This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

Beamforming and Power Splitting Designs for

AN-aided Secure Multi-user MIMO SWIPT
Systems
Zhengyu Zhu, Student Member, IEEE, Zheng Chu, Ning Wang, Member, IEEE,
Sai Huang, Zhongyong Wang, and Inkyu Lee, Fellow, IEEE

Abstract—In this paper, an energy harvesting scheme for a splitter for EH and information decoding (ID) has been studied
multi-user multiple-input-multiple-output (MIMO) secrecy chan- [9], [10].
nel with artificial noise (AN) transmission is investigated. Join- On the other hand, in the literature we see increasing
t optimization of the transmit beamforming matrix, the AN
covariance matrix, and the power splitting ratio is conducted research interest in secrecy transmission through physical layer
to minimize the transmit power under the target secrecy rate, (PHY) security designs [11]. Unlike conventional cryptograph-
the total transmit power, and the harvested energy constraints. ic methods which are normally adopted in the network layer
The original problem is shown to be non-convex, which is and rely on computational security, PHY security approaches
tackled by a two-layer decomposition approach. The inner layer are developed from the information-theoretic perspective such
problem is solved through semi-definite relaxation, and the
outer problem, on the other hand, is shown to be a single- that provable secrecy capacity can be achieved [12]–[16].
variable optimization that can be solved by one-dimensional (1- PHY security techniques have been proposed to enhance
D) line search. To reduce computational complexity, a sequential information security of multiple antenna systems by casting
parametric convex approximation (SPCA) method is proposed more interference to potential eavesdroppers. By adding artifi-
to find a near-optimal solution. The work is then extended to cial noise (AN) and projecting it onto the null space of infor-
the imperfect channel state information case with norm-bounded
channel errors. Furthermore, tightness of the relaxation for the mation user channels in transmit beamforming, the potential
proposed schemes are validated by showing that the optimal eavesdroppers would experience a higher noise floor and thus
solution of the relaxed problem is rank-one. Simulation results obtain less information about the messages transmitted to the
demonstrate that the proposed SPCA method achieves the same legitimate receivers [17], [18]. In SWIPT systems, AN injec-
performance as the scheme based on 1-D but with much lower tion can improve secrecy capacity of information transmission
complexity.
while not affecting simultaneous power transfer [19]–[28].
The AN-aided beamforming for SWIPT operation has been
investigated in various multiple-input multiple-output (MIMO)
I. I NTRODUCTION
channels [21]–[27]. More recently, robust AN-aided transmit
In recent years, the idea of energy harvesting (EH) has been beamforming with unknown eavesdroppers was studied for
introduced to power electronic devices by energy captured multiple-input single-output (MISO) cognitive radio systems
from the environment. However, harvesting from natural en- based on different channel uncertainty models [28].
ergy sources such as solar and wind depends on many factors In SWIPT systems, when information receivers (IRs) and
and thus introduces severe reliability issue. Radio frequency energy-harvesting receivers (ERs) are in the same cell, the ERs
(RF) signal can be utilized as an alternative to more reliably are normally closer to the transmitter, compared with the IRs,
deliver energy to EH devices while simultaneously transmit- because the power sensitivity level of ER is typically low. This
ting information [1]–[4]. Based on this idea, simultaneous raises a new information security issue for SWIPT systems
wireless information and power transfer (SWIPT) schemes because the ERs can potentially eavesdrop the information
have been proposed to extend the lifetime of wireless networks transmission to the IRs with relatively higher received signal
[5]–[10]. For SWIPT operation in multiple antenna systems strength [22], [25], [29]. In order to guarantee information
[7], [8], co-located receiver architecture employing a power security for the IRs, it is desirable to implement some mech-
anism to prevent the ERs from recovering the confidential
Part of this work was presented at the 2016 IEEE International Conference message from their observations.
on Communications (ICC’16), May 23-27, 2016, Kuala Lumpur, Malaysia. Motivated by the aforementioned observations, in this paper,
Z. Zhu is with the School of Information Engineering, Zhengzhou Uni-
versity, Zhengzhou, China. He was with the School of Electrical Engineer- we study secrecy transmission over a multi-user MIMO secre-
ing, Korea University, Seoul, Korea (e-mail: [email protected]). Z. cy channel which consists of one multi-antenna transmitter,
Chu is with the Faculty of Science and Technology, Middlesex University, multiple legitimate single antenna co-located receivers (CRs)
London, NW4 4BT, U.K. (Email: [email protected]) N. Wang and Z. Wang
are with the School of Information Engineering, Zhengzhou University, and multiple multi-antenna ERs. We employ an AN injection
Zhengzhou 450001, China (e-mail: {ienwang, iezywang}@zzu.edu.cn). S. scheme to mask the desired information-bearing signals for
Huang is with the School of Information and Communication Engineering, secrecy consideration without imposing any structural restric-
Beijing University of Posts and Telecommunications, Beijing, China (e-mail:
[email protected]). I. Lee is with School of Electrical Engineering, Korea tion on the AN. In comparison with existing works which do
University, Seoul, Korea (e-mail: [email protected]). not consider power splitter at the legitimate receivers [19],

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

[20], [23], [24], in this paper, each CR is assumed to adopt beamforming based PM problem with perfect CSI. Section
a power splitter to collect energy from both the information- IV extends the PM results to the imperfect CSI case. Section
bearing signal and the AN. The design objective is to jointly V illustrates the computational complexity of the proposed
optimize the transmit beamforming matrix, the AN covariance algorithms. The numerical results are shown in Section VI.
matrix, and the power splitting (PS) ratio such that the AN Finally, we conclude the paper in Section VII.
transmit power is maximized1 subject to constraints on the Notation: Vectors and matrices are denoted by bold low-
secrecy rate, the total transmit power, and the energy harvested ercase and bold uppercase letters, respectively. (·)T and (·)H
by both the CRs and the ERs. Because of the coupling effect represent matrix transpose and Hermitian transpose. The oper-
in the joint optimization problem, determination of the AN ator ⊗ represents the Kronecker product. For a vector x, ∥x∥
covariance matrix and the PS ratio makes the derivation of indicates the Euclidean norm. CM ×L and HM ×L denote the
the secrecy rate and the harvested energy at the CRs more space of M × L complex matrices and Hermitian matrices,
complicated. respectively. H+ represents the set of positive semi-definite
The formulated power minimization (PM) problem for Hermitian matrices, and R+ denotes the set of all nonnegative
AN-aided secrecy transmission is shown to be non-convex, real numbers. For a matrix A, A ≽ 0 means that A is positive
which cannot be solved directly [31]. The PM problem is semi-definite, and ∥A∥F , tr(A), |A| and rank(A) denote the
thus transformed into a two-layer optimization problem and Frobenius norm, trace, determinant, and the rank, respectively.
solved accordingly through semi-definite relaxation (SDR) and vec(A) stacks the elements of A in a column vector. 0M ×L is
one-dimensional (1D) line search. We first propose a joint a zero matrix of size M × L. E{·} is the expectation operator,
optimization design for the case with perfect channel state and ℜ{·} stands for the real part of a complex number. [x]+
information (CSI). The framework is then extended to robust represents max{x, 0} and λmax (A) denotes the maximum
designs for systems having deterministic or statistical CSI eigenvalue of A.
uncertainties. The contributions of this work are summarized
as follows: II. S YSTEM M ODEL
• For the case with perfect CSI at the transmitter, the inner In this section, we consider a multi-user MIMO SWIPT se-
loop of the PM problem is solved through SDR, while the crecy system which consists of one multi-antenna transmitter,
outer loop is shown to be a single-variable optimization L single-antenna CRs and K multi-antenna ERs, as shown in
problem, where a one-dimensional line search algorithm Fig. 1. We assume that each CR employs the PS scheme to
is employed to find the optima. To reduce computational receive the information and harvest power simultaneously. It
complexity, a sequential parametric convex approxima- is assumed that the transmitter is equipped with NT transmit
tion (SPCA) method is also investigated [15], [26], [35]– antennas, and each ER has NR receive antennas.
[39].
• For the imperfect CSI case with deterministic chan-
nel uncertainties, we consider a worst-case robust PM
(WCR-PM) problem. By exploiting the S-procedure [31],
the semi-infinite constraints are transformed into linear
matrix inequalities (LMIs) and the inner loop can be
relaxed into an SDP by employing the SDR method.
The corresponding robust optimal design is proposed.
Furthermore, an SPCA-based iterative algorithm is also
addressed with low complexity.
• For both the perfect CSI and imperfect CSI cases, the
tightness of the SDP relaxation is verified by showing
that the optimal solution is rank-one.
Fig. 1: A multi-user MIMO SWIPT secrecy system
Compared with our preliminary work [25], major additional
work and results incorporated in this paper are summarized in We denote by hc,l ∈ CNT the channel vector between the
the following. 1) This paper has extended the problem of AN transmitter and the l-th CR, and He,k ∈ CNT ×NR the channel
power maximization to both perfect and imperfect CSI cases, matrix between the transmitter and the k-th ER. The received
which introduces substantial changes in the analyses. 2) An signal at the l-th CR and the k-th ER are given by
SPCA-based iterative algorithm has been proposed to solve
the problem such that the computational complexity is largely c,l x+nc,l , ∀l,
yc,l = hH
reduced compared with the 1-D search method used in the e,k x+ne,k , ∀k,
ye,k = HH
previous work.
where x ∈ CNT is the transmitted signal vector, and nc,l ∼
The rest of this paper is organized as follows: The system
CN (0, σc,l
2
) and ne,k ∼ CN (0, σk2 I) are the additive Gaussian
model of a multi-user MIMO secrecy channel with SWIPT is
noise at the l-th CR and the k-th ER, respectively.
presented in Section II. Section III investigates the transmit
In order to achieve secure transmission, the transmitter
1 AN power maximization is equivalent to minimizing the transmit power employs transmit beamforming with AN, which acts as in-
of the information signal [21]. terference to the ERs, and provides energy to the CRs and

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

ERs. The transmit signal vector x can be written as formulated as

x = qs + w, (1) min
tr(Q)
Q, W, ρc,l
( )
ρc,l hH c,l Qhc,l
where q ∈ C NT
defines the transmit beamforming vector, s s.t. log 1+ 2 +hH Wh )+σ 2
ρc,l (σc,l
with E{s2 } = 1 is the information-bearing signal intended for  c,l c,l p,l

the CRs, and w ∈ C NT
represents the energy-carrying AN,  
− max log I + (σk I+He,k WHe,k ) He,k QHe,k  ≥ R̄c,l ,(4a)
2 H −1 H
which can also be composed by multiple energy beams. k
As the CR adopts PS to perform ID and EH simultaneously, tr(Q + W) ≤ P, (4b)
the received signal at the l-th CR is divided into ID and EH Ēc,l
hHc,l (Q + W)hc,l + σc,l ≥
2
components by the PS ratio ρc,l ∈ (0, 1]. Therefore, the signal ηc,l (1 − ρc,l )
, (4c)
for information detection at the l-th CR is given by
( ) Ēe,k
min tr HH e,k (Q+W)He,k +NR σk ≥ , ∀k, (4d)
2
√ √ ηe,k
c,l x+nc,l )+np,l , ∀l,
ID
yc,l = ρc,l yc,l +np,l = ρc,l (hH k
Q ≽ 0, W ≽ 0, 0 < ρc,l ≤ 1, ∀l, rank(Q) = 1, (4e)
where np,l ∼ CN (0, σp,l ) is the additive Gaussian noise at the
2
where R̄c,l is the target secrecy rate, P is the total transmit
l-th CR.
power, and Ēc,l and Ēe,k denote the predefined harvested
Denoting Q = E{qqH } as the transmit covariance matrix
power at the l-th CR and the k-th ER, respectively. The
and W = E{wwH } as the AN covariance matrix, the
constraint (4d) guarantees that a minimum energy harvested
achieved secrecy rate at the l-th CR is given by
power should be achieved by the k-th ER.
[ ( H )
ρc,l hc,l Qhc,l
R̂c,l = log 1+ B. One-Dimensional Line Search Method (1-D Search)
ρc,l (σc,l +hH
2 2
c,l Whc,l )+σp,l
 ]+
  Problem (4) is non-convex due to the secrecy rate constraint
− max log I+(HH e,k WHe,k +σk I)
2 −1 H
He,k QHe,k  , ∀l. (4a), and thus cannot be solved directly. In order to circumvent
k
(2) this issue, we convert the original problem by introducing a
slack variable t for the k-th ER’s rate. Then we have
The harvested power at the l-th CR and the k-th ER is therefore
min tr(Q)
( H ) Q, W, ρc,l , t
Ec,l = ηc,l (1−ρc,l ) hc,l (Q+W)hc,l +σc,l , ∀l,
2 ( )
( ) tρc,l hHc,l Qhc,l
( ) (3) s.t. log t+ 2 +hH Wh )+σ 2
≥ R̄c,l , ∀l, (5a)
Ee,k = ηe,k tr HH e,k (Q+W)He,k +NR σk , ∀k,
2 ρc,l (σc,l c,l
 c,l p,l

  1
I+(σk2 I+HH e,k WHe,k )
−1 H
He,k QHe,k  ≤ , ∀k, (5b)
where ηc,l and ηe,k represent the EH efficiency of the l-  t
th CR and the EH efficiency of the k-th ER, respectively. (4b) − (4e).
In this paper we set ηc,l = ηe,k = 0.3 for simplicity. The
results can be easily extended to scenarios with different ηc,l Problem (5) is still non-convex in constraints (5a) and (5b),
and ηe,k values. In the following section, we consider the which can be addressed by reformulating (5) into a two-layer
transmit beamforming based PM problem to jointly optimize problem. For the inner layer, we solve problem (5) for a given
the transmit covariance matrix Q, the AN covariance matrix t, which is relaxed as
W, and the PS ratio ρc,l . f (t) = min tr(Q)
Q, W, ρc,l , t

( ) ( 2 σp,l 2
)
III. M ASKED B EAMFORMING BASED P OWER s.t. hH
c,l tQ−(2 R̄c,l
−t)W hc,l ≥ (2 R̄c,l
−t) σc,l + , (6a)
ρc,l
M INIMIZATION WITH P ERFECT CSI
( 1t −1)(σk2 I+HH e,k WHe,k ) ≽ He,k QHe,k , ∀k,
H
(6b)
In this section, we study transmit beamforming optimization (4b) − (4e),
under the assumption that perfect CSI of all the channels is
available at the transmitter. where f (t) is defined as the optimal value of problem (6),
which is a function of t. Even though the function f (t) cannot
be expressed in closed-form, numerical evaluation of f (t) is
feasible.
A. Problem Formulation
Remark 1: It is noted that the LMI constraint (6b) is ob-
In this problem, the transmit power of the information signal tained from [18, Proposition 1], and is based on the assumption
is minimized subject to the total transmit power constraint, that rank(Q) ≤ 1, which will be shown later.
the secrecy rate constraint, and the EH constraints of the CRs By ignoring the non-convex constraint rank(Q) = 1, prob-
and the ERs such that the AN transmit power is maximized lem (6) becomes convex and thus can be solved efficiently by
for secrecy consideration. The AN-aided PM problem is thus an interior-point method for any given t [31]. The outer layer

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

problem, whose objective is to find the optimal value of t, is can be rewritten as

then formulated as
log(r1 r2 ) ≥ R̄c,l , ∀l, (9a)
min f (t) ρc,l hH
t (7) c,l Qhc,l
1+ ≥ r1 , ∀l, (9b)
s.t. tmin ≤ t ≤ tmax , 2 H 2
ρc,l (σc,l +hc,l Whc,l )+σp,l
where tmax and tmin are the upper and lower bounds of t, tr(HH
e,k QHe,k ) 1
respectively. The solution to problem (7) can be found by 1+ ≤ , ∀k, (9c)
σk2 + tr(HH
e,k WHe,k ) r2
one-dimensional line search. For the line search algorithm, we
need to determine the lower and upper bounds of the searching which can be further simplified as
interval for t. It is straightforward that tmax = 1 can be used as r1 r2 ≥ 2R̄c,l , ∀l, (10a)
the upper bound due to the feasibility of (5b), while a lower hH
c,l Qhc,l
bound is calculated as
σ2
≥ r1 − 1, ∀l, (10b)
( )−1 2 +hH Wh + p,l
σc,l
ρc,l hHc,l Qhc,l
c,l c,l ρc,l
t ≥ min 1 + σk2 + tr(HH
l 2 + hH Wh ) + σ 2
ρc,l (σc,l e,k WHe,k )
(
c,l c,l p,l
)−1 ≥ r2 , ∀k. (10c)
H
hc,l Qhc,l σk2 + tr(HH
e,k (Q + W)He,k )
≥ min 1 + 2 2 + hH Wh
(8)
l σc,l + σp,l c,l c,l The inequality constraint (10a) is equivalent to 2R̄c,l +2 +(r1 −
( )−1 r2 )2 ≤ (r1 + r2 )2 , which can be converted into a conic
P ∥hc,l ∥2
≥ min 1 + 2 2 , tmin . quadratic-representable function form as
l σc,l + σp,l [√ ]

where the first inequality is based on the secrecy rate R̄c,l ≥ 0, 2R̄c,l +2 r1 − r2 ≤ r1 + r2 , ∀l. (11)
the third inequality follows from (4b). In the following theo- By transforming inequality constraints (10b) and (10c) into
rem, we prove the equivalence of problem (7) and the original
problem (4). 2
Theorem 1: The transmit beamforming PM problem (4) is σp,l qH Hc,l q
2
σc,l +wH Hc,l w+ ≤ , ∀l, (12a)
equivalent to problem (7) when rank(Q) ≤ 1. ρc,l r1 − 1
Proof: Let us denote the optimal solutions of (4) and (7) as σk2 + wH Ĥe,k w
f ∗ and f opt , respectively. First, we show that f opt is a feasible σk2 + wH Ĥe,k w + qH Ĥe,k q ≤ , ∀k, (12b)
r2
point of problem (7), i.e. f opt ≤ f ∗ . It is noted that (4) and
(6) have the same objective function and the optimal solution where Hc,l = hc,l hH H
c,l and Ĥe,k = He,k He,k , we observe
of (4) satisfies the constraints of (6) given the assumption that that these two constraints are non-convex, but the right-hand
rank(Q) ≤ 1 [18], which gives rise to f (t∗ ) = f ∗ , where t∗ side (RHS) of both (12a) and (12b) have the function form of
is the optimal value of t. In addition, it follows f opt ≤ f (t∗ ). quadratic-over-linear, which are convex functions [31]. Based
Next, we prove that the solution to problem (6) is achievable on the idea of the constrained convex procedure [40], these
in problem (4), i.e. f ∗ ≤ f opt . From (4a), (6a) and (6b), quadratic-over-linear functions can be replaced by their first-
we can show that the optimal solutions of (6) are feasible order expansions, which transforms the problem into convex
solutions of (4a) when rank(Q) ≤ 1. Therefore, we conclude programming. Specifically, we define
that f ∗ = f opt .  wH Aw
Utilizing the results in Remark 1 and Theorem 1, next we fA,a (w, t) = , (13)
t−a
show the tightness of the AN-aided PM problem (4) by the
following theorem. where A ≽ 0 and t ≥ a. At a certain point (w̃, t̃), the first-
Theorem 2: Provided that problem (6) is feasible for a given order Taylor expansion of (13) is given by
t > 0, there exists an optimal solution to (4) such that the rank 2ℜ{w̃H Aw} w̃H Aw̃
of Q is always equal to 1. FA,a (w, t, w̃, t̃) = − (t − a). (14)
t̃ − a (t̃ − a)2
Proof: See Appendix A. 
Problem (7) can be solved by conducting one-dimensional By using the above results of Taylor expansion, for the
line search for f (t) over t and choosing the minimum f (t) points (q̃, r̃1 ) and (w̃, r̃2 ), we can transform constraints (12a)
as the optimal solution. Solving the SDP problem (6) with and (12b) into convex forms, respectively, as
the optimal f (t), we can obtain the optimal design variables 2
σp,l
(Q∗ , W∗ , ρ∗c,l ). The optimal beamforming vector q∗ is then
2
σc,l + wH Hc,l w + ≤ FHc,l ,1 (q, r1 , q̃, r̃1 ), ∀l, (15a)
ρc,l
computed by eigenvalue decomposition Q∗ = q∗ q∗H . 2 r2
σk2 + wH Ĥe,k w + qH Ĥe,k q ≤ σk2 ( − 2 )
r̃2 r̃2
C. Low-Complexity SPCA Algorithm + FĤe,k ,0 (w, r2 , w̃, r̃2 ), ∀k. (15b)
In this subsection, we propose an SPCA based iterative σ2
method to reduce the computational complexity. By introduc- Denoting gr1 ,l = FHc,l ,1 (q, r1 , q̃, r̃1 ) − σc,l2
− ρc,l
p,l
and gr2 ,k =
ing two slack variables r1 > 0 and r2 > 0, the constraint (4a) σk ( r̃2 −r̃2 )+FĤe,k ,0 (w, r2 , w̃, r̃2 )−σk , (15a) and (15b) can be
2 2 r2 2
2

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

recast as the following second-order cone (SOC) constraints IV. M ASKED B EAMFOMRING BASED ROBUST PM FOR
I MPERFECT CSI
[2wH hc,l , gr1 ,l − 1]T ≤ gr1 ,l + 1, ∀l, (16a)
Due to channel estimation and quantization errors, it may
[2w He,k ; 2q He,k ; gr2 ,k − 1] ≤ gr2 ,k + 1, ∀k. (16b)
H H T
not be possible to have perfect CSI in practice. In this section,
we extend the PM optimization method to more practical
Next we employ the SPCA technique for the SOC constraints scenarios with imperfect CSI. Specifically, we consider an
(4c) and (4d) [42] to obtain convex approximations. By AN-aided WCR-PM formulation under norm-bounded channel
substituting q , q̃ + ∆q and w , w̃ + ∆w into the left- uncertainty.
hand side (LHS) of (4c), we obtain

qH Hc,l q + wH Hc,l w + σc,l

2 A. Worst-Case Based Robust PM Problem
= (q̃+∆q)H Hc,l (q̃+∆q)+(w̃+∆w)H Hc,l (w̃+∆w) + σc,l
2 Now, we adopt imperfect CSI based on the deterministic
model [30]. Specifically, the actual channel between the trans-
≥ q̃H Hc,l q̃ + 2ℜ{q̃H Hc,l ∆q} + w̃H Hc,l w̃ mitter and the l-th CR, denoted by hc,l , and the actual channel
+ 2ℜ{w̃H Hc,l ∆w} + σc,l
2
, between the transmitter and the k-th ER, denoted by He,k , can
(17) be modeled as
hc,l = h̄c,l + ec,l , ∀l,
where the inequality is given by dropping the quadratic terms (23)
∆qH Hc,l ∆q and ∆wH Hc,l ∆w. Similarly, in the LHS of He,k = H̄e,k + Ee,k , ∀k,
(4d), we have where h̄c,l and H̄e,k denote the estimated channel available at
( ) the transmitter, and ec,l and Ee,k are the bounded CSI errors
tr HH
e,k (Q+W)He,k
with ∥ec,l ∥ ≤ εc,l and ∥Ee,k ∥F ≤ εe,k , respectively.
=qH Ĥe,k q + wH Ĥe,k w Ēc,l
We define Êc,l , ηc,l
Ēe,k
and Êe,k , ηe,k . By taking the CSI
=(q̃+∆q)H Ĥe,k (q̃+∆q)+(w̃+∆w)H Ĥe,k (w̃+∆w) model (23) into account, the AN-aided WCR-PM problem can
≥q̃H Ĥe,k q̃+2ℜ{q̃H Ĥe,k ∆q} + w̃H Ĥe,k w̃ be rewritten as
+ 2ℜ{w̃H Ĥe,k ∆w}. min tr(Q)
Q, W, ρc,l
(18) ( )
ρc,l (h̄c,l +ec,l )H Q(h̄c,l +ec,l )
s.t. log 1+ 2 +(h̄ +e )H W(h̄ +e ))+σ 2
According to (17) and (18), we obtain linear approximations ρc,l (σc,l c,l c,l c,l c,l
  p,l
of the concave constraints (4c) and (4d) as  
− max log I+ H̄E (H̄e,k +Ee,k ) Q(H̄e,k +Ee,k ) ≥ R̄c,l ,(24a)
H
k
q̃H Hc,l q̃+2ℜ{q̃H Hc,l ∆q}+ w̃H Hc,l w̃
(19) tr(Q+W) ≤ P, (24b)
Ēc,l
+ 2ℜ{w̃H Hc,l ∆w} + σc,l2
≥ , ∀l, Êc,l
ηc,l (1 − ρc,l ) (h̄c,l +ec,l )H (Q+W)(h̄c,l +ec,l )+σc,l
2
≥ , (24c)
1−ρc,l
and ( )
tr (H̄e,k +Ee,k )H (Q+W)(H̄e,k +Ee,k )
q̃H Ĥe,k q̃+2ℜ{q̃H Ĥe,k ∆q}+ w̃H Ĥe,k w̃ ≥ Êe,k −NR σk2 , (24d)
Ēe,k (20) Q ≽ 0, W ≽ 0, 0 < ρc,l ≤ 1, rank(Q) = 1, (24e)
+ 2ℜ{w̃H Ĥe,k ∆w}+NR σk2 ≥ , ∀k.
ηe,k ( )−1
where H̄E = σk2 I+(H̄e,k +Ee,k )H W(H̄e,k +Ee,k ) .
Finally, by rearranging (4b) as
√
∥[qT wT ]∥ ≤ P, (21) B. One-Dimensional Line Search Method
The above robust PM problem is not convex in terms of
the original problem (4) is transformed into
the channel uncertainties. We therefore consider relaxation for
min ∥q∥ constraint (24a) by introducing a slack variable t1 similar to
q, w, ρc,l , r1 , r2 , gr1 ,l , gr2 ,k
(22) the previous section. The constraint (24a) is then transformed
s.t. (11), (16), (19), (20), (21), 0 < ρc,l ≤ 1, ∀l. into
( 2 σp,l 2
)
Given q̃, w̃, r̃1 , and r̃2 , problem (22) is convex and can be (h̄c,l +ec,l )H T̃l (h̄c,l +ec,l ) ≥ (2R̄c,l −t1 ) σc,l + , (25a)
ρc,l
efficiently solved by convex optimization software tools such ( )
as CVX [34]. Based on the SPCA method, an approximation ( t11 −1) σk2 I+(He,k +Ee,k )H W(He,k +Ee,k )
with the current optimal solution can be updated iteratively, ( )H ( )
≽ He,k +Ee,k Q He,k +Ee,k , ∀k, (25b)
which implies that (4) is optimally solved. In Section VI, we
will show that the proposed SPCA method achieves the same where T̃l = t1 Q − (2R̄c,l − t1 )W. The constraint (25b)
performance as the 1-D search scheme, but with much lower is obtained under the assumption that rank(Q) ≤ 1 [18,
complexity. Proposition 1]. Problem (24) has semi-infinite constraint (24c),

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

(24d), (25a), and (25b). In order to make the problem tractable, as

we exploit the S-procedure [31] to transform the constraints min tr(Q)
(24c), (24d), (25a), and (25b) into LMIs. For completeness, Q, W, t1 , al , bl

the S-procedure is presented in Lemma 1 in the following. 1 1

s.t. tr(Q + W) ≤ P, + ≤ 1, rank(Q) = 1,
Lemma 1: (S-Procedure [31, Appendix B.2]) Let a function al bl
(31)
f m (x) with x ∈ CN ×1 (m = 1, 2) be defined as Q ≽ 0, W ≽ 0, λc,l ≥ 0, λe,k ≥ 0,
{ } αc,l ≥ 0, αe,k ≥ 0, al ≥ 1, bl ≥ 1, ∀l, ∀k,
f m (x) = xH Am x + 2Re bH m x + cm (26)
(28), (29a), (29b), (30).
where Am ∈ HN ×N , bm ∈ CN ×1 and cm ∈ RN ×1 . Then, By removing the nonconvex constraint rank(Q) = 1, the above
f m (x) ≤ 0 holds if and only if there exists θ ≥ 0 such that problem (31) becomes convex and can be solved by applying
[ ] [ ] a solver in [34] given t1 . Tightness of the relaxation of (24)
A1 b1 A2 b2 is shown by the following theorem.
θ − ≽ 0,
bH
1 c1 bH2 c2 Theorem 3: Provided that the robust problem (24) is feasible
for a given t1 , there always exists an optimal solution Q with
provided that there is a point x̂ which satisfies fm (x̂) < 0.  rank(Q) = 1.
To employ the S-procedure, we rewrite the first constraint Proof: See Appendix B. 
in (25) as So far, we have tackled the WCR-PM problem (31) by
2
σp,l deriving a tight rank relation in Theorem 3. Note that problem
eH T̃ e
c,l l c,l +2R{ h̄H
T̃ e
c,l l c,l }+h̄ H
T̃ h̄
c,l l c,l ≥ (2 R̄c,l
−t1 )(σ 2
c,l+ ). (24) can also be solved by applying one-dimensional line
ρc,l search over t as in Section III.
(27) 1
1
In addition, we introduce al = ρc,l and bl = 1−ρ1 c,l to convert
the non-convex constraints into convex ones. According to C. Low-Complexity SPCA Algorithm
Lemma 1, by using a slack variable λc , (27) can be expressed Now, let us consider another reformulation of the WCR-PM
as problem (24) based on the SPCA algorithm. The optimization
[ ] framework can also be recast as a convex form by incorporat-
λc,l I+ T̃l T̃l h̄c,l
≽ ing channel uncertainties. First, the robust secrecy rate (24a)
0.
h̄H c,l T̃l h̄c,l −(2
h̄H −t1 )(σc,l
R̄c,l 2 2
c,l T̃l +al σp,l )−λc,l ε2c,l can be rewritten as
(28)
Let us define h̄e,k , vec(He,k ). Using Lemma 1 again and log(r3 r4 ) ≥ R̄c,l , ∀l, (32a)
the property vec(M1 M2 M3 ) = (M3 ⊗ M1 )vec(M2 ) [33],
T H
(h̄c,l +ec,l ) Q(h̄c,l +ec,l )
constraints (24c) and (24d) become 1+ 2 ≥ r3 , (32b)
σ 2 +(h̄ +e )H W(h̄ +e )+ σp,l
[ ] c,l
(
c,l c,l c,l c,l ρc,l
)
αc,l I+QW QW h̄c,l H
≽ 0,(29a) tr (H̄
(
e,k + E e,k ) Q( H̄ e,k + E e,k ) 1
) ≤ ,(32c)
h̄c,l QW h̄c,l QW h̄c,l +σc,l −bl Êc,l −αc,l εc,l
H H 2 2 1+ 2
[ ( ) ( ) ] σ k + tr (H̄ e,k + E e,k )H W(H̄
e,k + E e,k ) r 4
αe,k I+ ( I ⊗ QW ) I ⊗ QW h̄e,k
≽ 0, ∀k, (29b) where r3 > 0 and r4 > 0 are slack variables. The inequalities
e,k I ⊗ QW
h̄H θe,l
in (32) can be rearranged, which gives
where αc,l and (αe,k are slack)variables, and QW = Q+W,
r3 r4 ≥ 2R̄c,l , ∀l, (33a)
and θe,l = h̄H e,k I ⊗ (Q+W) h̄e,k − Ê e,k +N R σ 2
k −α e,k ε 2
e,k . 2 H
σ p,l q (Hc,l +∆c,l )q
To transform the constraint (25b) into a tractable convex LMI, 2
σc,l +wH (Hc,l +∆c,l )w + ≤ , (33b)
we exploit the following lemma. ρ c,l r3 − 1
Lemma 2 [18]: For F1 , F2 , F3 ∈ CM ×M , we denote σk2 + wH (Ĥe,k +∆e,k )w + qH (Ĥe,k +∆e,k )q
} 2 Z+F3 , satisfying g(Z) ≽ 0 for
H H H
g(Z) = { Z F1 Z+Z F2 +F σ 2 + wH (Ĥe,k +∆e,k )w
∀Z ∈ Z|tr(ZH F4 Z) ≤ 1 with F4 ≽ 0. Then, the following ≤ k , (33c)
r4
LMI holds:
[ H
] [ ] where ∆c,l = h̄c,l eH H H
c,l + ec,l h̄c,l + ec,l ec,l and ∆e,k =
F3 F2 I 0 H̄e,k EH H H
−α ≽ 0, e,k +Ee,k H̄e,k +Ee,k Ee,k stand for the CSI uncertainty.
F2 F1 0 −F4 It is straightforward to show that
where α ≥ 0.  ∥∆c,l ∥F ≤ ∥h̄c,l eH
c,l ∥F +∥ec,l h̄c,l ∥F +∥ec,l ec,l ∥F
H H

By Lemma 2, the constraint (25b) can be equivalently given ≤ ∥h̄c,l ∥∥eH

c,l ∥ + ∥ec,l ∥∥h̄c,l ∥+∥ec,l ∥
H 2
(34)
as
[( 1 ) ] = ε2c,l + 2εc,l ∥h̄c,l ∥,
( t1 −1)σk2 −λe,k I+ H̄H
e,k WQ H̄e,k H̄H
e,k WQ
λ ≽ 0, (30) ∥∆e,k ∥F ≤ ∥H̄e,k EH
e,k ∥F +∥Ee,k H̄e,k ∥F +∥Ee,k Ee,k ∥F
H H
WQ H̄e,k WQ + ε2e,k I
e,k
≤ ∥H̄e,k ∥F ∥EH
e,k ∥F + ∥Ee,k ∥F ∥H̄e,k ∥F +∥Ee,k ∥F
H 2
( )
where λe,k is a slack variable and WQ = ( t11 −1)W−Q . = ε2e,k + 2εe,k ∥H̄e,k ∥F .
According to (25)-(30), the WCR-PM problem is now given (35)

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

Note that ∆c,l and ∆e,k are norm-bounded matrices as Substituting q = q̃ + ∆q and w = w̃ + ∆w into the LHS
∥∆c,l ∥F ≤ ξc,l and ∥∆e,k ∥F ≤ ξe,k where ξc,l = ε2c,l + of (45a) and (45b), the SPCA technique can be applied to
2εc,l ∥h̄c,l ∥ and ξe,k = ε2e,k + 2εe,k ∥H̄e,k ∥F . Similarly, we approximate (45a) and (45b), respectively, as
equivalently recast (33a) as
[√ ] q̃H Ȟc,l q̃+2ℜ{q̃H Ȟc,l ∆q}+ w̃H Ȟc,l w̃

2R̄c,l +2 r3 − r4 ≤ r3 + r4 , ∀l. (36)
Êc,l
+2ℜ{w̃H Ȟc,l ∆w} ≥ − σc,l
2
, (46a)
According to [9], we can minimize constraint (32b) by maxi- 1−ρc,l
mizing the LHS of (33b) while minimizing its the RHS. Then q̃H Ȟe,k q̃+2ℜ{q̃H Ȟe,k ∆q}+ w̃H Ȟe,k w̃
the constraints (33b) and (33c) can be approximately rewritten
+2ℜ{w̃H Ȟe,k ∆w} ≥ Êe,k −NR σk2 . (46b)
as, respectively,
2
2
σp,l Eventually, the WCR-PM problem is converted into the
max σc,l +wH Ac,l w +
∥∆c,l ∥F ≤ξc,l ρc,l following convex form as
(37)
qH Ac,l q
≤ min , min ∥q∥
∥∆c,l ∥F ≤ξc,l r3 − 1 q, w, ρc,l , r3 , r4
(47)
max σk2 H
+ w Be,k w + q Be,k q H s.t. (36), (44a), (44b), (46a), (46b), (21), 0 < ρc,l ≤ 1.
∥∆e,k ∥F ≤ξe,k

σk2 + wH Be,k w (38) Given q̃, w̃, r̃3 , and r̃4 , problem (47) is convex and can
≤ min , be solved by employing an interior-point method to update
∥∆e,k ∥F ≤ξe,k r4
iteratively until convergence.
where Ac,l = Hc,l +∆c,l and Be,k = Ĥe,k +∆e,k .
In order to minimize the RHS of (37) and (38), a loose
approximation [41] is applied, which gives V. C OMPUTATIONAL C OMPLEXITY
H H
q Ac,l q q Ȟc,l q
min ≥ , In this section, we evaluate the computational complexity
∥∆c,l ∥F ≤ξc,l r3 − 1 r3 − 1 of the proposed robust methods. As will be shown in Section
(39)
σk2 + wH Be,k w σ 2 + wH Ȟe,k w VI, the proposed SPCA algorithm achieves substantial im-
min ≥ k ,
∥∆e,k ∥F ≤ξe,k r4 r4 provement in complexity for the same performance compared
with the method based on 1-D search. Now we compare
where Ȟc,l = Hc,l−ξc,l INT and Ȟe,k = Ĥe,k−ξe,k INT . Using
complexity of the algorithms through analyses similar to that
similar technique to the LHS of (37) and (38) yields
in [32] and [42]. The complexity of the proposed algorithms
max wH Ac,l w ≤ wH H̄c,l w, (40) are shown in Table I on the top of next page. We denote
∥∆c,l ∥F ≤ξc,l
n, D = log2 tmax −t η
min
, and Q as the number of decision
max σk2 + wH Be,k w + qH Be,k q variables, the 1-D search size, and the SPCA iteration number,
∥∆e,k ∥F ≤ξe,k respectively. The complexity analysis is given in the following.
(41)
≤ σk2 + wH H̄e,k w + qH H̄e,k q, 1) PM with 1-D Search in problem (6) involves K LMI
constraints of size NR + 1, two LMI constraints of size NT ,
where H̄c,l = Hc,l +ξc,l INT and H̄e,k = Ĥe,k +ξe,k INT .
and 4L + K + 1 linear constraints.
From (37)-(41), (33b) and (33c) can be given as
2) PM with SPCA in problem (22) has L SOC constraints
2
σp,l qH Ȟc,l q of dimension 2, L SOC constraints of dimension NT + 1, K
2
σc,l + wH H̄c,l w + ≤ , (42)
ρc,l r3 − 1 SOC constraints of dimension 2NT + 1, one SOC constraints
of dimension 2NT , and L + 3K linear constraints.
σk2 + wH Ȟe,k w
σk2 + wH H̄e,k w + qH H̄e,k q ≤ . (43) 3) WCR-PM with 1-D Search in problem (31) consists of
r4 2L LMI constraints of size NT +1, K LMI constraints of size
Exploiting the same method in (13)-(14), we obtain NR NT + 1, two LMI constraints of size NT , and 5L + 2K + 1
2
σp,l linear constraints.
2
σc,l + wH H̄c,l w + ≤ FȞc,l ,1 (q, r3 , q̃, r̃3 ), (44a) 4) WCR-PM with SPCA in problem (47) contains L SOC
ρc,l
constraints of dimension 2, L SOC constraints of dimension
σk2 + wH H̄e,k w + qH H̄e,k q NT + 1, K SOC constraints of dimension 2NT + 1, one SOC
2 r4 constraints of dimension 2NT , and L + 3K linear constraints.
≤ σk2 ( − 2 ) + FȞe,k ,0 (w, r4 , w̃, r̃4 ). (44b)
r̃4 r̃4 For example, for a system with L = 2, K = 3, NT =
By using a loose approximation approach for constraints 4, NR = 2, D = 15, and Q = 8, the complexity of the PM
(24c) and (24d), we have with 1-D search, the PM with SPCA, the WCR-PM with 1-
D search, and the WCR-PM with SPCA, are O(1.04 × 107 ),
Êc,l O(3.70×105 ), O(1.17×108 ), and O(1.45×105 ), respectively.
qH Ȟc,l q + wH Ȟc,l w ≥ − σc,l
2
, (45a)
1−ρc,l Thus, the complexity of the proposed SPCA method is only
wH Ȟe,k w + qH Ȟe,k q ≥ Êe,k −NR σk2 . (45b) 1% compared to the scheme based on 1-D search.

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

TABLE I: Complexity analysis of the proposed algorithms

Algorithms ( √ { Complexity Order })
PM with O nD KNR +2K+2NT +4L+1 K(NR +1)3 +2NT3 +n[K(NR +1)2 +2NT2 +4L+K+1]+n2
1-D Search where n = O(2NT +L)
2

PM with ( √ { })
O nQ 5K+5L+2 (2K+L+2)NT +3L+K+n(3K+L)+n2 where n = O(2NT +2L+K+2)
SPCA
WCR-PM ( √ {
O nD (KNR +2L+2)NT +7L+3K+1 }) 2L(NT +1)3 +K(NR NT +1)3 +2NT3 +n[2L(NT +1)2
with 1-D
Search +K(NR NT +1)2 +2NT2 +5L+2K+1]+n2 where n = O(2NT2 +4L+K)
WCR-PM ( √ { })
O nQ 5K+5L+2 (2K+L+2)NT +3L+K+n(3K+L)+n2 where n = O(2NT +L+2)
with SPCA

VI. N UMERICAL R ESULTS cases within just 8 iterations. In addition, the Golden Section
Algorithm is used as the 1-D search method in the simulations.
In this section, we present numerical results to validate
In Fig.3, we can see that the 1-D method can converge within
performance of the proposed transmit beamforming schemes.
15 iterations for in all cases.
In the simulations, we consider a system where the transmitter It is easily seen from the plots that the convergence of all
is equipped with NT = 4 transmit antennas, two CRs are only cases can be quickly achieved within 14 iterations, and the
equipped with single antenna, and three ERs have NR = 2 convergence speed becomes fast as E increases. It is observed
receive antennas. Both large-scale and small-scale fading are that the PCU schemes converges slower than the FCU schemes
considered in the channel model.( The )−αsimplified large-scale for the SOCP-SCA regardless of E. This is due to the fact that
fading model is given by DL = dd0 , where d represents
the distance between the transmitter and the receiver, d0 is
a reference distance equal to 10 m in this work, and α = 3
45
is the path loss exponent [43]. We define dc = 40 m as the Average harvested power of information signal (dBm) WCR−PM with SPCA (E = 30 dBm)
distance between the transmitter and the CRs, and de = 20 40
PM with SPCA (E = 30 dBm)
WCR−PM with SPCA (E = 20 dBm)
m as the distance between the transmitter and the ERs, unless PM with SPCA (E = 20 dBm)
35 WCR−PM with SPCA (E = 10 dBm)
otherwise specified. PM with SPCA (E = 10 dBm)

Because all the receivers are are expected to harvest en- 30

ergy from the RF signal, we consider line-of-sight (LOS) 25

communication scenario where the Rician fading model is
20
adopted for small scale fading
√ coefficients. The
√channel vector
KR LOS 1 N LOS
hc,l is expressed as hc,l = 1+KR hc,l + 1+K R
hc,l , 15

LOS
where hc,l indicates the LOS deterministic component with 10

c,l ∥2 = DL , hc,l
∥hLOS 2 N LOS
represents the Rayleigh fading 5
component as hc,lN LOS
∼ CN (0, DL I), and KR = 3 is the
0
Rician factor. It is noted that for the LOS component, we 1 2 3 4 5 6 7 8 9 10
Iteration Number
use the far-field uniform linear antenna array model [44].
In addition, the noise power at the l-th CR is set to be Fig. 2: Average transmit power of information signal versus
2
σc,l = −60 dBm for information transfer and σp,l
2
= −50 dBm iteration numbers for the SPCA method
for power transfer. The noise power at all the ERs is σk2 =
−50 dBm, ∀k. The channel error bound for the deterministic Fig. 4 illustrates the average transmit power of the infor-
model is set to εc,l = εe,k = ε, ∀l, k. Consequently, the mation signal in terms of different target secrecy rates with
channel error covariance matrices are given as Nc,l = ε2 INT P = 30 dBm and Ēc,l = Ēe,k = 10 dBm, ∀k. It is
and Dk = ε2 INT NR . The EH efficiency coefficients are set to observed that the transmit power increases with the secrecy
ηc,l = ηe,k = 0.3. rate target. In addition, the SPCA algorithm achieves the same
For the perfect CSI case, we compare the PM with SPCA performance as the 1-D search method, but with much lower
algorithm and the PM with 1-D search method. For the case complexity. Compared with the no-AN PM and null-AN PM,
with imperfect CSI, we show the performance of the WCR-PM the power consumption of the proposed AN-aided scheme is 9
with SPCA algorithm, the WCR-PM with 1-D search method, dB and 6 dB lower. Moreover, we can check that the proposed
the no-AN PM with W = 0, the non-robust method which scheme performs better than the scheme with ρc = ρc,l = 0.5,
computes a solution without considering channel uncertainties, and the performance gap becomes larger as the target secrecy
and the null-AN PM which means AN lie in the null space of rate increases. This indicates that optimizing the PS ratio ρc,l
the CRs. is important, especially when the target secrecy rate is high.
Fig. 2 and Fig. 3 illustrates the convergence of the SPCA In Fig. 5, we compare the average transmit power with
method and the 1-D method with respect to iteration numbers respect to different numbers of ER antennas by fixing NT = 8,
for P = 50 dBm, Ēc,l = Ēe,k = E, R = 1 bps/Hz, P = 40 dBm, Ēc,l = Ēe,k = 10 dBm, and R̄c,l = 1 bps/Hz.
and ε = 0.01, respectively. It is easily seen from the Fig.2 In this figure, one can observe that the performance of the
that convergence of the SPCA method is achieved for all 1-D search method and that of the proposed SPCA algorithm

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

remains the same regardless of the value of NR . This is due

50
WCR−PM with 1−D (E = 30 dBm)
to the fact that all the harvested power at the ERs can be
PM with 1−D (E = 30 dBm) provided by the AN signal.
Average harvested power of Information signal (dBm)

45
WCR−PM with 1−D (E = 20 dBm)
PM with 1−D (E = 20 dBm)
40 WCR−PM with 1−D (E = 10 dBm)
PM with 1−D (E = 10 dBm)
35 25

30

Average transmit power of information signal (dBm)

20

25
15

20
10
15
5
10
0
5
no−AN PM
−5
0 null−AN PM
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 PM (fixed ρc = 0.5)
Iteration Number −10
WCR−PM with 1−D
WCR−PM with SPCA
−15
Fig. 3: Average transmit power of information signal versus PM with 1−D
PM with SPCA
iteration numbers for the 1-D method −20
−10 −5 0 5 10 15 20
Target harvested power at the CR (dBm)

20 Fig. 6: Average transmit power of information signal versus the

target harvested power at the CR
18
Average transmit power of information signal (dBm)

16
In Fig. 6, we plot the average transmit power in terms of
14 different target harvested power at the CR with P = 40 dBm,
12 Ēe,k = 10 dBm and R̄c,l = 0.5 bps/Hz. We can check that
the curves of the PM and the WCR-PM schemes increase
10
with the same slope. Moveover, when the harvested power
8
no−AN PM
target decreases, the performance gap between the no-AN PM
6 PM (fixed ρc = 0.5) scheme and the proposed PM scheme becomes wider. This
null−AN PM indicates that AN is essential in achieving the performance
4 WCR−PM with 1−D
WCR−PM with SPCA gains. Furthermore, the PM scheme and the WCR-PM scheme
2 PM with 1−D
PM with SPCA
require 6 dB and 4.5 dB lower power than the PM scheme with
0 fixed ρc , respectively.
0.5 1 1.5 2 2.5 3
Target secrecy rate (bps/Hz)

Fig. 4: Average transmit power of information signal versus target 12

no−AN PM
Average transmit power of information signal (dBm)

secrecy rate null−AN PM

PM (fixed ρc = 0.5)
10
WCR−PM with 1−D
WCR−PM with SPCA
PM with 1−D
22 8
PM with SPCA
no−AN PM
20 null−AN PM
Average transmit power of information signal (dBm)

PM (fixed ρc = 0.5) 6
18
WCR−PM with 1−D
16 WCR−PMwith SPCA
4
PM with 1−D
14 PM with SPCA

12 2

10
0
8 1 1.5 2 2.5 3 3.5 4 4.5 5
Number of ERs
6
Fig. 7: Average transmit power of information signal versus the
4
number of ERs
2

0 Fig. 7 evaluates the average transmit power of information

2 3 4 5 6
Number of receive antennas, NR signal with respect to different number of ER with NT = 6,
P = 30 dBm, Ēc,l = Ēe,k = 5 dBm, and R̄c,l = 1 bps/Hz. It
Fig. 5: Average transmit power of information signal versus the is observed that both the proposed SPCA algorithms and the 1-
number of receive antenna at the ERs
D search method achieved the same performance, and the PM

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

with ρc = 0.5 exhibits a 2.8 dBm loss over the scheme with Open Research Fund of National Mobile Communications
the optimal ρ∗c,l regardless of the number of ERs. In addition, Research Laboratory, Southeast University (No. 2015D04),
we can find that the curve of the no-An PM scheme increases Specialized Research Fund for the Doctoral Program of Higher
slowly with the number of ERs. Education (SRFDP): 20134101120001; Outstanding Young
Talent Research Fund of Zhengzhou University (1521318003),
the China Postdoctoral Science Foundation (2015T80779),
1
and by the National Research Foundation of Korea (NR-
0.9 F) funded by the Korea Government (MSIP) under Grant
0.8 2014R1A2A1A10049769. The work of Zheng Chu was sup-
0.7
ported by the Newton Fund/British Council Institutional Links
under Grant ID 216429427, Project code 101977.
0.6

A PPENDIX A
CDF

0.5

0.4
P ROOF OF T HEOREM 2
Non−robust method We first consider the Lagrange dual function of (6) as
0.3 PM with ρ = 0.5
c
PM with SPCA L(Q, W, Z, Y, ξl , Ae,k , γ, µl , θk ) = tr(Q) − tr(ZQ) − tr(YW)
[ ( ) 2 ]
0.2
PM with 1−D
WCR−PM with SPCA σp,l
0.1
WCR−PM with 1−D −ξl tr hc,l hH
c,l [tQ−(2 R̄c,l
−t)W] − (2R̄c,l
− t)(σ 2
c,l + )
0
ρc,l
0 5 10 15 20 25 30 35 40
∑ [ ( )]
Minimum harvested power at the ER (dBm)
K
( 1 ) 1
− tr Ae,k He,k ( −1)W − Q He,k +( −1)σk I
H 2
t t
k=1
Fig. 8: CDF of the minimum harvested power at the ERs [ ]
( ) ( ) Ēc,l
+γ tr(Q+W)−P −µl tr hc,l hc,l (Q+W) − H 2
+σc,l
Finally, in Fig. 8, we plot the cumulative density function 1−ρc,l
(CDF) of the minimum harvested power at the ERs with NT = ∑ ( )
K
( )
4, P = 30 dBm, Ēc,l = 10 dBm, Ēe,k = 10 dBm, R̄c,l = 0.5 − θk tr HH e,k (Q+W)H e,k − Ēe,k +N 2
R k ,
σ
bps/Hz, and ε = 0.1. It is observed that the proposed robust k=1
schemes always satisfy the predefined harvested power target where Z ∈ HN + , Y ∈ H+ , ξl ∈ R+ , Ae,k ∈ H+ , γ ∈ R+ ,
T NT NT
(Ēe,k = 10 dBm), whereas the non-robust method achieves µl ∈ R+ , and θk are the dual variables of Q, W, (6a), (6b),
only 25% of the predefined harvested power at the ERs. (4b), (4c), and (4d), respectively. Then, some of the related
KKT conditions are listed as
VII. C ONCLUSION
∂L ∑
K

In this paper, we have proposed a transmit beamforming = I − Z − (ξl t + µl )hc,l hH

c,l + He,k Ae,k HH
e,k
∂Q
k=1
power minimization scheme for a multi-user MIMO SWIP-
T secrecy communication system where power splitters are ∑
K
+ γI − θk He,k HH
e,k = 0, (48a)
employed by the receivers for SWIPT operation. The original
k=1
problem, which was shown to be non-convex, was relaxed to
formulate a two-layer problem. The inner layer problem was ∂L ∑ K
= −Y+[ξl (2R̄c,l −t)−µl ]hc,l hH
c,l + γI− θk He,k HH
e,k
recast as a sequence of SDPs and solved accordingly. Then the ∂W
k=1
optimal solution to the outer problem, on the other hand, has
∑
K
1
been obtained through one-dimensional line search. This opti- − ( −1)He,k Ae,k HH
e,k = 0, (48b)
mization framework has also been extended to robust secrecy t
k=1
transmission designs by incorporating deterministic channel ZQ = 0, Y ≽ 0, Ae,k ≽ 0, ξl ≥ 0, µl ≥ 0, ∀k. (48c)
uncertainties. Moreover, tightness of the relaxation scheme has
been investigated for both the perfect and imperfect CSI cases From the Lagrangian function and the KKT conditions, we
by showing that the optimal solution is rank-one. To reduce the have 0 < ρc,l ≤ 1 and the KKT condition ξl > 0 and µl > 0.
computational complexity, an SPCA based iterative algorithm Now, we will show these conditions via the dual problem of
has been proposed, which achieved near-optimal solution in (6) as
both the perfect and imperfect CSI cases. Finally, numerical max min L(Q, W, Z, Y, ξl , Ae,k , γ, µl )
Z,Y,Ae,k ,ξl ,γ,µl ,θk Q,W,ρc,l
results have been provided to validate the performance of the [
proposed transmit beamforming schemes. = max min − tr(ZQ)−tr(YW)
Z,Y,Ae,k ,ξl ,γ,µl ,θk Q,W,ρc,l

ξl (2R̄c,l −t)σp,l µl Ēc,l ∑ 1

2 K
ACKNOWLEDGMENT
+ + − ( −1)σe2 tr(Ae,k ) (49)
This work was supported in part by the China Nation- ρc,l 1−ρc,l t
k=1
al 863 Project (2014AA01A705), the National Natural Sci-
∑K ]
ence Foundation of China under grant (61401402, 61571402, +[ξl (2 R̄c,l
−t)−µl ]σc,l −γP +
2
θk (Ēe,k − NR σk ) .
2
61571401, 61640003, 61401401, 61501404, 61601516), an k=1

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

Since problem (6) is convex and satisfies the Slater’s con- where Z ∈ HN + , Y ∈ H+ , ξl ∈ R+ , Tc,l ∈ H+ , Te,k ∈
T NT NT

dition, the duality gap between (6) and (49) is zero, and the H+ , Rc,l ∈ H+ , and Re,k ∈ H+
NT NT NT NR
are the dual variables
strong duality holds. Therefore solving problem (6) is equiv- of Q, W, (28), (30), (29a), and (29b), respectively, and
alent to solving (49). In addition, the constraint 0 < ρc,l ≤ 1 [ ]
λc,l I 0NT ×1
can be satisfied as Ac,l = H ,
0 −(2R̄c,l −t1 )(σc,l
2 2
+aσp,l )−λc,l ε2c,l ,
ξl (2R̄c,l −t)σp,l
2
µl Ēc,l [ 1×N T
]
min + . (( t11 − 1)σk2 −λe,k )I 0NT ×1 [ ]
0<ρc,l ≤1 ρc,l 1−ρc,l Ae,k = H λe,k , Ge,k = H̄e,k I ,
01×NT ε2e,k
I
Also the optimal variable ρ∗c,l , and the dual variables ξl∗ , µ∗l [ ]
αc,l I 0NT ×1 [ ]
are related by Bc,l = , Ĥe,k = I h̄e,k ,
√ 0H σ 2
−b Ē c,l −α ε2
c,l c,l
[ 1×NT c,l ]
ξl∗ (2R̄c,l −t)σp,l
2
αe,k I 0NT2 ×1 [ ]
ρ∗c,l = √ √ . Be,k = H , Hc,l = I h̄c,l ,
01×N 2 −Ēe,k +NR σk −αe,k εe,k
2 2
ξl∗ (2R̄c,l −t)σp,l2 + µ∗l Ēc,l T

(j,j)
From the above inequality, we will show that ξl∗ > 0 and and De,k ∈ HN H
+ is a block submatrix of Ĥe,k Re,k Ĥe,k as
T

µl > 0 by contradiction. Suppose that ξl∗ = 0 and/or µ∗l = 0.

∗  (1,1) (1,N )

De,k · · · De,k R
Then there are two cases (i.e., ρ∗c,l = 0 or 1), which violate the  
Ĥe,k Re,k ĤH 
e,k = 
.. .. ..  , ∀k.
constraints (4a) and (4c). Thus, it follows ξl > 0 and µl > 0. . . . 
Now, subtracting (48b) from (48a) yields (NR ,1) (NR ,NR )
De,k · · · De,k
1 ∑
K
Now, we consider the following related KKT conditions as
Z+ξl∗ 2R̄c,l hc,l hH
c,l = I+Y+ He,k Ae,k HH
e,k . (50)
t
k=1 ∂L ∑ K
= I − Z + ξl I − t1 Hc,l Tc,l HH
c,l + Ge,k Te,k GH
e,k
We post-multiply Q by both sides of (50) and use (48c) as ∂Q
k=1
( ) ∑
K ∑
1∑
K NR
∗ R̄c,l (j,j)
H
ξl 2 hc,l hc,l Q = I+Y+ H
He,k Ae,k He,k Q. − Hc,l Rc,l HH
c,l − De,k = 0, (51a)
t k=1 j=1
k=1

Then, it becomes ∂L
= −Y + ξl I + (2R̄c,l − t1 )Hc,l Tc,l HH
c,l
( )−1 ∂W
1∑
K
∗ R̄c,l
ξl 2 I+Y+ H
He,k Ae,k He,k hc,l hH ∑ 1
K
c,l Q = Q. − ( − 1)Ge,k Te,k GH e,k − Hc,l Rc,l Hc,l
H
t
k=1 t1
k=1
Due to ξl∗ > 0, we have ∑K ∑NT
(j,j)
rank(Q) − De,k = 0, (51b)
[ ( )−1 ] k=1 j=1
1∑
K
QZ = 0, W ≽ 0, (51c)
=rank ξl∗ 2R̄c,l I+Y+ He,k Ae,k HH
e,k h c,l hH
c,l Q ( ( ) )
t Ac,l + Hc,l t1 Q − (2
H R̄c,l
− t1 )W Hc,l Tc,l = 0. (51d)
k=1

c,l ) ≤ 1.
=rank(hc,l hH Subtracting (51b) from (51a) generates
This completes the theorem.  ∑
K
1
I−Z+Y−2R̄c,l Hc,l Tc,l HH
c,l + Ge,k Te,k GH
e,k = 0,
t1
A PPENDIX B k=1

P ROOF OF T HEOREM 3 and it follows

First we write the Lagrange dual function of (31) as ∑
K
1
Z+2R̄c,l Hc,l Tc,l HH
c,l = I + Y + Ge,k Te,k GH
e,k .
L(Q, W, Z, Y, ξl , Tc,l , Te,k , Rc,l , Re,k ) = tr(Q) − tr(ZQ) t1
k=1
( )
−tr(YW)+ξl tr(Q + W)−P − tr(Tc,l Ac,l ) Pre-multiplying both sides of the above equality by Q and
( ) ∑ applying the inverse yields
( ) K
−tr Tc,l Hc,l t1 Q−(2
H R̄c,l
−t1 )W Hc,l − tr(Te,k Ae,k ) ( ∑K )−1
R̄ H 1 H
k=1 2 QHc,l Tc,l Hc,l I+Y+
c,l
Ge,k Te,k Ge,k = Q.
∑ ( ) t1
K
( 1 ) k=1
− tr Te,k GH e,k ( −1)W− Q Ge,k −tr(Rc,l Bc,l ) Hence, we have the rank relation as
t1
k=1
rank(Q)
( ) ∑ K ∑NR
( )
−tr Rc,l Hc,l (Q+W)Hc,l −
H
tr(Re,k Be,k ) ( ∑K
1 )
R̄c,l H H −1
k=1 j=1 = rank 2 QHc,l Tc,l Hc,l I+Y+ Ge,k Te,k Ge,k
t1
∑∑
K NR ( ) k=1

− tr (Q+W)De,k ,
(j,j)
≤ rank(Hc,l Tc,l HHc,l ).
k=1 j=1 (52)

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

Now, we will compute the rank of the matrix Hc,l Tc,l HHc,l . [12] Z. Chu, H. Xing, M. Johnston, and S. Le Goff, “Secrecy rate op-
To this end, we consider the following two equalities as timizations for a MISO secrecy channel with multiple multi-antenna
eavesdroppers,” IEEE Trans. Wireless Commun., vol. 15, no. 1, pp. 283-
[ ]
I 0 HH c,l = I,
297, Jan. 2016.
( ) [13] A. Khisti and G. W. Wornell, “Secure transmission with multiple
[ ] [ ] antennas I: The MISOME wiretap channel,” IEEE Trans. Inform. Theory,
I 0 Ac,l = λc,l Hc,l − 0 h̄c,l . (53) vol. 56, no. 7, pp. 3088-3104, Nov. 2010.
[14] Z. Chu, K. Cumanan, Z. Ding, M. Johnston, and S. Y. Le Goff, “Secrecy
[ ] rate optimizations for a MIMO secrecy channel with a cooperative
Pre-multiplying I 0 and post-multiplying HH c,l to (51d), jammer,” IEEE Trans. Vehicular Technol., vol. 64, no. 5, pp. 1833-1847,
it follows May 2015.
[ ] [15] Q. Li, Q. Zhang, and J. Qin, “Secure relay beamforming for SWIPT
I 0 Ac,l Tc,l HH
c,l in amplify-and-forward two-way relay networks,” IEEE Trans. Vehicular
[ ] H( )
− I 0 Hc,l t1 Q−(2R̄c,l −t1 )W Hc,l Tc,l HH c,l = 0,
Technol., vol. 65, no. 11, pp. 9006-9019, Nov. 2016.
[16] Z. Chu, K. Cumanan, Z. Ding, M. Johnston and S. Le Goff, “Robust
and thus we get outage secrecy rate optimizations for a MIMO secrecy channel,” IEEE
( ) Wireless Commun. Lett., vol. 4, no. 1, pp. 86-89, Feb. 2015.
λc,l I+t1 Q − (2R̄c,l − t1 )W Hc,l Tc,l HH
c,l
[17] S. Goel and R. Negi, “Guaranteeing secrecy using artificial noise,” IEEE
[ ] Trans. Wireless Commun., vol. 7, no. 6, pp. 2180-2189, Jun. 2008.
= λc,l 0 h̄c,l Tc,l HH c,l . [18] Q. Li and W.-K. Ma, “Spatially selective artificial-noise aided transmit
optimization for MISO multi-eves secrecy rate maximization,” IEEE
It is easily verified that λc,l I + t1 Q − (2R̄c,l − t1 )W ≽ 0 Trans. Signal Process., vol. 61, no. 10, pp. 2704-2717, May 2013.
[19] L. Liu, R. Zhang, and K.-C. Chua, “Secrecy wireless information and
and it is non-singular. Thus multiplying this matrix will not power transfer with MISO beamforming,” IEEE Trans. Signal Process.,
change the rank of the resulting matrix, and we obtain vol. 62, no. 7, pp. 1850-1863, Apr. 2014.
[20] M. Khandaker and K. Wong, “Masked beamforming in the presence of
rank(Hc,l Tc,l HHc,l ) energy-harvesting eavesdroppers,” IEEE Trans. Inf. Forensics Security,
( )
( ) vol. 10, pp. 40-54, Jan. 2015.
= rank λc,l I+t1 Q−(2 R̄c,l
−t1 )W Hc,l Tc,l Hc,l
H [21] Q. Shi, W. Xu, J. Wu; E. Song, and Y. Wang, “Secure beamforming for
MIMO broadcasting with wireless information and power transfer,” IEEE
( )
[ ] H
(54) Trans. Wireless Commun., vol. 14, no. 5, pp. 2841-2853, May 2015.
= rank λc,l 0 h̄c,l Tc,l Hc,l [22] D. W. K. Ng, E. S. Lo, and R. Schober, “Robust beamforming for secure
communication in systems with wireless information and power transfer,”
( )
[ ] IEEE Trans. Wireless Commun., vol. 13, no. 8, pp. 4599-4615, Aug. 2014.
≤ rank 0 h̄c,l ≤ 1. [23] M. Tian, X. Huang, Q. Zhang, and J. Qin, “Robust AN-aided secure
transmission scheme in MISO channels with simultaneous wireless in-
formation and power transfer,” IEEE Signal Process. Lett., vol. 22, no.
This completes the proof of Theorem 3.  6, pp. 723-727, Jun. 2015
[24] Z. Chu, Z. Zhu, M. Johnston, and S. L. Goff, “Simultaneous wireless
R EFERENCES information power transfer for MISO secrecy channel,” IEEE Trans.
Vehicular Technol., vol. 15, no. 1, pp. 283-297, Jan. 2016.
[1] L. R. Varshney, “Transporting information and energy simultaneously,” [25] Z. Zhu, Z. Chu, Z. Wang, I. Lee, “Joint optimization of AN-aided
in Proc. IEEE ISIT, pp. 1612-1616, Jul. 2008. beamforming and power splitting designs for MISO secrecy channel with
[2] S. Bi, C. K. Ho, and R. Zhang, “Wireless powered communication: SWIPT,” in Proc. IEEE ICC, pp. 1-6, May. 2016.
opportunities and challenges,” IEEE Commun. Mag., vol. 53, no. 4, pp. [26] Y. Wang, R. Sun, and X. Wang, “Transceiver design to maximize the
117-125, Apr. 2015. weighted sum secrecy rate in full-duplex SWIPT systems,” IEEE Signal
[3] Z. Ding, C. Zhong, D. W. K. Ng, M. Peng, H. A. Suraweera, R. Process. Lett., vol. 23, no. 6, pp. 883-887, Jun. 2016.
Schober, and H. V. Poor, “Application of smart antenna technologies in [27] Z. Chu, Z. Zhu and J. Hussein, “Robust optimization for AN-Aided
simultaneous wireless information and power transfer,” IEEE Commun. transmission and power splitting for secure MISO SWIPT system,” IEEE
Mag., vol. 53, no. 4, pp. 86-93, Apr. 2015. Commun. Lett., vol. 20, no. 8, pp. 1571-1574, Aug. 2016.
[4] D. W. K. Ng, E. S. Lo, and R. Schober, “Energy-effcient resource allo- [28] J. Xiong, D. Ma, K. Wong, and J. Wei, “Robust masked beamforming
cation in OFDMA systems with hybrid energy harvesting base station,” for MISO cognitive radio networks with unknown eavesdroppers,” IEEE
IEEE Trans. Wireless Commun., vol. 12, no. 7, pp. 3412-3427, Jul. 2013. Trans. Vehicular Technol., vol. 65, no. 2, pp. 744-755, Feb. 2016.
[5] R. Zhang and C. Ho, “MIMO broadcasting for simultaneous wireless [29] S. Leng, D. Ng, and R. Schober, “Power efficient and secure multiuser
information and power transfer,” IEEE Trans. Wireless Commun., vol. communication systems with wireless information and power transfer,”
12, no. 5, pp. 1989-2001, May 2013. in Proc. IEEE ICC, pp. 800-806, Jun. 2014.
[6] Z. Zhu, K.-J. Lee, Z. Wang, and I. Lee. “Robust beamforming and power [30] S. Mohammadkhani, S. M. Razavizadeh, and I. Lee, “Robust filter
splitting design in distributed antenna system with SWIPT under bounded and forward relay beamforming with spherical channel state information
channel uncertainty,” in Proc. IEEE VTC spring, pp. 1-5, May 2015. uncertainties,” in Proc. IEEE ICC, pp. 5023-5028. Jun. 2014.
[7] H. Lee, S.-R. Lee, K.-J. Lee, and I. Lee, “Optimal beamforming designs [31] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, UK:
for wireless information and power transfer in MISO interference chan- Cambridge University Press, 2004.
nels,” IEEE Trans. Wireless Commun., vol. 14, no. 9, pp. 4810-4821, Sep. [32] K.-Y. Wang, A. Man-Cho So, T.-H. Chang, W.-K. Ma, and C.-Y. Chi,
2015. “Outage constrained robust transmit optimization for multiuser MISO
[8] Z. Zhu, Z. Chu, Z. Wang, and I. Lee, “Robust beamforming design for downlinks: tractable approximations by conic optimization,” IEEE Trans.
MISO secrecy multicasting systems with energy harvesting,” in Proc. Signal Process., vol. 62, no. 21, pp. 5690-5705, Nov. 2014.
IEEE VTC Spring, pp. 1-5, May 2016. [33] K. B. Petersern and M. S. Pedersern, The Matrix Cookbook, Nov. 2008
[9] Z. Zhu, Z. Wang, K.-J. Lee, Z. Chu, and I. Lee, “Robust transceiver [Online]. Available: http://matrixcookbook.com
designs in multiuser MISO broadcasting with simultaneous wireless [34] M. Grant and S. Boyd., “CVX: Matlab software for disciplined convex
information and power transmission,” Journal of Commun. and Networks, programming,” version 2.0 beta, Available: http://cvxr.com/cvx, Sep.
vol. 18, no. 2, pp. 173-181, Apr. 2016. 2012.
[10] Z. Zhu, Z. Wang, Z. Chu, X. Gao, Y. Zhang, and J. Cui,“Robust [35] A. Beck, A. Ben-Tal, and L. Tetruashvili, “A sequential parametric
beamforming based on transmit power analysis for multiuser multiple- convex approximation method with applications to nonconvex truss
input-single-output interference channels with energy harvesting,” IET topology design problems,” J. Global Optim., vol. 47, no. 1, pp. 29-51,
Commun., vol. 10, no. 10, pp. 1221-1228, Jul. 2016. May 2010.
[11] L. Dong, Z. Han, A. P. Petropulu, and H. V. Poor, “Improving wireless [36] Q. Yang, H.-M. Wang, Y. Zhang, and Z. Han “Physical layer security
physical layer security via cooperating relays,” IEEE Trans. Signal in MIMO backscatter wireless systems,” IEEE Trans. Wireless Commun.,
Process., vol. 58, no. 3, pp. 1875-1888, Mar. 2010. vol. 15. no. 11, pp. 7547-7560, Nov. 2016.

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

[37] C. Wang, H.-M. Wang, D. W. K. Ng, X.-G. Xia, and C. Liu “Joint Ning Wang received the B.E. degree in Communi-
beamforming and power allocation for secrecy in peer-to-peer relay cation Engineering from Tianjin University in 2004,
networks,” IEEE Trans Wireless Commun., vol. 14, no. 6, pp. 3280-3293, the M.A.Sc. degree in Electrical Engineering from
Jun. 2015. the University of British Columbia in 2010, and the
[38] C. Wang and H.-M. Wang “Robust joint beamforming and jamming Ph.D. degree in Electrical Engineering from the U-
for secure AF networks: low complexity design,” IEEE Trans. Vehicular niversity of Victoria in 2013. He was a Postdoctoral
Technol., vol. 64, no. 5, pp. 2192-2198, May 2015. Research Fellow with the Department of Electrical
[39] C. Wang, H.-M. Wang, X.-G. Xia “Hybrid opportunistic relaying and and Computer Engineering, the University of British
jamming with power allocation for secure cooperative networks,” IEEE Columbia. From 2004 to 2008, he was with China
Trans. Wireless Commun., vol. 14, no. 2, pp. 589-605, Feb. 2015. Information Technology Design and Consulting In-
[40] A. S. Vishwanathan, A. J. Smola, and S. V. N. Vishwanathan, “Kernel stitute as a mobile communication system engineer,
methods for missing variables,” in Proc. 10th Int. Workshop Artif. Intell. specializing in large-scale commercial mobile communication system design,
Stat., pp. 325-332, Jan. 2005. network traffic analysis, and radio network optimization. Currently he is an
[41] M. Bengtsson and B. Ottersten, “Optimum and suboptimum transmit Associate Professor with the School of Information Engineering at Zhengzhou
beamforming,” in Handbook of Antennas in Wireless Communications, University, Zhengzhou, China. His research interests include resource allo-
L. C. Godara, Ed., CRC Press, Aug. 2001. cation and security design of future cellular networks, channel modeling
[42] Z. Zhu, Z. Chu, Z. Wang, and I. Lee, “Outage constrained robust for wireless communications, statistical signal processing, and cooperative
beamforming for secure broadcasting systems with energy harvesting,” wireless communications.
IEEE Trans. Wireless Commun., vol. 15, no. 11, pp. 7610-7620, Nov.
2016.
[43] H. Wang, S. Ma, T.-S. Ng, and H. V. Poor, “A general analytical
approach for opportunistic cooperative systems with spatially random
relays,” IEEE Trans. Wireless Commun., vol. 10, no. 12, pp. 4122-4129,
Dec. 2011.
[44] E. Karipidis, N. D. Sidiropoulos, and Z.-Q. Luo, “Far-field multicast
beamforming for uniform linear antenna arrays,” IEEE Trans. Signal
Process., vol. 55, no. 10, pp. 4916-4927, Oct. 2007.

Sai Huang is an assistant professor with the Key

Laboratory of Universal Wireless Communications,
Ministry of Education and a lecturer at Beijing U-
niversity of Posts and Telecommunications (BUPT),
Beijing, China. He holds Ph.D. degrees in Informa-
tion and Communication Engineering from BUPT.
His research interests include spectrum sensing and
Zhengyu zhu received B.S. degree from Henan dynamic spectrum management in cognitive wireless
University in 2010, received the Ph.D. degree from networks, universal signal detection and identifica-
Zhengzhou University, China, in 2017. Currently, tion, cooperative wireless communications.
He is with the School of Information Engineering,
Zhengzhou University, China. His research interests
include information theory and signal processing for
wireless communications such as MIMO wireless
network, physical layer security, wireless coopera-
tive networks, convex optimization techniques, and
energy harvesting communication systems.

Zhongyong Wang received his B.S. and M.S. de-

grees in Automatic Control from Harbin Shipbuild-
ing Engineering Institute, Harbin, China, in 1986
and 1988, respectively, and received his Ph.D. de-
Zheng Chu received the Ph.D. degree from the gree in Automatic Control Theory and Application
School of Electrical and Electronic Engineering, from Xi’an Jiaotong University, Xi’an, China, in
Newcastle University, U.K., in 2016. He is cur- 1998. Since 1988, Zhongyong Wang has been with
rently with the Faculty of Science and Technology, Zhengzhou University, Zhengzhou, China, as a lec-
Middlesex University. His research interests include turer in the Department of Electronics. From 1999
physical layer security, wireless cooperative net- to 2002, he was an associate professor, and in 2002
works, wireless power transfer, convex optimization he was promoted to professor in the Department of
techniques, and game theory. Communication Engineering. Prof. Wang’s general fields of interest cover
numerous aspects within embedded systems, signal processing and commu-
nication theory.

1556-6013 (c) 2017 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIFS.2017.2721908, IEEE
Transactions on Information Forensics and Security

Inkyu Lee received the B.S. (Hons.) degree in

control and instrumentation engineering from Seoul
National University, Seoul, Korea, in 1990 and the
M.S. and Ph.D. degrees in electrical engineering
from Stanford University, Stanford, CA, USA, in
1992 and 1995, respectively. From 1995 to 2001,
he was a Member of Technical Staff with Bell
Laboratories, Lucent Technologies, where he stud-
ied high-speed wireless system designs. From 2001
to 2002, he worked for Agere Systems (formerly
Microelectronics Group of Lucent Technologies),
Murray Hill, NJ, USA, as a Distinguished Member of Technical Staff. Since
September 2002, he has been with Korea University, Seoul, where he is
currently a Professor at the School of Electrical Engineering. During 2009,
he visited the University of Southern California, Los Angeles, CA, as a
Visiting Professor. He has published over 120 journal papers in IEEE and
has 30 U.S. patents granted or pending. His research interests include digital
communications, signal processing, and coding techniques applied for next-
generation wireless systems. Dr. Lee has served as an Associate Editor of the
IEEE TRANSACTIONS ON COMMUNICATIONS from 2001 to 2011 and
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS from 2007
to 2011. In addition, he has been a Chief Guest Editor of the IEEE JOURNAL
ON SELECTED AREAS IN COMMUNICATIONS (Special Issue on 4G
Wireless Systems) in 2006. He was a recipient of the IT Young Engineer
Award at the IEEE/IEEK Joint Award in 2006 and of the Best Paper Award
at APCC in 2006, IEEE VTC in 2009, and ISPACS in 2013. He was also
a recipient of the Best Research Award from the Korea Information and
Communications Society in 2011 and the Best Young Engineer Award from
the National Academy of Engineering in Korea (NAEK) in 2013. He has been
elected as a member of NAEK in 2015.

1556-6013 (c) 2017 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.