum band. Then, the energy received at the ith CR user at the kth sensing area is written as given below: From the above Equation (9) E ik ( j ) represents the energy sample received by the kth sensing area for the corresponding ith CR user with N representing the total number of samples. With the above energy detection function given in Equation (9), this study employs NC, an LNCC, that categorises the observations into diverse classes by means of quantitative variables. NC classifies a test instance, in LNCC, the current sensing report as obtained from Equation (9) into one of several NC classes by majority voting. The vote is then updated for estimating the distance among sensing results. As far as CRN is concerned, it is highly impossible to say which results are more similar. So, similarity between the two sensing reports has to be obtained. The classification surface is split into diverse centroids and the distance of the present sensing result to each of the centroids is identified.
Let Dis (a i , C ) represent the distance where C denotes the centroids given by C ∈ C 1 , C 2 , … , C n . The distance is measured to each of the centroids representing either H 0 or H 1 . Based on the measured distance, the current sensing result is categorised either to H 0 or H 1 . First, the distance calculation is made by using the Levenshtein distance and then NC classifier is employed for classification. The Levenshtein distance computes exact distance within vectors. The collaborative virtual sensing used by [2] differs from each other due to the deadline. So, in this study, focus is made on evaluating the similarity among sensing results through the Levenshtein distance where it analyses the distance of current sensing result and the sensing classes. The Levenshtein distance between sensing results u and v (length |u|and |v|) is mathematically expressed as given below: From the above Equation (10), 1 u p ≠v q corresponds to the function equivalent to 0 when u p = v q and equivalent to 1. On the other hand, Lev uv (|p||q|) refers to the distance between the first p sensing areas of u and first q sensing areas of v. As explained above, a sensing result has p elements according to u and q elements belong to v, respectively. Levenshtein distance function between the test and training sample sensing reports is computed. If the distance is lower, this result is robust classifier result. Otherwise, the result is not robust. Then, for example, the sensing report a p has to be categorised as one of the sensing classes that considered centroids to thea p . The candidate set of centroids for a p is denoted as N (a p ) such that N (a p ) ∈ C 1 , C 2 , … , C n .
The current sensing result is evaluated in conjunction through all the member of each of the sensing classes according to the N (a p ). The membership function is denoted as Ma p (l ). Each element of Ma p (l ) represents the result of comparing a p with the qth member of lth sensing class that are obtained using Equation (10). Let h l 0 be the event with sensing result a p belonging to class l and h l 1 be the event with sensing result a p not based on the class l . In addition, let EV l w be the event where the w elements in Ma p (l ) are higher than a threshold. After that, the per-class centroids with the current sensing result a p related to the class l is mathematically expressed as given below: From the above Equation (11), a p refers to the set of indices of samples belonging to class l ∈ B. According to the predicted function, the local decision for the pth CR user for lth class represented by S pl is given below: According to the predicted function as provided in Equation (12), the activity of PUs are observed and the attackers detected so that the SUs use the licensed spectrum whenever idle. The pseudo code representation of node behaviour monitoring using LNCC is given below.
A robust algorithm against PUEA using LNCC is presented above. In the algorithm as stated above, two main things are carried out. To start with, in the local decision-making process, the sensed spectrum based on the STATLOG function is first initialised to discover the energy for predicting the users on the spectrum band. In next step, energy sample received at the ith CR user at the kth sensing area is analysed to classify between the normal and MUs, energy sample received is analysed. Levenshtein distance between sensing results u and v is measured. Subsequently, Levenshtein distance function among the test and training sample sensing reports is measured in next step. If the distance between the test and training sample sensing reports is minimum, this result is robust classifier result. Otherwise, the result is not classified result. Finally, actual classification process is carried out using per-class centroids and prediction function. Here, actual decision is given to the FC , and the CR user (i.e. PU

EXPERIMENTAL SETUP
The proposed OS-LNCC is compared with existing methods [1,2,5,16]. To evaluate OS-LNCC, network simulator (NS2) is used. The metric used for this evaluation is SD, percentage of error in prediction and throughput. The evaluation is conducted on a network size of 1000 *1000 m area with the packet size set as 2500 bytes. Number of packets broadcasted per session is set to 500. The total available spectrum is set to be 54-72 MHz. The bandwidth usable by CRs is constrained to be 2, 4 and 6 MHz [2]. The bandwidth of the CCC is set as 2 MHz. The simulation results are analysed by using probable density optimal logical sensing and LNCC models in OS-LNCC. Initially a visual interface for network topology is presented. Then, the sensing model is employed to offer the optimal decision about the spectrum sensing. This aids to lessen the SSDF attack. In addition, LNCC model is applied to observe the behaviour of PU and thus detect the MUs or attackers in CRN. This is performed with the help of computing Levenshtein distance function where the distance between test and training sample sensing reports is calculated. Therefore, OS-LNCC alleviates attack against PUEA and SSDF in CRN.

Performance evaluation of SD
A multi-antenna BS was introduced [40] to improve the secure transmission. In the BS, the imperfect channel state information (CSI) and statistical CSI of the link among the BS and satellite user are presented. A multi-antenna BS with two beamforming (BF) algorithms such as hybrid zero-forcing (HZF) and partial zero-forcing (PZF) was designed for statistical CSI cases for  overcoming the optimisation issue. SD refers to the time delay occurrence in the channel sensing phase. In other words, SD is referring the number of channels sensed through an SU prior to data broadcasting: From the above Equation (13), SD is measured based on the time occurrence in the channel sensing phase, that is, Prob FA , Prob D and the number of CRN nodes N . It is measured in terms of milliseconds (ms). Table 1 given below shows the SD values obtained by applying the Equation (13) for five different methods such as OS-LNCC, existing methods in [1, 2, 5 16]. Figure 4 given above shows the graphical representation of SD using OS-LNCC, existing methods in [1, 2, 5 16]. Ten different simulation runs were conducted and similar number of CRN nodes in the range of 50 to 500 for fair comparison. Here, the CRN nodes refer to the inclusion of PU and SU in CRN. From the figure, it is illustrative that the SD is in the increasing trend for all the five methods. With 50 number of CRN nodes considered for experimentation, the time consumed in channel sensing phase using OS-LNCC was found to be 1.  [16]. From this, it is inferred that the SD was identified to be lesser using OS-LNCC method when compared to [1, 2, 5 16]. The existing method in [1] was used to detect the attack. But, it failed to provide better SD results. This is because of the incorporation of probable density optimal logical sensing algorithm. By applying this algorithm, the SD is even found to be better in the presence of multi-hop CRN. This is because of identifying two hypotheses, first, by applying the STALOG function, and then measuring the false alarm and detection by means of Logical OR rule. With this, the SD is said to be reduced using OS-LNCC method by 20% compared to [1], 52% compared to [2], 55% compared to [5] and 59% compared to [16].

Performance evaluation of percentage of error in prediction
Cross-layer attack first refers to the attack occurring at more than one layer. In OS-LNCC, the attack occurring in the physical layer called PUEA and attack occurring at the data link layer called SSDF are considered and defence mechanism is provided accordingly. During the provisioning of defence mechanism, certain amount of wrong predictions are also said to have taken place. Hence, the percentage of wrong detection is measured as given below: From the above Equation (14), the percentage of error in prediction (% of Err p ) is measured based on the percentage ratio of samples of CRN nodes provided as input N and the wrong predictions made Wrong p by the network. It is measured in terms of percentage (%). Table 2 given below shows the percentage of error detection values arrived at by applying Equation (14) for five different methods, OS-LNCC, existing methods in [1, 2, 5 16]. Figure 5 given above shows the graphical representation of percentage (%) of error detection with respect to different number of CRN nodes considered at different time intervals and different sessions. As already mentioned that the CRN nodes used in the simulation refers to both the PU and SU in the CRN network, the percentage of error detection measured also differs and hence not said to be proportionately increasing. However, as the number of CRN node increases, percentage of error detection also increased. But it is found to be comparatively lesser when applied with OS-LNCC when compared to the existing methods [1, 2, 5 16]. This is evident from the simulation. When 50 number of nodes was considered for simulation, the wrong predictions made during the overall defence architecture using OS-LNCC was found to be 4, 7 using [1], 11 using [2], 13 using [5] and 19 using [16]. The existing method [1] was used to find the bounds for prediction errors with Chernoff 's bounds. But, the error rate was not minimised. The reason is because of the application of probable density logical-OR rule, that also considers the departure rate dr of PU with which the final decision is taken by the fusion centre regarding spectrum falsification. In addition, by using Levenshtein distance between two sensing reports, classification of PU and MU are made significantly. With this the percentage of error prediction is reduced using OS-LNCC method by 31% compared to [1], 46% compared to [2], 51% compared to [5] and 56% compared to [16].

Performance evaluation of throughput
Finally, the efficiency of the cross-layer defence architecture is measured based on the throughput rate. The throughput of CRNs is closely related to the spectrum sensing frame structure. Spectrum sensing frame structure is separated into local sensing time, cooperative time, and transmission time. The throughput of the CRN is depending on sensing method and it is offered as follows.
From the above Equation (15), the throughput rate T is measured based on the data rate DR of each cognitive user, U unused channels as perceived by the SU, Time CP corresponding to the time period of the competing phase and Time TS referring to the time duration of the entire time slot. It is estimated in the unit of bits per second. Table 3 given below shows the throughput rate by applying Equation (15)   Graphical representation of throughput Figure 6 given above shows the graphical representation of throughput with respect to 500 different CRN nodes. In throughput measure, time period of the slot is taken as 0.050 ms and the time period of the contending phase for proposed method is 0.035 ms. The PUs channel utilisation factor is considered as 4 for all the five methods. The data rate of each channel is found to 8 ms for the proposed method. Hence, the throughput is observed as 22.4 bps for OS-LNCC method, 16.8 bps for [1], 14 bps for [2], 12.01 bps for [5] and 10 bps for [16]. Here, throughput is referred as the total data rate supposed by the CRN. From the figure, it is precise with wrong detection. The users in the CRN obtain a false notion that acquiring larger throughput rate. The existing method [2] was applied to higher resource utilisation. But, the method does not maximise the effective throughput. From the simulation values provided above, it proves that throughput estimation without considering optimal sensing and robust classification results offers a false result of performance. Evaluating the throughput rate of stateof-the-art methods, [1, 2, 5 16], the throughput is improved in OS-LNCC method compared to [1, 2, 5 16] as OS-LNCC FIGURE 7 Graphical representation of computational complexity method makes sure that a primary channel is sensed by only one SU and not by the MU by effective classification based on the LNCC. With this, the throughput rate using OS-LNCC method is found to be better by 21% compared to [1], 38% compared to [2], 51% compared to [5] and 68% compared to [16].

Computational complexity
The amount of time taken to detect the MUs from the normal CRN nodes is referred as computational complexity. Computational complexity is calculated in terms of milliseconds (ms) with the mathematical formula shown in below.
Computational Complexity = Number of CRN node * Time (16) From Equation (16), computational complexity is measured. Table 4 given below shows the computational complexity values arrived at by using the Equation (16) for five different methods, OS-LNCC, existing methods in [1, 2, 5 16]. Figure 7 given above presents the graphical representation of computational complexity with respect to 500 different CRN The results reported above confirm that with the increase in the time, computational complexity also increases. It minimises the time taken for the MUs. Hence, proposed OS-LNCC method attains reduced computation complexity by 30% compared to [1], and 18% compared to [2], 17% compared to [5] and 20% compared to [16].

CONCLUSION
In this study, a novel OS-LNCC for multi-hop CRN is proposed as defence architecture for CRN with probable density optimal logical sensing model and LNCC are examined. Sensing a channel in the existence of multi-hop is examined to influence the throughput of the CRN. The algorithm for mitigating the spectrum falsification by means of STATLOG function along with the algorithm for significant classification between the CRN normal and MU by means of novel machine learning for reducing PUE is designed and thereby throughput of CRN is enhanced. In addition, the effectiveness of OS-LNCC method is analysed in terms of SD, percentage of error prediction and throughput. The simulation outcomes are clearly evident that the performance of OS-LNCC compared with conventional methods. Thus, OS-LNCC method enhances the percentage of error prediction, throughput, which increases the defencemechanism of CRN with minimum SD.
In this study, only the probability of attack detection and false alarm is considered. In future, we can consider the other network parameters such as data rate and overall network throughput at the presence of attackers. In addition, a more efficient method for considering jointly SSDF and PUEA attacks in a CR is still a foremost issue needing to be addressed. SSDF attack still poses significant performance degradation in CRN. In future, we can use some network's parameters and classification methods to separate the MUs from the normal users to enhance the performance of the CRN.