See

discussions, stats, and author profiles for this publication at: http://www.researchgate.net/publication/283351905

Identification and Punishment Policies for

Spectrum Sensing Data Falsification Attackers
Using Delivery-based Assessment
ARTICLE in IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY NOVEMBER 2015
Impact Factor: 1.98 DOI: 10.1109/TVT.2015.2497349

READS

53

3 AUTHORS, INCLUDING:
Saud Althunibat

Fabrizio Granelli

Al-Hussein Bin Talal University

Universit degli Studi di Trento

31 PUBLICATIONS 93 CITATIONS

248 PUBLICATIONS 1,025 CITATIONS

SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate,

letting you access and read them immediately.

SEE PROFILE

Available from: Saud Althunibat

Retrieved on: 29 December 2015

Identification and Punishment Policies for

Spectrum Sensing Data Falsification Attackers
Using Delivery-based Assessment
Saud Althunibat
Al-Hussein Bin Talal University, Maan, Jordan
saud.althunibat@ahu.edu.jo
Birabwa J. Denise
Kyambogo University, Kampala, Uganda
denise.birabwa@gmail.com
Fabrizio Granelli
University of Trento, Trento, Italy
granelli@disi.unitn.it

AbstractSpectrum sensing data falsification (SSDF) at-

data to evaluate the made global decision and the individual

tack represents a major challenge for cooperative spectrum

reports. Mathematical analysis and simulation results show

sensing (CSS) in cognitive radio networks. In SSDF attack,

promising performance of both algorithms compared to

a malicious user or more send false sensing results to

previous works especially when then the number of attackers

the fusion center in order to mislead the global decision

is very large.

about spectrum occupancy. Thus, SSDF attack degrades

the achievable detection accuracy, throughput and energy

I. I NTRODUCTION

efficiency of cognitive radio networks. In this paper, a novel

attacker-identification algorithm is proposed that is able to

The increase of wireless services is accompanied with

skillfully detect attackers and neglect their reported results.

an increase in demand for the radio spectrum, a resource

Moreover, we provide a novel attacker-punishment algorithm

that cannot be expanded. Most useful radio spectrum has

that aims at punishing attackers by lowering their individual

energy efficiency, motivating them either to quit sending false
results or to leave the network. Both algorithms are based

This

work is funded by the Research Project GREENET (PITN-GA-

2010-264759).

on a novel assessment strategy of the sensing performance

Copyright (c) 2015 IEEE. Personal use of this material is permit-

of each user. The proposed strategy is called delivery-based

ted. However, permission to use this material for any other purposes

assessment, which relies on the delivery of the transmitted

must be obtained from the IEEE by sending a request to pubspermissions@ieee.org.

November 1, 2015

DRAFT

already been allocated, and thus, it becomes extremely

fading, noisy reporting channel or malfunctioning sensor

hard to find vacant bands for new services. However,

[15]. Such type of CUs is called unintentional attacker

measurements show that licensed spectrum is rarely used

[16]. Never the less, both intentional and unintentional

at full capacity all time by its licensed users [1]. Aiming at

attackers degrade the detection accuracy, which, in turn,

solving the problems of spectrum scarcity and inefficient

influences throughput and energy efficiency of the other

spectrum utilization, cognitive radio (CR) technology has

honest CUs. Therefore, it is of paramount importance to

been proposed [2], [3]. In CR, the unlicensed users, also

eliminate these attackers from the network.

called cognitive users (CUs), can opportunistically utilize

the temporarily unused portions of the licensed spectrum.
CR has enabled and supported many emerging application
[4].
In CR, as an initial step, CUs must sense the spectrum
for available opportunities, to avoid any collision or interference with the licensed users [5]. However, individual
spectrum sensing suffers from shadowing and multipath
fading, leading to degraded performance represented by
inducing interference at the licensed users and inefficient
utilization of the spectrum opportunities [6]. Therefore,
cooperative spectrum sensing (CSS) is proposed to improve the sensing performance [7], [8]. In CSS, all CUs
send their local sensing results, to a central entity, called
fusion center (FC), which combines all results and makes
a global decision about spectrum availability.

The two well-known approaches, Bayesian detection

[17] and Neyman-Person test [18], for signal detection are
no longer optimal in the presence of SSDF attack [19]. In
addition, both approaches require a prior knowledge about
the local sensing performance. Several works have investigated the defense against SSDF attack. For example, [14]
proposes an algorithm to identify attackers by counting the
number of mismatches between each CUs local decisions
and the global decision at the FC. Once, the number of
mismatches exceeds a given threshold, the corresponding
CU will be considered as an attacker, and thus, its reports
will be ignored. This approach however becomes unreliable when the the number of attackers is large, giving an
unreliable final decision. An outlier detection method is
presented in [20], where the report history of each CU is
represented in a high-dimensional space in order to detect

Although CSS improves the reliability of spectrum

any abnormality. A detection scheme is proposed in [21],

sensing process, it introduces extra energy consumption

where it calculates a trust value as well as a consistency

[9], time delay [10] and security threats [11]. In this

value for each CU based on its past reports. Once both

paper we handle the security threat that is so called

values fall below predefined thresholds, the received re-

spectrum sensing data falsification (SSDF) attack [12].

ports from the corresponding CU are no longer considered

SSDF attacker is represented by a CU that sends false

in the fusion process. However, the algorithm is valid

spectrum sensing reports, trying to cause a wrong global

only for one attacker. In [22], an algorithm that involves

decision about spectrum availability at the FC [13]. The

setting randomly distributed evaluation frames is proposed.

motivation of SSDF attackers is to prevent other CUs

In each evaluation frame, the FC decides the spectrum is

from exploiting the spectrum, such that they can increase

free irrespective of the reported local decisions. A CU is

their own transmission opportunities [14]. However, some

then scheduled for data transmission, and depending on

honest CUs may appear like attackers because of their

its success, the actual status of the spectrum is defined,

bad sensing performance caused by either shadowing and

giving the ability for the FC to assess local decisions

November 1, 2015

DRAFT

in that frame and assign to each CU a weight related

obtained spectrum status, all the individual sensing results

to its actual performance. A drawback of this algorithm

can be evaluated accordingly. Based on the evaluated

is that it causes interference to the licensed users during

performance of each CU, attackers can be seamlessly

evaluation frames. Recently, an adaptive reputation-based

detected and removed from the fusion process at the FC.

clustering against collaborative attackers is proposed in

Identifying attackers possess an initial step to alleviate

[23]. It is based on clustering CUs into multiple clusters

their effects on the network performance. However, a

according to the sensing history and the reputation of each

further action should be taken against identified attack-

CU. Such a step separates attackers into one cluster (or

ers in the subsequent data transmission phase. Depriving

more), alleviating their influence on the global decision

attackers of scheduling opportunity in data transmission

since each cluster casts only one vote in global voting at

phase is a bad choice. This is due to the fact that the

the FC. The algorithm is developed to handle different

attacker-identification is an imperfect process, where a

scenarios of collaboration among attackers. Although a

false identification of an honest CU as an attacker is

high performance has been shown, the adaptive clustering,

probable. Moreover, an identified attacker could be an

internal voting, and reputation updating phases may induce

honest CU that suffers from poor sensing performance.

high complexity and consume a significant amount of time

On the other hand, keeping all CUs, honest and attackers,

and energy resources. It is worth mentioning that there are

equal in scheduling probability is unfair with respect to

other promising algorithms against SSDF attacks in non-

the honest CUs. In this paper, we propose a scheduling

centralized networks. For example, authors of [24] and

policy based on assigning a scheduling probability to each

[25] have proposed a biologically-inspired algorithm to

CU related to its sensing performance. For attackers, such

detect attackers in Ad-Hoc cognitive radio networks. The

policy establishes a punishment strategy, where a low

algorithm implies that, after exchanging the sensing results

scheduling probability is assigned to them, and hence,

with the neighbors, each CU should identify the neighbor

reduced individual throughput and energy efficiency. Thus,

with the maximum deviation as an attacker. The algorithm

the proposed punishment policy is aiming at motivating

is iteratively repeated until a consensus is reached.

attackers to quit reporting false reports. On the other

Identifying attackers is a very crucial process that should

be carefully carried out to avoid detecting honest CUs as
attackers. Thus, attacker-identification should be built on

hand, honest CUs will gain proportional fair distribution

of data transmission corresponding to their local sensing
performance.

a reliable base that cannot be affected if the number of

Although the considered setup is challenging, as it will

attackers is large. In this paper, we consider the delivery of

be described later, both proposed policies show promising

the transmitted data as a base of evaluating the individual

results even in the worst case scenario where the number

performance, and consequently, identifying attackers. No-

of attackers is very large. Mathematical analysis and

tice that in infrastructure-based cognitive radio networks,

simulation results explore the significant improvement in

the data transmission is performed through the base station

the overall performance achieved by the proposed policies

(BS) [26]. Thus, it is easy to ensure if the transmitted

compared to previous works. The contributions of this

data is successfully delivered or not, and hence, the actual

work can be summarized as follows

spectrum status will be known at the FC. Using the

November 1, 2015

Introducing data delivery as a base for evaluating the

DRAFT

performance of the individuals in infrastructure-based

the identification and punishment policy has been

CRNs. Delivery-based assessment is a novel strategy

presented in terms of the detection accuracy and

and has never been proposed before to the best of our

energy efficiency.

knowledge.

Proposing a novel attacker-identification algorithm

that is able to skillfully detect attackers and com-

A very related work is [14]. However, there are several

differences that should be highlighted as follows

pletely eliminate their influence on the CRN.

The work in [14] presents an identification algorithm

for attackers by evaluating their sensing performance

Proposing attacker-punishment algorithm that is

based on the majority decision. Such an algorithm can

based on lowering the energy efficiency of the at-

work well in presence of a low number of attackers.

tacker, motivating it either to quit attacking or to leave

However, when the number of attacker is large, the

the CRN.

reliability of majority decision is highly degraded

The initial idea of this paper has been proposed earlier

as the majority is attackers. Such a drawback has

in our work [27]. However, besides the expanded literature

motivated us to find an alternative evaluation base

review, introduction, and motivations, there are several

rather than the majority decision. Thus, in our work,

differences/increments over our previous work [27], sum-

the data delivery has been used in order to assess

marized as follows

the sensing performance of users. Employing data

The proposed identification policy in [27] is based on

delivery in such a purpose is a novel contribution that

instantaneous check, while in this paper the mismatch

should be accounted for our work. Employing data

counters are checked after T sensing rounds. Such a

delivery has shown very good performance results

difference results in a totally different performance

even in the case of the large number of attackers

between the two polices.

(worst-case scenario).

In this paper, an extensive mathematical analysis

The optimization of the removal (ignoring) threshold

of performance of the proposed identification and

in [14] has not yielded a closed form expression of the

punishment polices has been presented, while [27]

optimal threshold, while a closed form mathematical

lacks the mathematical analysis.

expression has been presented in our work of the

Unlike this paper, the optimization of the identi-

optimal removal threshold that maximizes the differ-

fication threshold has not been addressed in [27]

ence between the ignoring probability of attackers and

neither mathematically nor by simulations. Moreover,

honest users.

the worst-case scenario has been investigated in the

The work in [14] is only an identification algorithm,

current work for both: the identification algorithm and

while our work includes a punishment policy for at-

the punishment policy

tackers. Punishing attackers by lowering their energy

Simulation results in [27] have been focused on the

efficiency is a novel contribution has not be presented

energy efficiency performance of the attacker/ honest

before. The mathematical and simulation results have

users. It means that the attention was mostly paid

proved the effectiveness of the proposed punishment

for the punishment policy performance. However,

policy.

in the current work, a detailed evaluation of both

November 1, 2015

The rest of the paper is organized as follows: SecDRAFT

tion II describes the system model and the attacker model,

idle spectrum as used.

followed by the employed evaluation metrics, while sec-

As CSS demands, all CUs report their local decisions to

tion III presents the proposed delivery-based assessment

the FC which combines and issues a final decision about

approach. The proposed attacker-identification algorithm

spectrum occupancy according to a specific fusion rule

is discussed in IV along with the necessary mathematical

(FR). The general FR for binary local decisions is called K-

framework and the analysis of the worst case scenario.

out-of-N rule [30]. Based on this FR, if the number of local

Section V proposes the attacker-punishment algorithm.

decisions of 1 is larger or equal to the threshold K, the

Performance evaluation and simulation results are pre-

global decision should be 1 (used). Otherwise, the global

sented in Section VI, while conclusions are drawn in

decision is 0 (unused). If we denote the local decision

Section VII.

in the ith round by un,i , then the global decision of that

II. S YSTEM M ODEL

Consider a CRN consisting of N CUs cooperating

in order to opportunistically access the licensed spectrum whenever it is free. The CRN is considered of
infrastructure-based type [13], where the CSS and data
transmission is coordinated by the BS. An example of
such network is IEEE 802.22 [28]. The adopted CR model
in this paper is Interweave model, where both CUs and
licensed users coexist on the same geographical area,
and CUs can use the spectrum only if it is unoccupied
by the licensed users [29]. For the sake of simplicity,
the licensed spectrum is modeled as a single channel

round Ui is made as follows

1 used,
Ui =

0 unused,

if

PN

if

PN

n=1

un,i K

(1)

n=1 un,i <K

Three popular FRs are derived for this rule, namely,

OR-rule (K = 1), AND-rule (K = N ) and Majority
rule (K = N/2) [31]. Similar to the local decision, the
reliability of the final decision is measured by two metrics,
the overall detection probability (PD ) and the overall falsealarm probability (PF ). Both are defined as at the local
level but regarding the final decision rather than the local
decision. Both PD and PF can be combined to describe
the global detection accuracy in one metric called error
probability (Pe ) given as follows [30]

although it can be easily extended to multiple-channel

scenario. In each CSS round, each CU senses the licensed

Pe = P0 PF + P1 (1 PD )

spectrum and, depending on its sensing result, it solves

where P0 and P1 are the probabilities that the spectrum is

a hypothesis testing problem deciding on one of two

unused or used, respectively.

(2)

hypotheses, either H0 that implies spectrum is unused, or

Upon issuing the final decision, a CU will be scheduled

H1 for spectrum is used. It then reports its binary local

for data transmission only if the final decision is unused,

decision un = {1 used, 0 unused} to the FC

while in case of identifying the spectrum as used, the

that is located at the BS.

FC will not schedule any of the CUs in order to avoid

The reliability of the local decision of a CU is evalu-

interference to the licensed users.

ated by two indicators: local detection probability (Pdn )

and local false-alarm probability (Pdf ). While the former

A. Attacker Model

represents the probability of identifying a used spectrum

As in other wireless networks, CRNs are usually vul-

as used, the latter denotes the probability of identifying an

nerable to different security threats. One of these threats,

November 1, 2015

DRAFT

which is not typical in the other wireless networks, is

SSDF attack. In SSDF attack, a malicious CU sends false
reports about the spectrum availability to the FC in order
to mislead the final decision. The motivation behind such
attack is to exploit the spectrum holes for their own
transmission. To satisfy this motivation, the optimal attack
strategy is to always report the spectrum as used, also
called Always-Yes attack [32]. However, such strategy is
easy-to-detect at the FC. Thus, smarter attackers usually
follow a different strategy in order to elude the FC and

Fig. 1.

avoid detection and negligence. The smart strategy is based

attackers

An example of a cognitive radio network in presence of SSDF

on inverting the actual local sensing result in a selective

manner. Specifically, an attacker decides in each CSS
round to attack or not with a probability, denoted as Pm . If

attackers will cooperate among themselves to make their

the attacker decides to attack in a specific round, it simply

own global decision based on their honest performance.

flips its own local decision and reports it to the FC. Such

Accordingly, once a false-alarm occurs at the FC, if their

attacker model is usually termed as Byzantine attackers

own global decision does not agree with the decision of

[32][34]. The sensing performance, i.e., Pdn and Pf n , of

the FC, the attackers will select one of them randomly

an attacker as it appears at the FC based on such strategy

to transmit its own data individually. From now on, we

can be mathematically modeled as follows [14]

denote the detection and false-alarm probabilities of the

Pdn = Pm (1

ac
Pdn
)

Pf n = Pm (1

Pfacn )

A
and PFA , respectively.
global decision of attackers by PD

+ (1

ac
Pm )Pdn

(3)

+ (1

Pm )Pfacn

(4)

The following steps summarize the function of the

ac
where Pdn
and Pfacn represent the actual (honest) detection

attacker model considered in this paper

1) At each sensing round, all attackers will sense the

and false-alarm probabilities, respectively. Notice that this

spectrum (as the honest users do), and each attacker

model is valid for an honest CU if we set Pm to zero.

will individually make a local decision regarding the

For the sake of simplicity, let us assume that all honest CUs are identical in their sensing performance, i.e.,
Pdn = Pdh and Pf n = Pf h . Likewise, the attackers are
considered to have identical performance, i.e., Pdn = Pda
and Pf n = Pf a .

spectrum occupancy.
2) Each attacker will individually decide to send a false
report or not (attack or not) with a probability Pm .
a) If an attacker has decided to attack, it will
invert its local decision and report it to the FC.

Since the main motivation of attackers is to increase

b) Otherwise (if the attacker has decided not to

their achievable throughput while degrading the through-

attack), it will send its actual (honest) local

put of the honest CUs, the attacker will exploit the case of

decision to the FC.

false-alarm to perform individual transmission without co-

3) Directly, attackers will share their actual (honest) lo-

ordination from the BS. Specifically, we consider that the

cal decisions, and decide internally a global decision

November 1, 2015

DRAFT

(let us call it the global attackers decision).

more energy consumption. Therefore, for evaluation pur-

4) If the FC has made a global decision that the

pose, we use the individual energy efficiency of the CU

spectrum is unused, one of the users (it could be

as comparison metric between attackers and honest CUs.

an attacker) will be scheduled for data transmission

Individual energy efficiency of a CU is defined as the

in this round.

ratio of the individual throughput achieved in bits to

5) If the FC has made a global decision that the

the individual energy consumed in Joule. According to

spectrum is used, then attackers will check their own

the considered setup, it is expected that the individual

global decision (global attackers decision). If it is

achievable throughput, the individual energy consumption

different from the global decision of the FC, one of

and the individual energy efficiency will be different for

the attackers will be scheduled for data transmission

an honest CU and an attacker.

in this round.
Notice that the cooperation among attackers assumed
in this paper is different from other assumptions in the

C. An example

literature. The cooperation assumed here includes sharing

Let us consider a CRN of 5 honest CUs with identical

the local decisions among attackers in order to exploit the

detection and false-alarm probabilities equal to 0.8 and

spectrum hole missed by the FC, if any. Other assumptions

0.1, respectively. The final decision is made based on

may imply sharing the local decisions before reporting

Majority rule. In Fig. 2, we plot the effects on the detection

them to the FC, aiming at deciding if local decisions

accuracy and the achievable throughput if a number of

should be changed or not [23].

attackers has joined the CRN. The local detection and

false-alarm probabilities of attackers are identical and

B. Throughput and Energy Efficiency

equal to 0.1 and 0.8, respectively. Fig. 2-a shows the error

According to the considered CRN model, an honest CU

probability of the final decision, as an indicator of the

has the chance to transmit only if it has been legitimately

detection accuracy, versus the number of joined attackers,

scheduled by the FC. On the other hand, an attacker can

while Fig. 2-b shows the achievable throughput of an

get a transmission opportunity in two cases: (i) if it has

attacker and an honest CU versus the number of joined

been legitimately scheduled by the FC, and (ii) if it has

attackers. The achievable throughput is divided into two

been selected by the other attackers to transmit in the

parts: legitimate throughput resulting from scheduling by

case of a false-alarm at the FC. We call the achievable

the BS, and illegitimate throughput achieved by individual

throughput in the first case the legitimate throughput, while

transmission without coordination of the BS. Clearly, the

the illegitimate throughput is the throughput achieved in

increase in the error probability and the degradation in

the second case.

the achievable throughput of honest CUs increase as the

Notice that increasing the false-alarm probability, which

number of attackers increases. On the other hand, the

is a result of SSDF attackers, will increase the illegit-

throughput of attackers increases due to the high false-

imate throughput of attackers, which, in turn, degrades

alarm probability that they can cause. Such a simple exam-

the achievable throughput of the honest CUs. However,

ple explores the importance of encountering the attackers

increasing the throughput is always accompanied with

in CRNs.

November 1, 2015

DRAFT

8
(a)

Error probability (Pe)

10

mission phase in order to formalize a performance indicator for each CU, which can be further employed to identify

10

attackers and honest CUs. The reader should note that

considering data delivery as an evaluation base is much

10

10

The number of attackers (M)

more reliable than the global decision even in the case of

(b)

2500

Legitimate throughput of an honest CU

Legitimate throughput of an attacker
Illegitimate throughput of an attacker

Throughput [bits]

2000

large number of attackers.

1500

From implementation point of view, the delivery-

1000

based assessment approach can be easily applied in

500

infrastructure-based CRNs with a BS coordinating the

4
6
The number of attackers (M)

10

data transmission, as assumed in this paper. However, for

Fig. 2. Example: (a) the error probability versus the number of attackers,

centralized CRNs without a BS, where CUs individually

and (b) the throughput versus the number of attackers

access the spectrum, the data delivery can be verified by

an additional monitoring process during data transmission

III. D ELIVERY- BASED A SSESSMENT

performed by the FC itself or another delegated trusted

CU. Notice that the monitoring process is much easier than

Most of the previous work depends either on a prior

spectrum sensing since the transmitting user is known at

knowledge about the local performance of the CUs or the

the FC. Another option which can verify the data delivery

final decision reliability to detect attackers and remove

is requesting a feedback from the scheduled CU. However,

them. The prior knowledge is not always available and

it should be taken into account the probability that the

the global decision lacks reliability in the presence of large

scheduled CU is an attacker providing false feedback. In

number of attackers. Instead, in this paper, we propose a

order to avoid any induced drawback in the delivery-based

novel approach that can seamlessly evaluate the sensing

assessment approach, we consider only infrastructure-

performance of each CU, and consequently, identify at-

based CRNs in this paper, which has been widely adopted

tackers. The proposed approach is based on the delivery

in the literature [26], [35][40], while the applicability of

of the transmitted data of the scheduled CU. Specifically,

delivery-based approach on other mentioned CRN types is

if the licensed channel has been decided as unused and

left as future work.

one of the CUs has been scheduled for data transmission,

the successful delivery of the transmitted data reveals that
the global decision was correct and the channel is actually
unused. In the other case, if the transmitted data cannot be
successfully delivered, the global decision is identified as
incorrect and the channel is actually occupied. Notice that
in both cases, the FC has doubtlessly realized the actual
channel status, which can be used to assess all the received
local decisions as correct or not.
Delivery-based assessment continues in each data transNovember 1, 2015

In the following sections, we describe two novel policies, attacker-identification policy and attacker-punishment
policy. Both of them are developed based on the deliverybased assessment approach. While attacker-identification
policy aims at detecting attackers and ignoring their reported local decision in the fusion process, the attackerpunishment policy is basically a scheduling policy that
leads to a proportional resource distribution according to
the evaluated individual performance of each CU. Such a
fair scheduling policy acts as a punishment for attackers
DRAFT

after T rounds Bn,T follows a binomial distribution func-

and a reward for honest CUs.

tion, as follows
IV. ATTACKER -I DENTIFICATION P OLICY

P rob.{Bn,T = b} =

 
T b
(1 n )T b
b n

(6)

Attacker-identification is a key factor to improve the

where b = 0, 1, 2..., T , and n denotes the probability that

overall performance of the CRNs either in terms of detec-

the counter B will be incremented by one (the probability

tion accuracy or energy efficiency. Attacker-identification

that the local decision of nth user is wrong given that

should be carefully carried out in order to avoid incorrectly

the global decision is unused), which can be derived as

identifying honest CUs as attackers. Once an attacker is

follows

identified, it should be removed from the fusion process

at the FC, where its reports should be ignored. In this

n = P (Bn,i = Bn,i1 + 1)
= P (H0 un,i = 1 Ui = 0)

section we propose a novel attacker-identification policy

that is able to identify the attackers whatever their number

+ P (H1 un,i = 0 Ui = 0) (7)

in the network is.

The proposed policy is based on assessing the local
decisions according to the delivery of the transmitted
data of the scheduled CU. In detail, once the spectrum
is identified as unused, a CU will be scheduled for

Using the following theorem on conditional probability

[41]
P (A1 A2 A3 ) = P (A1 )P (A2 |A1 )P (A3 |A1 A2 ) (8)
the first term in (7) can be expanded as follows

data transmission. Consequently, based on the success of

delivering the transmitted data, the actual spectrum status
can be correctly defined and used to evaluate the local
decisions. Thus, the local decisions reported in that round

P (H0 un,i = 1 Ui = 0) =
= P (H0 )P (un,i = 1|H0 )P (Ui = 0|un,i = 1 H0 )
= P0 Pf n P (Ui = 0|un,i = 1H0 )

can be classified false or correct. If the local decision

(9)

is false, a corresponding counter will be incremented by

one. After a sufficient amount of time, let us say T CSS
rounds, if a counter of a specific CU exceeds a predefined
threshold, it will be considered as an attacker, and hence,
its reports will be ignored at the fusion process.

Likewise, the second term in (7) can be expanded as

follows
P (H1 un,i = 0 Ui = 0) =
= P (H1 )P (un,i = 0|H1 )P (Ui = 0|un,i = 0 H1 )

Following the proposed policy, a zero-initialized

counter, denoted by Bn,i , for each CU is updated each
CSS round as follows

Bn,i1 + 1, if Ui = 0 & Si 6= un,i

Bn,i =

Bn,i1 ,
Otherwise

= P1 (1 Pdn )P (Ui = 0|un,i = 0 H1 )

(10)
by substituting (9) and (10) in (7), n can be rewritten as

(5)

where the subscript n refers to the CU index, the subscript

follows

n = P0 Pf n P (Ui = 0|un,i = 1 H0 )

i refers to the sensing round index, and Si represents the

actual status of the spectrum. The final value of the counter
November 1, 2015

+ P1 (1 Pdn )P (Ui = 0|un,i = 0 H1 ) (11)

DRAFT

10

The probability n can be found for an honest CU,

which can be simplified using the binomial law as follows

denoted by h , by substituting the following probabilities

Bn,T = T n

(18)

in (11)
Also, if we denote the ignoring threshold by , the
ignoring probability of the nth CU can be expressed as
P (Ui = 0|un,i = 1 H0 )
1

N
1
X

a2
X

follows:

honest

f (j, M, Pf a )f (k j, H 1, Pf h ) (12)

Pign,n P rob.{Bn,T } =

b=

k=K1 j=a1

bn (1 n )T b
(19)

P (Ui = 0|un,i = 0 H1 )
a2
N
1 X
X

T  
X
T

Accordingly, the average number of the remaining CUs

=
honest

after T CSS rounds, i.e. those CUs that have not been

f (j, M, Pda )f (k j, H 1, Pdh ) (13)

ignored, can be given as follows

k=K j=a1

where a1 = max(0, k H + 1), a2 = min(k, M ), H is

the number of honest CUs, M is the number of attackers

NT = N

N
X

Pign,n = H(1 Pign,h ) + M (1 Pign,a )

n=1

(20)

and the function f (, , ) denotes the binomial function

where Pign,h and Pign,a are the ignoring probabilities for

[41], as follows

an honest CU and an attacker, which can be obtained by

 

f (, , ) =
(1 )

(14)

By the same way, the probability n can be found for

an attacker, denoted by a , by substituting the following

substituting h and a instead of n in (19), respectively.

A. Optimizing of
It is worth noting that has a significant role in the

probabilities in (11)

proposed policy. Low values of may result in identifying

some honest CUs as attackers, whereas some attackers can
P (Ui = 0|un,i = 1 H0 )
1

N
1
X

a4
X

not be identified at high values of . Therefore, should be

attacker

carefully optimized. An approach to optimize the threshold

f (j, M 1, Pf a )f (k j, H, Pf h ) (15)

k=K1 j=a3

is to maximize the difference between the ignoring

probability of attackers and the ignoring probability of

P (Ui = 0|un,i = 0 H1 )
1

a4
N
1 X
X

honest CUs. Mathematically, the maximization problem

attacker

can be expressed as follows

f (j, M 1, Pda )f (k j, H, Pdh ) (16)
max Pigna Pignh

k=K j=a3

where a3 = max(0, k H), a4 = min(k, M 1).

Now, from (6), the average value of Bn,T of the nth

T
X

b P rob.{Bn,T = b} =

b=0

T
X
b=0

November 1, 2015

 
T b
(1 n )T b
b n

by substituting the values of Pigna and Pignh using (19),

the maximization problem can be rewritten as follows
T  
T  
X
X
T b
T b
max
a (1 a )T b
(1 h )T b

b
b h
b=
b=
(22)

CU, denoted by Bn,T , can be derived as follows

Bn,T =

(21)

(17)

The optimal value of can be computed using Lagrange

method, where the derivative of the function with respect
DRAFT

11

to is equalized to zero. Since is an integer, the

rule as the employed FR, notice that both probabilities

derivative of Pigna and Pignh are respectively given as

given in (11) can be respectively approximated in such

follows
Pigna

Pignh

scenario as follows
 
T
= Pigna (+1)Pigna () =
(1a )T
(28)
P (Ui = 0|un,i = 1 H0 ) 0
a
wc
(23)
 
(29)
P (Ui = 0|un,i = 0 H1 ) 1
T
wc
= Pignh (+1)Pignh () =
h (1h )T

(24) These approximations are valid since in the case of M 

Accordingly, the first derivative of the function under

H the probability of making a correct final decision (as

optimization in (21) can be given as follows

in (28)) is almost absent and the probability of making a

(Pigna Pignh ) =

 
 
T
T

a (1 a )T +
(1 h )T = 0

false final decision (as in (29)) is almost one.

Now, by substituting (28) and (29) in (11), the probabilities h and a can be computed as follows

(25)
The binomial coefficients can be canceled, and the equation can be rearranged as follows

 
T
a (1 h )
1 h
=
h (1 a )
1 a

wc

wc

P1 (1 Pdh )

(30)

P1 (1 Pda )

(31)

Consequently, since Pdh 1 and Pda 0, then h

(26)

0 and a P1 . Using (19), it is easy to show that Pignh

Now, by applying the natural logarithm to both sides, the

0 while Pigna is still high, and hence, attackers can be

optimal values of the ignoring threshold that maximizes

easily detected with a proper choice of even in the worst-

the difference between the ignoring probabilities of attack-

case scenario.

ers and honest CUs, denoted by , can be given as follows



h
ln 1
1a

= T
(27)


h)
ln ah(1
(1a )

)
(wc

where de is the ceiling operator that should be applied to

to make it an integer.
B. Worst-case scenario

The optimal ignoring threshold in the worst-scenario

can be also approximated by substituting (30) and

(31) in (27), as follows



1 Pdh
ln PP00 +P
+P1 Pda


wc T 
da )(P0 +P1 Pdh )
ln (1P
(1Pdh )(P0 +P1 Pda )

(32)

V. ATTACKER -P UNISHMENT P OLICY

For the purpose of exploring the high performance of

Ignoring the reports received from the CUs identified

the proposed attacker-identification policy, we consider the

as attackers helps to improve the overall performance of

worst case scenario. The worst case scenario is represented

the network. However, a false identification is probable,

when a large number of attackers is present confronted by

where some honest CUs might be identified as attackers

a low number of honest CUs (i.e., M  H).

by error. Moreover, as stated earlier, not all of attackers

The performance can be clearly shown in terms of the

intentionally send false reports to the FC. Some honest

ignoring probability of attackers and honest CUs. From

CUs suffer from multi-path fading and shadowing during

(19), the ignoring probability of a CU mainly depends on

sensing or noisy reporting channels, leading to a bad

its corresponding n probability. Considering the majority

sensing performance. This type of honest CUs will appear

November 1, 2015

DRAFT

12

like attackers at the FC side. Thus, depriving CUs that are

identified as attackers from data transmission represents

be identified as unused at the FC, expressed as follows

P (Ui = 0) = P0 (1PF )+P1 (1PD ) = 1P0 PF P1 PD

a harmful action towards the unintentional attackers. On

the other hand, providing the same transmission chance
among all CUs does not attain fairness from honest CUs
point of view. Instead, in this section, we provide a novel

(36)
Consequently, using the average value of Bn,i given in
(18), the average value of Psn at the ith round can be
easily derived as follows

scheduling policy that distributes the spectrum resources

among CUs in a proportional fair manner. The proposed

Psn = PN

j=1

scheduling policy allocates scheduling probability to each

attackers, while it provides a fair resource distribution for

honest CUs.

i P (Ui = 0) i j

=

P (Ui = 0) n
PN
N P (Ui = 0) j=1 j

CU based on its sensing performance that appears at

the FC. Such policy can be deemed as punishment for

i P (Ui = 0) i n

(37)

The reader should note that the computation of P (Ui =

0) and n before T are different from those after T . This
is because after T , some of the users will be identified as

The proposed policy is also based on delivery-based as-

attackers, and hence, their reports will be ignored while

sessment as in the proposed attacker-identification policy.

making the global decision at the FC. Also, it is worth

Therefore, the assigned scheduling probability for each CU

mentioning that scheduling probabilities are computed

depends on the instantaneous value of the counter B. The

based on the accumulated counters B and x which should

scheduling probability of the nth CU is computed each

be kept updated as long as the CRN lasts.

CSS round as follows

According to the proposed punishment policy, the aver-

xi Bn,i
Psn = PN
j=1 (xi Bj,i )

(33)

age achievable throughput for an honest CU, denoted by

Dh , can be expressed as follows

where xi represents the number of times in which the

spectrum was identified as unused by the final decision
until the ith CSS round, expressed as follows

xi =

xi1 + 1,

if Ui = 0

xi1 ,

if Ui = 1

Dh = P0 (1 PF )R Tt Psh

(38)

where R is the data rate, Tt is the transmission time and

Psh is the average scheduling probability for an honest
CU. The factor P0 (1PF ) represents the case of no false-

(34)

alarm at the FC. On the other hand, the average achievable

throughput for an attacker, denoted by Da , is divided into

According to (33), an increase in the counter Bn,i for

a CU implies a magnified punishment through reducing
th

the scheduling probability. At i

CSS round, the value of

xi follows a binomial distribution, where its average value

can be given as follows
xi = i P (Ui = 0)

two parts; legitimate and illegitimate, and can be expressed

as follows
Da = P0 (1PF )RTt Psa +P0 PF (1PFA )RTt (1/M )
(39)
Notice that the first term (legitimate throughput) is iden-

(35)

tical to the honest CU except the difference in the

scheduling probability, while the second term includes

where P (Ui = 0) is the probability that the spectrum will

November 1, 2015

the illegitimate throughput. The factor P0 PF (1 PFA )

DRAFT

13

represents the case that a false-alarm occurs at the FC

Case I : i T : as the number of attackers is very large,

and no false-alarm made by the attackers global decision.

then both PD and PF approximately equal to 0 and 1,

Likewise, the average energy consumption for an honest

respectively. Substituting that in (36), it can be simplified

as follows

CU, denoted by Eh , is expressed as follows

Eh = ess + P (Ui = 0)et Psh

P (Ui = 0)

(40)

where ess and et are the energy consumed in spectrum

sensing and data transmission, respectively. For an attacker, the average energy consumed (Ea ) is given as

P1

(43)

wcI

Using (43) and the approximated values of h and a ,

given in (30) and (31), the scheduling probability for an
honest CU in the worst-case scenario before removing
identified attackers can be approximated as follows

follows
Ea = ess + P (Ui = 0)et Psa + P0 PF (1 PFA )


A
+ P1 PD (1 PD
) et (1/M ) (41)
where the first, the second and the third terms refer
to the energy consumed in spectrum sensing, legitimate
transmission and illegitimate transmission, respectively.

P1 P1 (1 Pdh )
N
P

M
P
1
1 (1 Pda ) HP1 (1 Pdh )
wcI
Pdh

M Pda + HPdh
(44)

Psh 

Likewise, the scheduling probability for an attacker in

the worst-case scenario before removing the identified
attackers can be approximated as follows

As a comprehensive metric, the individual energy efficiency can be introduced as the ratio of the average
achievable throughput to the average energy consumption,
as follows

Psa 

wcI

Pda
M Pda + HPdh

(45)

As Pdh is usually much larger than Pda , the scheduling

probability for an honest CU should be larger than an

D
=
E

(42)

It is obvious from the proposed attacker-punishment

policy that an attacker will be punished by reducing its
scheduling probability that yields in lowering the achievable throughput, and consequently poor energy efficiency.
Such punishment can generate a reaction at the attacker
side if its energy efficiency falls below a specific threshold.
The expected reaction is represented by either leaving the
CR or quiting attacking and switching to an honest mode.

attacker according to (44) and (45).

Case II: i > T : The analysis of this case is different
form the previous one since the ignored attackers are no
longer affecting the global decision. For sake of simplification, we consider that all attackers have been removed
and none of the honest CUs are incorrectly removed.
This assumption is reasonable and can be attained by
the proposed attacker-identification policy with a proper
adjustment of . Also, we consider that the CRN contains
a sufficient number of honest CUs that can attain high

A. Worst-case scenario

global detection probability ( 1) and low global falsealarm probability ( 0) after removing attackers. By

Considering the worst case scenario (M  H), the

analysis can be divided into two cases, (i) before removing

applying these assumptions to (11) and (36), the following

approximations can be obtained

the identified attackers (i T ), and (ii) after removing the

identified attackers (i > T ):
November 1, 2015

P0 P f h

(46)

wcII

DRAFT

14

P0 Pf a

TABLE I

(47)

S IMULATION PARAMETERS

wcII

P (Ui = 0)

P0

(48)

Parameter

Value

Parameter

Value

wcII

P0

0.5

64 Kbps

However, these approximations cannot be directly ap-

Pdh

0.8

Tt

0.3 sec

plied to (37) since the counters are affected by the

Pf h

0.1

ess

11 mJ

first case (i T ). Instead, it can be applied to (33)

Pda

0.1

et

0.5 J

Pf a

0.8

FR

Majority

with taking into account the effect of the first case.

Accordingly, the scheduling probability for an honest CU
in the worst-case scenario after removing the identified
attackers can be seamlessly obtained by subtituting the

A. Attacker-Identification Policy

above approximations in (37). The reader can notice that

The probability of incrementing the Bn counter, n ,

the scheduling probability for an honest CU is larger than

plays a key role in the proposed attacker-identification

the scheduling probability for an attacker since Pdh > Pda

policy. Fig. 3 plots n for honest CUs and attackers versus

and Pf h < Pf a .

the total number of attackers present in the CRN. The

large difference between h and a even for the whole

VI. P ERFORMANCE E VALUATION AND S IMULATION

range of M is due to the reliable evaluation base, i.e., the

R ESULTS

data delivery, by which the counters are updated. Notice

that even in the case of large number of attackers, the

This section provides a comprehensive evaluation of the

honest CUs still have low probability of incrementing their
two proposed polices. Particularly, we show the perforcounters compared to the attackers. The initial fluctuation
mance of the proposed attacker-identification policy comin both curves is due to the FR and odd-even of the
pared to the proposed policy in [14]. Briefly, the proposed
attacker-identification in [14] has the same procedure as
ours except that the evaluation is based on the agreement
with the global decision taken at the FC. Regarding
the proposed attacker-punishment policy, as there is no
similar policy in the literature, we explore the performance
by comparing the individual energy efficiency between

total number of CUs (N ). For example, at M = 2 and

M = 3, the total numbers of CUs are N = 7 and
N = 8, respectively, while the FR in both cases is
K = 4. However, the induced fluctuation diminishes as
M increases. Another important note is on the range of
M  H, where both h and a stay constant and to
the values obtained in (30) and (31), respectively, which

attackers and honest CUs.

verifies the approximations we made in the worst-case
A CRN of a fixed number of honest CUs (H = 5)
is considered. The number of attackers is left variable in
order to show its influence on the different system parameters and probabilities. The simulation parameters regarding
the licensed spectrum occupancy, energy consumption, and
local sensing performance are kept fixed as shown in Table
I. Other parameters which differ among figures are listed

scenario.
The ignoring probability of attackers and honest CUs
versus the ignoring threshold for the proposed policy and
[14] is shown in Fig. 4 at M = 1 and in Fig. 5 at
M = 10. In both figures and for both types of CUs,
the ignoring probability is a decreasing function of .
Considering our proposal in both figures, at low values of

in the caption of the corresponding figure.

November 1, 2015

DRAFT

15
1

0.7

for an honest CU ( ) (Simulation)

h

0.9

for an attacker ( ) (Simulation)

a

0.6

for an honest CU ( ) (Theory using Eq.(11))

h

0.8

for an attacker ( ) (Theory usig Eq.(11))

proposed policy

proposed policy

ignh

0.7

0.5
Ignoring probability

CounterIncrementing probability

0.4

0.3

igna

Pignh proposed in [14]

0.6

Pigna proposed in [14]

0.5
0.4
0.3

0.2
0.2
0.1

0.1

0
0

10

15

20

10

The total number of attackers (M)

Fig. 3. The counters incrementing probability for honest CUs (h ) and

15

20

25

30

Ignoring threshold ()

25

Fig. 4.

The ignoring probability for honest CUs and attackers versus

the ignoring threshold (). (T = 30, M = 1)

attackers (a ) versus the total number of attackers (M ). (T = 30)

(sensitive). For example, as shown in Fig. 4, can take the

(less than 3), both attackers and honest users have a high

values from 4 to 9 while keeping the ignoring probability

ignoring probability. This is because is low, which is the

of an attacker above 90% and the ignoring probability of

number of mismatches, and any normal user can exceed

an honest user is less than 10%.

it. At high values of (more than 15), both attackers and

1

honest users will not be able to exceed the threshold, and

0.9

thus, they will not be ignored. At medium values of ,

0.8

which is the critical range, honest users will not exceed it,

notice that when the honest CUs represent the majority,

Fig. 4, both policies present a good performance and all

Pignh proposed in [14]

0.5
0.4
0.3

honest CUs when is properly adjusted. However, when

0.2
0.1

the attackers pose the majority of the CUs, Fig. 5, the

ignoring probability of honest CUs is more than that of

proposal is still able to provide Pigna = 1 and Pignh = 0

Pigna proposed in [14]

0.6

attackers can be identified without ignoring any of the

the attackers in the policy proposed in [14], whereas our

Pigna proposed policy

0.7
Ignoring probability

while attackers will exceed the ignoring threshold. Also,

Pignh proposed policy

10

15

20

25

30

Ignoring threshold ()

Fig. 5.

The ignoring probability for honest CUs and attackers versus

the ignoring threshold (). (T = 30, M = 10)

with a proper choice of . This is due to the fact that the

One of the major problems of attackers is increasing

global decision is used in [14] as evaluation base, which is

the interference at the licensed users, which is caused by

mainly affected by the majority of CUs, while our proposal

increasing the missed-detection probability at the global

is approximately unaffected by the majority of CUs.

decision. In Fig. 6, we show the performance of the pro-

An interesting property of the proposed policy is that the

posed attacker-identification policy in terms of the missed-

proper is not only one value, whereas it can take a wider

detection and false-alarm probabilities versus the ignoring

range. In other words, the selection of is not very critical

threshold . It can be noted that the missed-detection

November 1, 2015

DRAFT

16
1

can be hugely reduced by employing the proposed policy.

0.8

alarm probability since it represents an important perfor-

0.7

mance metric. Fortunately, our proposal can achieve a very

0.6

Pigna Pignh

However, an eye should be kept on the resulting false-

low missed-detection probability, and simultaneously, keep

T=10
T=20
T=30

0.9

0.5
0.4

a low false alarm probability for a wide range of (from

0.3

4 to 11). Also, the superiority of our proposal with respect

0.2

to [14] is evident, which proves the high performance of

0.1

the proposed policy even if the attackers represent the

0

majority.

10

15

20

25

30

Ignoring threshold ()

Fig. 7.

The difference between ignoring probability for attackers

(Pigna ) and honest CUs (Pigna ) versus the ignoring threshold () for

different values of T . ( M = 1)

0.9
0.8

25

0.7
0.6
0.5

Falsealarm probability Proposed

Misseddetection probability Proposed
Falsealarm probability [14]
Misseddetection probability [14]

0.4

Average no. of ignored honest CUs Proposed

Average no. of ignored attackers Proposed
Average no. of ignored honest CUs Proposed in [14]
Average no. of ignored attackers Proposed in [14]

20

0.3

15

0.2
0.1
0

Fig. 6.

10

12
14
16
18
Ignoring threhold ( )

20

22

24

26

28

30

10

The missed-detection and false-alarm probabilities versus the

5

ignoring threshold () for the proposed attacker-identification policy and

the proposal of [14]. (T = 30, M = 10)

10

15

20

25

Number of attackers (M)

The difference between the ignoring probabilities for

attackers and honest CUs, which is used as optimization
objective, is shown versus at different durations of the

Fig. 8.

The average number of ignored honest CUs and attackers at

the optimal ignoring threshold ( ) versus the total number of attackers

(M ) for the proposed attacker-identification policy and the one proposed
in [14]. (T = 30, = )

evaluation time window (T ) in Fig. 7. The curve show a

convex shape that achieves its maximum at the optimal

attacker. On the other hand, the proposal of [14] works

ignoring threshold ( ).

well only when the majority of CUs are honest. In the

From Fig. 4, Fig. 5 and Fig. 7, the importance of

case of the majority being attackers, the proposal of [14]

optimizing is clear. Thus, we use the optimal that

either identifies all CUs as attackers or does not identify

maximizes the difference between Pigna and Pignh for the

any of the CUs as attackers.

two policies to find the number of ignored attackers and

honest CUs versus the total number of attackers, as shown

B. Attacker-Punishment Policy

in Fig. 8. Regarding our proposal, almost all attackers can

As we have shown the performance of the proposed

be identified whatever their number, and, at the same time,

attacker-identification policy in the previous results, we

none of the honest CUs will be incorrectly identified as

now investigate on the performance of the attacker-

November 1, 2015

DRAFT

17

punishment policy. Particularly, the influence on the in-

x 10

Individual energy efficiency for an honest CU ( h)

dividual energy efficiency of attackers and honest CUs

Individual energy efficiency for an attacker ( )

a

will be shown before and after removing the identified

attackers from the fusion process. Notice that, as the
energy efficiency combines both the throughput and energy
consumption together, there is no need to show them
individually.
Fig. 9 shows the individual energy efficiency of an at-

Individual energy efficiency () [bits/Joule]

2.5

1.5

0.5

tacker and honest CU versus the total number of attackers

before removing the identified attackers, i.e. when i T .

11

13

15

17

19

The total number of attackers (M)

The individual energy efficiency of honest CUs decreases

Fig. 9. The individual energy efficiency of an honest CU and an attacker

as the number of attackers increases due to the increase in

versus the total number of attackers (M ) before removing the identified

the false alarm and the missed detection rates. Increasing

attackers (i T ). (T = 30)

the false-alarm rate degrades the achievable throughput,

while increasing the missed-detection rate wastes the en-

number of attackers the proposed policy can not provide

ergy consumption. The individual energy efficiency of an

the desired performance unless the attackers are removed.

attacker initially increases and then starts decreasing as the

Fig. 10 and Fig. 11 plot the individual energy efficiency of

number of attacker increases, as shown in Fig. 9. There are

an attacker and an honest CU versus the ignoring threshold

two reasons of the initial improvement. The first reason is

() after removing the identified attackers at M = 1 and

that increasing the number of attackers will increase the

M = 10, respectively. Apparently, has a significant

false-alarm rate in the global decision taken at the FC,

role in the performance of the attacker-punishment after

which increases their chances to exploit the unoccupied

removing the identified attackers (i > T ). A proper choice

channel in an illegitimate transmission. The second reason

of can remove all attackers from the fusion process and

is decreasing the false alarm rate in the decision made

leave only the honest CUs. Hence, the former effect of

cooperatively by the attackers themselves. However, at

the attackers on the sensing performance (PD and PF )

large number of attackers, the individual energy efficiency

will be totally eliminated, which consequently, reduces the

degrades as they equally share the illegitimate transmis-

illegitimate throughput of attackers. Notice that at = T ,

sion. An important note is that if we equally distribute the

none of the attackers nor the honest CUs will be removed

legitimate transmission opportunities among all CUs, i.e.

and thus the obtained values will be exactly as in the case

without punishment, an attacker will legitimately achieve

of i T .

the same energy efficiency as an honest CUs, and due to

the illegitimate transmission attackers will achieve higher
energy efficiency than honest CUs.

The optimization of should be carried out in order

to avoid punishing honest CUs rather than attackers. In
Fig. 12, is set to the optimal value, and the individual

From Fig. 9, the proposed attacker-punishment policy

energy efficiency of an attacker and an honest CU are

succeeds in reducing the energy efficiency of attackers at

found versus the number of attackers. The high perfor-

low number of attackers. However, in the presence of large

mance of the proposed attacker-punishment policy clearly

November 1, 2015

DRAFT

18
4

x 10

3.5

x 10

Individual energy efficiency for an honest CU ( h )

Individual energy efficiency for an attacker ( )

a

Individual energy efficiency for an honest CU ( h )

Individual energy efficiency

2.5

Individual energy efficiency for an attacker ( a )

1.5

Individual energy efficincy (/mu) [bits/Joule]

2.5

1.5

0.5

0.5

Fig. 10.

10

15

20

The ignoring threshold ( )

25

30

11

13

15

17

19

The total number of attackers (M)

The individual energy efficiency of an honest CU and an

Fig. 12.

The individual energy efficiency of an honest CUs and an

attacker versus the ignoring threshold () after removing the identified

attacker at the optimal ignoring threshold ( ) versus the total number of

attackers (i > T ). ( M = 1, T = 30).

attackers (M ) after removing the identified attackers (i > T ). (T = 30,

= )

x 10

Individual energy efficiency for an honest CU ( h )

Individual energy efficiency () [bit/Joule]

Individual energy efficiency for an attacker ( a )

VII. C ONCLUSIONS

2.5

Two policies to combat spectrum sensing data falsifica2

tion attackers in infrastructure-based cognitive radio networks have been proposed. The first policy is an attacker-

1.5

identification policy that aims at detecting attackers and

ignoring their reported sensing results, whereas the second

is an attacker-punishment policy that redistributes the

transmission opportunities among users based on their

0.5

Fig. 11.

10

15

20

Ignoring threshold ( )

25

30

local performance. Both policies are developed based on

The individual energy efficiency of an honest CU and an

attacker versus the ignoring threshold () after removing the identified

a novel approach for assessing the local performance ac-

attackers (i > T ). ( M = 10, T = 30).

cording to the delivery of the transmitted data. Analytical

and simulation results have shown that the attacker identification policy is able to identify attackers whatever

appears in the difference in the energy efficiency even

their number in the network, and the attacker-punishment

in the case of large number of attackers. The individual

policy is able to punish attackers by degrading their

energy efficiency of an honest CU slightly decreases as the

individual energy efficiency compared to the honest users.

number of attackers increases due to the increase in the

probability of not-detecting some of the attacker as their
number increases. However, the energy efficiency of an
honest CU is still more than twice the energy efficiency
of an attacker.

November 1, 2015

Future work will include the evaluation of the performance of the proposed policies in presence of different
attackers strategies. Indeed, an open challenge for any
security policy is to consider the case when attackers may
learn from the outcome of their previous decisions and act
DRAFT

19

adaptively.

[14] A. S. Rawat, P. Anand, C. Hao, and P. K. Varshney, Collaborative

Spectrum Sensing in the Presence of Byzantine Attacks in Cognitive
Radio Networks, IEEE Transactions on Signal Processing, vol.59,

R EFERENCES

no.2, pp.774,786, Feb. 2011.

[15] S. Mishra, A. Sahai and R. W. Brodersen, Cooperative sensing

[1] FCC, Spectrum Policy Task Force Report (ET Docket no.02-135),
Nov. 2002.

among cognitive radios, IEEE International Conference on Communications (ICC), 2006.

[2] J. Mitola and G. Q. Maguire, Cognitive radio: Making software

[16] J. Vartiainen, Always one/zero malicious user detection in coop-

radios more personal, IEEE Personal Communications, vol. 6, no.

erative sensing using the FCME method, 7th International ICST

4, pp. 13-18, August 1999.

Conference on Cognitive Radio Oriented Wireless Networks and

[3] S. Haykin, Cognitive radio: brain-empowered wireless communications, IEEE Journal on Selected Areas in Communications, vol. 23,
no. 2, 2005, pp. 201-220.
[4] J. Wang, M. Ghosh, and K. Challapali, Emerging cognitive radio
applications: A survey, IEEE Communications Magazine, vol.49,
no.3, pp.74-81, March 2011.
[5] A. Ghasemi and E. S. Sousa, Spectrum sensing in cognitive radio
networks: requirements, challenges and design trade-offs, IEEE
Communications Magazine, vol.46, no.4, pp.32-39, April 2008.
[6] M. Di Renzo, L. Imbriglio, F. Graziosi and F. Santucci, Cooperative Spectrum Sensing over Correlated Log-Normal Sensing and
Reporting Channels, IEEE Global Telecommunications Conference
(GLOBECOM), 2009.
[7] A. Ghasemi and E. S. Sousa, Collaborative spectrum sensing for
opportunistic access in fading environments, IEEE International
Symposium on Dynamic Spectrum Access Network, Baltimore- USA,
November 2005.
[8] E. Axell, G. Leus, E. G. Larsson and H. V. Poor, Spectrum Sensing
for Cognitive Radio : State-of-the-Art and Recent Advances, IEEE
Signal Processing Magazine, vol.29, no.3, pp.101-116, May 2012.
[9] S. Althunibat, S. Narayanan, M. Di Renzo and F. Granelli,EnergyEfficient Partial-Cooperative Spectrum Sensing in Cognitive Radio
over Fading Channels, IEEE VTC-Spring, 2013.
[10] S. Wang, Y. Wang, J. P. Coon and A. Doufexi, Energy-efficient
spectrum sensing and access for cognitive radio networks, IEEE
Transactions on Vehicular Technology, vol. 61, no. 2, 2012, pp. 906912.
[11] A. G. Fragkiadakis, E. Z. Tragos and I. G. Askoxylakis, A Survey

Communications (CROWNCOM), 2012.

[17] L. Lu, S. Y. Chang, J. Zhang, L. Qian, J. Wen, V. K. N. Lau, R. S.
Cheng, R. D. Murch, W. H. Mow, and K. B. Letaief, Technology
Proposal Clarifications for IEEE 802.22 WRAN Systems, IEEE
802.22 WG on WRANs, March, 2006.
[18] J. Hillenbrand, T. Weiss and F. K. Jondral, Calculation of Detection and False Alarm Probabilities in Spectrum Pooling Systems,
IEEE Communication Letters, Vol.9, no.4, 2005, pp. 349-351.
[19] W. El-Hajj, H. Safa and M. Guizani, Survey of Security issues
in Cognitive Radio Networks, Journal of Internet Technology, vol.
12, no. 2 , 2011, pp. 181-198.
[20] H. Li, and Z. Han, Catching attacker(s) for collaborative spectrum
sensing in cognitive radio systems: an abnormality detection approach, IEEE Symposium on New Frontiers in Dynamic Spectrum,
2010.
[21] W. Wang, H. Li, Y. Sun, and Z. Han, Attack-proof collaborative
spectrum sensing in cognitive radio networks, in Proceeding of
CISS, 2009, pp. 130134.
[22] S. Althunibat, M. Di Renzo and F. Granelli, Robust Algorithm
Against Spectrum Sensing Data Falsification Attack in Cognitive
Radio Networks, IEEE VTC-Spring, 2014.
[23] Ch. S. Hyder, B. Grebur, L. Xiao and M. Ellison, ARC: Adaptive
Reputation based Clustering Against Spectrum Sensing Data Falsification Attacks, IEEE Transactions on Mobile Computing, vol. 13,
no. 8, 2014, pp. 1707-1719.
[24] F. R. Yu, M. Huang, and H. Tang, Biologically Inspired
Consensus-based Spectrum Sensing in Mobile Ad Hoc Networks
with Cognitive Radios, IEEE Network, pp. 26-30, June 2010.

on Security Threats and Detection Techniques in Cognitive Radio

[25] Z. Li, F. R. Yu, and M. Huang, A Distributed Consensus-Based

Networks, IEEE Communications Surveys and Tutorials, vol. 15,

Cooperative Spectrum Sensing in Cognitive Radios, IEEE Trans.

no.1, 2013.

Vehicular Technology, vol. 59, no. 1, pp. 383-393, Jan. 2010.

[12] R. Chen, J. M. Park, Y. T. Hou and J. H. Reed, Toward secure

[26] R. Zhang, Y. C. Liang, S. Cui, Dynamic Resource Allocation

distributed spectrum sensing in cognitive radio networks, IEEE

in Cognitive Radio Networks, IEEE Signal Processing Magazine,

Communications Magazine, vol. 46, no. 4, pp. 50-55, April 2008.

vol.27, no.3, pp.102-114, May 2010.

[13] A. Attar, H. Tang, A. V. Vasilakos, F. R. Yu and V. Leung, A

[27] S. althunibat, B. J. Denise and F. Granelli, A Punishment Policy

survey of security challenges in cognitive radio networks: Solutions

for Spectrum Sensing Data Falsification Attackers in Cognitive Radio

and future research directions, Proceedings of the IEEE, vol. 100,

Networks, IEEE VTC-fall,2014.

no. 12, 2012, pp. 3172-3186.

November 1, 2015

[28] C. Stevenson, G. Chouinard, Z. Lei, W. Hu, S. J. Shellhammer, W.

DRAFT

20

Caldwell, IEEE 802.22: The first cognitive radio wireless regional

area network standard, IEEE Communications Magazine, vol.47,

[41] J. J. Schiller, R. A. Srinivasan and M. R. Spiegel, Schaums Outline

of Probability and Statistics. New York, NY: McGraw-Hill, 2008.

no.1, pp.130-138, January 2009.

[29] A. Goldsmith, S. A. Jafar, I. Maric and S. Srinivasa, Breaking
spectrum gridlock with cognitive radios: An information theoretic
perspective, Proceedings of the IEEE, vol. 97, no. 5, pp. 894-914,
2009.
[30] S. Althunibat, M. Di Renzo and F. Granelli, Optimizing the K-outof-N Rule for Cooperative Spectrum Sensing in Cognitive Radio Networks, IEEE Global Communications Conference (GLOBECOM),
2013.
[31] W. Zhang, R. K. Mallik, and K. Letaief, Optimization of cooperative spectrum sensing with energy detection in cognitive radio
networks, IEEE Transactions on Wireless Communications, vol. 8,
no. 8, pp. 5761-5766, December 2009.
[32] P. Kaligineedi, M. Khabbazian and V. K. Bhargava, Secure
Cooperative Sensing Techniques for Cognitive Radio Systems,
IEEE International Conference on Communications (ICC 08), 2008,
pp.3406-3410.
[33] L. Lamport, R. Shostak, and M. Pease, The byzantine generals
problem, ACM Transactions on Programming Languages and Systems, vol. 4, no. 3, p. 382-401, July 1982.
[34] R. Chen, J. M. Park, and K. Bian, Robust distributed spectrum
sensing in cognitive radio networks, IEEE INFOCOM, Apr. 2008,
pp. 1876-1884.
[35] I. F. Akyildiz, W.Y. Lee, M. C. Vuran and S. Mohanty, A
survey on spectrum management in cognitive radio networks, IEEE
Communications Magazine, vol.46, no.4, pp.40-48, April 2008.
[36] W. Han, J. Li, Z. Tian and Y. Zhang, Efficient Cooperative
Spectrum Sensing with Minimum Overhead in Cognitive Radio,
IEEE Transactions on Wireless Communications, vol.9, no.10, pp.
3006-3011, October 2010.
[37] E. Peh and Y. C. Liang, Optimization for Cooperative Sensing
in Cognitive Radio Networks, IEEE Wireless Communications and
Networking Conference (WCNC), pp.27-32, 2007.
[38] W. Saad, Z. Han, R. Zheng, A. Hjrungnes, T. Baar, and H. V. Poor,
Coalitional Games in Partition Form for Joint Spectrum Sensing
and Access in Cognitive Radio Networks, IEEE Journal of Selected
Topics in Signal Processing, vol. 6, no. 2, pp.195-209, April 2012.
[39] E. C. Y. Peh, Y. C. Liang, Y. L. Guan, and Y. Zeng, Optimization
of Cooperative Sensing in Cognitive Radio Networks: A SensingThroughput Tradeoff View, IEEE Transactions on Vehicular Technology, vol. 58, no. 9, pp. 5294-5299, Nov. 2009.
[40] X. Gelabert, O. Sallent, J. Perez-Romero, R. Agusti, Flexible Spectrum Access for Opportunistic Secondary Operation in Cognitive
Radio Networks, IEEE Transactions on Communications, vol. 59,
no. 10, pp. 2659-2664, October 2011.

November 1, 2015

DRAFT