A survey on security issues in cognitive radio based cooperative sensing

Cognitive radio based cooperative spectrum sensing (CSS) is severely affected when some secondary users maliciously attack it. Two attacks regarded as key adversaries to the success of CSS are spectrum sensing data falsiﬁcation (SSDF) and primary user emulation attack (PUEA). Defending SSDF and PUEAs has received signiﬁcant attention in research in the past decade globally. This paper performs a state-of-the-art comprehensive survey of the researches on defending SSDF and PUEAs. First, the preliminaries like Hypothesis testing for detecting the primary user and different models of CSS are discussed brieﬂy. Then a categorization of the defence mechanisms for defending both the attacks has been proposed as active and passive. Active mechanisms are suitable for an immediate defence in a limited time span, while passive mechanisms are suitable for ﬂexible CSS systems that are ready to detect the attacks over a period of time and suppress them permanently by bringing changes in their underlying operations. An in-depth tutorial on both the defence mechanisms is provided from the perspectives of the secondary users throughput and the interference to the primary user. Finally, a detailed survey on the open research problems in this area and some possible solutions has been performed.

by using CR technology for the unlicensed users. An unlicensed user was termed as a secondary user (SU). It was realized that a CR can be installed in an SU to sense the presence of the PU signal on its spectrum. This function was termed as spectrum sensing (SS). If the PU signal was found to be absent, its spectrum can be used by the SU. However, this usage was subject to a stringent condition that there should be no interference to the PU. If the PU signal is present, the SU should refrain from any kind of transmission.
Since [6] has been published, SS has emerged as a significant area of research [7,8]. The most important objective of SS has been identified as accuracy, so that the PU should not suffer interference. However, the SS performed by an individual SU can be highly inaccurate in deep fading environments. To overcome this problem, a cooperation between more than one SU to sense the spectrum has been proposed [9][10][11]. This scheme is known as cooperative spectrum sensing(CSS). CSS is classified as centralized and decentralized [12]. In centralized CSS, the cooperating SUs report their data or decisions to a central body known as a fusion centre (FC) or a base station (BS). Based on these data or decisions, the FC forms its decision about the presence or the absence of the PU and broadcasts it to all the SUs. On the other hand, decentralized CSS is based on the sharing of data or decisions among all the cooperating SUs to arrive at a final decision. The focus here is mainly on centralized CSS.
CSS significantly improves the SS performance in deep fading environments. However, its performance depends on the data or decision of an individual SU. Any flaw in an individual SU's data or decision can cause a significant error at the FC [13]. There can be various reasons for these flaws in the SU data or decision, like hardware malfunction, deep fading etc. A major reason for such a flaw is when an unlicensed transmitter misleads the SU ecosystem about the status of the PU spectrum for selfish reasons, by manipulating its air interfaces and transmitting fake signals in the environment. Such an unlicensed user is termed as a malicious user (MU) [11]. If the MU is a part of CSS, it can mislead the CSS system by deliberately reporting wrongly to the FC. Such security issues from MUs constitute one of the most significant concerns in CR systems [13]. The sensing capability enabled by the CRs gives way to the MUs to cause attacks. The openness of the lower layers adds advantage to the MUs. The MUs can sense the spectrum and transmit fake signals according to the spectrum status. If the honest SUs are misdirected and cease their transmissions, the MUs take the advantage and start using the spectrum.
There have been extensive researches on the possible malicious attacks [14][15][16][17][18]. These attacks have been named based on their strategies, like the receiver jamming attack, eavesdropping, authentication attacks, intruding attack, exogenous attack, and greedy attack. The receiver jamming attacks are caused when the MU affects the received signal-to-noise ratio (SNR) of the SU severely by transmitting absurd signals or noise over the channel. Eavesdropping is caused when the MUs overhear the SU and PU data and attack accordingly. Authentication attacks are caused when the MUs transmit fake signals to deceive the SUs, unlike jamming where the purpose of the MUs is to obstruct the channels of the SUs. Intruding attacks are caused when the MUs pose as legitimate SU nodes and mislead the system. Exogenous attacks are nearly similar to jamming attacks, while greedy attack can be seen as a combination of authentication and eavesdropping attacks.
The domain of attack strategies is reasonably broad. More specific attacks can be identified further and terminologies can be given to them based on their strategies. Broadly, most of these attacks can been classified under spectrum sensing data falsification (SSDF) attacks and the primary user emulation attacks (PUEAs). The attacks where the MUs misdirect the FC regarding the spectrum status by reporting to it wrongly are categorized as SSDF attacks. These attacks may cause the FC to declare the vacant spectrum as occupied or the occupied spectrum as vacant. On the other hand, the attacks where an MU misdirects other SUs and the FC regarding spectrum occupancy by transmitting emulated PU signals are categorized as PUEAs [19]. An MU causing PUEA is termed as a primary user emulator (PUE). Several researches propose counter mechanisms for these attacks. The objective of this paper is to perform a survey of the researches on SSDF attacks and PUEAs. Both the attacks have been studied for a CSS system, as they are more vulnerable to the attacks than individual CR systems. There are various counter mechanisms in the existing literature for these attacks [14][15][16][17][18]. We categorize the existing counter mechanisms as active and passive approaches. The active approach involves directly detecting the SSDF attacks and PUEAs with the help of some actions by the SU network towards the MUs. These actions involve detecting the location of the MUs by transmitting training signals, or installing hardware to detect the honesty and maliciousness of the incoming signals from the SUs by differentiating their strength, features or statistical properties. These approaches are advantageous where quickness is the primary need. However, the cost paid for quickness is temporariness of their impact. To counter SSDF attacks, they directly eliminate the malicious data reported by the attackers, by identifying anomalies in the statistical properties of the malicious data. They can eliminate the malicious data also by differentiating the location of its source from the locations of the PU and the SUs. Active schemes thwart the PUEAs by cancelling the PUE signal. They check the PUE signals for a tag already embedded in the PU signal. The PUE signals are eliminated if they do not possess the tag. The PUE signals are also identified based on the location of their source.
The passive approaches on the other hand resist the SSDF attacks and PUEAs with some changes in the underlying operations of CSS. These changes involve designing their spectrum sensing mechanism in a manner to mitigate the effects of the SSDF attacks on a particular parameter like the probability of false alarm or the probability of detection, and avoid the PUE signal. For defending SSDF attacks, they use the reputation of an SU based on this history of its reports and using game theoretic mechanism to maximize the utility parameter aimed at increasing the probability of detection. They counter the PUEAs by either suppressing the harmful effects on PUEAs or avoiding the PUE signals. The passive approaches do not counter the MUs head-on. However, they provide a more permanent defence to CSS than the active mechanisms. They are advantageous for flexible systems, which can easily redesign their architecture.
The motivation of this paper is to present a state-of-theart comprehensive survey of security aspects of the CR technology. Precisely, the most relevant security threats like SSDF and PUEAs, the related countermeasures like active and passive strategies, their working principles, and methodologies have been discussed in detail. Both the attacks and protection techniques apply to all three domains viz., commercial, public safety and military. The major motivations of the present work are as follows: • Discuss in detail the taxonomies of the SSDF and PUEAs, their mechanisms and the corresponding attack parameters, which determine where, who, how, and when these attacks are launched. • Propose a classification of existing defence algorithms to typical SSDF and PUEAs and provide an in-depth tutorial on the state-of-the-art defence schemes from the perspectives of throughput of the SU and interference to the PU, respectively. • Analyse the spear-and-shield relation between the SSDF and PUEAs and defence from an interactive location based, statistical parameter based, attack signal avoidance based, damage suppression based, game-theoretical based perspectives. • An analytical comprehensive survey of the current state-ofthe-art researches for the detection of the SSDF and PUEAs.
• Discuss the open research challenges, present some possible solutions, and present future research directions in these research areas.
The rest of the paper is organized as follows: Section II illustrates the fundamental system model for spectrum sensing, CSS, types of CSS, and how the security attacks affect CSS. The fundamental models for SSDF and PUEAs are also explained here. Section III performs a survey on the counter strategies against SSDF attacks. Section IV presents the counter strategies against the PUEAs. Various categorizations like active and passive mechanisms for thwarting SSDF attacks and PUEAs are done within the respective sections. Section V provides a detailed explanation on the open research directions and Section VI concludes the paper.

SSDF attacks
SSDF attacks are caused when the MUs misdirect the FC about the spectrum occupancy by reporting wrongly to it. When the PU spectrum is vacant, the MUs report to the FC that the PU spectrum is occupied. The FC broadcasts the honest SUs to cease their transmissions. The MUs may take the advantage of this situation by using the band for their own purpose. In the other scenario, the MUs may misdirect the FC to declare an occupied spectrum as unoccupied. Following this decision, the honest SUs may start transmitting which would cause a severe interference to the PU. The PU may impose penalty on the honest SUs and can also restrict them from further using its spectrum [20]. The MUs may take the advantage of this situation by getting the honest SUs punished.

PUEAs
PUEAs are caused when the PUEs deceive the honest SUs and the FC by transmitting emulated PU signals. PUEs are classified as always-on and smart [21]. Always-on PUEs do not know about the status of the licensed spectrum. They seamlessly transmit emulated PU signals on the PU spectrum band. On the other hand, smart PUEs are capable of knowing the PU spectrum status by individual decision making or listening to other SUs. They plan and attack depending on the presence of the PU. As a result of the PUEAs, the honest SUs may get misdirected in deciding the spectrum status and report wrong decisions to the FC. If the PU is absent, the SUs mistake the PUE's signal to be coming from the PU and cease their transmissions. The PUE takes the advantage of this situation by transmitting on the vacant band. On the other hand, when the PU is present, the PUEs cause interference to it. Smart PUEs cause interference deliberately while always on PUEs do it in anyway. As a result, the PU may penalize the honest SUs and the PUE may take advantage of the situation.

Mathematical models of SSDF Attacks and PUEAs
The mathematical models of SSDF attacks and PUEAs require the understanding of CSS. A CSS model has SUs placed randomly in the environment as shown in Figure 1. Each SU performs spectrum sensing on the frequency band allocated to the PU and reports its data or decision at the FC. The FC uses these data or decision values and makes a decision regarding spectrum occupancy. The decisions at individual SUs or at the FC are formed with the help of binary hypothesis testing problem.
Suppose an SU is lying in the vicinity of a PU base station. As the PU base station broadcasts wireless signals in the environment, the CR equipment installed in this SU receives an incoming signal sample and performs binary hypothesis testing to detect the presence of PU signal in it [6]. The process is explained as follows:

Hypothesis testing problem in CR
Let hypothesis H 0 denote the absence while H 1 denote the presence of the PU. The CRs make their decisions in favour of H 0 or H 1 based on the signal they receive from the PU base station. Let there be an SU in the vicinity of the PU and let y(k) denote the signal it receives from the PU at the kth instant of time. The presence of the PU can be expressed in the binary hypothesis form as where s(k) denotes the PU signal, h(k) denotes the multipath channel fading component and w(k) is the additive white Gaussian noise (AWGN) at the SU receiver. Now the CR installed in the SU uses an spectrum sensing mechanism to perform the binary hypothesis testing. There are many spectrum sensing techniques used for this purpose. The most famous primary user detection technique is energy detection. The focus in this survey is on the security attacks in energy detection based CR systems. Energy detection is a blind signal detection technique proposed in [22]. It does not need any knowledge of the characteristics of the signal. The CRs compute the energy of the incoming signal. Then, the computed energy is compared to a predefined threshold. If it is higher than the threshold, the decision is made in favour of H 1 otherwise H 0 is concluded. The calculation of the threshold is done using the Neyman-Pearson method [23]. The decision statistic in energy detection systems is the energy of y(k) given as where K is the number of samples considered for evaluating the energy. The SU compares T E in (2) to a predefined threshold E and decides H 1 or H 0 as follows The probability of false alarm of the FC (P f ) is the probability of an SU making a false decision when the PU signal is absent. This means that the FC thinks that the PU is transmitting when in reality, it is not transmitting. Mathematically, P f is written as The closed form expression for P f obtained after solving (11) is obtained as where is the signal to noise ratio of the SU with the PU given as = where A 2 is the PU signal power. In (5), Q m (., .) is the generalized Marcum Q-function given as (6) where I m−1 is the modified Bessel's function of (m − 1)th order. Similarly, the probability of misdetection of the FC (P m ) is the probability of the FC making error in detecting the presence of the PU signal. This means that the FC thinks that the PU is not transmitting, when in reality it is transmitting. P m can be expressed mathematically as The closed form expression for P m obtained after solving (7) is given as where Γ(., .) and Γ(.) are the incomplete and complete gamma functions respectively. Individual spectrum sensing can be inefficient due to presence of environmental obstacles. To overcome the limitation, cooperative spectrum sensing is carried out with the help of multiple SUs.

Cooperative spectrum sensing
The pathloss and multipath fading caused to the PU signal result into inaccurate sensing. Consequently, the SU can make a wrong decision about spectrum occupancy. This can cause a severe interference to the PU. As a counter measure, a cooperation among multiple SUs for sensing the spectrum has been proposed [24]. The process is called cooperative spectrum sensing (CSS). CSS is more prone to security attacks compared to individual spectrum sensing, as attacks can be launched more efficiently when attackers exist within the system. The main focus of this paper is on the effects of security issues on CSS and their counter mechanisms. CSS is necessitated for the CR sensing. The idea of CSS is to exploit the spatial diversity of the signal propagation environment. Exploiting spatial diversity was first proposed in refs. [25] and [26]. In CRs, diversity is obtained when different copies of the same signal are received by multiple antennas or multiple SUs [12] as a result of multipath diversity. The gain obtained by exploiting multipath diversity is called diversity gain. The idea of CSS is that out of the various copies of the PU signal received by different SUs, some copies would have suffered more attenuation while some copies would have suffered lesser attenuation. The SU receiving a highly attenuated signal is likely to make a wrong decision regarding spectrum occupancy. In this regard, if a decision is made by combining the individual decisions of the SUs, the bad decisions will be compensated by the decisions from the good paths, and the effects of attenuation can be mitigated. CSS significantly reduces the inaccuracy in spectrum sensing by a single user. Each SU shares its data or decisions with other SUs and forms its decision about the spectrum occupancy. It is mainly categorized as centralized and decentralized CSS. In centralized CSS, the SUs fuse their individual data or decisions to a central body known as the FC, as shown in Figure 1. In decentralized CSS, there is no FC [27]. It is based on sharing of data or decisions among the SUs as shown in Figure 2. The focus of this work is more on centralized CSS systems as they have higher application due to their well-structured architecture.
In data fusion, the individual data of each SU is fused at the FC. For instance, if energy detection mechanism is followed by the SUs, the T E value of each SU is transmitted to the FC. The FC adds all the incoming T E values and forms a new global statistic E G . The global statistic E G is compared to a threshold and a decision is formed by the FC on spectrum occupancy, which is known as the global decision. Let there be N SUs participating in CSS and T Ei denotes the energy value of the ith SU, the calculation of E G has been shown mathematically as The global statistic E G is compared to the global threshold value G and the decision on spectrum occupancy is obtained as The calculation of G is done with the help of Neyman-Pearson method. Similar to P f for individual SUs, the probability of false alarm of the FC (P fa ) is the probability of the FC making a false decision when the PU signal is absent. This means that the FC thinks that the PU is transmitting when in reality, it is not transmitting. Mathematically, P fa is written as Similarly, the probability of misdetection of the FC (P md ) is the probability of the FC making error in detecting the presence of the PU signal. This means that the FC thinks that the PU is not transmitting, when in reality it is transmitting. P md can be expressed mathematically as In decision fusion CSS, each SU transmits its binary decision to the FC in the form of 1s and 0s, where 1 indicates the presence and 0 indicates the absence of the PU. The FC follows AND, OR or the majority rule to form its decision regarding spectrum occupancy, where the PU presence is decided if N , 1, or N ∕2 SUs transmit 1 to the FC respectively. The formulation of P fa and P md is similar to (11) and (12), with E G replaced by the sum of the binary decisions from the SUs [12].
where n = 1 for OR, n = N for AND, and n = N ∕2 for the majority rule.
P md can be expressed mathematically as

Selection of the FC
The selection of the FC is based on the principle of trusted nodes in the network. We consider a mechanism where one of the cooperating SUs are equipped to function as the FC. The choice of that SU is based on the merit of its decisions. In this regard, a trusted cooperating SU which has a history of reporting correct decisions is chosen. At any instant of time, when the decision of a cooperating SU is consistent with the global network decision, its reputation value will be increased. On the other hand, if it is not consistent with the global network decision, it will be decreased. The reputation value associated with any network is calculated as where d (k) is the value representing the global decision. If an SU is repeatedly reporting wrong values, it is discarded. The decision to discard an SU is decided by a discard threshold value . The ith SU is discarded for repeatedly making wrong decisions if r i becomes lesser than . However, in fading environments, the possibility of an SU receiving attenuated information from the PU spectrum is high. Thus it may fail to detect the PU spectrum correctly which will affect its merit which is based on the history of its decisions. In this regard, a threshold Δ is additionally used to prevent the environment affected SUs. The reputation value of an SU is lowered only if it reaches a value lesser than . The node which has the highest reputation value is assigned as the FC. Generally devices like the server, the controlling unit, or the SU base station are able to maintain such a high reputation value. Thus, these devices are made the FC, where are the SUs report their individual data or decisions.

2.5.2
Election of spectrum head with the help of the cooperative spectrum sensing and transmission energy levels After making the decision on the availability of the spectrum for SUs, the next step for the FC is allocation of the spectrum. The part of the spectrum used by the PU is wide and can be shared by multiple SUs. The FC divides the PU spectrum into various sub-carriers of different bandwidths. One cooperating SU is allotted one sub-carrier each. The significant issue is what strategy to follow the sub-carrier allocation. The objective of the FC is to allot the sub-carriers to the SUs in a manner that the achievable sum-rate of the SU system is maximized. In this regard, the transmission energy levels of the SUs becomes significant, as the data-rate of the SU is a function of the transmission energy level. Also the correct detection of the available subcarrier is crucial for the maximization of the achievable sum-rate of the SU system. Thus, efficient usage of the energy levels of CSS plays a significant role in spectrum allocation.
To explain this mathematically, we consider a comprehensive scenario of n 1 = 1, 2, … , N 1 networks available where CSS is being performed. Let the kth sub-carrier be assigned to the mth SU under the coverage of sth base station within the n 1 th network. The variable x n 1 smk indicates the happening of this situation by assuming a value 1 and when it assumes a value 0, it indicates that this event has not happened. The data transmission rate achieved by the mth SU while performing its transmission through the sub-carrier k within the network n and base station s is given as for n 1 = 1, 2, … , N 1 , s ∈  n , m ∈  ns , k ∈  ns (17) where D FC n 1 sk denotes the decision taken by the FC. D FC n 1 sk = 1 means that the sub-carrier k is detected as a spectrum hole (a vacant part of the spectrum) in the network n 1 base station s. The correctness of D FC n 1 smk largely depends on the energy levels of the CSS. Thus, the data rate of an SU is highly dependent on the energy levels of CSS as derived earlier in expression (9). Further, P n 1 smk represents the transmission energy level of the SUs.  n 1 s ,  n 1 s , and  n 1 are the sets of SUs, sub-carriers, and base stations within the network n 1 . The allotment of the kth sub-carrier to the mth SU is indicated by the variable x n 1 msk . The achievable sum-rate of the SU network is given as The allotment of the sub-carrier (or the frequency sub-bands) is done to the SUs by optimizing the transmission energy level of the SUs P n 1 smk and the sub-carrier allocation variable x n 1 msk as max x n 1 msk ,P n 1 smk R (19) subject to x n 1 smk ∈ 1 (22)

Effects of security attacks in CSS
Apparently in CSS, the final belief of an SU about the spectrum occupancy depends on the data or decisions of other SUs. The belief may be formed with the help of the FC or by the sharing of the data or decisions among the SUs. In either case, it depends on other SUs. As mentioned earlier, it is possible that some SU in the network indulges in malpractices to misguide the spectrum sensing system for its selfish interests and such an SU is termed as a malicious user (MU) [28]. The MUs severely affect the signal processing and propagation characteristics of the sensing system. Their presence necessitates the redesigning of the sensing strategy of the SUs. As a result, the throughput of the SUs is degraded and the interference on the PU increases severely. There are different ways in which the MUs can launch attacks. These attacks have been classified into two categories which have been widely studied in researches: spectrum sensing data falsification (SSDF) attacks and primary user emulation attacks (PUEAs). The following section describes the attack mechanism and mathematical models of SSDF and PUEAs, and how they affect CSS.

2.6.1
Mechanism of SSDF attacks These attacks are caused in centralized CSS systems, as shown in Figure 3. They are the outcome of the Byzantine failure problem which is caused when a cooperating SU reports wrongly to the FC [29]. This false reporting may be due a hardware fault or a selfish intention of an MU. The false reporting due to selfish intentions of the MUs is known as the SSDF attack [29,30]. The MUs report falsely to misguide the FC regarding the spectrum status. On getting misdirected, the FC broadcasts wrong decisions to the cooperating SUs. Hence, the entire network is misdirected regarding the spectrum status. If the FC gets misguided regarding the presence of the PU and broadcasts it as absent to the cooperating SUs, the SUs will start transmitting resulting into a strong interference to the PU. As a consequence, the PU may impose penalty on all the SUs. On the other hand, if the FC gets misdirected about the absence of the PU and broadcasts it as present, the SUs will cease their transmissions. Hence, their throughput will be severely affected. These attacks can be launched individually by one or more the MUs or the MUs can collude and cooperatively launch attacks [20,31,32]. Another consequence of these attacks is that the PU becomes reluctant to share its spectrum with the SUs. When the MUs misguide the CSS about the absence of the PU, they cause increase in the probability of false alarm (P fa ). The MUs cause a false increment in the value of E G which increases P fa and severely affects the SU throughput. On the other hand, when the MUs misguide the CSS about the presence of the PU, they lead to increase the probability of misdetection (P md ). Here, the MUs cause false decrements in the value of E G which lead to an increase in P md and interference to the PU. There are mainly two approaches to counter individual SSDF attacks in literature: active approaches and passive approaches [14,15,[33][34][35][36][37][38]. The major active approaches in literature are location identification based and statistical approaches. The major passive approaches are reputation based mechanisms, game theoretic, and damage suppression based approaches.
This paper uses an audit set mechanism to identify the SSDF attacks. As the sensing reports collected at the FC in the form of temporal data forms a rich audi data set, it can be analysed to identify malicious nodes. Using the audit data set to identify the malicious attacks has been widely used in the literature [109][110][111][112]. This mechanism fundamentally exploits the richness of the information available in the audit data set. The rich audit data set mechanism is based on the assumption that the number of malicious nodes is less than the total number of nodes. However, some existing works do not follow this assumption [113,114]. The idea behind this mechanism is that the sensing reports which get accumulated at the FC reflect the temporal behaviour of the SUs. When the SUs transmit their data to the FC, the sensing reports start getting accumulated at the FC at different time slots. To use the richness of the audit data set, first a set of random variables (RVs) X 1 , X 2 , … , X is formed. Each of these RVs represent the frequency of occurrence of different events. Let us define an event as (g, l i ) where g denotes the global decision of the FC and l i denotes the local decision of the ith SU, then there are four possible events for the pair of each SU and FC (00,01,10,11). The four events here will be expressed as X 1 , X 2 , X 3 , X 4 As mentioned earlier, the SSDF attacks are caused when an MU inverts its true decision on the spectrum to make the FC make a wrong decision. Suppose, the SU senses that the spectrum is vacant, so it inverts its decision and reports to the FC that the spectrum is occupied. This is done by sending a high energy value or sending a wrong decision to the FC. For these types of attacks, the interest of the attack detector will be in detecting the event X 3 . Let the SU invert its value with a probability , then the frequency of occurrence of the event X 3 , denoted as X n is expressed as The first term in (25) represents the frequency of event type 10 for the case when the PU signal is present. Multiplying h denotes honest SUs and the term p m shows that the event 1 is due to the misdetection by honest SUs. On the other hand, h and 1 − p m show that 10 happens even when there is no misdetection by the honest SUs. The terms (1 − h) and p m show that the event 10 happens due to the MUs wile misdetection is happening while and (1 − h) happen together show that 10 is happening because the MUs are inverting their decisions with probability . On the other hand for detecting the SSDF attacks which make the FC change its decision to 0 when the spectrum is actually occupied, the event type 01 is to be singled out. The frequency of these attacks is expressed as Taking advantage of the rich audit data set, the authors in [109] present two anomaly detection scheme for detecting the malicious nodes. The first scheme exploits the frequency property of such statistical data and develops a lightweight intrusion detection scheme. The second scheme uses Markov chain model based on the ordering property of the data to develop another intrusion detection scheme. The audit data set has been used to identify the malicious attacks in [111,112]. After identifying the malicious attacks, the improvement in the system security performance is studied with the help of audit bits. Further in [110] Wu et al. use the audit sets to perform a comprehensive investigation into attack cost and attack benefit from the malicious perspective to identify the tendency of an SU for becoming an attacker. In [115], Fu et al. develop a Bayesian-inferencebased sliding window trust model to identity the SSDF attacks without any prior knowledge about the attackers. The attacker is decided on the basis of a trust model, where a trust value is assigned to each of the SUs.

2.6.2
Mechanism of PUEAs These attacks are caused when an MU deliberately transmits emulated PU signals [39], as shown in Figure 4. As a result of this, the honest SUs sense the PU to be present and cease their transmission. MU takes advantage of this situation and starts transmitting on the licensed band. Such attacks are known as primary user emulation attacks (PUEAs) and the MUs causing them are called primary user emulators (PUEs).
PUEAs are more damaging than SSDF attacks, so much that they change the basic hypothesis testing model in (1). PUEs have been classified as always-on and smart. Always-on PUEs always transmit the emulated signals while smart PUEs transmit the emulated signals only when the PU is not transmitting. The hypothesis testing model in (1) for PUEAs becomes where h p (k) and h e (k) denote the channel fading component from the PU transmitter and the PUE transmitter respectively. s(k) is the PU signal which is emulated by the PUE. For smart PUEAs, the hypothesis testing model is modified as respectively. Hypotheses H 00 and H 11 denote the presence of neither PU signal nor PUE signal and the presence of both the signals respectively, while hypotheses H 01 and H 10 denote the presence of PUE signal but not the PU signal, presence of PU signal but not the PUE signal respectively. The mathematical definitions for P fa and P md for the above hypothesis testing problems in (27) and (28) remain the same as defined in (11) and (12). However, the closed form expressions obtained after solving (11) and (12) for an honest SU system in (1) change completely when PUE is present. PUEAs severely affect the throughput of the SUs. They also increase interference to the PU. We categorize the counter mechanisms to PUEAs also as active and passive mechanisms. Active mechanisms aim at cancelling or suppressing the PUE signal by differentiating it from the PU signal. This differentiation is facilitated by differentiating the received signal strength (RSS), some features, location of the PU and the PUE [39][40][41]. The second approach is passive where instead of detecting the PUE signal, a specific damage done by the PUE on the SUs is suppressed [21,[42][43][44]. The counter strategies proposed in the existing literature to counter both the attacks have been detailed as follows.
For detecting the PUEs, a Strong Stackelberg equilibrium (SSE) based surveillance strategy has been followed [116]. The network manager performs the surveillance of the PU spectrum. The attacker is assumed to be rational enough to adapt to the surveillance strategy. The task of the network manager is to observe the spectrum and decide whether the signal occupying it is from the PU or from the PUE. A multi-channel is considered for surveillance. A game theoretic approach is formulated where the attacker and the network monitor are designated as the players. The attacker emulates the PUE signals while the network monitor, regarded as the defender, monitors the spectrum. The strategy set of the attacker over K 1 bands is denoted by Similarly, the strategy set of the defender is written as Based on these actions, the attacker and the defender receives certain pay-offs. Based on the pay-offs obtained by the defender, the primary user emulation attacks are detected. The variables that enable the calculation of the pay-offs are illustrated as follows: Attacker A is the benefit of using the channel for any CR user at one data frame.
• P t A is the penalty value for being captured by the defender.
For the defender • C t S is the cost for implementing the surveillance process. • G t S is the benefit for detecting the attack.
The parameters C t A , G t A , P t A , C t S and G t S help in the formulation of the pay-off. After performing rigorous calculations, pay offs are obtained in the form of these parameters for the various conditions on the actions pf the PU and the PUE. For example if the defender is performing surveillance when the attacker is attacking when the PU is not transmitting, the defender receives a reward −C t S + G t S . This reward is received by the defender only when it is correctly able to detect the attack. Similarly when the defender is performing surveillance when the attacker is attacking while the PU is transmitting, the reward obtained by the defender on correct detection is −C t A . Similarly the attacks are detected in the other scenarios.
Further in [117], the mechanism has been extended to detecting the PUEAs with the help of the rewards obtained by the SUs. It is assumed that a penalty is imposed on the SUs when the PU suffers interference. The formulation of the reward obtained by an SU is given as The expected reward for the SUs is obtained as where represents the discount factor (0 < ≤ 1). The status of an SU is represented with a Markov chain model and the presence or the absence of the PUEA is determined. The utilities and payoffs defined can be used to identify PUEAs by identify-ing their effects the SU throughput under a Rayleigh flat fading channel [118] Karimi et al. also derive a test statistic based on the generalized likelihood ratio test to differentiate the PUEs from the actual PU. An investigation on the damages caused by the smartness of the PUEAs has also been conducted and significant damage is found. Rathee et al. in [119] use the payoffs for the network manager, the PU and the SUs to propose a secure hand-off mechanism for identifying the PUEAs. It is achieved by introducing a coordinating CR user which computes the level of trust of each SU based on its behavioural characteristics. This method is further improved with the help of artificial neural network (ANN) in [120].

COUNTER STRATEGIES AGAINST SSDF ATTACKS
As discussed earlier, the SSDF attacks are caused when one or more MUs transmit fake signals to misdirect the SUs performing CSS [11,34]. These attacks can be caused by a single MU, multiple MUs independently, or multiple MUs in collusion. As mentioned earlier, the counter mechanisms for these attacks have been categorized as active and passive approaches. Active approaches include location identification based and statistical approaches while passive approaches include reputation based mechanisms, game theoretic, and damage suppression based approaches. Location based approaches counter the SSDF attacks by identifying the locations of the PU, the SUs, or MU. Statistical approaches differentiate the statistical properties of the MUs' data with the honest SUs' data. The statistical tests are based on outlier rejection from a sample of data. An outlier is a value in a data-set having different statistical properties. The incoming data to the FC from the cooperating SUs follow a specific statistical distribution. A random low or high value transmitted by an MU belongs to a different distribution and behaves as an outlier. Using statistical tests, these outliers can be identified and discarded. Different statistical tests have been reported in the literature [45,46]. Reputation based mechanisms depend on the previous history of the SU node. The reporting of the SUs is compared with a trusted SU. Based on the previous history of the reported decisions, the SU nodes are assigned weight values. The SU nodes repeatedly giving values similar to the values of the trusted node are weighed high. The main challenge in these mechanisms is to identify the trusted SU. In game theoretic approaches, the SUs, the PU, and the FC are regarded as players. Their actions are termed as strategies and the outcome of their actions are termed as utilities or payoffs. To counter the security threats with game theoretic approaches, the benefits obtained by an SU in reporting honestly or maliciously are termed as honest and malicious utilities. The objective in these approaches is to maximize the honest utilities and minimize the malicious utilities, so that the SUs do not turn malicious. In damage suppression, instead of detecting the MUs, a particular damage of the SSDF attacks is identified and is minimized. The damage can be caused on the PU or the SUs. The various mechanisms to counter the active attacks in the research literature are explained next. Avoid the PUE signal or its interference by making appropriate changes in it's own sensing mechanism.

Active methods
Active methods aim at directly eliminating the malicious readings or detecting the MUs. The prominent active methods are location identification based approaches and statistical approaches. We brief about the existing works in each of these approaches as follows:

Location identification based approaches
Location identification based mechanisms counter the SSDF attacks by identifying the locations of the PU, the SU or MU. The idea of using localization based approaches for countering MUs can be attributed to the Ma et al. in [14], where the localization of PUs has been carried out. They have aimed to locate the PU which helps in reducing interference to it and in detecting it. It can be seen from (7) that P m is a function of channel gain h. The channel gain h varies inversely with the square of the distance due to the pass loss exponent. Thus, P m of an SU is a function of its distance from the PU. The authors in [14] use the knowledge of P m s to detect the distances l 1, l 2 and l 3 in Figure 5 and formulate a weighted least squares mechanism and an iterative mechanism to design an algorithm to localize the PU. The PU lies at the intersection of the circles drawn with each SU at the centre and its distance from the PU taken as the radius. The objective of the proposed algorithm is to identify the intersection point. Let t = t 1 t 2 … t N , 0 ≤ t n ≤ T, n = 0, 1, 2, … , N be the vector signifying the time instants, when the PU signal is present among the T instants of sensing time and = [ x y ] T be the mean estimated position of the PU, which is the expected value of the estimated position of the PÛ. Suppose, p(t, ) is the joint probability of t and . The Cramer-Rao lower bound (CRLB) of the mean square error (MSE) in the estimation of position vector is given as where where q(t, ) = ln(p(t, )). Further this method is modified for detecting MUs. Ma et al. also design location aware mechanism access control protocol and show significant throughput gains. There are two kinds of algorithms for localization: range based algorithms and range free algorithms. The range free algorithms are less accurate but simple to implement, whereas range based algorithms are accurate but complicated to implement. Ma et al. have obtained a simple to implement and accurate algorithm and term it as semi-range algorithm. They have mainly focused on how MUs are affecting the localization of the PU. The above concept can be used by the MUs can be used to geo-locate the position of the SUs from their sensing reports [15]. This leakage of the location of the SUs is considered as the privacy leakage of the SUs. The defence against privacy leakage is carried out in two steps viz., the privacy preserving sensing report aggregation (PPSRA) and the distributed dummy report injection (DDRI) protocols to prevent the privacy leakage. These two steps are together named as privacy preserving CSS. The PPSRA uses cryptographic techniques with which the FC can get the final result from each SU without needing its calculated individual value. Subsequently, the DDRI algorithm provides differential location privacy for SUs. A single service provider (SP) has been considered as a central entity serving a group of SUs.
When multiple SPs are considered, they are used for performing CSS. Each SP has the perfect knowledge of the group of SUs it is serving. In [31], the attackers constitute of the SUs and the SPs. The information of the SUs is shared with every entity in the network. The malicious SPs and SUs can misuse this information to geo-locate the honest SUs and compromise with their location privacy. To counter such privacy leakage, a privacy preservation framework termed as PrimCos is proposed. In this mechanism, before executing CSS, the original data is transferred to a privacy preserving form that pre-

Reputation based
A weighted decision rule is followed to implement CSS. The data from each SU is multiplied by a weight value.

Game theoretic approach
The SUs, the PU, and the FC are players. Their actions are strategies and the outcome of their actions are utilities or payoffs.
References [64][65][66][67][68] The malicious utility seen by an SU can go higher than the honest utility and can result into an inefficient MU detection.

Damage suppression
A particular damage caused by the MU to a specific entity of the SU network is detected and is reduced References [69][70][71][72][73] Heavily depend on the estimation of the channel gains between the PU transmitter and the attacker and the SU transmitter and the attacker.
serves the statistical information. Additionally, the SPs share perturbed sensing data by adding a random noise to it. The perturbation hides the exactness of the data and prevents its leakage. However, the reconstruction of the perturbed data is challenging and remains an open research problem. The complexity of this system is another challenge, as sensing is a time bound process. The proposed mechanism gives the SUs sufficient protection for privacy and incentivizes them to share their data. Nevertheless, the accuracy in the localization of the PUs is severely affected, particularly when they are mobile. In [32], such effects are studied for a system facing SSDF attacks by tracking the mobile PUs. The accuracy in the localization of PU is severely affected under the attacks. The objective is to suppress the attacks from the MUs by tracking the mobile PUs. However, it becomes challenging when MUs manipulate the data, because localization largely depends on the sensing results of the SUs. To face this challenge, the authors propose sequential Monte Carlo combined with shadow fading estimation (SOLID) for attack/fault tolerant mobile PUs. This mechanism integrates sequentially Monte Carlo estimation based target tracking with shadow fading estimation. The sequential MonteCarlo estimation based tracking uses the sensing reports of the SUs iteratively to track the location of the PU. The shadow fading estimation part of the mechanisms refers to estimating the shadow fading correlation in the sensing reports induced by the PU's mobility. When the shadow fading estimation is integrated with sequential Monte Carlo estimation based tracking, the effects of MUs on localizing the PU is signifi- References [17,18,41,[78][79][80][81][82][83][84][85][86][87][88][89] Successful implementation of this mechanism is complex as it involves the insertion of the tag in the PU signal. References [21,42,44,96,97] The suppressing harmful effects of PUEAs based defence mechanisms focus on a very narrow subset of damages. The other damages remain untouched or may even result in deteriorating them. Though a constraint may be put on the deteriorating parameters, but still these mechanisms fail to make a comprehensive defence strategy.
Avoiding the signals from the PUE These approaches aim to avoid the PUE signals, rather than detecting them or suppressing the harmful effects done by them. Generally they are based on wideband channels which facilitate hopping over different channels. The challenge addressed in these mechanisms for the SUs to identify the presence of the PUE signal on a particular channel without latency and hop on a different channel.
References [98][99][100][101][102][103][104][105][106] The mechanism of avoiding PUE signals is difficult to be implemented on a narrow channel. Generally they need hopping on multiple number of channels to avoid the PUE signal. They are also subject to a prior detection of the PUE signal.
cantly reduced. In [47], the authors aim to detect the honest SUs in a CSS rather than MUs. The authors aim to form a cluster with no MUs with a fast searching algorithm and propose a scheme based on clustering the honest SUs with a fast sensing algorithm. The proposed fast sensing algorithm groups the SUs into different clusters. The clusters transmit their decisions about the presence of the PU to the FC. From the decisions of various clusters, their distances from the PU is estimated. This estimated distance is compared to the actual distance and the corresponding error is calculated. If the error associated with a cluster is below a predefined threshold, it is regarded as honest with no malicious sensor. Such a cluster is categorized as a trusted cluster. Next, the FC decides the status of the PU channel only with the information shared by the trusted clusters.

Statistical test based approaches
The signals transmitted by MUs are likely to follow different distributions from that of the signals transmitted by the SUs. This difference in the statistical distribution can be exploited to detect the signals of the MUs. In statistical attack prevention mechanisms, the difference in the statistical properties of the honest SUs and the MUs is observed [33,34]. These attack prevention mechanisms were first introduced in [33] where the authors consider distributed detection in the presence of cooperative Byzantine attacks. These attacks are launched by Byzantine sensors which have been reprogrammed deliberately to send fictitious observations to the FC. To counter these attacks, the authors propose to obtain the optimal attacking distributions for the Byzantine sensors that minimizes the detection error exponent. Subsequently, the smallest error exponent is obtained that signifies the power of the attack. The smallest error exponent is further used to obtain the minimum fraction of Byzantine attackers that are capable of damaging successful PU detection at the FC. Using this idea for detecting MUs in CSS can be attributed to the work in [46]. As mentioned earlier, MUs transmit fake data to the FC which can be a constant or any other random value. Hence, the data from the MUs is likely to follow a distribution different from the SUs' data. The fundamental idea of [46] is outlier detection, where the data from MUs are treated as outliers [46]. The basic notion followed here to detect outliers is done on the basis of outlier factor o i , which is the difference with the mean of the data ( E ) divided by the standard deviation ( E ) of the data, given as If the o i associated with the ith data is greater than a threshold value, then that data value is regarded as an outlier. However, this method has a limitation that it is dependent on E and E , which are contaminated by the outliers. To overcome this limitation, a bi-variate estimate based outlier factor calculation is proposed. The bivariate estimate is based on assigning penalty values to each cooperating SU. It depends on the number of iterations where positive and negative changes in the data values of the SUs happen. The penalty assigned to the ith SU P i [k] is given as where S + [k] and S − [k] are the sets of iterations for which Δ̂[k] = Δ̂[k] − Δ̂[k − 1] is positive and negative respectively.
The above mechanism is subject to the challenging requirement of a priori knowledge of the environment, as the incoming SUs' data at the FC depends on the channel gain between the respective SU and the FC. This limitation can be overcome by modifying the statistical test to detecting the statistical differences between the CSS decision and the individual SU decisions. In [48], the authors propose an algorithm based on the Kruskal-Wallis test to detect the MUs without any a priori knowledge. This test is a non-parametric method for testing whether samples originate from the same distribution. The proposed algorithm uses the statistical differences between the cooperative decision and the individual SU decisions. The statistical mechanisms can use Bayesian approach if the FC has some knowledge about the strategy of the MUs. In [49], a Bayesian approach is followed to detect the MUs. The authors assume some knowledge about the strategy of MU's data to be available with the FC. Since, it is assumed that the MUs' strategy is known, the a posteriori probability of their attacks can be known with reverse engineering on some captured MUs and hence all MUs are detected subsequently. The Bayesian approach is an efficient mechanism to detect the MUs, however, depends on the a priori knowledge of the SUs. This dependence is overcome in [16], where it is shown that it is always possible to detect the MUs when they attack individually, even when the information on their strategy is not available. Attacking individually means that the MUs are not aware about the responses of other SUs. When the MUs know or can hear other SUs, a strategy has been formulated such that the SUs are never detected by them. Then, a strategy is formulated for detecting the MUs. The prior condition for this strategy is availability of the knowledge of P fa and P d . The condition for this strategy is to have the knowledge P fa and P d . The authors point that practically the strategies of the MUs' cannot be known a priori, particularly for a system when the MUs can hear other SUs and modify their attacks accordingly. Hence, the mechanisms proposed will fail to detect the MUs. To overcome the challenge of the unknown strategies of the MUs, a data mining technique to detect abnormalities has been proposed in [16]. This technique is fundamentally the application of the outlier detection method in (35)-(37) without the requirement of the a priori information on the MUs. Finally, the sensing reports of the SUs are kept in high dimension space and the possible abnormalities are detected. The idea of this work is to place the sensing reports of the SUs in a high dimension space and detect possible abnormalities.
CR systems are generally based on wideband sensing as PU is likely to hold license over a wide spectrum band. The study of security issues on widebands is thus significant. In [50], Srinu et al. study wide band sensing under SSDF attacks. The wide band is divided into multiple bands. They propose to predict the PU signal status on the multiple bands by integrating the entropy and cyclic properties of the received signal. The presence of MUs complicates accomplishing this proposal. The MUs are eliminated by designing by designing the generalized extreme statistical deviate and adjust boxplot methods by using uncertainty and autocorrelation properties of the signals. It is found that the proposed mechanism is efficient in eliminating the MUs. In [51], Tang et al. have aimed to save the CR-mobile ad hoc networks (MANETs) from security risks. The proposed scheme is a bio-inspired consensus algorithm. It takes the inspiration from self-organising behaviour of the birds, fishes, ants, honeybees or other animals in groups. Engineering algorithms have been formulated on the basis of the behaviour of animals in groups in the existing researches [52][53][54]. Studying the difference between the communication at the higher level and the animals at the lower level, the authors generate fundamental insights at the group decision at the higher level and individual communication. The algorithm makes the use of the local communication between the SUs. The major advantage of the algorithm is its low cost. Also the failure of a local SU node does not lead to the failure of the complete system. The basic requirement of this algorithm is to use the local communication and the group level communication model, to strike out the unwanted signals inserted by the SSDF attacks, and then make the correct decision about the presence of the PU which can be viewed as the multi agent coordination system. The summary of [51] is that it is local communication added to the global decision. He et al. in [55] propose a statistical approach based on identifying the abnormal statistical spectrum sensing behaviours to suppress the malicious activities. Two hidden Markov models (HMM) have been adopted for the behaviours of the honest SUs and the MUs. The detection of the MUs is achieved by differentiating between the models. Due to the abnormal statistical characteristics, the HMM of attackers will be different from that of the honest SUs. Some features of the HMM will depict the difference. The algorithm is about inferring those different features regarded as the HMM parameters to facilitate the detection of the MUs. Active mechanisms are efficient for performing moment-to-moment MU detection and elimination. However, as they require complex hardware designs for implementation, their application may be difficult in certain systems. The next study is performed on passive attack prevention mechanisms for defending SSDF attacks.

Passive methods
In passive methods, the SUs or the FC make changes in their own detection strategies which minimize the damages caused by the MUs [35][36][37][38]. Passive methods include reputation based approaches, game theoretic approaches, and damage suppression. We briefly describe the major works done in each of these approaches below:

Reputation based mechanisms
The reputation based mechanisms are mainly followed in centralized CSS systems [56,74]. The honesty is decided after observing the alignment of its decisions with the decisions of the FC over a period of time. In [35], existing data fusion techniques are investigated for their resistance towards the Byzantine Failure problem. Then, they propose a new algorithm which uses a variable number of samples rather than using a fixed number of samples. The reputation based mechanisms are based on a "reputation value" r i . The ith SU participating in CSS is given a reputation r i , which depends on the history of the truthfulness of its decisions. The history is recorded over a period of time, where r i is updated as where D is the final decision of the spectrum sensing system. The weighted decision rule followed for applying CSS is given as where the variable g fulfils the requirement that enough weight is allocated to the SU that has a slightly negative reputation value, which can happen due to the environmental attenuations. A general reputation based algorithm known as the sequential probability ratio test (SPRT) [57], is given as SPRT has been extended by introducing a reputation based mechanism to develop the weighted sequentially probability ratio test (WSPRT). The SPRT supports sampling a variable number of observations. If a sensing terminal has low SS accuracy, the SPRT can collect more local SS reports. The reputation mechanism is introduced by associating weights with the likelihood function of each SU. The weights depend on the consistency of the sensing terminals with decision of the FC. At a time instant, if a sensing terminal is consistent with the FC, its reputation is incremented by 1. If it is not consistent, it is decremented by 1. A weight associated with an SU is a non-decreasing function of its reputation. It is seen that the proposed WSPRT is highly efficient in detecting the MUs.
The accuracy brought by considering more number of local SS reports can be further enhanced by considering more num-ber of samples. The authors in [36] aim to restrict the damage caused by the MUs by considering more samples. They proposed the enhanced WSPRT (EWSPRT) and enhanced weighted sequential zero/one test (EWZOT) which are robust against SSDF attacks. The EWSPRT and the EWZOT reduce the sample requirements for suppressing the effects of the MUs. However, large number of samples causes high data overhead. In [37], a reputation based approach is followed to detect the MUs. The system model contains multiple MUs and it is assumed that the number of MUs are known. An onion peeling approach, where the suspicious levels are calculated for each SU, is applied to detect the MUs. If the suspicious level is less than a threshold then the SU is regarded as honest. If the suspicious level crosses the threshold, its report is eliminated from decision making. The test is applied till there are no MUs left in the network. However, having the knowledge of the MUs is a strong assumption.
In [38], the performance limits of CSS under the SSDF attacks have been analysed when the information on the number MUs is not available. The authors show that above certain fraction of MUs, it is impossible for any reputation based attack strategy to counter them. The authors propose a new reputation based scheme. It is assumed that the attackers do not have the knowledge other SUs in the network and are dependent only on their own decisions. The attack prevention strategies also aim at countering collaborative attacks. The proposed schemes are basically reputation based countering schemes designed by counting the mismatches between an SU's decisions and the FC's decision. After performing this operation for a period of time, the Byzantines are removed from the data.
The efficiency of the defence mechanisms deteriorate in mobile CRN systems. The fluctuations in the channel gains caused to the mobility of SUs causes reporting errors at the FC, which affect the respective SU's reputation. In [74] and [56], the mobile CRNs under attacks have been considered. The existing researches to mitigate the attacks were proposed for static CRNs. These algorithms for static CRNs cannot be applied for mobile CRNs, because they need localization information. If those algorithms are applied to mobile CRNs, the honest SUs having bad path-loss would be charged. For mobile CRNs, the authors propose to use two trust parameters: location reliability and malicious intention (LRMI). A high LR signifies the location of an SU having less path-loss whereas MI signifies to the intention of the SUs. It is observed that the proposed scheme improves both the MU detection rate and the PU detection.
In [58], the effect of multiple MUs on the energy efficiency along with the detection accuracy of the CR has been studied. A low overhead symmetric cryptographic mechanism based protocol to address SSDF attacks under a trade-off between the energy efficiency and the security of the CRN is given. An optimal number of security bits needed to maximize the energy efficiency has been derived. The authors show that the effects of MUs on energy efficiency depend on the fusion rule, which is used to form the global decision. Detecting the cause of abnormality in the data is challenging. The patterns of distortion brought by environmental obstacles and maliciousness are not always differentiable. The authors in [59] rely on the consistency of the system states estimated by each SU. The proposed scheme removes the abnormalities in the data of the various sensors irrespective of the cause of the deviations in the abnormal reports. System states refers to the path loss exponent [60] and the PU signal power. The MUs do not have control over these parameters. The authors propose iterative state estimation (IRIS) scheme to detect the abnormalities in the data. The IRIS scheme does two examinations: the received signal power and the pathloss exponent of the path. By examining each sensor's result, the scheme finds the measurement residual, which indicates the deviation of the PU signal power and the path loss from their normal values. In [61], the authors use combinatorial optimization identification (COI) [62] to counter the SSDF attacks. The proposed technique is a complement to the previously proposed IRIS technique. The authors have studied the attack scenario where IRIS does not work. The IRIS may also remove a good data sample having a large residual while keeping a compromised data sample. As a prevention mechanism, the authors propose COI to complement the IRIS. The limitation has also been overcome in [63] where the authors design a reputation based classifier to classify the SUs and the MUs. The classification is based on more than the traditional two divisions of honest and malicious. Such classification helps to handle the case when different honest SUs have different individual P f 's and P d 's and there are multiple types of MUs. An iterative expectation maximization algorithm is used to make this classification, calculate global P fa and P d , and reliably detect the PU for the sensing system under the attacks. Their assumption is that the MUs are less in number than the SUs. Except this assumption, no prior knowledge at the FC has been assumed. The MUs can be classified on various grounds like having bad signal and channel characteristics. The MUs have been divided on causing more than one type of attacks. This is done with only few divisions from the CRs while in traditional reputation based classifiers large number of SUs cannot classify the SUs perfectly. The proposed mechanism outperforms other reputation based mechanisms particularly for large number of SUs.
Reputation based methods also help to identity the SU node near to the FC. The detection system near to the decision node is identified with the consistency in the correctness of its decisions. When the decisions of a node suffer lesser pathloss, they remain uncorrupted. Pathloss varies inversely with the distance. The relationship between the transmit and received powers as per the pathloss model is given as [122] where √ G l is a constant and is the product of the transmit and receive antennas field radiation patterns in the line of sight (LOS) direction and is the signal wavelength. Thus, a node constantly reporting correct decisions is quite likely near to the decision node as it is suffering less path loss.
The major challenge here is to recognize that which node has a lesser path loss with the decision node. In this regard, the rep-utation of the cooperating nodes is observed. A node having a high reputation value quite likely is located near to the decision node. The reputation value is based on the trust of a node. In this regard, sliding window based trust mechanism has been used. It observes the consistency in the correctness of the decisions from an SU for a duration of T time slots. A sliding window accommodates the time slots 1 to T . A time period of K such windows is considered. As the sliding window slides on the time scale, its indexing changes from sliding window 1 to sliding window K . The sliding window K accommodates the time slots (K − 1)T to KT . Within the time period T slots of each sliding window, the number of matches between the decision of an SU and that of the FC is denotes as k m . Each window has a trust value associated with it and is given as where V T is an index showing the match between the local SU decision and the global decision of the FC. For K sliding windows, there are K trust values. The weighing factor of the l th sliding window, denoted by l , is given as The trust value of the mth SU is given as At the sliding window K , the long term accumulated weighted trust of the mth SU, denoted by R c m is given as A cumulative weighted average is helpful to characterize long term behaviours or strategies of an SU. A number of research works can be found on the above concept. It has been used for cluster based reputation formulation to detect intrusion attacks in [123], generalized Byzantine attack and defence in [124,125] for the analysis of attack strategies in the absence of defence and implementing suitable defence strategy in [126], for the removal of anomalous values from the data fusion process in [127], for the removal of malicious data in heterogeneous and ultra-dense networks in [128], for implementing a Bayesian inference based sliding-window trust mechanism in [129] analysing a cost benefit trade-off in Byzantine attacks in CSS in [130] and detecting MUs in non-orthogonal multiple access based systems in [131].

Game theoretic approach
In the game theoretic approaches, the CR communication systems are analysed with game theory [64][65][66]. The FC and the SUs make changes in their detection strategies as per the outcome of the game. In a general game theoretic set up, the SUs, the PU, and the FC are regarded as players. Their actions are termed as strategies and the outcome of their actions are termed as utilities or payoffs. The benefits obtained by an SU in reporting honestly and maliciously are termed as honest and malicious utilities respectively. A general objective in this approach is to maximize the honest utility and minimize the malicious utility, so that the SUs refrain from turning malicious. The honest utility for an SU is defined as where P e PU is the probability of error between the PU transmitter and the PU receiver when collisions happen, P e PU is the probability of error when there is no collision, U SU is the utility of the SU when it performs successful transmission, C p is the collision penalty imposed by the PU on the FC which is divided among the N cooperating SUs. The next step is the formulation of the malicious utility G U i (m, n) seen by the MUs as where S p = 1 − P e PU and m is the number of MUs. The general condition for an MU to attack is when it sees, Based on the above condition, the bounds on C p are derived for defending the attacks from the MUs under different scenarios. The m MUs form an attack strategy to get the highest possible damage to the honest SUs. The attack strategy is made with the help of game theory. Designing efficient defence mechanisms with Game Theory requires a thorough understanding of the behaviour of the MUs. In [64], colluded denial of service (DoS) attacks have been studied. The authors perform the security study from the perspective of the MUs. The colluded malicious activity has been studied in single and multiple stages. These attacks are performed with a game theoretic approach. The attackers maximize their utility in a group rather than maximizing their individual utilities. In a single stage scenario, a cooperative game is formulated where the malicious nodes aim to attack as many secondary networks as possible while putting a minimum constraint on the cost or penalty they have to pay. The theoretical expression for the net pay off has been formulated for the MUs and the optimal strategy is derived for them. In a multistage discrete time scenario, the Markov chain model for the dynamic behaviour of MUs has been formulated. The Markov chain models the change of the states. It is proved that as the system reaches the steady state, the net payoff becomes independent of the switching probability and becomes dependent on the number of SUs. Finally the authors obtain the optimal number of malicious users to carry out attacks successfully. The colluded DoS attacks have been countered in [67], where the effects of coordinated DoS on identity management (IdM) systems have been minimized. It is a damage suppression approach and will be explained in the next section. In [65], the authors have investigated an attack scenario where the SUs can access a single frequency band at a time and they hop among the different channels. They study the communication between the SUs and the MUs to find the optimum hopping strategy for the SUs using the Markov decision process. The hopping strategy is aimed towards countering the jamming attacks. Initially, the analysis is done for a perfect knowledge of the channels. Subsequently, two schemes have been proposed for the SUs to gain the knowledge of the adversaries. This knowledge enables them to handle attacks. The authors have also considered the case when the SUs can use all the channels simultaneously. The anti-jamming game is redefined for this case and the problem is modelled as Colonel Blotto game where random power allocation is the defence strategy. These games are zero sum games where players are tasked to distribute limited resources over several objects. The Nash Equilibrium has been established for this game which minimizes the worst-case damage.
In [66], the authors have modelled the jamming and antijamming processes as Markov decision processes. Following this approach, the SUs are able to counter the attacks from the MUs and maximize their pay-off function. First the authors use first policy iteration [68] method. However, the method is computationally expensive. Hence, the authors use Q function method to solve the problem.

Damage suppression approach
In this approach, a particular damage caused by the MU to a specific entity of the SU network is detected and is reduced [69][70][71][72][73]. Unlike the above approaches, it does not depend on detection, localization or down weighing of the MU. The affected entity can be the PU, the energy harvesting scheme, the channel coordination, the secrecy of the data etc. Damage suppression approaches have been widely studied in the literature. In [69], the damage caused by the MUs to the PU has been reduced. The SUs get the favour from the PU to use its spectrum only when the SUs secure the PU communication from the MUs present in the system. Hence, the SUs behave as "jammer" for the MUs. The mechanism has been carried out in two hops. In the first hop the SU transmitter sends the information to the relay set and the SU receiver acts as a friendly jammer to disturb the overhearing of the MUs. In the second hop, the relay transmits the information to the SU receiver and the SU transmitter acts as a friendly jammer to disturb the overhearing of the MUs.
A clear advantage offered by this mechanism is that the possibly disturbing SU service from the perspective of the PU is brought into a rather beneficial service. The possibly disturbing SU service from the perspective of the PU is brought into a rather beneficial service. The damage done by the MUs to the PU is minimized with the help of the SUs, however, the main objective of this work is to maintain any compromise in the secrecy of the SU. Optimal relay selection, time allocation, and power allocation are obtained for protecting the secrecy rate of the SUs while at the same time maintaining the minimum secrecy rate of the PU. Note that the damage caused due to secrecy leakage is reflected in the outage probability of the SU network. The secrecy rate is defined as the difference between the transmitter-receiver data rate and the transmittereavesdropper data rate. The normalized secrecy rates of the PU transmitter-receiver and the PU transmitter-eavesdropper are given as where subscripts PT, PR, and e denote the primary transmitter, the primary receiver, and the eavesdropper respectively, and PT−PR and PT−e are the SINRs at the PR and at the eavesdropper respectively, given as where P PT is the transmit power of the PU transmitter, h a→b is the channel gain on the link a → b link, and I a is the interference on the node a. a and b are dummy variables. Similarly, the secrecy rate of the SU transmitter-receiver link is given as where Depending on the damage to be reduced, c PT−PR and c ST−SR can be maximized or can be used for applying constraints over them.
Enhancing the secrecy performance in terms of data rate is a robust way to reduce the damage caused by the MUs. An outage probability analysis can be another way to reduce these damages. In [72], the secrecy performance is investigated in terms of an outage probability. A relay scheme is proposed to prevent eavesdropping and assist transmission from the CR transmitter to CR receiver. It is shown that the physical layer secrecy is significantly improved by increasing the number of relays. Another security issue addressed here is that a CR enabled SU turning malicious as it can sense the environment. This capability can be used to launch attacks. This increases the chances of a CRN to have more attacks than a normal communication system. The broadcast system also increases the chances to overhear secure communications.
The capability of the MUs to launch attacks considerably increases when they have energy harvesting capability. This issue is addressed in [70], where the MUs harvest energy for their operations like the honest SUs. The issue addressed is that MUs can also harvest energy for their operations like the honest SUs. The authors have formulated a throughput optimization problem under MU attacks as a Markov Decision process. Then a new solution based on deception tactics to stop smart MUs has been presented. Further, a learning algorithm for multiple SUs to find an optimal transmission policy has been designed. It is shown that these learning algorithms are effective in stopping smart MUs. In damage control mechanisms, the reliability in the coordination of channels plays a significant role. The damages caused by the attacks cause disruptions in the coordination of the channels. A preventive mechanism in this regard exists in [71], where the focus is on the coordination of channels in spectrum sharing systems in the presence of MUs. Reliable coordination of the channels is very important in CRNs. This coordination may be smooth until all the CRs in the SU network are honest. However, if the SUs start behaving contrary to their channel assignment, then such benefits associated with the channel coordination are negated. The authors in [71] focus on the dilemmas that occur in the coordination because of the attacks. The first dilemma is with the PU when it thinks that the SU is honest. On the other hand, if it thinks that the SU is malicious then it has to choose a band from a larger set of bands which invites more interferences. Similarly, there are dilemmas associated with the SUs about what the PU thinks about it. If the PU has a positive opinion about the SU, then the SU is allowed to transmit on the PU band. However, if the SU has a malicious image, then it will transmit on the SU band only. Corresponding to these beliefs, a game has been formulated and a game theoretic approach has been followed to solve the problem accordingly. The players are the PU, the SU, and the SP. The authors consider an example of two secondary service providers (SPs) with their own set of SUs. Suppose, each SP has been allotted a channel. The SPs will be benefitted in protecting its SUs from any possible interference. Similarly, they will get other benefits in harming the other channels. So, there are a set of strategies which the SUs can play. This reference also comes in the category of game theoretic approaches. However, since the authors aim to suppress the damage of SSDF attacks on the channel coordination, we have explained it in the damage suppression approaches category. In [67], the damage caused by the coordinated DoS attacks on the identity management (IdM) systems [64] has been minimized. IdM systems provide authentication authorities critical data. Through IdM systems, the SUs can authenticate at the single and multiple domains without providing additional information. IdM systems also help in reducing the complexity in the management of the data. They benefit many large systems like cyber-physical systems, smartgrids, and vehicular ad hoc networks. However, coordinated DoS attacks can severely affect IdM systems. The authors in [67] present scheme for DoS attacks mitigation by their organization and optimization of the IdM system (SAMOS). This scheme is based on adaptive reorganizations and optimization of the IdM systems' components to thwart DoS attacks. The coordinated DoS attack strategy has been considered from [64] and SAMOS aims at thwarting it. It is found that SAMOS is efficient in this purpose. A significant issue for the damage suppression attack prevention mechanisms is the optimal decision fusion rule for minimizing a damage caused by the MUs. In decision fusion rule based CSS, the optimal value of n has been obtained by obtaining and using the attack strength [73]. This ratio comes out to be the probability of an SU of being malicious. The optimal value of n minimizes the Bayes risk. However for data fusion, this issue remains unstudied.

COUNTER STRATEGIES AGAINST PUEAS
PUEAs are launched when the MUs attack by transmitting the emulated PU signal. The PUEAs severely compromise the functioning of the SU network. There are many existing and ongoing researches to defend the SU network from PUEAs. We categorize the major proposed mechanisms as active and passive approaches.

Active methods
The active methods aim at cancelling or suppressing the PUE signal by differentiating the power or a feature of the PUE signal from the PU signal. The active methods are mainly categorized into three mechanisms based on the nature of implementation as the Location based defence mechanisms, Finger print detection mechanism, and Game theoretic approach. We discuss here the major studies performed in the literature and possibilities offered by them for further research.

Location based defence mechanism
The first widely used mechanism for detecting PUEAs is location based defence mechanism. It is based on transmitter verification performed by identifying the location of the respective transmitter. The most fundamental work in this category has been done on detecting a PUE is in [39]. Here, the authors have focused on the vacant TV bands, since the FCC has recognized the TV bands suitable for the initial application for CR technology. These TV bands experience lower dynamic traffic compared to the cellular networks. In this paper, a transmitter verification scheme termed as LocDef (localization based defence) has been proposed. In this scheme, the source of the signal is verified for being the PU by observing the signal characteristics. The location information of the PU has been used to differentiate between the PU and the PUE signal characteristics. The authors assume a stationary and known PU location. To implement this scheme, LocDef uses a non-iterative localiza-tion scheme. This method primarily depends on the modelling of the received signal from the PUE. The LocDef approach is based on calculating the mean RSS at two different points. For illustrating the basic LocDef model, let us consider two SUs SU 1 and SU 2 . The location of the PU transmitter is assumed to be known to both the SUs. They also consist self equipped localization units which help them to identify their own locations. Let their location coordinates be (x 1 , y 1 ) and (x 2 , y 2 ). Each SU receives the signal from the attacker and can record its respective received signal strength (RSS), which is modelled by statistical log-loss signal propagation model as where P r is the RSS in dBm, P t is the transmission signal strength in dBm, d is the transmitter receiver distance, and is the path loss constant. s represents the standard deviation associated with the degree of shadow fading, and it is assumed that it is the same for both the SUs as they are very close to each other.
where d attacker,SU 1 and d attacker,SU 2 represent the distance from the attacker to the SU 1 and SU 2 . If the position of the attacker is (x, y), then The trace of the attacker is calculated as radius = d 12 2 , centre = In line with this mechanism, the RSS can be exploited in different ways to identify the PUE. Received signal strength indicators (RSSI) based learning mechanisms have been found efficient in detecting the PUEs [75]. The authors propose a technique based on particle swarm optimization algorithm and the RSSI for the PU/PUE position detection to increase the detection accuracy and decrease the probability of false alarms. The accuracy of the location for detecting the PUEAs can be improved with estimation techniques. Two criteria have been used to obtain total number of co-channel transmitters in the primary system [76]. The first criterion is the net MMSE criterion, which uses Cramer-Rao lower bound on location accuracy. In [76], the authors present two criteria to determine the total number of co-channel transmitters in the primary system. The first criterion is called the net MMSE criterion, which uses the Cramer-Rao lower bound on localization accuracy. The second criterion is the information theoretic criterion, the minimum description length. Although only signal strength measurements have been considered, the approach can be generalized to include other The security attacks can advance to the next level by making their own malicious base stations. As a result, they can subvert the cell association process of the genuine base stations. A stochastic geometry framework can be used to efficiently he random locations of the honest base stations and malicious base stations and analyse the impact of these attacks [77]. The conditional probability of the RLF given a suboptimal association triggered by MBSs has been derived analytically and the theoretical results using simulations have been verified. This paper also presents insights on the appropriate ranges of thresholds for MTs seeking association with LBSs and claiming RLF using the conditional RLF probability triggered by MBS attacks.
Location based PUEA detection can prove inefficient if the MUs further manipulate their signals and mislead the detector about their locations. To overcome these limitations, fingerprint based mechanisms have been widely used as prevention mechanisms for PUEA.

4.1.2
Finger print based authentication Fingerprint based mechanisms use signature embedded in the PU signal to detect its authentication. The SUs check the embedded signature in the incoming signal and discard it if they find discrepancy in it. The signature or the fingerprint can be a feature of the signal or an external tag embedded in the signal. Thus the fingerprint can be dependent on the PU signal [17], or independent of it [78]. A fingerprint dependent on the PU signal is transmitted concurrently with its data. It is added in the PU signal by superimposing a secret modulation in the transmitted PU signal without requiring additional bandwidth. The designing of the authentication finger print system has to be performed at the physical layer, as performing it at the higher layers will require the PU and the SUs to follow the same protocol. A PU signal-independent finger print is better than a dependent one. Here, a fingerprint termed as an authentication tag is developed in the first stage and in the second stage, the developed tag is embedded with the PU signal as shown in Figure 6.
In the first stage, a tag is developed and in the second, the developed tag is embedded with the PU signal. At the receiver, the PUE signals are checked for the tag and if they do not possess it, they are cancelled out. The fundamental finger print authentication model is shown in A tag is inserted in the PU signal at the transmitter. This tag has to be detected by the SU to know that the incoming signal is from the PU or the PUE. The PU transmit signal from (1) with tag embedded can be written as s(k) = g(t, x) where x is the transmit symbol and t is the embedded tag bit, given as t ∈ {0, 1}. Thus, g(0, x) and g(1, x) denote the transmitted signal for t ∈ {0, 1}. It is assumed that g(0, x) ∈  0 and g(1, x) ∈  1 , where  0 and  1 denote sets. Ift is the tag bit detected by the SU, the detection error probability is For successfully detecting the PUE, involves P e must be minimized at the SU which is done with the help of a Bayesian detector. The corresponding detection rule is formulated aŝ The above ML detection is solved depending on the distribution of the fading gain h(k) in (1). This method explores possibilities of using device specific features like device ID, MAC address, radio metrics like amplitude, frequency, and bandwidth as the fingerprints of the PUs. These features can also be used to determine the total number of transmitting devices in the PU spectrum [18]. The method proposed in [18] is non-parametric since the number of devices needs not be known a priori. After forming the fingerprints, the authors adopt the infinite Gaussian mixture model(IGMM) and propose a modified collapsed Gibbs sampling method to classify the extracted finger prints based on Bayesian classification. In the first module, the authors first check the phase of the signal and the carrier frequency of the signal because these properties are unique for the transmitter. Then they check the MAC address and from the MAC ID database, they differentiate it. The second module checks the signal ID. If the PUE transmits from the same ID but a different finger print, it is detected. It is found that the proposed classification efficiently detects the PUEs. The proposed mechanism is independent of the number of transmitters, hence it needs not to be estimated.
The efficiency of the PUEAs is enhanced to the next level if they are capable of detecting the channels which are to be used by the SUs and attack accordingly. A fingerprint based mechanism proposed to counter such PUEAs is based on allowing the SUs to share the messages among themselves [41]. Here, the RSS from the PU is considered as the signature or the fingerprint. The PUEs are assumed to have the capability to detect channels which are going to be used by the SUs and attack accordingly. The authors propose a cooperative mechanism based on sharing the messages among all the SUs. Each SU calculates a local function, a compatibility function, and exchanges the information with the neighbouring SUs. It calculates its own belief with the help of the information from all other SUs. The distribution of the RSS from the PU and the suspect are used to perform these operations. The belief calculation is performed iteratively until the SUs converge to a common belief about the PUE. Then, the PUE is detected and its signal characteristics are broadcast to the CRs in the network. Finally, the SUs are able to avoid the PUE's signal. In [79], a game theoretic framework has been proposed to discard the PUEAs. The authors formulate a non-cooperative multistage game between the SUs and PUEs. The strategy of the SUs and PUEs are different. The SUs leave the spectrum on the arrival of the PU. But the PUE tries to interfere with it while also trying to keep the SUs away. A novel belief updating system for updating the SUs belief about the PUs has been proposed as the game evolves. It is shown that the proposed belief updating system achieves satisfactory performance for the SUs in terms of recognizing the true PUs and cancelling out the signals from the PUEs. The relationship between the energy level of the RF signal and the acoustic information of the primary signal can be exploited to detect the PUEs, particularly when the primary signal is the TV signal. This relationship is very difficult to mimic. The approach has been exploited in [80], where microphones and TVs are the designated PUs. Further, RF fingerprinting has been studied experimentally to detect the PUEs [90]. Low cost CRs have been considered for detecting the PUEs. Investigations have been performed on the distortions made by low-end CR receivers in the RF fingerprints. It is concluded that RF fingerprint can be concluded only when the CRs have high end receivers.
In fingerprint based approaches, the a priori knowledge on the PU signal is an obvious limitation. A direct approach in this regard could be observing the externally visible characteristics of the PU signals and using them as fingerprints. However, the characteristic should not be mimicable. In [81], the PUE detection is performed by the acquisition and the reconstruction of the PU signal activity pattern. This technique acquires the activity pattern of the signal through SS. For example the ON and OFF periods of the signal may be acquired as the activity pattern. Then, it reconstructs the signal with a signal reconstruction model and checks it for any error. By examining a reconstruction error, the signals from a PU or a PUE are differentiated. The paper aims to overcome the limitations of location based identification and finger-print methods which need knowledge about the signal characteristics or the location of the PUE. Similarly in [85], a fingerprint mechanism independent of any previous knowledge about the system. The authors use the statistical characteristics of the users to address one of the important kinds of attacks called primary user emulation attack (PUEA). In the proposed scheme, the received power statistics of the CR users is used to detect the PUEA. Pu et el. in [82] detect the motion-related feature vectors of PUs in the FFT of its signal to differentiate PUE from it. A database where the signal from the PU is given is referred. The authors record the data from the incoming signal and compare it to the data from the database. For further classification, they feed this data to a covariance descriptor with the help of a neural algorithm. The covariance descriptor classifies the incoming signal database with the existing signal database to detect the PUEs. In [83], two strategies have been proposed for the detection of the PUEs. The first strategy is a channel surveillance process to check for the presence of the PUEs. The second strategy is aimed at detecting new transmission opportunities for the PUEs. An extra sensing process has been implemented to carry out the second strategy. The PUEs have been categorized as selfish, malicious, and mixed. The selfish PUEs aim for transmitting whereas malicious PUEs only aim to cause problem for the SUs, like obstructing the operation of the CR network. The mixed PUEs launch attacks in both the forms. In this process, the channel has been sensed and if an SU signal is found on the channel, it indicates that it is being used by a selfish PUE. After the sensing time, if the FC sees a channel is free which was declared busy after the sensing process, it was attacked by a PUE. It is found that it is possible to retrieve the opportunity of using the attacked channel by sensing it again after the sensing time finishes. Hence, an extra sensing process has been proposed. Essentially, the selfish PUE keeps transmitting throughout the sensing and the transmission periods of the SU whereas the smart PUEs transmit only for the sensing time. The authors have studied this problem with a game theoretic approach. The PUEs and the defenders are considered players and the Nash equilibrium has been derived for all the three attackers (malicious, selfish and mixed PUEs).
Certain inherent parameters in the PU signal like the minimum eigenvalue (EME) [86] can be used to mitigate PUEA through less spatial correlated secondary users (SUs). The SUs are selected based on their spatial correlation characteristics in log-normal shadow fading environment. Hence, the detection performance is improved. Simulation results show that the proposed method mitigates the PUEA effectively with lesser number of SUs.
Another significant mechanism has been proposed by employing an underlay RF fingerprint in the PU signal [87]. The fingerprint is introduced as an underlay waveform on the top of the header of the legitimate PU signal. This underlay waveform is consisted of wider bandwidth and higher baud rate than those of the original primary users signal. As a direct result, the underlay waveform exhibits a unique and different cyclostationary feature than that of the primary user signal. At the CR, a sophisticated cyclostationary analysis algorithm, combined with a primary user signal cancellation, is capable of revealing this unique cyclostationary feature to authenticate and validate the identity of the legitimate PU. The PUE attack signal, on the other hand, will not reveal this cyclostationary feature and fail to pass the authentication.
Further, a hybrid PUEA detection method has been proposed [88], where two kinds of wireless channel characteristics between the transmitter and the receiver, the Doppler spread and the variance of the received signal power, are utilized to infer the source of the received signal. In the CRN with mobile SU, the Doppler spread between the SU and the primary user (PU) may differ from that between the SU and the primary user emulator (PUE) because of different relative velocities. However, the performance of the PUEA detection method based on Doppler spread may be affected the relative position between the SU and the PUE. In order to make the detection performance be unaffected the relative position between the SU and PUE, the variance of the received signal power is also used as the signature of the transmitter.
In [89], the authors propose a reliable AES-encrypted DTV scheme, in which an AES-encrypted reference signal is generated at the TV transmitter and used as the sync bytes of each DTV data frame. By allowing a shared secret between the transmitter and the receiver, the reference signal can be regenerated at the receiver and be used to achieve accurate identification of authorized primary users. The effectiveness of the proposed approach through both theoretical derivation and simulation examples has been analysed. It is shown that with the AES-encrypted DTV scheme, the primary user can be detected with high accuracy and low false alarm rate under primary user emulation attacks. It should be emphasized that the proposed scheme requires no changes in hardware or system structure except of a plug-in AES chip. Potentially, it can be applied to today's DTV system directly to mitigate primary user emulation attacks, and achieve efficient spectrum sharing.

Game theoretic approach
In these approaches, the PUEs are actively cornered by formulating optimal strategies with the help of game theory. Generally, the PU, the SUs and the PUEs act as the players. They formulate strategies based on maximizing the rewards received by them respectively. In a typical setup, the PUEs may design strategies to maximize their rewards while the SUs and the PU play strategies to thwart these attacks. In these approaches, the perfectness of information shared between players having common interest players a significant role. The evolution of the imperfect information exchange to perfect information exchange between the SUs viz. players has been studied in [91], where a game theoretical framework is proposed to study the primary user emulation attack (PUEA) on cognitive radio nodes as a game of imperfect information between the SUs, who do not exchange game information between them against the adversaries generating the PUEA and to define optimal strategies with minor computational demands. When the SU challenges the PU emulator successfully, updating the information on a cloud-based database enables the rest of the network to know the identity of PUE. As the game evolves, the grand coalition of the SUs acts as the one without collaboration against the PU emulator playing a winning strategy. The performance of the game for optimal strategies is equal to the performance of the collaborative methods for PUEA detection. A possibility to be explored with game theoretic approaches is that whether the PUEAs can be mitigated by allowing an interaction between the SUs and the PUEs. In [92], the interaction between the PUE attacker and the secondary user is modelled as a constant-sum differential game which is called PUE attack game. The secondary user's objective is to find the optimal sensing strategy so as to maximize its overall channel usability, while the attacker's objective is to minimize the secondary user's overall channel usability. The Nash equilibrium solution of this PUE attack game is deprived, and the optimal anti-PUE attack strategy is obtained. Numerical results demonstrate the trajectories of the secondary user's optimal channel sensing strategies over time, and also shows that: by following the differential game solution, the secondary user can always optimize its channel usability when confronting PUE attacks. For illustrating the model considered for the game theoretic approach, let the number of channels not used by the PU be K . Let the number of channels which can be at most be tested by the SU be M and the number of channels which can be attacked by the PUE attacker be N . The strategy of the SU u(t ) is defined as the portion of M channels, denoted as u(t ) ∈ [ 1 M , 1]. The strategy of the attacker v(t ) is denoted as v(t ) ∈ [0, 1]. The left boundary of v(t ) is 0, as the attacker can decide whether to attack or not. The probability with which each PU channel will be sensed is given as . On the other hand, the probability with which each channel will be attacked will be . At a time instant t , the total number of available channels which are sensed by the SU but not attacked by the attacker is denoted as The number of channels which are successfully attacked (sensed and attacked), is denoted as: At the time instant t , the pure usability of channels for the SU is defined as For the time period of the cognitive radio network [0, T ], the total pure usability for an SU will be ∫ T 0 (x s (t ) − x a (t ))dt . The total power consumption for an SU is defined as where is the unit power consumption for sensing one channel. Thus, the overall utility for an SU is given as Similarly, total pure usability for the PUE and the overall utility are given as ∫ T 0 (x a − x s )dt and The SUs and the attackers fulfil their own objectives by maximizing or minimizing the expressions in (65) and (66) or their derivatives. In [93], a dynamic secure routing game framework to effectively combat jamming attacks in distributed cognitive radio networks has been proposed. First, a stochastic multistage zero-sum game framework based on the directional exploration of ad hoc on-demand distance vector (AODV) algorithms has been proposed. The zero-sum game captures the conflicting goals between malicious attackers and honest nodes and considers packet error probability and delay as performance metrics. The game-theoretic routing protocol guarantees a performance level given by the value of the game. Distributed Boltzmann-Gibbs learning is used for an on-line routing algorithm, in which the users do not have the knowledge of the attackers and the utility function. Instead, the users learn the payoffs based on their past observations. Simulations are used to illustrate the proposed routing mechanism and compare the algorithm with fictitious-play learning. Unlike typical distributed routing algorithms such as AODV routing, the proposed secure routing algorithm supports a novel recovery of routing path failure against unknown attackers. The game theoretic approach has been extended to formulate a multi-channel surveillance game between the selfish attack and the surveillance process in multi-channel CRNs [94]. The sequence-form representation method is adopted to determine the Nash equilibrium (NE) of the game. It is shown that performing the obtained NE surveillance strategy significantly mitigates selfish PUEA. Later it has been pointed that the network managers could adopt a surveillance process on the disallowed channels for identifying illegal channel occupation of selfish PUEA [94]. and hence mitigating selfish PUEA. Determining surveillance strategies, particularly in multi-channel context, is necessary for ensuring network operation fairness. In this paper, the authors formulate a game, called multi-channel surveillance game, between the selfish attack and the surveillance process in multi-channel CRNs. The sequence-form representation method is adopted to determine the NE of the game.
In [95], the problem is analysed within a Bayesian game framework, in which users are unsure of the legitimacy of the claimed type of other users. It is shown that depending on radios' beliefs about the fraction of PUs in the system, a policy maker can control the occurrence of emulation attacks by adjusting the gains and costs associated with performing or checking for emulation attacks.

Passive methods
The passive approaches have been applied in two ways: 1. Suppressing one or more harmful effects of PUEAs on the SU network rather than directly cancelling its signal with the help of the channel state information (CSI) of the PUE with the PU or the SU with the PU. 2. Avoiding the signals from the PUE using the CSI of the PU or the PUE. Passive approaches largely depend on the channel based signatures of the PU-SU link. It is defining the path between the PUE and the FC [120]. Let us consider a part of Figure 4 containing the PUE, one SU and the PU, as shown in Figure 7.
Detecting PUEAs can be done with the help of transmitter specific features or with the help of channel-specific features. The latter method detects the PUEs by estimating the statistical parameters of the channel gain of the PUEs with the SUs and the PU. Chen et al. in [120] exploit the fundamental characteristics of the communication channels with the path loss and the log-normal shadowing in CR networks. It is based on identifying PUEs by estimating the received signal power at the SU. The received signal power after the signal traverses a path of distance r, denoted by P r , will be proportional to r − , where is the path loss coefficient. P r is also proportional to a shadowing random variable G , which is given as, G = e . Thus, where follows a normal distribution  (0, 2 ). An SU receives signals from the PU with E 1 , E 2 , … , E n , where n is the number of observed signals. Then E 1 , E 2 , … , E n is i.i.d. and follow the same distribution as P ( ) As shown in Figure 7, r 1 is the distance of the PU from the SU. The next step will be to estimate the statistical parameters of P ( ) r . It is done as The estimation of u r and 2 r will be performed aŝ A large n gives accurate estimation of u r and 2 r . The detection of PUE is done as The difference between the passive approaches and the active approaches is that the passive approaches are independent of the features, the RSS, the external tagging, or the classification of the PU or the PUE. Secondly, the active approaches may have to make some modifications in their sensing strategies after detecting the PUE, like cancelling its signals. However, the passive approach follow the normal sensing integrated with a mechanism for the suppression of a particular damage caused by the PUE signal or its avoidance with the help of CSI. Following are some of the major works with passive approaches with the above applications:

Suppressing the harmful effects of the PUEAs
In these approaches, a particular harm to the SU system done by the PUEAs is countered. In approach, these mechanisms are identical to the Damage suppression based approaches for defending the SSDF attacks. The harmful effects can be in the form of throughput degradation, increased interference to the PU, or the combination of both. In [21], the channel information is used to minimize the damage caused by the PUEAs to the robustness of a CSS system. The paper considers a basic cooperative model with an always-on PUE. The expressions for P d and P fa are obtained for a CSS system where a PUE is present. Then, P d is maximized with a bound on P fa to obtain the optimal weights for performing CSS. The optimal weights are functions of the channel gains of the PU and the PUE with the SU. This leads to the minimization of the damages caused by PUEAs on the detection performance of CSS. The authors in [42] and [44] investigate the harmful effects of smart PUEAs on the CSS network. Smart PUEs attack only when the PU is absent. The degradation caused by the PUEAs is accounted as probability of error P e , which is defined as the sum of P fa and 1 − P d . Decision fusion CSS has been considered and both aim at minimization of P e . The parameter optimized for achieving the minimum P e is the transmission time. The difference between the works is that the authors in [42] perform this optimization for n out of N decision fusion systems whereas, the authors in [44] perform it for OR and AND rule based decision fusion systems. Alahmadi et al.in [96] aim at countering the damages caused by PUEAs on energy harvesting in an OFDM modulation based CR system. A DTV signal has been considered and the P2 pilot symbols are generated using an advance algorithm for encryption. The main aim of this research is sumrate-maximization of the energy harvesting based SU system facing PUEAs. The sum-rate is the sum of the downlink transmission and the uplink transmission of the SUs. Alahmadi et al. minimize the interference of the PUEAs which affect the sum rate for the SUs. The basic hypothesis model shown in (27) is considered in these mechanisms. First the particular damaged caused by the PUEAs. Next, this damage is suppressed with the help of weighted CSS. The weight values are designed in a manner to suppress the damages. Ideally, the weighted CSS scheme is carried out by making the weighted sum of the individual energy values transmitted by the SUs to the FC. However, here the weighted CSS scheme is slightly changed for mathematical simplicity. The individual signals from the SUs y i (k), i ∈  are added with weights assigned to each of them as Now the energy is calculated as The probability of false alarm and mis detection are obtained as where Γ(., .) and Γ() are the upper incomplete Gamma function and Gamma function respectively. The damage considered here is that the decrement caused in P d because of PUEAs. To suppress this damage, the maximization of P d has been carried out with a constraint on P fa . The Neymann-Pearson method now has dual tasks: the calculation of the decision threshold and the calculation of beamforming vector values w i s. The objective here is to minimize the effects of PUEAs as max P d FC ( , w) (78) subject to where P f FC = Pr (E beam > |H 0 ) and P f FC = Pr (E beam > |H 1 ) are the probabilities of false alarm and detection respectively. These values were functions of the channel gains on the PU-SU and the PUE-SU link, h p i and h e i , respectively. The optimal beamforming vector values were obtained as  [131], to study the statistical characteristics of the incoming data at the FC for PUEA avoidance in [132], for adaptive cooperative sensing to reduce sensing overhead while not affecting the sensing accuracy requirements in [133], to enable the assistance of the SUs in CSS for malicious attack suppression in [134] for obtaining the optimal voting rule and the optimal number of samples for performing CSS in the presence of MUs in [135], for the application of the maximum ratio combining in CSS for MU detection in [136], for the application of a heuristic algorithm, which chooses the optimal subset of cognitive sensors with minimum average energy consumption for CSS in [137], to use physical layer fingerprints in a multi-path Rayleigh fading channel for the mitigation of MUs in CSS [138,139], and for the application of generalized likelihood ratio test and Rao test for suppressing the effects of MUs in CSS in [140].
The expressions in (76) and (77) can be used to control the damages on different aspects of CSS. A significant aspect of CSS is its overall throughput. The lesser the P fa higher is the throughput of the system. Minimization of P fa can lead to maximization of the throughput. In [97], the focus is on the damage caused by the PUEA on the throughput of the CSS system. The throughput of a CSS system is reflected in it P fa . Lower the P fa , higher is the CSS throughput. Thus, the objective in [97] was the minimization of P fa in (76) with a constraint on P d in (77). Total error minimization is another objective widely studied within this category of attack defence mechanism. Total error is formulated with the sum of P fa and 1 − P d .
A significant requirement is reducing the damages of the PUEAs on the throughput of the SU at the network level. This study is performed in [107] by analysing the call blocking and dropping performance, where the authors develop the first centralized protocol. The authors developed the first centralized protocol to help SUs fuse their individual detection results with the help of a centralized controller, to better mitigate PUEA. Here, a distributed spectrum decision protocol in which SUs make individual spectrum decisions is developed and then individual sensing results with their onehop neighbours to increase resilience to PUEA are exchanged. This protocol has been termed as NEAT: Neighbour sssisted spectrum decision protocol. A detailed analysis of the protocol in terms of the probability of successful PUEA under Byzantine attacks has also been presented, where the malicious users can lie about their individual sensing results. It is shown that with negligible communication overhead, the proposed protocol reduces the probability of successful PUEA by up to three orders of magnitude in the presence of Byzantine attacks.

4.4
Avoiding the PUE signal conditions which throw light of some important facts such as whether the honest SUs will exist with PUEAs, with what probability PUEAs launch attacks, under what conditions the PUEAs are discouraged, with what probability the radios will challenge the license of other radios etc. These results can be used to determine appropriate policies to keep the rate of PUEAs low. Another approach for PUE signal avoidance has been proposed in [103], where the PUE attacks only when no one is sensing the channel. A new room of vulnerability in the conventional sensing approaches where the PUE attacks only when no one is sensing the channel has been introduced in [103]. This attack will decrease the channel utilization by SUs and create a Denial of Service (DoS) situation for victim SUs. This attack has been named as off-sensing attack. The authors also propose an analytical model to analyse the impact of the off-sensing attack in a CRN.
In [104], the PUE signal avoidance has been studied without any prior knowledge on the primary user activity characteristics and the secondary user access strategies. They formulate the problem as a non-stochastic online learning problem where the PUE attacker needs to dynamically decide the attacking channel in each time slot based on its attacking experience in previous slots. The challenge in the considered problem is that the PUE attacker cannot observe the reward on the attacked channel because it never knows if a secondary user ever tries to access it. To solve this challenge, they propose an attack-but observeanother (ABOA) scheme, in which the PUE attacker attacks one channel in the spectrum sensing phase, but observes at least one other channel in the data transmission phase. The authors propose two online learning algorithms, EXP3-DO and OPT-RO, to dynamically decide the attacking and observing channels.
In [105], before avoidance, the authors attempt to characterize the PUEA. This is done by estimating the channel parameter and transmission power of the primary user or the BS with MLE. The PUE then transmits the false signal by emulating these characters to spoof other honest cognitive users making the wrong decisions. For PUEA avoidance, an expectation maximization based algorithm has been proposed by estimating the channel parameters. In the proposed approach, the difference of the parameters of primary channel and attack channel to achieve the accurate detection through a simple iterative algorithm has been utilized.
In [106], PUEA avoidance is carried out by first characterizing received power at good secondary user. This is done by using a flexible log-normal sum approximation. The received power thus characterized is used to determine the probability of successful PUEA on each secondary user, which is used to develop the proposed protocol. Simulation results demonstrate that the proposed protocol can significantly reduce the probability of successful PUEA under Byzantine attacks (i.e., when the malicious users intentionally provide false spectrum decisions), while still following the spectrum evacuation etiquette.
It can be concluded that the most fundamental impact of the security attacks is on the throughput of the SUs. The interference on the PU signals also increases severely. The impact on the SU throughput occurs as security attacks can cause the SUs to think that the PU signal is present when the spectrum is vacant from the PU signal. Consequently, the SUs cease their transmissions. In the other case, security attacks make the SUs believe that the PU signal is absent on a frequency band, when it is actually present. The SUs transmit on the respective band and cause interference to the PU. As a result, the PU signal suffers interference. Due to this interference, it can also impose a penalty on the SUs. It can be seen that both the active and passive mechanisms are being followed to counter the SSDF and PUE attacks. Both these mechanisms have certain advantages and disadvantages in countering SSDF and PUE attacks respectively. The active mechanisms counter the attacks headon. They ensure a nearly instantaneous elimination of the corrupt data. However, the FC has to constantly be active in identifying the type of anomaly and its elimination. This may lead to an increased functional load on the FC. Passive mechanisms, on the other hand, are based on gradual suppression of the corrupt data over a time span. They are achieved with changes in the existing FC circuit which ensures malicious data suppression in the long term. However, operational latency can at times prove to be a disadvantage with them. In summary, for quick action, active mechanisms are recommended, whereas for robustness, passive mechanisms are recommended. The limitations with these approaches open a wide scope for further researches.

OPEN RESEARCH DIRECTIONS
As mentioned above, the approaches surveyed have certain shortcomings which opens many new research problems. The problems are based on increasing efficiency of their fundamental mechanisms and rectifying the problems created with their application respectively. These research problems have been explained as follows:

Active counter mechanisms for SSDF attacks
Several open research issues exist in active counter mechanisms for SSDF attacks. Location based defence mechanisms are effective but depend on the location identification of the PU for detecting the attacker. The location of the PU from each SU is transmitted to the FC via. their P m s. During this process, the MUs can overhear the SUs and can make the estimate of the locations of the PU and the SU. This information leakage is known as location privacy leakage.
The primary limitation in statistical data based detection mechanisms is caused due to the confidence level of the statistical mechanism followed. The confidence level is the threshold till which a data value is considered to be an outlier. A narrow confidence level ensures that a data value is regarded as outlier even with a slight deviation from the distribution of the other data values. This ensures robustness in detecting the outlying data from the MUs. However, in case when an SU lies in a deep faded environment, the data value it transits to the FC can also be corrupted and be regarded as an outlying data from an MU. Thus the confidence level of statistical test needs to be increased. However, that may lead to failure in detecting real malicious outlying values. The same has to be addressed when erroneous reporting channels are considered.
Further, statistical test mechanisms are generally applicable on data fusion based CSS systems only. However, in limited bandwidth CSS systems, only decision transfer is possible. Modifying statistical based mechanisms for decision fusion based CSS is another important research direction that needs to be addressed.

Passive counter mechanisms for SSDF attacks
Several limitations exist in the application of passive counter mechanisms. Reputation based mechanisms depend on recording the history of the SUs. This can lead to inaccuracy in detecting the MUs, as the defence mechanism has to be performed within the sensing time. There is a possibility that the sensing time is not enough for completing the process of collecting the history. A moment-to-moment outlier detection is more preferred for detecting the outliers. The second major practical research issue is detecting the reliable SU node, with which the data from the SUs can be compared and reputation can be assigned to them.
In game theoretic approaches, the formulation of the expected utility of the honest SUs depends on the probability of successful communication from the SU transmitter to the SU receiver. This is decided on the basis of the number of collisions suffered by the transmitted data. If the SU is located in a deep faded environment, the data from the SU transmitter will suffer more collisions, which will reduce the expected utility of the honest SUs. Consequently, the malicious utility seen by an SU can go higher than the honest utility and can result into an inefficient MU detection.
Damage suppression approaches heavily depend on the estimation of the channel gains between the PU transmitter and the attacker and the SU transmitter and the attacker. This estimation is challenging because attackers generally attack from secret locations. There is a possibility of inaccurate estimation of these channel gains which affects the damage suppression approach severely.

Active counter mechanisms for PUEAs
Several open research problems exist in active counter mechanisms for detecting the PUEAs. Location based defence mechanisms depend on matching the RSS and direction of arrival (DOA) of the signals from the transmitter with those from the PU. In case of mismatch, the transmitter is regarded as PUE. However, the PUE can emulate the RSS and the DOA of the PU can as the transmit signal power and location of the PU can be easily known. Once the PUE knows the locations of the receiving SUs, it can adjust the transmit power and its position accordingly to misdirect the SUs.
Similar limitations exist with fingerprint based detection approaches. The SUs detect the authenticity of the PU signal with the help of its fingerprint inserted in the form of a tag. However, successful implementation of this mechanism is complex. It involves the insertion of the tag in the PU signal which is then communicated to the SU. When the SU receives the PU signal, it performs ML detection to check the similarity of the tag in the PU signal with the one communicated to it earlier. ML detection largely depends on the estimated channel gain between the PU transmitter and the SU receiver. Thus, the mechanism is susceptible to the channel estimation error. Efficient designing of low complex finger print mechanism without compromising on the accuracy remains an open research problem.

Passive counter mechanisms for PUEAs
Passive counter mechanisms have a wide scope of further research. The suppressing harmful effects of PUEAs based defence mechanisms focus on a very narrow subset of damages. The other damages remain untouched or may even result in deteriorating them. For example [19] focuses on the interference to the PU, but have to skip the effects on the SU throughput. Similarly, [52] focuses on the energy harvesting issues, while [97] focuses on the SU throughput leaving other issues untouched. Though a constraint has been put on the interference to the PU, but still these mechanisms fail to make a comprehensive defence strategy. The mechanism of avoiding PUE signals is difficult to be implemented on a narrow channel. Generally they need hopping on multiple number of channels to avoid the PUE signal. They are also subject to a prior detection of the PUE signal. It must be noted that these attack mechanisms offer a huge scope for the application of learning algorithms, as they are highly dependent on moment-to-moment updates on the channels.

Some possible solutions
Following are some possible solutions to the problems presented: • Prevention before attack. It aims to make a targeted reinforcement considering networks' vulnerability, which will enhance the system's inborn robustness to attack. For example, selectively improve critical nodes' security levels. In consequence, the attack cost is increased while attack effectiveness declines. • Appropriate counterattack. First of all, offending is coordinated with traditional defence schemes including bad data detection. After a certain period, the knowledge of malicious users' behaviours are gradually obtained and the network develops the capacity of launching counterattack. Here, the counterattack means that users identified as malicious ought to be in certain punishment, which will improve users' responsibility widely neglected by users due to the openness of the underlying protocols.
• The location aware schemes were studied which are capable of countering intelligent attacks taking advantage of the mobility of the PU and the SUs. Intelligent attackers will take advantage of the mobility characteristic to launch more intangible and powerful attack, which deserves considerable attention and is a challenging problem for defence. Besides, both PUs and SUs can be mobile and the problem can be more complex. This requires defence schemes to be location-aware. It will be a common thing in CRNs that the scenarios are time-varying due to, e.g. the mobility of PUs and/or SUs, the variation of attack behaviours and PU activities, etc. Correspondingly, efficient defence schemes should dynamically adapt to time-varying scenarios. Specifically, in mobile CRNs, every SU's sensing environment is time-varying and when evaluating the confidence level of a SU's sensing report, location diversity must be taken into consideration.

DISCUSSIONS AND LIMITATIONS
The work discussed various defence mechanisms for countering SSDF attacks like reputation based mechanisms and statistical test based mechanisms. For countering SSDF attacks, reputation based mechanism depend on the prior information of the environment as they depend on calculating the Likelihood function of the received signal. Also these mechanisms can suppress the colluded attacks only in a sequential manner. Consequently, their computational complexity is increased considerably. Statistical test based mechanisms have a limitation that the SUs transmitting abnormal sensing reports, due to the environmental losses are tagged as outliers. In heterogeneous networks, the presence of MUs affects the correct estimation of cell boundaries. Smaller the cells, higher is the disparity in the reports of the SUs. However, the presence of MUs can result into higher disparity and affects the correct estimation of the cell size. In case of PUEAs, the use of analytical models for the received power for attack detection can lead to large number of samples and long sensing times, which is difficult to achieve in highly dynamic environment. The RSS based mechanisms depend on separate sensor networks which increases the costs related to deployment and maintenance. Also they assume fixed transmission power for the signals transmitted from the MUs. In signature based mechanisms to detect the PU, modification of the PU signals and the need of a certification authority are necessary. Also, it can be affected to the vulnerabilities related to encryption and description of the PU data. In the mechanisms related to the modelling of PUEAs, the position of the attacker the distances between the PU, the SUs and the attacker have to be known in advance.

CONCLUSION
This paper performs a state-of-the-art survey of the defence mechanisms for mitigating security attacks in cognitive radio systems. In the beginning, the preliminaries of CSS have been briefly presented for general readers. Key CR operations like hypothesis testing and different types of CSS have been presented here. Next, the mechanisms of SSDF and PUEAs have been discussed. The different situations and the intentions of the attackers and their effects on CSS are discussed in detail. We showed the effects of SSDF and PUEAs on the throughput of the SU system and on the interference to the PUs. Then, a detailed study on the defence mechanisms for both these attacks was performed. The defence mechanisms were categorized as active and passive, based on the nature of their defence strategy. In a nutshell, active mechanisms directly detect and eliminate the MU, whereas passive mechanisms are based on bringing changes in the underlying operations of the CSS to mitigate the effects of the attacks. Finally, a detailed illustration of the open research problems and some possible solutions to these problems have been presented. It is envisioned that searching for new defence mechanisms for SSDF and PUEAs is a fruitful research area.