1548 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO.

3, MARCH 2019

A Unified QoS and Security Provisioning

Framework for Wiretap Cognitive Radio
Networks: A Statistical Queueing
Analysis Approach
Yichen Wang , Member, IEEE, Xiao Tang , and Tao Wang

Abstract— Due to the spectrum-sharing feature of cognitive also provided, which demonstrate the impact of the PU’s QoS and
radio networks (CRNs) and the broadcasting nature of wireless the SU’s security requirements on SU’s throughput as well as the
channels, providing quality-of-service (QoS) provisioning for advantage of our proposed optimal power allocation scheme over
primary users (PUs) and protecting information security for the fixed power allocation scheme and the conventional security-
secondary users (SUs) are two crucial and fundamental issues based water-filling policy.
for CRNs. Consequently, in this paper, we establish a unified
QoS and security provisioning framework for wiretap CRNs. Index Terms— Cognitive radio, statistical queueing analysis,
Specifically, different from the widely used deterministic QoS pro- QoS provisioning, information security protection, effective
visioning method and information-theoretical security protection capacity, effective bandwidth, power allocation.
approach, our established framework, which is built on the theory I. I NTRODUCTION
of statistical queueing analysis, can quantitatively characterize
the PU’s QoS and the SU’s security requirements. By adopting the
theories of effective capacity and effective bandwidth, we further
convert the QoS and security requirements to the equivalent
W ITH the rapid evolution of wireless technologies, cog-
nitive radio (CR) communication, a promising yet
challenging technology that is realized by dynamic spectrum-
PU’s effective capacity and SU’s effective bandwidth constraints. sharing, has been regard as an efficient approach for relieving
Following our developed framework, we formulate the nonconvex
optimization problem, which aims at maximizing the average the spectrum resources shortage problem [1]–[6]. As sharing
throughput of SU subject to PU’s QoS requirement, SU’s security with the licensed spectrum of primary users (PU), the data
constraint, as well as SU’s average and peak transmit power transmission of secondary users (SU) cannot impact the quality
limitations. Then, we adopt the techniques of convex hull and of PU’s communications, which makes it important to provide
probabilistic transmission to convert the original nonconvex prob- efficient Quality-of-Service (QoS) provisioning for PU in
lem to the equivalent convex problem and obtain the optimal
power allocation scheme through the Lagrangian method. More- cognitive radio networks (CRNs). Moveover, the broadcast
over, we also develop a fixed power allocation scheme which is nature of wireless communications makes it difficult to shield
suboptimal but has low complexity. The simulation results are transmitted signals from unintended recipients. Consequently,
the information security is also a critically important issue for
Manuscript received April 2, 2018; revised August 3, 2018 and wireless CRNs.
November 8, 2018; accepted January 2, 2019. Date of publication Janu- The traditional approaches for PU’s QoS provisioning can
ary 23, 2019; date of current version March 11, 2019. This work was be mainly divided into two categories, which are PU’s maxi-
supported in part by the National Natural Science Foundation of China under
Grant 61871314, in part by the National Science and Technology Major mum allowed interference power-based method and PU’s min-
Project under Grant 2018ZX03001003-004, and in part by the Key Laboratory imum required transmission rate-based approach, respectively.
of Wireless Sensor Network and Communication, Chinese Academy of The interference power-based method imposes average/peak
Sciences, under Grant 2015003. This paper was presented at the IEEE
Vehicular Technology Conference-Fall, Chicago, USA, in 2018. The associate interference power constraints at SU transmitters. Specifically,
editor coordinating the review of this paper and approving it for publication the average interference power constraint guarantees that the
was K. Zeng. (Corresponding author: Yichen Wang.) average interference power received by PU will not exceed
Y. Wang is with the Department of Information and Communications Engi-
neering, School of Electronic and Information Engineering, Xi’an Jiaotong the predefined threshold [7]–[9]. On the contrary, the peak
University, Xi’an 710049, China, and also with the Key Laboratory of Wireless interference power constraint requires that the interference
Sensor Network and Communications, Shanghai Institute of Microsystem and power has to keep below the threshold all the time [10]–[12].
Information Technology, Chinese Academy of Sciences, Shanghai 200050,
China (e-mail: [email protected]). By employing the interference power constraints, the received
X. Tang is with the Department of Communication Engineering, interference power at PU receiver can be efficiently lim-
Northwestern Polytechnical University, Xi’an 710129, China (e-mail: ited. Different from the interference power-based method, the
[email protected]).
T. Wang is with the Department of Information and Communications Engi- PU’s transmission rate-based approach usually determines
neering, School of Electronic and Information Engineering, Xi’an Jiaotong the PU’s minimum transmission rate and requires the
University, Xi’an 710049, China (e-mail: [email protected]). SU transmitter to adjust the transmit power such that the
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. required minimum average/instantaneous transmission rate
Digital Object Identifier 10.1109/TWC.2019.2893381 can still be satisfied under SU’s interference [13]–[16].
1536-1276 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
WANG et al.: UNIFIED QoS AND SECURITY PROVISIONING FRAMEWORK FOR WIRETAP CRNs 1549

Moreover, PU’s minimum instantaneous transmission rate a stringent delay QoS demand. On the other hand, main-
requirement can be converted to the equivalent signal-to- taining the minimum average transmission rate cannot
noise plus interference (SINR) requirement [17] and can provide a relatively stable transmission rate to ensure data
also be relaxed to PU’s transmission outage probability will not be accumulated at the buffer of transmitter for the
constraint [18]–[20]. delay-sensitive services. Furthermore, different users may
Currently, the information security protection technologies have distinct delay QoS requirements that would cause a
for wireless communications are mainly categorized by two possibility of spanning in a very wide range from loose
classes: the cryptography theory based encryption technique delay demand to stringent delay requirement, which is
and the physical-layer security method. The encryption is difficult to use the traditional approaches to guarantee.
usually applied in the high protocol layers [21]. However, • In physical layer, the system security performance is
physical-layer security approach is built on the information usually evaluated by secrecy capacity, which couples
theory for secure data transmissions [22], [23]. As keyless the reliability and security within the single metric.
secure transmission can be realized by taking advantages of In other words, such a metric requires perfect security and
the intrinsic characteristics of wireless channels, physical- reliability simultaneously. However, the metric cannot
layer security technology has been extensively researched accurately reflect the delay-sensitive security requirement
in recent years [24]–[26]. The specific research about over realistic wireless channels. Specifically, the highly
physical-layer security technology includes artificial noise inj- time-varying wireless environment cannot always ensure
ection [27]–[29], anti-eavesdropping signal design [30], [31], a positive secrecy capacity, thus the legal users may lose
secure beamforming and precoding scheme [32]–[34], cooper- some transmit opportunities which would lead to the
ation based secure transmission [35], [36], as well as security- transmission rate degradation. Moreover, since different
oriented power control and resource allocation [37]–[39]. wireless services may have different security levels, it is
To provide efficient information security protection for CRNs, difficult to only use the secrecy capacity to describe
CR-based physical-layer security techniques attract lots of distinct security requirements.
research attentions. The security-based resource allocation • A large portion of applications/services are delay-
schemes are studied in [40] and [41]. In [42] and [43], sensitive, which means the information will expire after
the cooperative jamming based secure transmission strate- a certain period. For example, dynamic password is
gies are investigated for cognitive relay networks, where a an efficient approach to guarantee the user information
portion of cognitive relays are selected to convey informa- security. Each dynamic password is usually a randomly
tion and the others cooperate in jamming the eavesdropper generated number sequence and has a certain valid period.
such that the system security can be enhanced. In [44], In this case, the user’s information security only requires
authors focus on providing security protection for PU, where that the generated dynamic password will not be obtained
SUs can play as friendly jammers to confound the eavesdrop- by the eavesdropper within the valid period. Even if
per and obtain spectrum opportunities as a reward. Except the eavesdropper breaks the dynamic password beyond
the above discussed information security issues caused by the valid period, the user’s information is still immune
the existence of passive eavesdroppers, wireless networks are to the security threat because the password has already
also vulnerable to malicious attacks. In [45], chance discovery expired. Moreover, the information can be extracted
and usage control-based hierarchical framework is proposed only if sufficient amount of data are accumulated. Thus,
for attack detection over wireless sensor networks. Authors the delay-sensitive secrecy requirement is satisfied if
in [46] study the physical-layer authentication to combat with we can guarantee the eavesdropper will not accumulate
spoofing attacks over dynamic wireless networks, where the enough data within the certain valid period. However,
authentication process is formulated as a zero-sum authenti- it is difficult to use the secrecy capacity to achieve delay-
cation game and the Nash equilibrium as well as the spoofing sensitive secrecy provisioning.
detector are obtained. References [47] and [48] investigate Consequently, in order to provide efficient PU’s QoS pro-
the possible malicious attacks for CRNs. The corresponding visioning and information security protection, there is an
spectrum sensing as well as joint resource allocation scheme urgent need to establish a framework for wiretap CRNs, which
are derived. can not only accurately characterize PU’s fine-grained delay
Although many works have been conducted towards requirements, but also decouple the information security and
PU’s QoS provisioning and information security protection transmission reliability as well as precisely describe different
for CRNs, the QoS and security protection are still facing delay-sensitive security requirements.
many challenges, which are described as follows. To achieve the above goals, in this paper we establish
• The highly-stochastic nature of wireless channels leads a unified framework for PU’s delay QoS and SU’s delay-
to the fluctuation of PU’s receiving power of desired sensitive secrecy protections. Specifically, based on the theory
and interference signals, which causes the difficulty to of statistical queueing analysis, we use queue-length bound
establish the explicit relationship between PU’s delay violation probability to characterize PU’s delay QoS and SU’s
QoS requirement and the corresponding maximum tol- delay-sensitive security requirements. By employing the theo-
erable interference power. Moreover, on the one hand, ries of effective capacity and effective bandwidth, we further
keeping the minimum instantaneous transmission rate convert PU’s transmission data queue-length bound violation
means PU cannot tolerate any delay and thus implies probability constraint to the equivalent maximum sustainable
1550 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 3, MARCH 2019

traffic load requirement and convert SU’s wiretap data queue-

length bound violation probability constraint to the equivalent
minimum necessary traffic load demand. Following our estab-
lished framework, we formulate the nonconvex optimization
problem which aims at maximizing the average throughput
of SU subject to PU’s delay QoS requirement, SU’s delay-
sensitive secrecy constraint as well as SU’s average and peak
transmit power constraints. Thanks to the theories of convex
hull and probabilistic transmission, we convert the original
nonconvex problem to an equivalent convex problem and
obtain the optimal power allocation strategy by using the Fig. 1. System model for underlay-based wiretap cognitive radio networks
Lagrangian method. Moveover, a suboptimal scheme named with statistical QoS provisioning of primary users and statistical delay-
fixed power allocation policy is also developed, which only sensitive security of secondary users.
adapts to PU’s delay QoS and SU’s delay-sensitive secrecy
requirements. II. S YSTEM M ODEL
To summarize, the contributions of our work are as We consider a wiretap CRN coexists with a primary network
follows. (PN) by sharing the same spectrum band with bandwidth W ,
1) Different from the widely-used physical-layer security as shown in Fig. 1. The PN consists of one primary sender
technology, our proposed statistical delay-sensitive secrecy (PS) and one primary receiver (PR). The wiretap CRN includes
protection approach aims to provide the information secrecy one secondary sender (SS), one secondary receiver (SR), and
provisioning within the predefined period by utilizing the one secondary eavesdropper (SE). Specifically, the PS and
delay-sensitive feature of applications/services. In other words, SS transmit information to their corresponding receivers PR
we guarantee the eavesdropper will not accumulate enough and SR, respectively. The SE eavesdrops the information sent
data to extract the legitimate user’s information before the by SS and will be interfered by the transmission of the
protected information expires. In this way, not only the delay- PS-PR link. The channel power gains between PS and PR,
sensitive secrecy requirement of arbitrarily given applica- PS and SR, PS and SE, SS and PR, SS and SR, as well
tion/service can be characterized, but also the legitimate user is as SS and SE, which are denoted by hpp , hps , hpe , hsp ,
enabled to adapt the resource allocation to the secrecy require- hss , and hse , respectively, follow the Rayleigh fading model.
ment of the application/service and the network performance We assume that all channel power gains are stationary, ergodic,
can be improved. independent and block fading processes,1 which implies that
2) Based on the theory of statistical queueing analysis, the channel gains keep unchanged during each frame, but vary
we develop the unified QoS and security protection framework independently from one frame to another. This assumption is
for wiretap CRNs. By employing the queue-length bound widely accepted in existing literatures [7], [15], [20], [27],
violation probability, both PU’s delay QoS and SU’s delay- [37], [41], [42], [44]. Furthermore, we denote the duration of
sensitive secrecy requirements can be quantitatively and accu- each frame as T and define the network gain vector (NGV)
rately characterized. Moreover, built upon the theories of as h  {hpp , hps , hpe , hsp , hss , hse }.2
effective bandwidth and effective capacity, the delay QoS and In this paper, as we mainly focus on protecting PU’s delay
delay-sensitive secrecy descriptions are further converted to QoS requirement and SU’s delay-sensitive secrecy demand,
the equivalent maximum sustainable traffic load and minimum
1 At each time epoch, the Rayleigh fading coefficient are composed by two
necessary traffic load requirements, respectively, which make
the QoS and secrecy requirements tractable. i.i.d. complex circularly symmetric complex Gaussian (CSCG) random vari-
ables implying that the Rayleigh fading model can be represented by discrete-
3) By employing the theories of convex hull and probabilis- time i.i.d. Gaussian process. As discrete-time i.i.d. Gaussian process can be
tic transmission, we develop an efficient method to convert proved to be stationary, we assume the channel power gains are stationary.
the nonconvex problem to the equivalent convex problem. The ergodicity means that the time average of the channel converges to the
same limit for almost all realizations of the fading process, which holds in
Based on the developed convexification approach, we obtain may models. Moreover, due to the spatial differences, different channels are
the optimal power allocation scheme, which can not only usually independent.
2 We assume that the channel information h can be perfectly known at SS.
dynamically vary with channel conditions in the network,
Specifically, hss and hps can be obtainde by using classic channel training,
but also adapt to PU’s QoS and SU’s secrecy requirements estimation, and feedback mechanisms. For hpp and hsp , one possible method
simultaneously. for deriving these information is to cooperate with PN. Then, by using the
The rest of this paper is organized as follows. Section II classic channel training and estimation mechanisms, PR can obtain hpp and
hsp and can feed them back to SS. For obtaining hse , one typical example
describes the system model. In Section III, we discuss the the- in practical wireless systems is that, SS-SR link and SE belong to the same
ory of statistical queueing analysis. In Section IV, the frame- CRN while the currently scheduled user SR has confidential messages to
works for PU’s statistical QoS and SU’s statistical security send which need protection from the unscheduled user SE. In this case,
the currently unscheduled user SE is regarded as the potential eavesdropper
provisioning are established. In Section V, the optimization and the legitimate transmitter SS has the information on the eavesdropper
problem is formulated and the optimal power allocation SE due to the previous scheduling. Moreover, if the SS-SR link and SE
scheme is derived. We develop the fixed power allocation belong to the same CRN, then SE can use similar method to get hse . The
elementary impact of imperfect channel estimations is analyzed via simultion
scheme in Section VI. The simulation results are provided in in Section VII. The further studies on how to combat with the imperfect
Section VII. The conclusions are drawn in Section VIII. channel estimations are left for our future works.
WANG et al.: UNIFIED QoS AND SECURITY PROVISIONING FRAMEWORK FOR WIRETAP CRNs 1551

TABLE I
S UMMARY OF N OTATIONS AND D EFINITIONS

we formulate two queues for the CRN, as shown in Fig. 1. III. P RELIMINARIES ON S TATISTICAL
Specifically, on the one hand, the queue at PS stores the Q UEUEING A NALYSIS
data that will transmit to PR. As the input of the queue Different from the traditional deterministic method, the sta-
comes from the upperlayer of PS, without loss of generality, tistical queueing analysis aims at meeting the queue-length
we assume that the queue at PS has a constant arrival rate bound violation probability requirement, which is given
process denoted by RA . Since the stored data will be conveyed by [50]
to PR via the stochastic wireless channel, the amount of data  
that can leave the queue within a given period dynamically Pr Q ≥ Qth ≤ Pth , (1)
varies. Thus, the queue at PS is modeled with a time-varying
data service process denoted by RP . On the other hand, where Q represents the queue-length, Qth denotes the prede-
the eavesdropped data from SS is stored in the queue at SE. fined queue-length threshold, and Pth is the maximum allowed
As the eavesdropped data, which experiences the stochastic violation probability. Specifically, the statistical provisioning
wireless channel between SS and SE, plays as the input approach requires that the probability of queue-length Q
of the queue, we assume that the queue at SE is with the exceeding the threshold Qth should be below the tolerable
stochastic data arrival process denoted by B. Moreover, lots violation probability Pth . Based on large deviation principle
of services are delay-sensitive, i.e., the information has a (LDP), Chang [50] showed that for a dynamic queueing system
certain valid period. The delay-sensitive feature implies that with stationary ergodic arrival and service processes, under
the eavesdropper has to accumulate the sufficient amount of sufficient conditions, the queue length process Q satisfies
  
data within a valid period in order to obtain the unexpired log Pr Q > Qth
effective information. The valid period of data becomes the − lim = θ. (2)
Qth →∞ Qth
key of this delay-sensitive secrecy protection. Even though
the eavesdropper might accumulate sufficient amount of data, To be more specific, the above result means that the probability
the effective information still cannot be obtained because the of the queue length exceeding a certain threshold decays
data has already been out of date. Take the dynamic password exponentially as the threshold. Then, the queue-length bound
that we have already discussed as the example. The dynamic violation probability can be approximately determined by
 
password, which is usually a randomly generated number
Pr Q ≥ Qth ≈ e−θQth , (3)
sequence, has a particular valid period. If enough data required
for deciphering the password is not accumulated within the where θ is defined as the queueing exponent. We can observe
valid period, the eavesdropper cannot obtain the effective from Eq. (3) that the queueing exponent θ describes the decay-
password as the current dynamic password expires. Due to ing speed of the violation probability versus the queue-length
the delay-sensitive feature, the eavesdropped data will become bound. Note that the deterministic exponential approximation
useless and be removed if sufficient amount of data cannot serves as a tight upperbound for the actual queue-length bound
be accumulated within a certain period. Therefore, the queue violation probability [51]. Consequently, based on Eqs. (1)
at SE is modeled with the constant data departure process and (3), meeting the queue-length bound violation probability
denoted by RE , which implies that the eavesdropped data will requirement implies that θ satisfies
become useless beyond the certain delay. By formulating the  
1
dynamic queueing system at PS and SE, we aim to provide θ≥− log Pth . (4)
statistical delay QoS provisioning for the PS-PR link and Qth
statistical delay-sensitive security protection for the SS-SR link To further analyze the queue-length bound violation prob-
which will be detailed in Section IV. The notations used in ability, we consider a general queueing system with a time-
this paper are listed in Table I. varying arrival process RA and stochastic service process RP
1552 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 3, MARCH 2019

the queue-length bound violation probability requirement

described by Eq. (1) can be satisfied. Therefore, on the one
hand, if the arrival rate and service rate processes are given,
we can use θ to analyze the queue-length violation probability.
On the other hand, under the given targeted queue-length
violation probability requirement, we can use the target θ to
design the arrival rate or service rate process over the wireless
channels.

IV. T HE S TATISTICAL Q UEUEING A NALYSIS BASED Q O S

Fig. 2. Decomposition of the queueing system with stochastic data arrival AND S ECURITY P ROVISIONING F RAMEWORK
and service processes.
A. PU’s Statistical Delay QoS Provisioning
The queue at PS in Fig. 1, which is used for PU’s statistical
as shown in Fig. 2. To analyze the statistical QoS provisioning delay QoS protection, can be viewed as the second virtual
requirement of the dynamic queueing system, we can divide queueing system in Fig. 2. Denote Qp and Qthp as the queue-
the queueing system into two virtual queueing systems. The length and the threshold of the queue at PS, respectively.
first virtual queue can be seemed as the system with the We rewrite Eq. (4) as the following inequality:
time-varying arrival process RA and a constant service rate.
The second virtual queue can be viewed as the system with a 1  
θp ≥ − log Pthp , (8)
constant arrival rate and the time-varying service process RP . Qthp
Then, we can determine the required constant service rate
where θp represents PU’s QoS exponent and Pthp denotes
of the first virtual queueing system and the constant arrival
maximum allowed violation probability. Based on Eq. (8),
rate of the second virtual queueing system by employing
we can obtain that smaller value of θp (i.e., larger values
the theories of effective bandwidth and effective capacity.
of Qthp and Pthp ) implies looser PU’s queueing-delay QoS
Specifically, effective bandwidth is defined as the required
requirement and larger value of θp (i.e., smaller values of Qthp
minimum constant service rate of a queueing system under
and Pthp ) implies more stringent PU’s delay QoS requirement.
the given stochastic arrival rate process subject to the speci-
Especially, when θp → 0, PU can tolerate an arbitrary long
fied queueing exponent θ requirement [50]. On the contrary,
delay; when θp → ∞, PU cannot allow any delay.
effective capacity, which can be regarded as the dual concept
According to Eq. (7), we can obtain that meeting the
of effective bandwidth, is defined as the maximum constant
predefined PU’s transmission rate queue-length bound viola-
arrival rate that a queueing system can support under the given
tion probability requirement Pr {Qp ≥ Qthp } ≤ Pthp only
stochastic service rate process subject to the specified queueing
requires satisfying the following inequality:
exponent θ requirement [51].
C (θp ) ≥ T W RA ,
Mathematically, the effective bandwidth of the time-varying EPU (9)
arrival process RA and the effective capacity of the time-
varying service process RP , denoted by EB (θ) and EC (θ), which implies that the effective capacity of the service rate
respectively, can be written as [50], [51] process RP , i.e., the maximum supportable arrival rate of PS,
⎧    cannot be smaller than the given constant arrival rate RA .
⎪
⎪ 1
⎨EB (θ) = lim log Eh eθT W RA
t→∞ θt
   (5)
⎪
⎪ 1 B. SU’s Statistical Delay-Sensitive Security Provisioning
⎩EC (θ) = − lim log Eh e−θT W RP
t→∞ θt
The queue at SE in Fig. 1, which is used for storing the
where Eh {·} denotes the expectation over h. If both RA eavesdropped data and guaranteeing SU’s statistical security
and RP are time-uncorrelated processes, the expressions for requirement, can be viewed as the first virtual queueing system
effective bandwidth and effective capacity reduce to in Fig. 2. Denote Qthe , θe , and Pthe as the queue-length
⎧   
⎪ 1 threshold, SU’s statistical delay-sensitive secrecy exponent,
⎪
⎨ B
E (θ) = log E h e θT W RA
θ and maximum allowed violation probability of this queue,
   (6)
⎪ 1 respectively. Then Eq. (4) can be rewritten as
⎪
⎩ C
E (θ) = − log E h e −θT W RP
θ 1  
θe ≥ − log Pthe . (10)
According to the theory of statistical queueing analysis, we can Qthe
obtain that if the effective bandwidth of the time-varying
We can observe from Eq. (10) that smaller value of θe implies
arrival process RA in the first virtual queueing system is no
looser SU’s statistical delay-sensitive secrecy requirement and
larger than the effective capacity of the time-varying service
larger value of θe implies more stringent SU’s statistical
process RP in the second virtual queueing, i.e., θ only needs
delay-sensitive secrecy requirement. Especially, when θe → 0,
to meet the following inequality
SU has no secrecy requirement; when θe → ∞, SU’s security
EB (θ) ≤ EC (θ), (7) requirement becomes very stringent.
WANG et al.: UNIFIED QoS AND SECURITY PROVISIONING FRAMEWORK FOR WIRETAP CRNs 1553

Based on Eq. (7), we can obtain that meeting the predefined (nats/s/Hz), is upper-bounded by
SU’s eavesdropped rate queue-length bound violation proba-
      
bility requirement Pr {Qe ≥ Qthe } ≤ Pthe only requires to B h, Ps (θp , θe , h) = min log 1+ γ SR , log 1 + γ SE
satisfy the following inequality:  
= log 1+ γ 0 , (14)
EEVE
B (θe ) ≤ T W RE , (11)
h P (θ ,θ ,h) h P (θ ,θ ,h)
which implies that the effective bandwidth of the arrival rate where γ SR = sshpss Ppp+σe2 and γ SE = sehpes Ppp+σe2 denote
process B, i.e., the minimum service rate of SE, cannot be the signal-to-interference plus noise ratio (SINR) at SR
larger than the given constant departure rate RE . and SE, respectively, and γ 0 = min γ SR , γ SE . Then,
Remark: Parameters Qthe and Pthe can be viewed as the we can formulate the corresponding optimization problem
amount of data that is required to accumulate to extract the which aims at maximizing the average throughput of SS sub-
information transmitted via the SS-SR link and the maximum ject to PU’s statistical delay QoS requirement, SU’s statistical
allowed probability that the unexpired effective information is delay-sensitive secrecy constraint as well as SS’s average and
obtained by SE, respectively. Then, Qthe and Pthe are both peak transmit power constraints. Specifically, the optimization
determined by the characteristics of the services/applications problem can be mathematically written as
carried by the SS-SR link and thus can be obtained by SS.  
Then, SS can further determine the required θe by Eq. (10). (P1) max Eh Rs (h, Ps (θp , θe , h)) (15)
Ps (θp ,θe ,h)≥0
Moreover, since the parameter RE is used to reflect the  
valid period of the application/service with the delay-sensitive s.t. EPU
C θp , Ps (θp , θe , h) ≥ T W RA , (16)
 
secrecy requirement, it can also be obtained by SS. EVE
EB θe , Ps (θp , θe , h) ≤ T W RE , (17)
 
Eh Ps (θp , θe , h) ≤ Pav , (18)
V. O PTIMAL P OWER A LLOCATION S TRATEGY FOR
PU’ S D ELAY Q O S AND SU’ S I NFORMATION Ps (θp , θe , h) ≤ Ppk , ∀ h, (19)
S ECURITY P ROTECTION
where Pav and Ppk denote the maximum allowed average and
In this section, we investigate the optimal power allocation
peak transmit power for SS, respectively. Eq. (16) denotes
scheme for the wiretap CRN based on our established frame-
PU’s statistical delay QoS requirement and Eq. (17) represents
work, which can meet PU’s statistical delay-QoS and SU’s
SU’s statistical delay-sensitive secrecy constraint. By plugging
delay-sensitive security requirements simultaneously.
Eq. (6) into Eqs. (16) and (17) respectively, and per-
forming some mathematical manipulations, the maximization
A. Optimization Problem Formulation problem (P1) can be transformed to the following equivalent
In our considered wiretap CRN, we assume PU transmits minimization problem:
with a constant power Pp ,3 but SS sends data to SR with  
variable transmit power denoted by Ps (θp , θe , h) which is (P2) min Eh − Rs (h, Ps (θp , θe , h)) (20)
Ps (θp ,θe ,h)≥0
dynamically adjusted according to PU’s statistical QoS con-  
straint specified by the QoS exponent θp , SU’s statistical s.t. Eh f (Ps (θp , θe , h)) − e−βp RA ≤ 0,
delay-sensitive secrecy requirement described by the secrecy (21)
exponent θe , and network channel conditions characterized  
by the NGV h. Based on Shannon’s theory, the maximum Eh g (Ps (θp , θe , h)) −e βe RE
≤ 0,
service rates of PS and SS, denoted by Rp (h, Ps (θp , θe , h)) (22)
(nats/s/Hz) and Rs (h, Ps (θp , θe , h)) (nats/s/Hz), respectively, Eqs. (18) and (19),
are determined by

hpp Pp where βp = θp T W , βe = θe T W , g (Ps (θp , θe , h)) =
Rp (h, Ps (θp , θe , h)) = log 1 +  βe
hsp Ps (θp , θe , h) + σ 2 1 + γ 0 , and
(12)
 −βp
hpp Pp
and f (Ps (θp , θe , h)) = 1 + .
 hsp Ps (θp , θe , h) + σ 2
hss Ps (θp , θe , h)
Rs (h, Ps (θp , θe , h)) = log 1 + , (13) (23)
hps Pp + σ 2
respectively, where σ 2 denotes the variance of additive white We can easily observe that the constraints given by Eqs. (18)
Gaussian noise (AWGN). Moreover, the information leak- and (19) are both affine. Moreover, both e−βp RA and eβe RE
ing rate from SS to SE, denoted by B h, Ps (θp , θe , h) are constants as well as the expectation operation Eh {·} is
linear. Consequently, the convexity of the objective function,
3 We can treat the cellular system as the PN. Then, for the downlink PU’s statistical delay QoS requirement, and SU’s statisti-
transmission, the base station serves as the PS and the mobile terminal cal delay-sensitive secrecy constraint, which are given by
plays as the PR. The based station usually transmits with a fixed power. For
example, LTE specifications require the base station to transmit with a fixed Eqs. (15), (16), and (17), respectively, are determined by
power 46dBm [52], [53]. Rs (h, Ps ), f (Ps ) and g (Ps ), respectively. As g (Ps ) is related
1554 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 3, MARCH 2019

 
to γ 0 = min γ SR , γ SE , we can obtain as follows:
⎧ β L (Ps (θp , θe , h) , λ, μ, ν)
⎪
⎪ hss Ps (θp , θe , h) e  
⎪
  ⎨ 1 + hps Pp + σ 2
SR SE
, if γ <γ , = Eh L (Ps (θp , θe , h) , λ, μ, ν)
g Ps =  βe
⎪
⎪ hse Ps (θp , θe , h) − λe−βp RA − μeβe RE − νPav , (28)
⎪
⎩ 1+ , if γ SR ≥ γ SE .
hpe Pp + σ 2
where
(24)
L (Ps , λ, μ, ν) = −Rs (h, Ps ) + λf (Ps ) + μg (Ps ) + νPs ,
SR SE
Without loss of generality, we assume γ < γ for the (29)
simplicity of analysis.
To analyze the convexity of Rs (h, Ps ), f (Ps ) and g (Ps ), λ, μ, and ν denote the Lagrangian multiplies associ-
we derive their second derivatives, which are given by ated with constraints given by Eq. (21), (22), and (18),
respectively. Since λe−βp RA , μeβe RE and νPav are all con-
 2 2 stants as well as the expectation operation Eh {·} is linear,
d2 Rs (h, Ps ) hss hss Ps (θp , θe , h)
=− 1+ , the non-convexity of L (Ps (θp , θe , h) , λ, μ, ν) causes that
dPs2 hps Pp +σ 2 hps Pp + σ 2
L (Ps (θp , θe , h) , λ, μ, ν) is nonconvex with respect to
(25) Ps (θp , θe , h) within [0, Ppk ]. To obtain the optimal solution,
 −βp −1
d2 f (Ps ) βp h2sp hpp Pp hpp Pp we will employ the theories of convex hull and probabilis-
dPs2
= 3 · 1 + h P + σ2 tic transmission to convert L (Ps (θp , θe , h) , λ, μ, ν)4 to the
(hsp Ps + σ 2 ) sp s
equivalent convex function, which will be detailed in the
(βp −1)hppPp −2 hsp Ps (θp , θe , h)−2σ 2
× , following Section V-B.
hpp Pp + hsp Ps (θp , θe , h)+σ 2
(26)
B. Convex Hull and Probabilistic Transmission
and According to the theory of convex hull [55], to deter-
mine the optimal solution of the nonconvex optimization
 β −2 problem (P2), we need to get the lower-boundary function of
d2 g (Ps ) βe (βe − 1)h2ss hss Ps (θp , θe , h) e
= 2 1 + , the convex hull for the two dimensional region C determined
dPs2 (hpsPp + σ 2 ) hps Pp + σ 2
by all points (Ps , L (Ps , λ, μ, ν)) with Ps ∈ [0, Ppk ]. Since the
(27) convex hull of the two-dimensional region C is the point set
formed by all convex combinations of two-dimensional points
respectively. We can easily derive from Eq. (25) that Rs (h, Ps ) in C, the lower-boundary of convex hull for region C can be
is concave with respect to Ps (θp , θe , h) within [0, Ppk ]. Based determined by the function denoted by L̃ (Ps , λ, μ, ν), which
on Eq. (26), we can get an inflexion point implying that is the minimum convex envelop of the original nonconvex
f (Ps ) is nonconvex over Ps (θp , θe , h) ∈ [0, Ppk ]. According function L (Ps , λ, μ, ν). Therefore, for the nonconvex part of
to Eq. (27), we can analyze the convexity of g (Ps ) from L (Ps , λ, μ, ν), we can tackle its counterpart of L̃ (Ps , λ, μ, ν)
two cases. On one hand, if 0 < βe < 1, we have that instead. Fig. 3(c) illustrates the original nonconvex function
d2 g (Ps ) /dPs2 < 0. On the other hand, when βe > 1, L (Ps , λ, μ, ν) in solid line and function L̃ (Ps , λ, μ, ν) in
the inequality d2 g (Ps ) /dPs2 > 0 holds. Thus, g (Ps ) is dashed line. Specifically, consider two consecutive tangent

also nonconvex with respect to Ps (θp , θe , h). Then, we have points of the nonconvex part of L (Ps , λ, μ, ν) i.e., the end-
that Eqs. (21) and (22) are both nonconvex functions over 
points
 of line-segment
 in L̃ (Ps , λ, μ,ν) , which are denoted
Ps (θp , θe , h) ∈ [0, Ppk ] and thus (P2) is a nonconvex opti-
by P̃s1 , L(P̃s1 ) and P̃s2 , L(P̃s2 ) , respectively. In this
mization problem. According to the theory of convex optimiza-
tion, we have known that the Karush-Kuhn-Tucker (K.K.T.) region, any point within the line-segment can be represented
conditions are only the necessary conditions for the optimal by the aforementioned two endpoints. This can be interpreted
solution of the nonconvex optimization problem and will cause by the probabilistic transmission
 strategy.
 In particular, any
the nonzero gap between the value of the Lagrangian function transmit power Ps within P̃s1 , P̃s2 can be denoted by
   
at the solution derived by solving the K.K.T. conditions
ξ P̃s1 , L(P̃s1 ) + (1 − ξ) P̃s2 , L(P̃s2 ) (ξ ∈ [0, 1]), which
and the maximum value of the corresponding dual problem.
Consequently, we cannot obtain the optimal solution by simply implies that Ps can be realized by using power P̃s1 with
using the Lagrangian approach. However, if we can convert probability ξ and using power P̃s2 with probability (1 − ξ)
the nonconvex Lagrangian function to the equivalent convex for transmission.
function, the Lagrangian method is a convenient approach to Based on the above discussions, it is critically important to
derive the optimal solution as zero gap between the global obtain the lower-boundary function L̃ (Ps , λ, μ, ν). To this end,
optima of the Lagrangian function and the dual problem can we need to get the inflexion points as well as the tangent points
be achieved. 4 In Section V-B, we generally express SU’s transmit power P (θ , θ , h)
s p e
To obtain the optimal power allocation scheme, we construct explicitly as Ps , and suppress (θp , θe , h) in order to avoid cumbersome
the Lagrangian function, denoted by L (Ps (θp , θe , h) , λ, μ, ν), expression.
WANG et al.: UNIFIED QoS AND SECURITY PROVISIONING FRAMEWORK FOR WIRETAP CRNs 1555

Fig. 3. Illustration of determining the lower-boundary function for the nonconvex function by using our proposed four-step procedures. (a) Determine the
local minimum points of the nonconvex function (Step 1). (b) Choose suitable points from all derived local minimum points (Step 2). (c) Determine all tangent
points and the lower-boundary function (Steps 3 and 4).

of L (Ps , λ, μ, ν). However, as the closed-form expression Algorithm 1: Select the Suitable Points From Point Set C1
of the L̃ (Ps , λ, μ, ν) cannot be obtained, we will determine 1 InPut: C1 = {Pi }(i = 1, · · · , |C1 |);
L̃ (Ps , λ, μ, ν) from the following three cases. 2 Initialization: C2 ← {P1 }, n ← 1;
2
Case 1: d L(PdP s ,λ,μ,ν)
2 ≥ 0 with respect to Ps within [0, Ppk ], 3 repeat
s
i.e., L (Ps , λ, μ, ν) is convex. 4 for n ← n + 1, n + 2, · · · , |C1 | do
In this case, L̃ (Ps , λ, μ, ν) is the same as L (Ps , λ, μ, ν) 5 s(n ) ← L(Pn ,λ,μ,ν)−L(P
Pn −Pn
n ,λ,μ,ν)
;
given by Eq. (29). n ← arg min

s(n 
);
2 6

Case 2: d L(PdP s ,λ,μ,ν)
2 < 0 versus Ps over [0, Ppk ], 7 C2 ← C2 {Pn };
s
i.e., L (Ps , λ, μ, ν) is concave. 8 n ← n ;
In this case, L̃ (Ps , λ, μ, ν) is determined by a line-segment
9 until n = |C1 |; 

with two endpoints (0, L(0, λ, μ, ν)) and (Ppk , L(Ppk ,
λ, μ, ν)), where (0, L(0, λ, μ, ν)) and (Ppk , L(Ppk , λ, μ, ν)) 10 OutPut: C2 = P̃j (j = 1, · · · , |C2 |).
are the beginning and ending points of L (Ps , λ, μ, ν), respec-
tively. Then, L̃ (Ps , λ, μ, ν) can be written as:
L̃ (Ps , λ, μ, ν) = L(0, λ, μ, ν) The convex optimization theory tells us that the first
L(Ppk , λ, μ, ν) − L(0, λ, μ, ν) derivative-function of a convex function is an increasing
+ Ps , Ps ∈ [0, Ppk ]. function. However, not all points in C1 satisfy such a
Ppk
(30) property. Therefore, to obtain the lower-boundary function
2 2
L̃ (Ps , λ, μ, ν), we further construct the set C2 with all points
Case 3: d L(PdP s ,λ,μ,ν)
 0 and d L(PdP s ,λ,μ,ν)
 0 with that satisfy the above property in C1 . To obtain the point
2
s  2
s 
respect to Ps within [0, Ppk ], i.e., L Ps , λ, μ, ν is neither set C2 , we develop Algorithm 1 shown in the next page.
convex nor concave. Moreover, we also obtain that the points in C2 obtained
In this case, it is essential to locate all the tangent points by Algorithm 1 satisfy the following theorem.
of L (Ps , λ, μ, ν), i.e., the endpoints of the line-segment of Theorem 1: Denote κi as the slope of the line-
L̃ (Ps , λ, μ, ν), to determine the expression of L̃ (Ps , λ, μ, ν). 
segment whose endpoints   are the two adjacent  points
To achieve this goal, we follow the four steps below. P̃i , L(P̃i , λ, μ, ν) and P̃i+1 , L(P̃i+1 , λ, μ, ν) of the point
Step 1: Determine the local minimum points of set C2 , where i = 1, · · · , |C2 | − 1. Then, ∀1 < i < j ≤
L (Ps , λ, μ, ν).    |C2 | − 1, we have κi ≤ κj .

 s ,λ,μ,ν) 
Denote C̃1  P  (Ps − P  ) dL(PdP s
 ≥ 0 , Proof: The proof is provided in Appendix A.
◦ ◦
Ps =P  Based on Theorem 1, we can declare that the points in C2
where Ps ∈U (P  , δ) and U (P  , δ) denotes P  ’s deleted satisfy the aforementioned property of the convex function.
neighborhood with radius δ. In other words, the elements of Take Fig. 3(a) as an example, in which the elements of
C̃1 are the local minimum points of L (Ps , λ, μ, ν). To facil- point set C1 are annotated by the cross mark. By employing
itateour discussions, we expand C̃1 to C1 such that C1 = Algorithm 1, we derive the corresponding point set C2 ,
C̃1 {0, Ppk }. Then, without loss of generality, we rewrite whose elements are showed in Fig. 3(b) with the dot mark.
C1 as {Pi }i∈{1,2,··· ,|C1 |} with Pi ≤ Pj , ∀1 ≤ i ≤ j ≤ |C1 |. By comparing Fig. 3(a) with Fig. 3(b), we can find that P2
It is obvious that the set C1 includes all local minimum points belongs to C1 , but is excluded from C2 .
and the two endpoints, which are illustrated in Fig. 3(a) and Step 3: Determine the tangent points based on point set C2 .
annotated by the cross mark. In this step, we need to get the tangent points cor-
Step 2: Select the suitable points from the point set C1 . responding to each P̃i in C2 within its δ neighborhood
1556 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 3, MARCH 2019

 
U (P̃i , δ), where i = 1, 2, · · · , |C2 |. Specifically,   based on our P̃|C2 |−1 , P̃|C2 |−1 + δ includes  no tangent  point when
obtained  point  set C , the     can be divided
interval 0, P L (Ps , λ, μ, ν) is convex over P̃|C2 |−1 , P̃|C2| or one tangent

2 pk  
as P̃1 , P̃2 ··· P̃i , P̃i+1 ··· P̃|C2 |−1 , P̃|C2 | . point when L (Ps , λ, μ, ν) is nonconvex over P̃|C2 |−1 , P̃|C2 | .
As function  L (Ps , λ, μ, ν) is nonconvex over Ps ∈  If χ = |C2 |, L (Ps , λ, μ, ν) is nonconvex over Ps ∈
P̃i , P̃i+1 (i = 2, · · · , |C2 | − 2), the tangent points, which P̃|C2 |−1 , P̃|C2 | . Then, the tangent points of L (Ps , λ, μ, ν)
are used for determining L̃ (Ps , λ, μ, ν), depend  on the depend on its convexity over subinterval P̃1 , P̃2 . In par-
 
convexity  of L (Ps, λ, μ, ν) over subintervals P̃1 , P̃2 ticular, if L (Ps , λ, μ, ν) is convex over P̃1 , P̃2 , each P̃i
and P̃|C2 |−1 , P̃|C2 | . In particular,  we define χ  corresponds
  to two tangent points within its left neighborhood
argmini=1,··· ,|C2 | L P̃i , λ, μ, ν implying that L (Ps , λ, μ, ν) P̃i − δ, P̃i for 3 ≤ i ≤ |C2 |− 1 and there  is onlyone tangent
achieves its minimum value at Ps = P̃χ . Then, the tangent point within P̃2 ’s left neighborhood P̃2 − δ,P̃2 . However,
points of L (Ps , λ, μ, ν) are determined as follows. if L (Ps , λ, μ, ν) is not convex over P̃1 , P̃2 , each P̃i cor-
If χ = 1, i.e., L (Ps , λ, μ, ν) achieves its minimum value responds to two tangent points within its left neighborhood

at Ps = P̃1 , L (Ps , λ, μ, ν) is the nonconvex function of P̃i , P̃i + δ for 2 ≤ i ≤ |C2 | − 1.
Ps over P̃1 , P̃2 . Then, the tangent points of L (Ps , λ, μ, ν)  If 3 ≤ χ ≤ |C2 | − 2, we can obtain the following
depend on its convexity over subinterval P̃|C2 |−1 , P̃|C2 | . conclusions:
 
Specifically,
 on one hand, if L (Ps , λ, μ, ν) is convex over • The left neighborhood P̃2 − δ, P̃2 of P̃2 includes
P̃|C2 |−1 , P̃|C2 | , each P̃i corresponds  to two tangent points one
 tangent
 point if L (Ps , λ, μ, ν) is convex over
within its right neighborhood P̃i , P̃i +δ for 2 ≤ i ≤ |C2 |−2 P̃1 , P̃2 or two tangent
 point if L (Ps , λ, μ, ν) is not
and there is only convex over P̃1 , P̃2 .
 one tangent point within P̃|C2 |−1 ’s right  
neighborhood P̃|C2 |−1 , P̃|C2 |−1 + δ . On the other hand, if • The right neighborhood P̃|C2 |−1 , P̃|C2 |−1 + δ of
L (Ps , λ, μ, ν) is also nonconvex over P̃|C2 |−1 , P̃|C2 | , we can P̃|C2 |−1 includes one tangent  point if L (Ps , λ, μ, ν)
obtain that each P̃i corresponds  to two  tangent points locating is convex over interval P̃|C2 |−1 , P̃|C2 | or two tan-
in its right neighborhood P̃i , P̃i + δ for 2 ≤ i ≤ |C2 | − 1. gent points if L (Ps , λ, μ, ν) is not convex over
If χ = 2, each P̃i corresponds  to two  tangent points P̃|C2 |−1 , P̃|C2 | .
within its right neighborhood P̃i , P̃i + δ for 3 ≤ i ≤ • Each P̃i corresponds to two tangent points within its left
 
|C2 | − 2. However, the tangent points corresponding to neighborhood P̃i − δ, P̃i for 3 ≤ i ≤ χ − 1 or its
 2 and P̃|C2 |−1
P̃  depend on the  convexity over subintervals right neighborhood P̃i , P̃i + δ for χ + 1 ≤ i ≤
P̃1 , P̃2 and P̃|C2 |−1 , P̃|C2 | . Specifically, for Ps = P̃2 , |C2 | − 2.  
there
 is a unique  tangent point within its right neighborhood • For Ps = P̃χ , both the left neighborhood P̃χ − δ, P̃χ
 
 P̃ 2 , P̃ 2  + δ . Moreover, if L (Ps , λ, μ, ν) is convex over and right neighborhood P̃χ , P̃χ + δ contain one tangent
 P̃1 , P̃ 2 , there
 is no tangent point within the left neighborhood
  point.
P̃2 − δ, P̃2 ; if L (Ps, λ, μ, ν) is not convex over P̃1 , P̃2 , Based on the above discussions, we develop Algorithm 2
the left neighborhood P̃2 − δ, P̃2 contains  one tangent point. shown in the next page to determine all tangent points corre-
For s P = P̃ |C2 |−1 , its right neighborhood P̃|C2 |−1 , P̃|C2 |−1 + sponding to the set C2 , where Δ in Algorithm 2 denotes the
δ contains  one tangent  points if L (Ps , λ, μ, ν) is con- step-length of each iteration. Fig. 3(c) illustrates the tangent
vex over P̃|C2 |−1 , P̃|C2 | or includes two tangent points if points annotated by the star mark. Then, we can construct the
    set C3 , which is composed by the tangent points obtained
P̃|C2 |−1 , P̃|C2 | is not convex over P̃|C2 |−1 , P̃|C2 | .
If χ = |C2 | − 1, each P̃i corresponds by Algorithm 2 as well as the two endpoints, i.e., Ps = 0 and
  to two tangent points
within its left neighborhood P̃i − δ, P̃i for 3 ≤ i ≤ |C2 | − 2. Ps = Ppk .
Similar with the discussions for χ = 2, the tangent points Step 4: Obtain the expression of L̃ (Ps , λ, μ, ν).
corresponding Based on our obtained set C3 , we can determine the lower-
  to P̃2 and P̃|C2 |−1 also  depend on the convexity boundary function L̃ (Ps , λ, μ, ν). Without loss of generality,
over P̃1 , P̃2 and P̃|C2|−1 , P̃|C2 | . For Ps = P̃2 , its left  |C3 |
neighborhood P̃2 − δ, P̃2 corresponds we denote C3  ψi i=1 and assume  that ψi < ψj
  to one tangent point if
L (Ps , λ, μ, ν) is convex over P̃1 ,P̃2 or two tangent points if if i < j.Then,  the interval  0, Ppk can be divided
L (Ps , λ, μ, ν) is not convex over P̃1 , P̃2 . For Ps = P̃|C2 |−1 , as ψ1 , ψ2 ··· ψ|C3 |−1 , ψ|C3 | . Consider two consecu-
   
there is a unique tangent tive subintervals ψi , ψi+1 and ψi+1 , ψi+2 , where i =
  point within its left neighborhood
P̃|C2 |−1 − δ, P̃|C2 |−1 . Moreover, its right neighborhood 1, 2, · · · , |C3 | − 2. On the above two consecutive subintervals,

⎧
⎨ L (ψ2n+1 , λ, μ, ν) − L (ψ2n , λ, μ, ν)    
L (ψ2n , λ, μ, ν) + Ps − ψ2n , Ps ∈ ψ2n , ψ2n+1
L̃ (Ps ,λ, μ, ν) = ψ2n+1 − ψ2n (31)
⎩
L (Ps , λ, μ, ν) , Otherwise
⎧
⎨ L (ψ2n , λ, μ, ν) − L (ψ2n−1 , λ, μ, ν)    
L (ψ2n−1 , λ, μ, ν) + Ps − ψ2n−1 , Ps ∈ ψ2n−1 , ψ2n
L̃ (Ps , λ, μ, ν) = ψ2n − ψ2n−1 (32)
⎩
L (Ps , λ, μ, ν) , Otherwise
WANG et al.: UNIFIED QoS AND SECURITY PROVISIONING FRAMEWORK FOR WIRETAP CRNs 1557

Algorithm 2: Find Out the Tangent Points Based on the must be convex over the latter subinterval. Consequently,
Point Set C2 the lower-boundary function L̃ (Ps , λ, μ, ν) can be determined
1 InPut: C2 = {P̃i }(i = 1, · · · , |C2 |); as follows.
2  
Initialization: C3 ← {(0, L(0, λ, μ, ν))}, If d L(PdP s ,λ,μ,ν)
2 ≥ 0 over ψ1 , ψ2 , i.e., L (Ps , λ, μ, ν) is the
2
 Δ > 0; s
convex function within the first subinterval,
3 if d2 L(Ps ,λ,μ,ν)
≥ 0 over Ps ∈ P̃1 , P̃2 then  then the lower-
 ψ1 , ψ2 → line-
dPs2 boundary function satisfies: convex over
 
4 for k ← 2 to |C2 | − 2 do segment over ψ2 ,ψ3 → convex over ψ3 , ψ4 → · · · →
◦ ◦
5 Pleft ← P̃k − δ, Pright ← P̃k+1 − δ; line-segment over ψ|C3 |−1 , ψ|C3 | if |C3 | is odd or convex
◦

6 while Pleft ≤ P̃k + δ do over ψ|C3 |−1 , ψ|C3 | if |C3 | is even. Consequently, the lower-
◦
7 while Pright ≤ P̃k+1 + δ do boundary function L̃ (Ps , λ, μ, ν) in this case can be deter-
8 if mined by (31), shown at the bottom of the previous page,
 
dL(Ps ,λ,μ,ν)  s ,λ,μ,ν) 
= dL(PdP = where n = 1, · · · , |C23 | − 1 if |C3 | is even and n =
dPs Ps =P ◦ Ps =P ◦
1, · · · , |C32|−1 if |C3 | is odd.
left s right
◦ ◦
L(Pleft ,λ,μ,ν)−L(Pright ,λ,μ,ν)
then  
◦ −P ◦
Pleft
 If L (Ps , λ, μ, ν) is nonconvex over ψ1 , ψ2 , then the lower-
right
◦ ◦

9 C3 ← C3 ∪ Pleft , L(Pleft , λ, μ, ν) , boundary function L̃ (Ps , λ, μ, ν) must be a line-segment  
 ◦ ◦
 within this subinterval  and
 satisfie: line-segment over
 ψ1 ,ψ2
Pright , L(Pright , λ, μ, ν) ; → convex over ψ2 ,ψ3 → line-segment over ψ3 , ψ4 →

10 else · · · → convex over ψ|C3 |−1 ,ψ|C3 | if |C3 | is odd or line-
11
◦
Pright ◦
← Pright + Δ; segment over ψ|C3 |−1 , ψ|C3 | if |C3 | is even. Therefore,
L̃ (Ps , λ, μ, ν) is given by (32), shown at the bottom of the
◦ ◦
12 Pleft ← Pleft + Δ; previous page, where n = 1, · · · , |C23 | − 1 if |C3 | is even and
n = 1, · · · , |C32|−1 if |C3 | is odd.
13 else Based on the above described four-step procedures, we can
◦
14 Pright ← P̃2 − δ; obtained the lower-boundary function L̃ (Ps , λ, μ, ν) of the
◦
15 while Pright ≤ P̃2 + δ do convex hull for the original nonconvex function L (Ps , λ, μ, ν),

dL(Ps ,λ,μ,ν) 
◦
L(Pright ,λ,μ,ν)−L(P̃1 ,λ,μ,ν) which is illustrated in Fig. 3(c) by the dashed line.
16 if dPs Ps =P ◦
= ◦ )−P̃
(Pright
right 1

then
  ◦ ◦

C. Optimal Solution
17 C3 ← C 3 Pright , L(Pright , λ, μ, ν) ;
By using the derived lower-boundary function L̃(Ps , λ,
18 else μ, ν), we can convert the original nonconvex Lagrangian
◦ ◦
19 Pright ← Pright + Δ; function L (Ps (θp , θe , h) , λ, μ, ν) given by Eq. (28) to
20 go to step 4 to step 12; the convexified Lagrangian function, which is denoted by
  L̃ (Ps (θp , θe , h) , λ, μ, ν) and can be written as
d2 L(Ps ,λ,μ,ν)
21 if dPs2 ≥ 0 over Ps ∈ P̃|C2 |−1 , P̃|C2 | then  
L̃ (Ps (θp , θe , h) , λ, μ, ν) = Eh L̃ (Ps (θp , θe , h) , λ, μ, ν)
22 else
23
◦
Pleft ← P̃|C2 |−1 − δ; −λe−βp RA − μeβe RE − νPav .
◦
24 while Pleft ≤ P̃|C2 |−1 + δ do (33)

dL(Ps ,λ,μ,ν) 
◦
L(P̃|C2 | ,λ,μ,ν)−L(Pleft ,λ,μ,ν)
25 if dPs ◦ =
Ps =Pleft ◦ (P̃|C2 | )−Pleft As the lower-boundary function L̃ (Ps , λ, μ, ν) is convex,
then the
 obtained
 L̃ (Ps (θp , θe , h) , λ, μ, ν) is also convex over
  ◦ ◦

0, Ppk . Then, the dual function of L̃ (Ps (θp , θe , h) , λ, μ, ν)
26 C3 ← C 3 Pleft , L(Pleft , λ, μ, ν)
is defined by
27 else  
28
◦
Pleft ◦
← Pleft + Δ; G (λ, μ, ν) = min L̃ (Ps (θp , θe , h) , λ, μ, ν)
0≤Ps (θp ,θe ,h)≤Ppk

   (34)
29 C3 ← C 3 P̃|C2 | , L(P̃|C2 | , λ, μ, ν) ;
 i  and the corresponding Lagrangian dual problem can be for-
30 OutPut: C3 = Ptan i
, L(Ptan , λ, μ, ν) i∈{1,2,··· ,|C3 |} .
mulated as follows:
 
(P3) max G (λ, μ, ν) (35)
λ≥0, μ≥0, ν≥0
the lower-boundary function L̃ (Ps , λ, μ, ν) only appear as one
of the following two cases. On the one hand,if L̃ (Ps , λ, μ, ν) According to the convex optimization theory, the dual func-
is convex over the former subinterval ψi , ψi+1 , then tion G (λ, μ, ν) is concave and the strong duality between
 μ, ν) must
L̃ (Ps , λ,  be a line-segment over the latter subin- L̃ (Ps (θp , θe , h) , λ, μ, ν) and G (λ, μ, ν) holds as the con-
terval ψi+1 , ψi+2 . On the other hand, if L̃ (Ps , λ, μ, ν) is vexified Lagrangian function L̃ (Ps (θp , θe , h) , λ, μ, ν) is con-
a line-segment over the former subinterval, L̃ (Ps , λ, μ, ν) vex. Consequently, by applying the Karush-Kuhn-Tucker
1558 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 3, MARCH 2019

(K.K.T.) conditions, the optimal power allocation denoted by • Case 3: In this case, L (Ps , λ, μ, ν) is neither con-
Ps∗ (θp , θe , h) must satisfy the following requirements: vex nor concave over 0, Ppk . Based on the analysis in
⎧  Section V-B, we know that L̃ (Ps (θp , θe , h) , λ, μ, ν) is deter-
⎪ dL̃ (Ps , λ∗ , μ∗ , ν ∗ ) 
⎪
⎪
⎪  = 0, Ps∗ (θp , θe , h) mined by Eqs. (31), (32) and (33). Since the convexi-
⎪
⎪ dPs (θp , θe , h) 
⎪
⎪ ∗ fied Lagrangian
  function L̃ (Ps (θp , θe , h) , λ, μ, ν) is convex
⎪
⎪
Ps =Ps (θp ,θe ,h)
⎪
⎪  ∈ (0, Ppk ) over 0, Ppk , the optimal solution Ps∗ (θp , θe , h) is also deter-
⎪
⎪
⎨ dL̃ (Ps , λ∗ , μ∗ , ν ∗ )  mined by Eq. (37).
 > 0, Ps∗ (θp , θe , h) = 0 Based on the above analysis, we can obtain the opti-
⎪ dPs (θp , θe , h) 
⎪
⎪
⎪ Ps=Ps (θp ,θe ,h)
∗
mal solution Ps∗ (θp , θe , h) corresponding to the convexified
⎪
⎪ ∗ ∗ ∗ 
⎪
⎪ dL̃ (P , λ , μ , ν )  Lagrangian function L̃ (Ps (θp , θe , h) , λ, μ, ν). Here we ana-
⎪ < 0, Ps∗ (θp , θe , h)
s
⎪
⎪  lyze the optimality of our derived Ps∗ (θp , θe , h) to the original
⎪
⎪ dP s (θ p , θ e , h) 
⎪
⎩ Ps =Ps∗ (θp ,θe ,h)
problem (P1).
= Ppk
In Case 1, L (P  s , λ, μ, ν) is a convex function of
(36) Ps (θp , θe , h) within 0, Ppk . Thus, the lower-boundary func-
where λ∗ , μ∗ , and ν ∗ are optimal Lagrangian multiplies, tion L̃ (Ps , λ, μ, ν) is the same as the original function
i.e., the optimal solution to the dual problem (P3), and can L (Ps , λ, μ, ν), which implies that the convexified Lagrangian
be numerically determined by the subgradient method [55]. function L̃ (Ps (θp , θe , h) , λ, μ, ν) is also identical with the
Then, the optimal power allocation strategy Ps∗ (θp , θe , h) can original Lagrangian function L (Ps (θp , θe , h) , λ, μ, ν). Con-
be determined as follows.   sequently, our obtained optimal power allocation strat-
• Case 1: If L (Ps , λ, μ, ν) is convex over 0, Ppk , egy Ps∗ (θp , θe , h) is also optimal solution to the original
L̃ (Ps (θp , θe , h) , λ, μ, ν) is determined by Eqs. (29) and (33). problem (P1).
Then, the optimal power allocation strategy Ps∗ (θp , θe , h) is In Case 2 and Case 3, L (Ps , λ,  μ, ν) is not a con-
given by vex function of Ps (θp , θe , h) over 0, Ppk , which causes
⎧  that the Lagrangian function L (Ps (θp , θe , h) , λ, μ, ν) is also
⎪
⎪ dL̃ (Ps , λ∗ , μ∗ , ν ∗ )  not convex. By employing the theories of convex hull and
⎪
⎪ 0, if dP (θ , θ , h)  >0
⎪
⎪ probabilistic transmission, we determine the lower-boundary
⎨ s p e
Ps =0
Ps∗ (θp , θe , h) = dL̃ (Ps , λ∗ , μ∗ , ν ∗ )  function L̃ (Ps , λ, μ, ν) of the convexhull for  L (Ps , λ, μ, ν),
⎪
⎪ P , if  <0 which is convex over Ps (θp , θe , h) ∈ 0, Ppk . Then, we con-
⎪
⎪
pk
dPs (θp , θe , h) 
⎪
⎪ vert the original nonconvex Lagrangian function to the con-
⎩ Ps =Ppk

P̃1 , Otherwise vexified Lagrangian function L̃ (Ps (θp , θe , h) , λ, μ, ν) and

(37) obtain the corresponding optimal solution Ps∗ (θp , θe , h) by
the Lagrangian approach. Since the lower-boundary func-
where P̃1 is the solution of the following equation:
tion L̃ (Ps , λ, μ, ν) is composed by the convex parts of
dL̃ (Ps (θp , θe , h) , λ∗ , μ∗ , ν ∗ ) L (Ps , λ, μ, ν) and line-segments, where the endpoints of each
dPs line-segment also belong to L (Ps , λ, μ, ν), our obtained opti-
λ∗ βp hsp hpp Pp mal solution Ps∗ (θp , θe , h) must belong to the range where
= the lower-boundary function L̃ (Ps , λ, μ, ν) overlaps with the
(hsp Ps + σ 2 )2
 −βp −1 original function L (Ps , λ, μ, ν). Consequently, the optimal
hpp Pp power allocation strategy Ps∗ (θp , θe , h) corresponding to the
· 1+ convexified Lagrangian function L̃ (Ps (θp , θe , h) , λ, μ, ν) is
hsp Ps + σ 2
 βe −1 also optimal to the original problem (P1).
μ∗ βe hss hss Ps Remark 1: In our proposed statistical queueing analysis
+ 1+ based QoS and security protection framework, we exploit
hps Pp + σ 2 hps Pp + σ 2
the delay-sensitive feature of wireless services/application for
hss security protection. In this way, we only need to guaran-
+ν ∗ − = 0. (38)
hps Pp + hss Ps + σ 2 tee that the eavesdropper cannot accumulate enough data
  from the legitimate link within the certain period. Thus,
• Case 2: If L (Ps , λ, μ, ν) is concave over 0, Ppk ,
the statistical delay-sensitive secrecy protection approach
L̃ (Ps (θp , θe , h) , λ, μ, ν) is determined by Eqs. (30)
dL̃(Ps (θp ,θe ,h),λ,μ,ν) can provide the possibility to improve the performance
and (33). We can easily obtain that dPs (θp ,θe ,h) = of legitimate link while meeting the given security
L(Ppk ,λ,μ,ν)−L(0,λ,μ,ν)
Ppk , which is a constant. Consequently, requirement.
the optimal power allocation strategy is determined by Remark 2: Different from the widely-used physical-layer
⎧ security technology, which uses the secrecy capacity as the
⎪
⎪ dL̃ (Ps , λ∗ , μ∗ , ν ∗ )
⎨ 0, if ≥ 0, main performance metric and thus couples the reliability and
Ps∗ (θp , θe , h) =  s (θp∗, θe∗, h)∗ 
dP security of wireless communications together, we employ the
⎪
⎪ dL̃ Ps , λ , μ , ν first-separation-then-coupling approach for security protec-
⎩ Ppk , if < 0.
dPs (θp , θe , h) tion. Specifically, we first separate the security from reliability.
(39) Then, by accurately characterize the delay-sensitive secrecy
WANG et al.: UNIFIED QoS AND SECURITY PROVISIONING FRAMEWORK FOR WIRETAP CRNs 1559

requirement, we formulate the optimization problem, where (1) Case 1: If the following two inequalities,
 −βp  
the given security requirement is treated as the constraint. 
hpp Pp −βp RA 
Thus, the reliability and security are coupled again. How- Eh 1+ − e  ≤0
ever, we can control the reliability and security requirements hsp Ps (θp , θe ) + σ 2 
Ps =P max
separately and thus achieve higher flexibility for resource (43)
utilization.
and
Remark 3: The convex hull and probabilistic transmissions   
 βe 
approach employed in this paper, where we mainly focus Eh 1 + γ1 − eβe RE  ≤ 0, (44)
on the single variable scenario, can also be extended to the Ps =P max
multiple variables scenario, such as the multi-user networks. hold simultaneously, which imply that both PU’s statistical
However, different from the inflexion point determined in delay QoS requirement and SU’s statistical delay-sensitive
this paper, we might need to construct the inflexion hyper- secrecy constraint can be satisfied even when SS transmits with
plane for the multiple variables case. Then, the arbitrary its maximum transmit power, then the optimal fixed power
non-convex optimization problem can be converted to the ∗
allocation scheme Pfixed (θp , θe ) = P max .
equivalent convex problem and easily solved by the well- (2) Case 2: If Eq. (43) and the inequality
 βe  
known Lagrangian method. Consequently, the utilized convex 
hull and probabilistic transmissions approach can be regarded Eh 1 + γ 1 − eβe RE  >0 (45)
Ps (θp ,θe )=P max
as an efficient and unified method for solving the non-convex are satisfied simultaneously, the optimal solution of the prob-
problem. lem (P4) only depends on the SU’s statistical delay-sensitive
secrecy constraint. Then, if the following inequality
 βe  

VI. T HE F IXED P OWER A LLOCATION S CHEME Eh 1 + γ 1 − eβe RE  ≤0
Ps (θp ,θe )=0
In Section V, we have obtained the optimal power allo- (46)
cation strategy Ps∗ (θp , θe , h), which adapts to PU’s delay
QoS requirements, SU’s information security constraint,
holds, there  a point P̃th1 ∈ [0, P max ) satis-
 must exist
βe
and channel conditions in the network. In this section, fying Eh 1 + γ 1 = eβe RE . Therefore, we can get
∗
we develop a suboptimal scheme named fixed power allo- Pfixed (θp , θe ) = P̃th1 .
cation, where SU’s transmit power only adjusts according (3) Case 3: If Eq. (44) and the inequality
 −βp  
to PU’s delay QoS and SU’s information security require- 
hpp Pp −βp RA 
ments. Then, the optimization problem can be formulated Eh 1 + − e  >0
as hsp Ps (θp , θe )+σ 2 
Ps =P max
(47)
(P4)
 are satisfied simultaneously, the optimal solution of problem
 
hss Ps (θp , θe ) (P4) only depends on PU’s statistical delay QoS requirement.
max Eh log 1 + Then, if the following inequality
hps Pp + σ 2  
−βp 
0≤Ps (θp ,θe )≤P max

(40) hpp Pp −βp RA 
 Eh 1+ −e  ≤0
−βp hsp Ps (θp , θe ) + σ 2 
hpp Pp Ps =0
s.t. Eh 1+ −e−βp RA ≤ 0 (48)
hsp Ps (θp , θe )+σ 2
(41) holds,
 there must exist apoint  P̃th2 ∈ [0, P max ) satisfying
  −βp
 βe = e−βp RA . Thus, the opti-
hpp Pp
Eh 1 + hsp Ps (θp ,θe )+σ2
Eh 1 + γ1 − eβe RE ≤ 0 (42)
∗
mal solution Pfixed (θp , θe ) = P̃th2 .
(4) Case 4: If Eqs. (45), (46), (47), and (48) hold
where P max = min {Pav , Ppk }, γ1 =
simultaneously, then we have that both PU’s statistical
hss Ps (θp ,θe ) hse Ps (θp ,θe )
min hps Pp +σ2 , hpe Pp +σ2 , Eqs. (41) and (42) delay QoS requirement and SU’s statistical delay-sensitive
denote PU’s statistical delay QoS requirement secrecy constraint are active. Thus, the optimal solution
and SU’s statistical delay-sensitive secrecy ∗
Pfixed (θp , θe ) =min{P̃th1, P̃th2 }.
constraint, respectively. Let t1 (Ps (θp , θe )) =
 −βp (5) Case 5: If the following inequality
hpp Pp −βp RA   
1 + hsp Ps (θp ,θe )+σ2 − e and t2 (Ps (θp , θe )) =  βe 
  β e
Eh 1 + γ1 − eβe RE  >0 (49)
1 + γ1 − eβe RE . Then, we have dt1 (Ps (θp , θe ))/ Ps (θp ,θe )=0
Ps (θp , θe ) > 0 and dt2 (Ps (θp , θe ))/Ps (θp , θe ) > 0, or
 −βp  
which imply that Eqs. (41) and (42) are both monotonically 
hpp Pp −βp RA 
increasing with SS’s transmit power Ps (θp , θe ). Consequently, Eh 1+ −e  >0
hsp Ps (θp , θe ) + σ 2 
the optimal fixed power allocation scheme, denoted by Ps =0
∗
Pfixed (θp , θe ), can be determined as follows. (50)
1560 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 3, MARCH 2019

when PU’s QoS and SU’s delay-sensitive secrecy requirements

are both loose, i.e., the value of QoS exponent θp and delay-
sensitive secrecy exponent θe are small, SU can get high
average throughput; when PU’s QoS and SU’s delay-sensitive
secrecy requirements are both stringent, i.e., the value of QoS
exponent θp and delay-sensitive secrecy exponent θe are large,
SU can only get low average throughput. Under the given θe ,
when PU’s QoS requirement becomes stringent, i.e., QoS
exponent θp increases, SU’s average throughput decreases.
Yet, when PU’s QoS requirement becomes too stringent that θp
is larger than a certain threshold, the average throughout of SU
vanishes, which implies that SU has to cease its transmission
to protect PU’s QoS requirement. Under the given θp , we can
observe the similar trend regarding SU’s throughput over θe ,
i.e., the secrecy rate decreases with respect to an increasingly
stringent secrecy exponent θe . As our proposed optimal power
allocation strategy can adapt to not only PU’s QoS and
SU’s secrecy requirements, but also the channel conditions
characterized by the NGV h, SU’s transmit power can be
more wisely allocated under the optimal strategy as compared
to the fixed power allocation scheme. Consequently, the opti-
mal power allocation strategy can achieve much higher SU’s
throughput than the fixed power allocation scheme, which can
be observed from Fig. 4(a).
To further verify the superiority of our proposed optimal
power allocation strategy, we employ another existing power
allocation scheme, namely, the conventional security-based
water-filling policy, which considers the secure transmission of
information over an ergodic fading channel in the presence of
an eavesdropper [37], [38]. Fig. 4(b) illustrates the normalized
average throughout of SU as the function of SU’s maximum
allowed average transmit power Pav under different values of
Fig. 4. (a) SU’s normalized average throughput of our proposed optimal
power allocation strategy and fixed power allocation scheme versus primary PU’s delay QoS demand θp and SU’s statistical delay-sensitive
user’s QoS exponent θp and secondary user’s security exponent θe . (b) SU’s secrecy requirement θe . We can observe from Fig. 4(b) that our
normalized average throughput of our proposed optimal power allocation, proposed strategy can achieve higher throughput as compared
the fixed power allocation, and the conventional security-based water-filling
policy under different values of PU’s QoS exponent θp and SU’s security to the conventional security-based water-filling policy, regard-
exponent θe versus SU’s average transmit power. less of SU’s maximum allowed average transmit power Pav .
Moreover, the higher Pav becomes, the greater advantage
holds, which implies that PU’s statistical delay QoS require- of our proposal achieves. This is because that our proposed
ment or SU’s statistical delay-sensitive secrecy constraint can- optimal strategy can adapt to PU’s delay QoS requirement,
not be satisfied even when SS gives up its own transmission, SU’s statistical delay-sensitive secrecy requirement, and the
∗
thus we can obtain Pfixed (θp , θe ) = 0. time-varying channel conditions. Our proposed strategy goes
beyond the throughput of the baseline approach. The conven-
tional security-based water-filling policy cannot satisfy SU’s
VII. S IMULATION R ESULTS
statistical delay-sensitive secrecy requirement, which stops
In this section, we evaluate the performance of our proposed transmission when the condition of SS-SR link is worse than
power allocation strategy through simulations. In our simula- that of SS-SE. However, by exploiting the stochastic feature
tions, we set the frame duration T = 2 ms, the bandwidth of wireless channels, our proposed optimal power allocation
W = 105 Hz, PU’s transmit power Pp = 10 dB, SU’s strategy can adapt to SU’s statistical delay-sensitive secrecy
maximum allowed average and peak transmit power Pav = requirement. our proposed optimal power allocation strategy
10 dB and Ppk = 15 dB, respectively. The constant data arrival has more transmit chances than the conventional security-
rate of PU RA = 1.5 nats/s/Hz and the constant data leaving based water-filling policy. Furthermore, we can also observe
rate of SE RE = 1 nats/s/Hz. from Fig. 4(b) that SU’s throughput can not be improved when
Figures 4(a) shows the normalized average throughput Pav is larger than a certain threshold for both our proposal and
of SU as the function of PU’s delay QoS exponent θp and the baseline schemes. The reason is that SU’s throughput is not
SU’s statistical delay-sensitive secrecy exponent θe under our only restrained by average transmit power Pav but also other
proposed optimal power allocation strategy and the fixed constraints, such as PU’s QoS and SU’s security demands.
power allocation scheme. We can observe from Fig. 4(a) that When Pav is larger than a certain threshold, SU’s throughput
WANG et al.: UNIFIED QoS AND SECURITY PROVISIONING FRAMEWORK FOR WIRETAP CRNs 1561

Fig. 5. (a) SU’s normalized average throughput of our proposed optimal Fig. 6. Performance evaluation of our proposed power allocation, the fixed
power allocation strategy, the fixed power allocation scheme, and the conven- power allocation scheme and the conventional security-based water-filling
tional security-based water-filling policy under different values of PU’s QoS policy versus transmit power Pp . (a) SU’s average utilized transmit power.
exponent θp and SU’s security exponent θe versus the data arrival rate RA (b) SU’s normalized average throughput.
of PS’s data queue. (b) SU’s normalized average throughput of our proposed
optimal power allocation strategy and fixed power allocation scheme under
different values of PU’s QoS exponent θp and SU’s security exponent θe as
a function of the data departure rate RE of SE’s eavesdropped data queue.
schemes versus SE’s data departure rate RE under different
PU’s delay QoS requirement θp and SU’s statistical delay-
sensitive secrecy demand θe . We can observe from Fig. 5(b)
is constrained by PU’s QoS and SU’s security demands, even that the average throughput of SU under the two schemes
increasing Pav , SU’s throughput is not increasing. are both increasing with SE’s data service rate RE under
Figures 5(a) shows the SU’s normalized average throughput any given θp and θe . This is because large value of RE
versus PU’s data arrival rate RA with different values of PU’s will cause the data eavesdropped by SE to become useless
delay QoS requirement θp and SU’s statistical delay-sensitive within a short period and thus lead to loose security constraint.
secrecy demand θe . We can observe from Fig. 5(b) that SU’s On the contrary, when RE gets smaller, the eavesdropped data
throughput achieved by all the schemes decrease with PU’s will be effective within the longer period. Thus, the smaller
data arrival rate RA under any given θp and θe . The reason value of RE means more stringent security constraint and
for this phenomenon can be explained as follows. In order results in the degradation of SU’s throughput. However, since
to support higher data arrival rate while meeting the given our proposed optimal power allocation strategy can smartly
statistical delay QoS requirement, PS needs to increase its allocate the upperbounded transmit power budget according
service rate. Thus, SS has to decrease its transmit power to channel conditions, the optimal strategy can achieve higher
to reduce the interference that is received by PR, which SU’s throughput as compared to the fixed power allocation
causes the degradation of SU’s throughput. Although SU can scheme.
only obtain lower throughput as RA increases, our proposed Figures 6(a) and 6(b) show SU’s average utilized transmit
optimal power allocation strategy still achieves the highest power and the normalized throughput versus PU’s transmit
SU’s throughput. power Pp under different PU’s delay QoS requirements and
Figures 5(b) shows the SU’s normalized average throughput SU’s delay-sensitive secrecy demands, respectively. We can
achieved by our proposed optimal and fixed power allocation observe from Fig. 6(a) and 6(b) that PU’s transmit power
1562 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 3, MARCH 2019

significantly affects SU’s performance. Specifically, when

PU transmits with low power, in order to meet PU’s sta-
tistical delay QoS requirement under the given data arrival
rate RA , SU has to give up its transmission to avoid the
degradation of PU’s data service rate and thus can only achieve
zero throughput. With the increase of PU’s transmit power,
SU is allowed to transmit. Then, SU’s average utilized transmit
power correspondingly increases and SU can achieve higher
throughput. When PU’s transmit power reaches a certain
threshold, SU’s average utilized power becomes flattened
due to the SU’s transmit power constraint. However, SU’s
throughput decreases while constantly increasing PU’s trans-
mit power, as shown in Fig. 6(b). This is because higher
PU’s transmit power results in larger interference to SU’s
transmission. However, SU cannot keep increasing its transmit
power due to the delay-sensitive secrecy requirement as well as
its average and peak transmit power constraints. Consequently,
SU’s throughput degrades with constantly increasing PU’s
transmit power. Although PU’s transmit power will cause
significant impact on SU’s performance, our proposed optimal
power allocation strategy can still outperforms the fixed power
allocation scheme and the conventional security-based water-
filling policy.
In this paper, we mainly focus on the scenario that
PS transmits with a constant power Pp . However, if advanced
power allocation scheme is available for PS and allowed to
share with the CRN, the average throughput achieved by
SU is expected to be improved. Consequently, we consider
two adaptive power allocation schemes for PU, which are
the well-known water-filling algorithm [56] and the effec-
tive capacity maximization based policy [57], respectively.
Specifically, the water-filling algorithm allows PS to adjust
the transmit power based on the channel power gain hpp and
the effective capacity maximization based power allocation Fig. 7. (a) SU’s normalized average throughput versus PU’s QoS exponent
θp under different power allocation schemes of PU, where PU’s maximum
policy enables PS to adapt the transmit power not only to the allowed average power for water-filling and effective capacity maximization
channel condition, but also to PU’s QoS requirement specified based schemes are set to be 10dB and SU’s secrecy exponent θe = 10−5 .
by the QoS exponent θp . Figure 7(a) shows SU’s normal- (b) SU’s normalized average throughput versus channel estimation error ρ
under different PU’s QoS and SU’s delay-sensitive secrecy requirements.
ized average throughput versus PU’s QoS exponent θp under
different power allocation schemes of PU. We can observe
from Fig. 7(b) that if adaptive power allocation schemes are channel information on the network performance via simu-
employed by PU and shared with the CRN, then CRN can lation. In particular, we employ  a widely-usedchannel esti-
achieve higher average throughput as compared to PU’s con- mation error model, i.e., g = 1 − ρ2 ĝ + ρ2 n, where
stant power transmission scheme. The performance advantage g denotes the true fading channel coefficient, ĝ denotes the
mainly comes from that, the adaptive power allocation scheme estimated fading channel coefficient, ρ2 ∈ [0, 1] represents the
can efficiently increase PU’s transmission rate through smartly variance for the effective channel estimation errors, and n is
allocate the upperbounded power budget and thus might toler- the circularly symmetric complex Gaussian (CSCG) random
ate higher interference power from SU’s transmission. Then, variable with zero mean and unit variance. Then, we have
SU may use larger transmit power and achieve better average the actual channel power gain h = |g|2 and the estimated
throughput performance. Moreover, as the effective capacity channel power gain ĥ = |ĝ|2 . Figure 7(b) shows the normal-
maximization based power allocation policy converges to ized average throughput achieved by SU versus the channel
the water-filling algorithm when θp → 0 [57], both water- estimation error ρ under different PU’s QoS requirements and
filling and effective capacity maximization based schemes SU’s delay-sensitive secrecy demands. We can observe from
achieve the same SU’s average throughput under small values Fig. 7(b) that the imperfect channel information will degrade
of θp . SU’s performance. To efficiently combat with the impact of
In the previous discussions, we assume that the perfect the obtained imperfect channel information, possible channel
channel information can be obtained, which usually can- estimation error should be integrated into the framework and
not be guaranteed in practical wireless networks. Conse- power allocation strategy designs, which are left for further
quently, we will evaluate the elementary impact of imperfect works.
WANG et al.: UNIFIED QoS AND SECURITY PROVISIONING FRAMEWORK FOR WIRETAP CRNs 1563

Fig. 8. The illustration of the proof for Theorem 1. (a) Case A: Ψ = i + 2. (b) Case B: Ψ = i + 1. (c) Case C: Ψ = i.

VIII. C ONCLUSIONS • Case B: Ψ = i + 1

In this paper, we developed a unified PU’s QoS and SU’s In this case, we have κi ≤ 0 and κi+1 ≥ 0 as illustrated
security provisioning framework for wiretap CRNs, which is in Fig. 8(b). Thus, κi ≤ κi+1 must hold.
• Case C: Ψ = i
based on statistical queueing analysis. Specifically, our estab-
lished framework is characterized by the queue-length bound The analysis in this case is similar as that in Case A.
violation probability, we further convert the QoS and security Specifically, as κi ≥ 0 and κi cannot be larger than the line-
requirements to the equivalent PU’s effective capacity and segment Li,i+2 , the line-segment Li,i+1 must be below the
SU’s effective bandwidth constraints. Based on the proposed line-segment Li,i+2 as shown in Fig. 8(c). Thus, according to
framework, we formulated the optimization problem aiming the geometrical relationship, we can also have κi ≤ κi+1 .
at maximizing the average throughput of SS subject to the
PU’s statistical delay QoS requirement, SU’s statistical delay- R EFERENCES
sensitive secrecy constraint as well as SS’s average and peak [1] S. Haykin, “Cognitive radio: Brain-empowered wireless communica-
transmit power constraints. To tackle the non-convexity in the tions,” IEEE J. Sel. Areas Commun., vol. 23, no. 2, pp. 201–220,
formulated problem, we convert the original nonconvex prob- Feb. 2005.
[2] J. Wang and X. Zhang, “Statistical QoS-driven cooperative power
lem to the equivalent convex problem by using the theories allocation game over wireless cognitive radio networks,” in Proc. IEEE
of convex hull and probabilistic transmission , and finally get Wireless Commun. Netw. Conf. (WCNC), San Francisco, CA, USA,
the optimal power allocation based on Lagrangian method. Mar. 2017, pp. 1–6.
[3] Y. Wang, P. Ren, Q. Du, and L. Sun, “Optimal power allocation for
Moreover, we also developed a fixed power allocation scheme, underlay-based cognitive radio networks with primary user’s statistical
which provides the suboptimal solution while features low delay QoS provisioning,” IEEE Trans. Wireless Commun., vol. 14,
complexity. Simulation results demonstrated the superiority of no. 12, pp. 6896–6910, Dec. 2015.
[4] Y. Wang, P. Ren, F. Gao, and Z. Su, “A hybrid underlay/overlay
our proposed optimal strategy as compared to the fixed power transmission mode for cognitive radio networks with statistical quality-
allocation scheme and the conventional security-based water- of-service provisioning,” IEEE Trans. Wireless Commun., vol. 13, no. 3,
filling policy. pp. 1482–1498, Mar. 2014.
[5] Y. Wang, P. Ren, and F. Gao, “Power allocation for statistical QoS
provisioning in opportunistic multi-relay DF cognitive networks,” IEEE
A PPENDIX A Signal Process. Lett., vol. 20, no. 1, pp. 43–46, Jan. 2013.
[6] A. Vosoughi, J. R. Cavallaro, and A. Marshall, “Trust-aware consensus-
T HE P ROOF OF T HEOREM 1 inspired distributed cooperative spectrum sensing for cognitive radio
ad hoc networks,” IIEEE Trans. Cogn. Commun. Netw., vol. 2, no. 1,
Proof: We denote Li,i+1 as the line-segment with pp. 24–37, Mar. 2016.
the two endpoints, which are P̃i , L(P̃i , λ, μ, ν) and [7] Y. Yang, S. Aissa, and K. N. Salama, “Spectrum band selection in delay-
  QoS constrained cognitive radio networks,” IEEE Trans. Veh. Technol.,
P̃j , L(P̃j , λ, μ, ν) that belong to the set C2 , and denote vol. 64, no. 7, pp. 2925–2937, Jul. 2015.
[8] M. R. Mili and L. Musavian, “Interference efficiency: A new
κi as the slope of of the line-segment Li,i+1 . Moreover, metric to analyze the performance of cognitive radio networks,”
for any given i (i = 1, 2, · · · , |C2 | − 2), we define Ψ  IEEE Trans. Wireless Commun., vol. 16, no. 4, pp. 2123–2138,
argmin L(P̃j , λ, μ, ν). Then, κi ≤ κi+1 can be proved by Apr. 2017.
j=i,i+1,i+2 [9] T. Aldalgamouni, M. C. Ilter, and H. Yanikomeroglu, “Joint power
the following three cases. allocation and constellation design for cognitive radio systems,” IEEE
Trans. Veh. Technol., vol. 67, no. 5, pp. 4661–4665, May 2018.
• Case A: Ψ = i + 2
[10] Z. Wang and W. Zhang, “Spectrum sharing with limited chan-
As Algorithm 1 tells us that κi is not larger than the nel feedback,” IEEE Trans. Wireless Commun., vol. 12, no. 5,
slope of the line-segment pp. 2524–2532, May 2013.
  Li,i+2 connected by two  endpoints [11] Z. Li, F. Xiao, S. Wang, T. Pei, and J. Li, “Achievable rate maximization
P̃i , L(P̃i , λ, μ, ν) and P̃i+2 , L(P̃i+2 , λ, μ, ν) , the line- for cognitive hybrid satellite-terrestrial networks with AF-relays,” IEEE
segment Li,i+1 must be below the line-segment Li,i+2 as J. Sel. Areas Commun., vol. 36, no. 2, pp. 304–313, Feb. 2018.
[12] L. Xu, A. Nallanathan, J. Yang, and W. Liao, “Power and bandwidth
shown in Fig. 8(a). Then, based on the geometrical relation- allocation for cognitive heterogeneous multi-homing networks,” IEEE
ship, we can have κi ≤ κi+1 . Trans. Commun., vol. 66, no. 1, pp. 394–403, Jan. 2018.
1564 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 18, NO. 3, MARCH 2019

[13] X. Kang, H. K. Garg, Y.-C. Liang, and R. Zhang, “Optimal power [35] X. Fang, N. Zhang, S. Zhang, D. Chen, X. Sha, and X. Shen,
allocation for OFDM-based cognitive radio with new primary transmis- “On physical layer security: Weighted fractional Fourier transform based
sion protection criteria,” IEEE Trans. Wireless Commun., vol. 9, no. 6, user cooperation,” IEEE Trans. Wireless Commun., vol. 16, no. 8,
pp. 2066–2075, Jun. 2010. pp. 5498–5510, Aug. 2017.
[14] Y. Li, N. Li, H. Li, W. Xie, M. Wang, and M. Peng, “Spectrum [36] S. Luo, J. Li, and A. P. Petropulu, “Uncoordinated cooperative jamming
sharing based on overlay cognitive full-duplex two-way OFDM relaying: for secret communications,” IEEE Trans. Inf. Forensics Security, vol. 8,
protocol design and resource allocation,” IEEE Trans. Veh. Technol., no. 7, pp. 1081–1090, Jul. 2013.
vol. 67, no. 3, pp. 2324–2334, Mar. 2018. [37] P. K. Gopala, L. Lai, and H. El Gamal, “On the secrecy capac-
[15] D. Darsena, G. Gelli, and F. Verde, “An opportunistic spectrum access ity of fading channels,” IEEE Trans. Inf. Theory, vol. 54, no. 10,
scheme for multicarrier cognitive sensor networks,” IEEE Sensors J., pp. 4687–4698, Oct. 2008.
vol. 17, no. 8, pp. 2596–2606, Apr. 2017. [38] X. Tang, P. Ren, and Z. Han, “Distributed power optimization for
[16] M.-G. Song, Y.-J. Kim, E.-Y. Park, and G.-H. Im, “Rate adaptation security-aware multi-channel full-duplex communications: A varia-
and power allocation for cognitive radio networks with HARQ-based tional inequality framework,” IEEE Trans. Commun., vol. 65, no. 9,
primary system,” IEEE Trans. Commun., vol. 62, no. 4, pp. 1178–1187, pp. 4065–4079, Sep. 2017.
Apr. 2014. [39] L. Xu, A. Nallanathan, X. Pan, J. Yang, and W. Liao, “Security-aware
[17] W. Hao, O. Muta, H. Gacanin, and H. Furukawa, “Power allocation for resource allocation with delay constraint for NOMA-based cognitive
massive MIMO cognitive radio networks with pilot sharing under SINR radio network,” IEEE Trans. Inf. Forensics Security, vol. 13, no. 2,
requirements of primary users,” IEEE Trans. Veh. Technol., vol. 67, no. 2, pp. 366–376, Feb. 2018.
pp. 1174–1186, Feb. 2018. [40] D. W. K. Ng, E. S. Lo, and R. Schober, “Multiobjective resource
[18] X. Kang, R. Zhang, Y.-C. Liang, and H. K. Garg, “Optimal power allocation for secure communication in cognitive radio networks with
allocation strategies for fading cognitive radio channels with primary wireless information and power transfer,” IEEE Trans. Veh. Technol.,
user outage constraint,” IEEE J. Sel. Areas Commun., vol. 29, no. 2, vol. 65, no. 5, pp. 3166–3184, May 2016.
pp. 374–383, Feb. 2011. [41] V.-D. Nguyen, T. Q. Duong, O.-S. Shin, A. Nallanathan, and
[19] B. Maham and P. Popovski, “Cognitive multiple-antenna network with G. K. Karagiannidis, “Enhancing PHY security of cooperative cognitive
outage and rate margins at the primary system,” IEEE Trans. Veh. radio multicast communications,” IEEE Trans. Cogn. Commun. Netw.,
Technol., vol. 64, no. 6, pp. 2409–2423, Jun. 2015. vol. 3, no. 4, pp. 599–613, Dec. 2017.
[20] W.-H. Chen, W.-R. Lin, H.-C. Tsao, and C. Lin, “Probabilistic power [42] S. Jia, J. Zhang, H. Zhao, and R. Zhang, “Relay selection for improved
allocation for cognitive radio networks with outage constraints and one- security in cognitive relay networks with jamming,” IEEE Wireless
bit side information,” IEEE Trans. Signal Process., vol. 64, no. 4, Commun. Lett., vol. 6, no. 5, pp. 662–665, Oct. 2017.
pp. 867–881, Feb. 2016. [43] Y. Liu, L. Wang, T. T. Duy, M. Elkashlan, and T. Q. Duong,
“Relay selection for security enhancement in cognitive relay net-
[21] W. Stallings, Cryptography and Network Security Principles
works,” IEEE Wireless Commun. Lett., vol. 4, no. 1, pp. 46–49,
and Practices. Upper Saddle River, NJ, USA: Prentice-Hall,
Feb. 2015.
2006.
[44] K. Tourki and M. O. Hasna, “Proactive spectrum sharing incen-
[22] C. E. Shannon, “Communication theory of secrecy systems,” Bell Labs
tive for physical layer security enhancement using outdated CSI,”
Tech. J., vol. 28, no. 4, pp. 656–715, Oct. 1949.
IEEE Trans. Wireless Commun., vol. 15, no. 9, pp. 6273–6283,
[23] A. D. Wyner, “The wire-tap channel,” Bell Syst. Tech. J., vol. 54, no. 8, Sep. 2016.
pp. 1355–1387, 1975. [45] J. Wu, K. OTA, M. Dong, and C. Li, “A hierarchical security frame-
[24] X. Chen, D. W. K. Ng, W. Gerstacker, and H. H. Chen, “A sur- work for defending against sophisticated attacks on wireless sen-
vey on multiple-antenna techniques for physical layer security,” IEEE sor networks in smart cities,” IEEE Access, vol. 4, pp. 416–424,
Commun. Surveys Tuts., vol. 19, no. 2, pp. 1027–1053, 2nd Quart., Mar. 2016.
2016. [46] L. Xiao, Y. Li, G. Han, G. Liu, and W. Zhuang, “PHY-layer spoofing
[25] A. Mukherjee, S. A. A. Fakoorian, J. Huang, and A. L. Swindlehurst, detection with reinforcement learning in wireless networks,” IEEE Trans.
“Principles of physical layer security in multiuser wireless networks: Veh. Technol., vol. 65, no. 12, pp. 10037–10047, Dec. 2016.
A survey,” IEEE Commun. Surveys Tuts., vol. 16, no. 3, pp. 1550–1573, [47] Z. Qin, Y. Gao, and M. D. Plumbley, “Malicious user detection based
3rd Quart., 2014. on low-rank matrix completion in wideband spectrum sensing,” IEEE
[26] Y. Wu, A. Khisti, C. Xiao, G. Caire, K.-K. Wong, and X. Gao, Trans. Signal Process., vol. 66, no. 1, pp. 5–17, Jan. 2018.
“A survey of physical layer security techniques for 5G wireless networks [48] H. Chen, M. Zhou, L. Xie, K. Wang, and J. Li, “Joint spectrum
and challenges ahead,” IEEE J. Sel. Areas Commun., vol. 36, no. 4, sensing and resource allocation scheme in cognitive radio networks with
pp. 679–695, Apr. 2018. spectrum sensing data falsification attack,” IEEE Trans. Veh. Technol.,
[27] S. Goel and R. Negi, “Guaranteeing secrecy using artificial noise,” vol. 65, no. 11, pp. 9181–9191, Nov. 2016.
IEEE Trans. Wireless Commun., vol. 7, no. 6, pp. 2180–2189, [49] Y. Wang, T. Wang, X. Tang, and P. Ren, “QoS and security aware
Jun. 2008. power allocation scheme for wiretap cognitive radio networks,” in Proc.
[28] Q. Li and L. Yang, “Artificial noise aided secure precoding for MIMO IEEE Veh. Technol. Conf. (VTC)-Fall, Chicago, IL, USA, Aug. 2018,
untrusted two-way relay systems with perfect and imperfect channel pp. 1–6.
state information,” IEEE Trans. Inf. Forensics Security, vol. 13, no. 10, [50] C.-S. Chang, “Stability, queue length, and delay of deterministic and
pp. 2628–2638, Oct. 2018. stochastic queueing networks,” IEEE Trans. Autom. Control, vol. 39,
[29] N.-P. Nguyen, H. Q. Ngo, T. Q. Duong, H. D. Tuan, and K. Tourki, no. 5, pp. 913–931, May 1994.
“Secure massive MIMO with the artificial noise-aided downlink train- [51] D. Wu and R. Negi, “Effective capacity: A wireless link model for
ing,” IEEE J. Sel. Areas Commun., vol. 36, no. 4, pp. 802–816, support of quality of service,” IEEE Trans. Wireless Commun., vol. 2,
Apr. 2018. no. 4, pp. 630–643, Jul. 2003.
[30] N. Zhao, F. R. Yu, M. Li, Q. Yan, and V. C. Leung, “Physical layer [52] Base Station (BS) Radio Transmission and Reception, Release 15,
security issues in interference-alignment-based wireless networks,” IEEE document 3GPP TS 38.104, 2018.
Commun. Mag., vol. 54, no. 8, pp. 162–168, Aug. 2016. [53] Cellular System Support for Ultra-Low Complexity and Low Through-
[31] R. Bassily and S. Ulukus, “Ergodic secret alignment,” IEEE Trans. Inf. put Internet of Things, Release 13, document 3GPP TR 45.820,
Theory, vol. 58, no. 3, pp. 1594–1611, Mar. 2012. 2015.
[32] J. Tang et al., “Associating MIMO beamforming with security codes to [54] D. Qiao and M. C. Gursoy, “Statistical delay tradeoffs in buffer-
achieve unconditional communication security,” IET Commun., vol. 10, aided two-hop wireless communication systems,” IEEE Trans. Commun.,
no. 12, pp. 1522–1531, Aug. 2016. vol. 64, no. 11, pp. 4563–4577, Nov. 2016.
[33] J. Tang, H. Wen, K. Zeng, L. Hu, and S. L. Chen, “Achieving uncondi- [55] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
tional security for MIMO-BAN under short blocklength wiretap code,” Cambridge Univ. Press, 2004.
in Proc. IEEE Veh. Technol. Conf. (VTC)-Fall, Toronto, ON, Canada, [56] A. J. Goldsmith and P. Varaiya, “Capacity of fading channels with
Sep. 2017, pp. 1–5. channel side information,” IEEE Trans. Inf. Theory, vol. 43, no. 6,
[34] J. Tang, H. Wen, H. H. Song, and F. Pan, “Combining MIMO beam- pp. 1986–1992, Nov. 1997.
forming with security codes to achieve unconditional communication [57] J. Tang and X. Zhang, “Quality-of-service driven power and rate
security,” in Proc. IEEE/CIC Int. Conf. Commun. China (CIC/ICCC), adaptation over wireless links,” IEEE Trans. Wireless Commun., vol. 6,
Shenzhen, China, Nov. 2015, pp. 105–109. no. 8, pp. 3058–3068, Aug. 2007.
WANG et al.: UNIFIED QoS AND SECURITY PROVISIONING FRAMEWORK FOR WIRETAP CRNs 1565

Yichen Wang (S’13–M’14) received the B.S. degree Xiao Tang received the B.S. degree in informa-
in information engineering and the Ph.D. degree in tion engineering (Elite Class named after Tsien
information and communications engineering from Hsue-Shen) and the Ph.D. degree in information
Xi’an Jiaotong University, China, in 2007 and 2013, and communication engineering from Xi’an Jiaotong
respectively. He is currently an Assistant Professor University in 2011 and 2018, respectively. From
with the Information and Communications Engineer- 2015 to 2016, he was a Visiting Student with the
ing Department, Xi’an Jiaotong University. From Department of Electrical and Computer Engineer-
2014 to 2015, he was a Visiting Scholar with the Sig- ing, University of Houston. He is currently with
nal and Information Group, Department of Electrical the Department of Communication Engineering,
and Computer Engineering, University of Maryland, Northwestern Polytechnical University. His research
College Park, MD, USA. He has published more interests include wireless communication and net-
than 80 technical papers in international journals and conferences. His research working, resource management, game theory, and physical layer security.
interests include mobile wireless communication and networks with emphasis
on cognitive radio techniques, ad hoc networks, MAC protocol design, statisti-
cal quality-of-service provisioning, resource allocation over wireless networks, Tao Wang received the B.S. degree in information
machine-type communication, device-to-device communication, and physical- engineering from Xi’an Jiaotong University, China,
layer security. in 2014, where he is currently pursuing the M.S.
Dr. Wang serves and has served as a Technical Program Committee Member degree with the Information and Communications
for many world-renowned conferences, including the IEEE GLOBECOM, Engineering Department. From 2014 to 2017, he was
ICC, and VTC. He is also a member of the IEEE Communications Society with Xi’an ThinkJoy Information Technology Co.,
and the IEEE Vehicular Technology Society. He received the Best Letter Ltd., China. His current research interests include
Award from the IEICE Communications Society in 2010 and the Exemplary cognitive radio techniques, statistical quality-of-
Reviewers Award from the IEEE C OMMUNICATIONS L ETTERS in 2014. He service provisioning, and resource allocation in wire-
is currently serving as an Editor for the KSII Transactions on Internet and less communication.
Information Systems.