2256 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO.

9, SEPTEMBER 2018

Constructive Interference Based Secure Precoding:

A New Dimension in Physical Layer Security
Muhammad R. A. Khandaker , Christos Masouros, and Kai-Kit Wong

Abstract— Conventionally, interference and noise are treated exploitation techniques have revolutionised this traditional way
as catastrophic elements in wireless communications. However, of dealing with known interferences [3], [4]. Constructive
it has been shown recently that exploiting known interfer- interference (CI) precoding approaches suggest that interfer-
ence constructively can contribute to signal detection ability at
the receiving end. This paper exploits this concept to design ence power can even contribute to the received signal power
artificial noise (AN) beamformers constructive to the intended if properly exploited [3]–[8]. This concept introduces a major
receiver (IR) yet keeping AN disruptive to possible eavesdroppers breakthrough in designing wireless communication precoding
(Eves). The scenario considered here is a multiple-input single- when the interference is known at the transmitter. In particular,
output wiretap channel with multiple Eves. This paper starts downlink beamforming design can be significantly improved
from AN design without any knowledge of Eve’s CSI, builds
with solutions with statistical CSI up to full CSI. Both perfect by symbol-level precoding of known interferences [7]–[10].
and imperfect channel information have been considered, in par- With the knowledge of the users’ data symbols and channel
ticular, with different extent of Eves’ channel responses. The state information (CSI), the interference can be classified
main objective is to improve the receive signal-to-interference as constructive and destructive interferences. The theory and
and noise ratio at IR through exploitation of AN power in an characterization criteria for constructive interference have been
attempt to minimize the total transmit power, while hindering
detection at the Eves. Numerical simulations demonstrate that extensively studied in [3]–[10]. More recently, the concept
the proposed constructive AN precoding approach yields superior has been exploited in energy harvesting systems [11], hybrid
performance over conventional AN schemes in terms of transmit beamforming [12], cognitive radio networks [13] and massive
power. Critically, they show that, while the statistical constraints MIMO systems [14]–[17]. The interference signals will be
of conventional approaches may lead to instantaneous IR out- constructive to the desired signal if that moves the received
ages and security breaches from the Eves, the instantaneous
constraints of our approach guarantee both IR performance and symbols away from the decision thresholds of the constellation
secrecy at every symbol period. (e.g. real and imaginary axes for QPSK symbols) towards
Index Terms— Interference exploitation, constructive interfer- the direction of the desired symbol. In order to keep the
ence, physical layer security, robust design. angle of the interference signals aligned with the angle of
the corresponding desired symbol, the transmit beamforming
I. I NTRODUCTION
vectors need to be carefully designed.

F IFTH-GENERATION (5G) wireless communication sys-

tems aim to achieve ultra-high spectral efficiency (SE)
and orders-of-magnitude improved energy efficiency (EE).
The broadcast nature of wireless channels makes the com-
munication naturally susceptible to various security threats.
However, the security of wireless data transmission has tra-
It is also expected that 5G networks will operate in mul- ditionally been entrusted to key-based cryptographic meth-
tiple tiers deploying ultra-dense small-cell base stations ods at the network layer. Recently, physical-layer security
(BSs), e.g., heterogeneous networks (HetNets). However, (PLS) approaches have attracted a great deal of attention
a major bottleneck in ultra-dense HetNets is the cross-tier and in the information-theoretic society since the accompany-
co-tier interference. In order to harvest the full potentials of ing techniques can afford an extra security layer on top
5G, developing sophisticated interference handling tools is a of the traditional cryptographic approaches [18]–[24]. PLS
crying need at the moment. exploits the channel-induced physical layer dynamics to pro-
Traditional approach to dealing with interference is to sup- vide information security. With appropriately designed coding
press the interference power in order to improve system perfor- and transmit precoding schemes in addition to the exploita-
mance [1], [2]. However, recent developments in interference tion of any available CSI, PLS schemes enable secret com-
Manuscript received August 31, 2017; revised January 24, 2018 and munication over a wireless medium without the aid of an
February 26, 2018; accepted February 26, 2018. Date of publication March 13, encryption key.
2018; date of current version April 26, 2018. This work was supported The extent of eavesdropper’s CSI available at the transmitter
in part by the Royal Academy of Engineering, U.K., and in part by the
Engineering and Physical Sciences Research Council Project under Grant plays a vital role in determining the corresponding optimal
EP/M014150/1 and Grant EP/R007934/1. The associate editor coordinat- transmission scheme. If full CSI of all the links is available
ing the review of this manuscript and approving it for publication was at the transmitter, then the spatial degrees of freedom (DoF)
Dr. Walid Saad. (Corresponding author: Muhammad R. A. Khandaker.)
The authors are with the Department of Electronic and Electrical Engi- can be fully exploited to block interception [21]. However,
neering, University College London, London WC1E 7JE, U.K. (e-mail: it is generally very unrealistic in practice. In particular, it is
m.khandaker@ucl.ac.uk; c.masouros@ucl.ac.uk; kai-kit.wong@ucl.ac.uk). almost impossible to obtain perfect eavesdroppers’ CSI since
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. eavesdroppers are often unknown malicious agents. The situa-
Digital Object Identifier 10.1109/TIFS.2018.2815541 tion can further worsen if multiple eavesdroppers cooperate
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/
KHANDAKER et al.: CONSTRUCTIVE INTERFERENCE BASED SECURE PRECODING: A NEW DIMENSION IN PLS 2257

Fig. 1. Exploiting AN to boost secrecy performance. (a) Conventional isotropic AN. (b) Conventional spatially selective AN. (c) Constructive AN to boost
received signal power.

in an attempt to maximize their interception through in wireless systems [30]. In this context, we redesign AN
joint receive beamforming. Hence Li and Ma [21] and signals in the form of constructive interference to the IR
Khandaker and Wong [22], [23] considered robust secrecy while keeping AN disruptive to potential Eves. We consider a
beamforming design based on deterministic channel uncer- multiple-input single-output (MISO) downlink system in the
tainty models whereas [25] considered probabilistic robust presence of multiple Eves as shown in Fig. 1c. We aim at min-
design. imizing the total transmit power while boosting the received
To make physical-layer secrecy viable, we usually need SINR at the IR as well as degrading the Eves’ SINR below
the legitimate user’s channel condition to be better than the certain threshold. The benefits of constructive interference-
eavesdroppers’. However, this may not always be guaranteed based AN precoding scheme is twofold compared to conven-
in practice. To alleviate the dependence on the channel con- tional AN-based physical-layer security schemes considered
ditions, recent studies showed that the spatial DoF provided in [20]–[23]. Firstly, the constructive AN will boost the
by multi-antenna technology can be exploited to degrade the receive SINR at the IR as opposed to the conventional AN-
reception of the eavesdroppers [20], [21]. A possible way to do based schemes which attempt to suppress AN signals along
this is transmit beamforming, which concentrates the transmit the direction of the IR. Secondly, to achieve a predefined
signal over the direction of the legitimate user while reducing level of SINR at the IR, constructive interference based
power leakage to the eavesdroppers at the same time. Apart precoding scheme requires lower power compared to con-
from this, a more operational approach is to send artificially ventional AN precoding, thus diminishing inter-user as well
generated noise signals to interfere the eavesdroppers delib- as inter-cell interferences. Depending on the extent of eaves-
erately [20]–[23]. Depending on the extent of eavesdroppers dropping information available at the transmitter, we pro-
CSI available at the transmitter, different strategies can be pose different constructive interference based secure precoding
applied to generate the optimal AN beams. If no eavesdrop- schemes. In particular, we consider different scenarios with
pers’ CSI is available, then a popular design is the isotropic i) no eavesdropping CSI, ii) Eves’ statistical CSI, and iii) Eves’
AN [20], where the message is transmitted in the direction full CSI, perfect and imperfect. Numerical simulations demon-
of the intended receiver’s channel, and spatio-temporal AN is strate that the proposed constructive AN precoding approaches
uniformly spread on the orthogonal subspace of the legitimate yield superior performance over conventional schemes in terms
channel (see Fig. 1a). This scheme guarantees that the intended of transmit power. For clarity, the contributions are summa-
receiver’s (IR’s) reception will be free from the interference by rized below:
the AN, while the Eves’ reception may be degraded by the AN. 1) We first consider the case when no information is
On the other hand, with knowledge of the eavesdroppers’ available about the eavesdroppers, with perfect IR CSI.
CSI to some extent, one can block the eavesdroppers’ inter- We propose constructive interference based AN design
ception more efficiently by generating spatially selective AN for the IR as opposed to the conventional isotropic AN
(see Fig. 1b) [21], [22]. More recently, an antenna array based design onto the null space of the IR.
directional modulation scheme (DM) has been studied which 2) Then, we design a secure precoding scheme with eaves-
enhances security through adjusting the amplitude and phase droppers’ CSI statistics available, such that the AN
of the transmit signal along a specific direction by varying is constructive to the IR while satisfying statistical
the length of the reflector antennas for each symbol while eavesdropping constraints thus reducing the required
scrambling the symbols in other directions [26]–[29]. Note that transmit power for given performance and secrecy
the AN is in general disruptive to the legitimate receivers as requirements.
well, in the above (conventional) schemes. This creates serious 3) Next, when full CSI is available, this allows the design
problems particularly when exact CSI can not be obtained. to move one step further to satisfy instantaneous SINR
This motivates us to rethink the role of interferences as well constraints at all nodes such that the AN is constructive
as the AN. to the IR and destructive to Eves, further impeding signal
In this paper, we exploit the knowledge of interference detection at the Eves and reducing the required transmit
readily available at the transmitter for improving security power to guarantee predefined security.
2258 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 9, SEPTEMBER 2018

4) We further develop a computationally efficient algorithm The transmit signal x can also be expressed as
for the constructive AN precoding scheme based on
projected gradient approach. x = bd sd + ze− j φd sd (6)
5) Finally, we design worst-case robust secure precoders in Assuming constant envelop1 d = 1, the instantaneous transmit
the presence of imperfect CSI of all the nodes. power is given by
In all cases, the proposed schemes outperform the conventional  2
 
AN-aided secure precoding schemes. Note that only the full PT = bd sd + z2 = bd + ze− j φd  . (7)
CSI case has been considered in [30] without proposing any
efficient solver. In the following, we design precoding schemes for instanta-
The rest of this paper is organized as follows. In Section II, neous transmit power minimization exploiting known interfer-
the model of a secret MISO downlink system is introduced. ence (AN in this case) power.
Section III demonstrates how constructive interference pre-
coding scheme boosts receive SINR. The SINR-constrained III. C ONSTRUCTIVE I NTERFERENCE
power minimization problems with a) no Eves’ CSI, b) Eves’ P RECODING T ECHNIQUE
statistical CSI, and c) all-perfect CSI have been studied in Recent advances in interference exploitation have demon-
Sections IV, V, and VI, respectively. In Section VII, strated that constructive interference precoding techniques
we develop an efficient solver for the constructive-destructive can significantly improve receive SINR thus reducing signal
precoding problem. On the other hand, robust constructive detection errors. The theory and characterization criteria for
AN precoding has been designed in Section VIII. Section IX constructive interference have been extensively studied first
presents the simulation results that justify the significance of in the context of code division multiple access (CDMA)
the proposed algorithms under various scenarios. Concluding systems [9], [32]–[34], and more recently to MIMO systems
remarks are provided in Section X. in [3]–[6] and [8]. To avoid repetition, we refer the readers to
II. S YSTEM M ODEL the above works for the details, while here we employ this
concept directly to design our new optimization problems.
Following [31], we consider a MISO downlink system
We will actively exploit interference (AN in this case) con-
where the transmitter (BS) equipped with NT transmit anten-
structively for the IR to reduce the required power for a given
nas intends to transmit a secret message to the IR in the
SINR threshold, while guaranteeing the secrecy constraint for
presence of K possible eavesdroppers. The IR and the Eves
the Eves. The AN signal will be constructive to the received
are all equipped with a single antenna for notational simplicity,
signal at the IR if that moves the received symbols away
while our techniques can be readily extended to multi-antenna
from the decision thresholds of the constellation (e.g. real
receivers. In order to confuse the Eves, the BS injects AN
and imaginary axes for QPSK symbols in Fig. 2a).2 Hence
signals into the secret message in an attempt to reduce the
we intend to keep the angle of that part aligned with the
receive SINRs at the Eves. Thus the received signal at the IR
angle of the corresponding desired symbol sd by appropriately
and those at the Eves are given, respectively, by yd and ye,k :
designing the transmit beamforming vectors. We can do so by
yd = hdT x + n d , (1) pushing the decision symbols towards the constructive regions
ye,k = he,k
T
x + n e,k , for k = 1, . . . , K , (2) of the modulation constellation, denoted by the green shaded
areas (cf. Fig. 2a).
where hd and he,k are the complex channel vectors between For constructive precoding, the AN signals received at
the BS and the IR and between the BS and the kth Eve, the IR are not suppressed or nullified in contrast to the
respectively, n d ∼ CN (0, σd2 ) and n e,k ∼ CN (0, σe2 ) are conventional use of AN [21], [22], [31], rather optimized
the additive Gaussian noises at the IR and the kth Eve, instantaneously such that they contribute to the received sig-
respectively. The BS chooses x as the sum of information nal power. If the AN signals can be aligned with the data
beamforming vector bd sd and the AN vector z such that the symbols sd by properly designing the beamforming precoding
baseband transmit signal vector is vectors, then the AN signals will contribute constructively.
x = bd sd + z, (3) Accordingly, it has been shown in [6] and [8] that the receive
SINR (4) at the IR can be rewritten as
where sd = de j φdis the confidential information-bearing
 T 
symbol for the IR, d indicates the constant amplitude and φd h bd + hT ze− j φd 2
d d
is the phase. γd = . (8)
σd2
Accordingly, the received SINR at the IR is given by
 T 2 Note that the receive SINR at the IR has actually become SNR
h bd 
γd =  d2 , (4) after constructive AN precoding. However, the SINR at the kth
h T z  + σ 2
d d 1 Without loss of generality, we assume d = 1 in this paper for notational
and that at the kth Eve is given by convenience. However, our analyses are valid for any value of d.
 2 2 Although we selected QPSK as a representative modulation scheme for
 T 
he,k bd  exposition, the proposed algorithms and our analyses apply to any PSK
γe,k =   , ∀k. (5) modulation scheme. Moreover, the proposed methodologies can be straight-
 T 2 forwardly adapted for multi-level modulation schemes like QAM following
he,k z + σe2 the analyses in [12].
KHANDAKER et al.: CONSTRUCTIVE INTERFERENCE BASED SECURE PRECODING: A NEW DIMENSION IN PLS 2259

Fig. 2. Exploiting constructive and destructive AN for QPSK symbols. (a) Constructive AN design for the legitimate receiver. Constructive interference
power pushes the decision symbols towards the constructive regions of the modulation constellation, denoted by the green shaded areas. (b) Destructive AN
design for the eavesdropper. Destructive AN pushes the received signal at the Eves away from the decision thresholds (red zone).

Eve remains the same as in (5) since no AN signal has been as

made constructive to the Eves.  
 hdT bd + ze− j φd =0 (10a)
Thus exploiting AN power constructively, the instantaneous  
SINR constraint at the IR can be formulated as the following  hdT bd + ze− j φd ≥ σd d , (10b)
system of constraints
where {x} indicates the imaginary part of the complex
 
number x. However, it can be observed from Fig. 2a that the
 hdT bd + hdT ze− j φd =  (sd ) (9a)
AN contaminated received signal ỹd does not necessarily need
   2
 hdT bd + ze− j φd to strictly align the angle of the desired signal. That is, ỹd lays
≥ d , (9b) on the constructive zone of the desired symbol sd as long as
σd2
the following condition is satisfied
where d is the SINR requirement for correct detection at |αI |
−θ ≤ φd ≤ θ, i.e., ≤ tan θ, (11)
the IR, {x} indicates the real part of the complex number αR − ˜ d
x and  x denotes the corresponding angle. Note that the √
where ˜ d  σd d and θ = π/M, M is the constellation
phases of the AN signals in (9b) have been shifted by the
size. Thus the strict angle constraint (10a) can be relaxed as
phase of the desired symbol sd . The constraint (9a) imposes    
 
that the AN fully aligns with the phase of the symbol of  hdT bd + ze− j φd 
interest sd at the IR, whereas the constraint (9b) guarantees    
that the constructively precoded AN signals can adequately ≤  hdT bd + ze− j φd − σd d tan θ. (12)
satisfy the SINR requirement at the IR. We note that this
signal alignment will only hold for the structure of the IR’s Note that the relaxed angular constraint (12) allows a larger
channel hd , while there will be no such alignment for the Eves’ feasibility region (entire green zone in Fig. 2a). Interestingly,
channels he,k . the QoS constraint d is embedded in (12). Hence we do
Essentially, the angular constraint (9a) is a very strict not need to explicitly impose it in the optimization procedure.
constraint. But exploiting the concept of constructive inter- In the following, we exploit this constructive interference
ference, we can actually relax this constraint without losing constraint in various scenarios depending on the extent of CSI
any optimality which results in available at the transmitter.
  a larger feasible region. Let us
denote ỹd  hdT bd + ze− j φd as the received signal ignoring
the AWGN at the IR, with constructive AN injected, and αR IV. U NKNOWN E AVESDROPPERS ’ CSI
and αI as the abscissa and the ordinate of the phase-adjusted In many practical scenarios, it is often difficult to obtain
signal ỹd , respectively. Applying basic geometric principles to any information about the eavesdroppers’ CSI, or it may
Fig. 2a, the constraints in (9) can be equivalently represented even be impractical to assume that the transmitter is aware
2260 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 9, SEPTEMBER 2018

of the presence of an Eve at all. However, communication given SINR threshold, thanks to the perfect knowledge of
secrecy can still be improved by transmitting AN. In order IR’s CSI. We do this by optimizing the transmitted signal part
to ensure secure communication in such cases, a reasonable (x in (1)), which comprises of the desired symbol and the AN
approach is to allocate minimum resources necessary to obtain signals. The direct benefit is that the IR’s SINR requirement
a certain level of quality-of-service (QoS) for the IR, and is satisfied to equality investing relatively lower power for
devote all other resources to making interception of the signal information transmission and the additionally saved power
more difficult. However, the downside is that the secrecy could be allocated to spreading the AN isotropically within
performance of such a scheme cannot be guaranteed. The given power budget. This should further help confusing any
eavesdropper’s received signal at a defined location can be potential eavesdropper. Thus considering the constructive form
of better quality than the IR’s thus allowing information of the IR’s SINR, as discussed in Section III, we formulate
leakage. In this section, we study conventional and interference the instantaneous total power minimization problem as
exploitation approaches to make the probability of such an  2
 
event as low as possible when no information is available about P1 : min  ρ Pt bd + (1 − ρ)Pt ze− j φd  (15a)
bd ,z
the potential eavesdroppers.    
 
s.t.  hdT ρ Pt bd + (1 − ρ)Pt ze− j φd 
A. Conventional Isotropic AN Design    
≤  hdT ρ Pt bd + (1 − ρ)Pt ze− j φd
A conventional approach is to allocate a fraction ρ of the 
available transmit power Pt for transmitting the confiden- − σd d tan θ, (15b)
tial message signal to achieve the minimum required SINR
at the IR such that the IR experiences no interference at z2 ≥ Pn . (15c)
all [31], [35]. The remaining power is distributed isotropically Note that problem (15) adopts the instantaneous transmit
onto the null space of the legitimate channel to yield as much power (including data symbols) as the objective to minimize,
interference as possible. Formally, the optimization problem as opposed to the average transmit power in conventional opti-
can be represented as mization framework (13). The relaxed angular constraint (15b)
min ρ Pt bd 2 + (1 − ρ)Pt z2 (13a) allows a larger feasibility region, which results in a lower
bd ,z minimum transmit power as we will observe in the simulation
 2
ρ Pt hdT bd  results of Section IX. It is also important to note that the con-
s.t.  2 ≥ d , (13b) straint (15c) guarantees the minimum AN transmitted power
(1 − ρ)Pt hdT z + σd2 and Pn is the guaranteed minimum noise transmit level.3 Since
= ⊥ hd , (13c) there is no information available about the eavesdropping
channels, the constraint (15c) plays an important role in secure
where ⊥ = I NT −hd hdH /hd 2 is the orthogonal complement
beamforming design. Since the optimization objective is to
projection matrix of hd . The optimal ρ is chosen such that the
minimize the total transmit power, the optimal solver would
legitimate IR’s SINR requirement in (13b) is just met, i.e.,
 2 allocate almost zero power to the AN signal without this
ρ Pt hdT bd  constraint. While this is desirable for saving power, it would
= d , (14) not disrupt the eavesdroppers’ reception as required. Thus
σd2
the constraint (15c) plays an important role in jamming the
σ 2
which yields ρ = dPt d , with bd = hhd2 , and Wn = eavesdroppers’ channel yet transmitting at a lower power
d
(1 − ρ)Pt ⊥ is the AN covariance matrix [31], [35]. compared to the conventional isotropic AN scheme introduced
Essentially, if the QoS requirements in problem (13) are too in the previous subsection. However, the problem (15) is still
demanding, then the problem will be infeasible. Hence the not convex due to the non-convex constraint (15c) and the
network designer must set the design parameters realistically coupling of the optimization variables. But we can convexify
such that the constraint (13b) is reachable within the given the constraint (15c) by reformulating it as a geometric mean
power budget Pt . However, this solution may not in general constraint (GMC) [36]. The problem is then solved for given ρ.
yield the best possible SINR for the IR. The optimal ρ can be obtained performing a one-dimensional
searching.
B. Constructive Isotropic AN Design
In practice, the conventional approach of allocating mini- V. S TATISTICAL E AVESDROPPER CSI
mum power for information transmission and maximum power Suppose that the transmitter does not know the instan-
for AN transmission may not always result in the maximum taneous CSI of the eavesdroppers, but can obtain the CSI
possible secrecy performance. Instead, allowing some extent statistics from long-term measures. Unlike traditional channel
of AN to leak to the IR in a constructive-interference fashion, training where pilot signals are transmitted to obtain CSI
will contribute to the received SINR at the IR [6], as discussed before actual data transmission begins, statistical CSI can be
in Section III. estimated based on historical transmissions. In this section,
In this section, we take the conventional isotropic beam- 3 It is assumed that the system designer can set this threshold such that the
forming approach one step forward by exploiting AN con- noise level makes correct decoding by the eavesdroppers extremely difficult.
structively for the IR to reduce the required power for a This may vary depending on the system requirements.
KHANDAKER et al.: CONSTRUCTIVE INTERFERENCE BASED SECURE PRECODING: A NEW DIMENSION IN PLS 2261

we assume that the time average can equivalently approxi- transmitter, one can design the transmit precoding and the
mate the ensemble average due to the ergodicity of random AN beamforming more effectively. In particular, we aim at
channels. For the legitimate IR’s MISO channel, we suppose designing the precoders such that the AN is constructive to
that the transmitter obtains the perfect CSI through feedback the IR while maintaining the conventional secrecy constraints
transmission from the receiver. Let us now define the kth Eve’s to the Eves. As such, the plain constructive interference based
channel correlation matrix as secure transmit precoding optimization problem with statistical
Eves’ CSI can be formulated as
Re,k = E he,k he,k
H
= μe,k μe,k
H
+ Qe,k , k = 1, · · · , K ,
 2
 
(16) min bd + ze− j φd  (19a)
bd ,z
   
where E{·} indicates statistical expectation, μe,k is the mean  
s.t.  hdT bd + ze− j φd 
and Qe,k is the covariance of he,k . In fact, the covariance    
Qe,k , ∀k, indicates the level of CSI uncertainty in second-order ≤  hdT × bd +ze− j φd − σd d tan θ,
statistics sense. For ease of exposition, let us now assume that
(19b)
the eavesdroppers’ channels have white covariances, i.e.,
bdH Re,k bd
Re,k = μe,k μe,k
H
+ σh,k
2
I NT , ∀k, (17) ≤ e,k , ∀k. (19c)
z Re,k z + σe2
H

with σh,k
2 ≥ 0. Obviously, σ 2 = 0 indicates the perfect CSI
h,k Note that the global optimal solution to the problem (19)
case which we elaborate in Section VI. The rest of the analyses can not be guaranteed due to the secrecy constraint (19c)
in this section is therefore based on the assumption that with statistical channel knowledge only. Manipulating this
2 > 0, i.e., the correlation matrix R
σh,k e,k is a nonsingular constraint, the problem (19) can be efficiently solved using
positive definite matrix. convex optimization toolboxes, e.g., CVX [37]. Using the
definition of Wd = bd bdH and Wn  zz H , one can express the
A. Statistical CSI Based Conventional Secure Precoding secrecy constraint (19c) as a linear matrix inequality (LMI).
With the knowledge of the eavesdroppers’ CSI to some Thus the problem (19) can be expressed as
extent, one can block the eavesdroppers’ interception more  2
 
efficiently by generating spatially selective AN. The design P2 : min bd + ze− j φd  (20a)
bd ,z
objective is still power minimization under SINR constraint    
 
at the IR, however, with additional secrecy constraints against s.t.  hdT bd + ze− j φd 
eavesdropping. In order to satisfy these secrecy requirements,    
conventional secrecy power minimization problem with Eves’ ≤  hdT × bd +ze− j φd − σd d tan θ,
CSI statistics is formulated as [31] (20b)
    
tr Re,k Wd − tr Re,k Wn σe
min Tr(Wd ) + Tr(Wn ) (18a) 0, ∀k,
Wd ,Wn σe 1
1 (20c)
s.t. Tr(Wd Rd ) − Tr(Rd Wn ) ≥ σd2 , (18b) 
d Wd bd
0 Wn 0. (20d)
1 bd 1
Tr(Wd Re,k )−Tr(Re,k Wn ) ≤ σe2 , ∀k, (18c)
e,k
Wd 0, Wn 0, rank(Wd ) = 1, (18d) Note that the constraint (20d) takes care of the rank constraint4
on Wd .
where Rd  hd hdH , Wd  bd bdH , Wn  zz H , and e,k is the
secrecy threshold for the k-th Eve. Conventionally, the non-
convex rank constraint is dropped so that the relaxed problem VI. S ECURE P RECODING W ITH F ULL CSI
can be solved using existing solvers [37]. Interestingly, it has
In this section, we assume that perfect CSI of all
been proven in [22] and [31] that for a practically represen-
the receivers (including potential eavesdroppers) is available
tative class of scenarios, the original problem can be solved
at the transmitter. This assumption is valid for scenarios where
optimally. Although the solutions proposed in [22] and [31]
the eavesdroppers are also active users of the system, possi-
are optimal from stochastic viewpoint, the hidden power in
bly for different services. In such cases, the transmitter can
the AN signals has been treated as harmful for the desired
estimate the CSI from the active eavesdroppers’ transmission.
information, and hence, either nullified or suppressed. In the
following subsection, we endeavour to develop a precoding
scheme exploiting the AN power constructively for the desired A. Conventional Secure Precoding With Full CSI
signal at the IR yet keeping it disruptive to the Eves.
With perfect CSI of both the IR and the Eves, the conven-
tional power minimization problem with QoS constraints is
B. Statistical CSI Based Constructive AN Precoding
With perfect CSI of the IR and statistical mean and 4 The problem (20) yields a unit-rank W in all Monte Carlo simulations
d
covariance of the eavesdroppers’ channels available at the we performed in Section IX.
2262 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 9, SEPTEMBER 2018

formulated as Since we aim at keeping the eavesdroppers’ received signal

outside the green (constructive) zone in Fig. 2b, i.e., θ ≤
P − Conv : min bd  + z 2 2
(21a)
bd ,z φe,k ≤ −θ, ∀k, we have the entire red zone to search the
 T 2
h bd  optimal point that minimizes the transmit power. That is,
s.t.  d2 ≥ d , (21b) the SINR restriction constraints at the Eves can be represented
h T z  + σ 2 by the following system of inequalities
d d
 2
 T   
he,k bd 
≤ e,k , ∀k. (21c) − he,kT
bd + ze− j φd
     
 T 2
he,k z + σe2 ≤  he,k T
bd + ze− j φd − σe e,k tan θ, ∀k,
The power minimization problem has been solved in many (24a)
 
existing works for different scenarios [21], [31]. One conven- T − j φd
 he,kbd + ze
tional approach is to reformulate problem (21) as the following    
semidefinite program (SDP) after relaxing the rank constraint ≥  he,k
T
bd + ze− j φd − σe e,k tan θ, ∀k,
min Tr(Wd ) + Tr(Wn ) (22a) (24b)
Wd ,Wn
1
s.t. Tr(Wd Rd ) − Tr(Rd Wn ) ≥ σd2 , (22b) where ˜ e,k  σe e,k . Thus exploiting the knowledge of
d the interfering signals (AN in this case), the constructive
1 AN-based precoding design problem with secrecy power min-
Tr(Wd Re,k )−Tr(Re,k Wn ) ≤ σe2 , ∀k, (22c)
e,k imization objective can be formulated as
Wd 0, Wn 0. (22d)  2
 
However, since the Eves’ CSI is now perfectly known, the cor- P3 : min bd + ze− j φd  (25a)
bd ,z
responding channel correlation matrices are obtained as Re,k =    
 
H . The reformulated problem (22) can be optimally
he,k he,k s.t.  hdT bd + ze− j φd 
solved using CVX [21], [31].    
≤  hdT bd + ze− j φd − σd d tan θ,
B. Constructive AN-Based Secure Precoding (25b)
In this section, our attempt is to further improve the secrecy (24a) and (24b) satisfied. (25c)
performance utilizing the full knowledge of the available CSI.
Since the perfect CSI of the eavesdroppers is also available, Problem (25) is a standard second-order cone program,
we can muddle the eavesdroppers reception more efficiently thus can be efficiently solved using interior-point based
than the correlation based CSI case in Section V-B. The solvers [37].
concept is that, we will design the AN beamformers such that Remark: It is important to note that, by the inclusion of
the AN signal is constructive to the IR while destructive to the data symbols in P1, P2 and P3, the IR’s SNR con-
the Eves. As long as some knowledge of the Eves’ channels straint is guaranteed on a symbol-by-sybmol basis, rather than
is available at the transmitter, one can do so by pushing the statistical secrecy of conventional approaches [20]–[23].
the received signal at the IR towards the decision thresholds In addition, Eves’ secrecy constraints in P3 are also guar-
(green zone in Fig. 2a) while pushing the received signal anteed during each symbol period. As will be shown in our
at the Eves away from the decision thresholds (red zone results, the statistical constraints of conventional formulation
in Fig. 2b). This makes correct detection more challenging for P − Conv allow a) for the IR’s SINR to instantaneously fall
the Eves and therefore reduces the receive SINR. The benefit below the required threshold, thus leading to an IR outage;
is that given secrecy thresholds can be guaranteed with lower b) for the Eves’ secrecy SINRs to be instantaneously higher
transmit power. More importantly, it will be shown in the than the statistical constraint, thus jeopardising the secrecy of
following optimization schemes that the secrecy constraints the useful data. By employing symbol-by-symbol constraints,
are guaranteed on a symbol-by-symbol basis, rather than the proposed approaches avoid this, and guarantee a continu-
the conventional statistical guarantees, which are prone to ous enforcement of the IR’s and Eves’ SINRs.
instantaneous outages.
By denoting αR,k and αI,k  as the real and imaginary parts
of ỹe,k  he,k
T bd + ze− j φd , respectively, ỹe,k , ∀k, will lay in VII. A N E FFICIENT S OLUTION FOR THE S ECURE
the red zone in Fig. 2b if either of the following two constraints C ONSTRUCTIVE P RECODING P ROBLEM
is satisfied In this section, we attempt to develop an efficient solver for
−αI,k the secure constructive AN-based precoding design problem.
φe,k ≤ −θ ⇒ ≤ tan θ, ∀k, if αI,k < 0,
αR,k − ˜ e,k For brevity, here we explore only the most challenging sce-
(23a) nario of constructive-destructive AN precoding problem (25).
αI,k However, the proposed solution can be downscaled to solve
φe,k ≥ θ ⇒ ≥ tan θ, ∀k, if αI,k > 0.
αR,k − ˜ e,k other problem formulations
 as well. Denoting x  bd +
T
(23b) ze− j φd and x̄  {x}T {x}T , the problem (25) can be
KHANDAKER et al.: CONSTRUCTIVE INTERFERENCE BASED SECURE PRECODING: A NEW DIMENSION IN PLS 2263

rewritten as Algorithm 1 Efficient Gradient Descent Algorithm to Solve

min x̄ 2
(26a) Problem (30)
x̄ 1: Input: A, c.
s.t. h̄dT x̄ + σd d tan θ ≤ h̄dT  x̄ tan θ, (26b) 2: Initialize λ(0) ≥ 0 and i = 0.
− h̄dT x̄ + σd d tan θ ≤ h̄dT  x̄ tan θ, (26c) 3: repeat
4: i:= i + 1.
T
h̄e,k x̄ + σe e,k tan θ ≥ T
h̄e,k  x̄ tan θ, ∀k,
5: Compute the direction of the gradient ∇ f (λ(i−1) ).
(26d) 6: Choose ai using backtracking linear search to update λ(i) :
− h̄e,k
T
x̄ + σe e,k tan θ ≤ T
h̄e,k  x̄ tan θ, ∀k,  
λ(i) = max λ(i−1) − ai ∇ f (λ(i−1) ), 02K +2 .
(26e)
 T 7: until convergence.
where h̄d  {hd }T {hd }T , h̄e,k 
  8: Output: Optimal dual variable λ∗ .
T 0 K ,K −I K
{he,k }T {he,k }T , and   . Now,
IK 0 K ,K
by defining the following notations
⎡ ⎤ ⎡ √ ⎤ VIII. ROBUST P RECODING W ITH I MPERFECT F ULL CSI
−h̄dT + tan θ h̄dT  σd √d
⎢ h̄T + tan θ h̄T  ⎥ ⎢ σd d ⎥ In the previous sections, secure precoding schemes have
⎢ d d ⎥ ⎢ ⎥
⎢ −h̄T − tan θ h̄T  ⎥ ⎢−σe e,1⎥ been developed assuming partial/statistical/full CSI available
⎢ e,1 e,1 ⎥ ⎢ ⎥
⎢ h̄T + tan θ h̄T  ⎥ at the transmitter. In this section, we consider a secure commu-
A  sec θ ⎢ ⎥, c ⎢
⎢ σe e,1 ⎥
⎥,
⎢ e,1
..
e,1
⎥ ⎢ .. ⎥ nication scenario where CSI of all nodes is obtainable through
⎢ ⎥ ⎢ ⎥
⎢ . ⎥ ⎢ . ⎥ channel training. However, the estimated CSI is imperfect due
⎢ T ⎥ ⎣
⎣−h̄e,K − tan θ h̄e,K ⎦
T −σe e,K⎦ to quantization and detection errors. Hence we study robust
T + tan θ h̄T 
h̄e,K e,K
σe e,K AN precoding design based on that imperfect CSI estimates.
We model the imperfect CSI considering the widely used
we can equivalently rewrite the problem (26) as Gaussian channel error model such that the channel error
min x̄2 (27a) vectors have circularly symmetric complex Gaussian (CSCG)
x̄ distribution. Thus, the actual channels between the BS and the
s.t. − Ax̄ + c ≤ 02K +2 , (27b) IR can be modeled as
where A is a (2K + 2) × 2NT matrix. The Lagrangian dual
hd = ĥd + ed , (34)
function of the problem (27) is given by
and that between the BS and the kth Eve can be modelled as
L (x̄, λ)  x̄2 + λT (−Ax̄ + c) , (28)
he,k = ĥe,k + ee,k , ∀k, (35)
where λ ≥ 0 is a (2K + 2) × 1 Lagrangian dual variable
associated with the constraint (27b). Setting ∂ L∂(x̄,λ)
x̄ = 02K +2 , where ĥd and ĥe,k , ∀k, denote the imperfect estimated CSI
we obtain the optimal solution to the problem (27) as available at the BS and ed , ee,k ∈ C NT ×1 , ∀k, represent the
1 channel uncertainties such that ed 2 ≤ εd2 , and ee,k 2 ≤
x̄∗ = AT λ. (29) εe2 , ∀k, respectively.
2
Thus the remaining task to find the optimal x̄∗ is to find
A. Conventional AN-Aided Robust Secure Precoding
the optimal dual variables λ∗ . Substituting x̄∗ into (28),
we formulate the dual problem of (27) as Conventional AN-aided downlink robust secrecy power
 T 2 minimization problem with SINR constraints is formulated as
A λ
min f (λ)  − cT λ. (30)
λ 4
min bd 2 + z2 (36a)
In general, it is difficult to derive the optimal solution to bd ,z
 T 2
the non-negative least-squares problem (30). In the following, h bd 
we propose a gradient descent algorithm to solve it. Note that s.t. min  d2 ≥ d , (36b)
ed ≤εd hT z + σ 2
the gradient of f (λ) is given by d d
 2
AAT λ  T 
he,k bd 
− c.
∇ f (λ) = (31) max  ≤ e,k , ∀k.
2  (36c)
ee,k ≤εe  T 2
Algorithm 1 summarizes the gradient descent method for he,k z + σe2
solving problem (30). Finally, we can obtain the optimal
Due to the spherical channel uncertainty model, constraints
beamforming vectors from x∗ as follows [8]:
(36b) and (36c) actually involve infinitely many constraints
x∗ which makes the problem (36) very difficult to solve. However,
b∗d = (32)
K +1 applying S-procedure [22, Lemma 2], the inequality con-
x∗ e− j φd straints in (36) can be transformed into convex LMI constraints
b∗n,i = , ∀i. (33)
K +1 and thus problem (36) can be readily solved using existing
2264 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 9, SEPTEMBER 2018

 2
 
min bd + ze− j φd  (37a)
bd ,z
       
 
s.t.  hdT bd + ze− j φd  ≤  hdT bd + ze− j φd − σd d tan θ, ∀ed  ≤ εd , (37b)
     
T
−  he,k bd + ze− j φd ≤  he,k
T
bd + ze− j φd − σe e,k tan θ, ∀ee,k  ≤ εe , ∀k, (37c)
     
 he,k
T
bd + ze− j φd ≥  he,k
T
bd + ze− j φd − σe e,k tan θ, ∀ee,k  ≤ εe , ∀k. (37d)

solvers. It has been proved in [23] that whenever problem expressed as

(36) is feasible, the corresponding transmit precoding solution  T 
is of rank-one hence optimal.  ĥd + ed b = ĥTR,d b R + ĥTI,d b I + eTR,d b R + eTI,d b I

= h̃dT b2 + ẽdT b2 , (42)

B. Constructive AN-Aided Robust Secure Precoding  T T T
In this section, we aim at constructive AN based robust with b2  b R b I . Thus the constraint (38b) can be
secure precoding design with imperfect knowledge of all CSI, explicitly expressed as the following two constraints
 
as opposed to its perfect CSI counterpart in Section VI. max h̃dT b2 + ẽdT b2 − h̃dT b1 + ẽdT b1 tan θ
With the deterministic channel uncertainty model described ed ≤εd
above, we consider worst-case based robust design. Thus the + σd d tan θ ≤ 0 (43)
constructive AN based robust power minimization problem can  
max − h̃dT b2 − ẽdT b2 − h̃dT b1 + ẽdT b1 tan θ
be formulated as given in (37) (at the top of the this page). ed ≤εd
Note that the information and the AN beamforming vectors + σd d tan θ ≤ 0. (44)
appear in identical form in the objective functions as well as
in the constraints in problem (37). Denoting b  bd + ze− j φd , Similarly, the constraints (38c) and (38d) can be, respectively,
the problem can thus be represented as rewritten as
 
min b2 (38a) max − h̃e,k T
b2 − ẽe,k
T
b2 − h̃e,k
T
b1 + ẽe,k
T
b1 tan θ
bd ,z ee,k ≤εe
  T     T 
  + σe e,k tan θ ≤ 0 (45)
s.t.  ĥd + ed b  ≤  ĥd + ed b  
 min T
h̃e,k + ẽe,k
b2 T
b2 − h̃e,k
T
b1 + ẽe,k
T
b1 tan θ
ee,k ≤εe
−σd d tan θ, ∀ed  ≤ εd , (38b)
 T    T + σe e,k tan θ ≥ 0, (46)
−  ĥe,k + ee,k b ≤  ĥe,k + ee,k  T
 where h̃e,k  hTRe,k hTIe,k . By replacing the CSI error
×b} − σe e,k tan θ, ∀ee,k  ≤ εe , ∀k, (38c) bounds in these constraints, the robust problem (38) can be
 T    T  reformulated as
 ĥe,k + ee,k b ≥  ĥe,k + ee,k b min b2 2 s.t. (47a)
 b1 ,b2
−σe e,k tan θ, ∀ee,k  ≤ εe , ∀k. (38d) h̃dT b2 − h̃dT b1 tan θ + εd b2 − b1 tan θ 
Considering the real and imaginary parts of each complex + σd d ≤ 0 (47b)
vector separately, we have − h̃dT b2 − h̃dT b1 tan θ + εd b2 + b1 tan θ 
+ σd d ≤ 0, (47c)
hd = ĥ R,d + j ĥ I,d + e R,d + j e I,d , (39)
T
− h̃e,k b2 T
− h̃e,k b1 tan θ − εe b2 + b1  tan θ
b = bR + j bI , (40)
+ σe e,k ≤ 0 (47d)
where the subscripts R and I indicate the real and imaginary T
h̃e,k b2 + h̃e,k
T
b1 tan θ − εe b2 + b1  tan θ
components of the corresponding complex notation, respec-
tively. As such, we have the real part, + σe e,k ≥ 0. (47e)
 T  The SOCP problem (47) can be efficiently solved using
 ĥd + ed b = ĥTR,d b R − ĥTI,d b I + eTR,d b R − eTI,d b I existing solvers [37].
Finally, we analyze the computational complexity of the
= h̃dT b1 + ẽdT b1 , (41)
problems P1, P2 and P3 based on interior-point method based
 T  T solvers. Note that in all three formulations, the number of
where h̃d  ĥTR,d ĥTI,d , ẽd  eTR,d eTI,d , and b1 
 T T decision variables is on the order of 2NT . Let us first examine
b R − bTI . Similarly, the imaginary component can be problem P1, which has 2 LMI constraints of size 1 (due to
KHANDAKER et al.: CONSTRUCTIVE INTERFERENCE BASED SECURE PRECODING: A NEW DIMENSION IN PLS 2265

TABLE I power required versus the SINR requirement at the IR for

C OMPLEXITY A NALYSIS OF THE P ROPOSED A PPROACHES the no Eves’ CSI case (Section IV), the statistical Eves’ CSI
case (Section V), the full CSI case (Section VI), and the
gradient descent method (Algorithm 1), as compared with
the corresponding conventional AN precoding schemes for
NT = 6, K = 3, and e = −5 (dB). For a fairer comparison,
we set Pt = NσT 2 d for the isotropic AN design with ρ = 21 .
d
It can be observed that the proposed constructive interference
algorithms achieve significant power gains compared to the
conventional AN precoding schemes. Interestingly, the con-
structive isotropic AN scheme (No CSI) requires lower power
compared to the statistical CSI counterpart, which is due to
the fact that the isotropic AN scheme does not impose eaves-
dropping constraints. However, the superiority of the statistical
CSI algorithm remains in the secrecy guarantee, which we
will observe in the next example. Note that although the
gradient descent algorithm requires higher power compared to
the constructive AN schemes, it requires much lower execution
time yet satisfying the instantaneous SINR constraints [8],
which we will observe in Figs. 5 and 6.
Next, we demonstrate the effects of the different extent
of available Eves’ CSI on the resulting Eve’s SINR. The
histograms of the instantaneously obtained SINRs at the
Fig. 3. Transmit power PT versus required SINR at IR d with NT = Eves normalized by the eavesdropping threshold e with
6, K = 3, and e = −5 (dB).
NT = 6, K = 4, for different CSI cases have been plotted
in Fig. 4. The red lines at position 1 indicate the normal-
the | · | operation) and 1 SOC constraint of size NT . Thus the ized threshold value of the corresponding constraint. It can
complexity of problem P1 is on the order shown in the first be observed that in many cases the instantaneous secrecy
row of Table I [25], [38]. Similarly, the complexity of problem thresholds are not satisfied under the conventional average
P2 and problem P3 can be quantified as shown in the second Eves’ SINR constraints, which jeopardizes the information
and the third row of Table I, respectively. The complexity of secrecy. For the constructive precoding schemes with no Eve’s
the robust problem (47) is shown in the last row. CSI, in line with the conventional precoding, no secrecy can
be guaranteed since there is no explicit secrecy constraint.
IX. S IMULATION R ESULTS However, the statistical Eves’ CSI significantly improves
This section presents numerical simulation results to eval- secrecy guarantee. The Eves’ SINR is perfectly constrained
uate the performance of the proposed constructive interfer- only with full CSI of all nodes. These results demonstrate
ence based PLS algorithms in a MISO wiretap channel. the importance of CSI accuracy for improving information
For comparison, conventional secure precoding performances secrecy.
have also been included. For simplicity, it was assumed that Fig. 5 shows the average execution time of the algorithms
e,k = e , ∀k and σd2 = σe2 = 1. Unless otherwise specified, per optimization versus the number of Eves for the full and
N = 3 and QPSK is the modulation scheme considered, perfect CSI case only, with NT = 6, d = 10 (dB), and
while it has been shown that the concept of constructive e = −5 (dB). Specifically, we denote the conventional
interference also offers benefits to larger scale systems and precoding schemes as ‘Conv Prec’, the constructive inter-
higher order PSK and QAM modulations [3], [39]. All the ference based precoding scheme developed in Section VI-B
estimated channel vectors are generated as independent and as ‘Const Prec’ with conventional eavesdropping constraints,
identically distributed complex Gaussian random variables and the destructive interference based scheme in the same
with mean zero and the TGn path-loss model for urban cel- section as ‘Const-Dest Prec’ in the figures below. The gradient
lular environment is adopted considering a path-loss exponent descent algorithm is denoted as ‘Gradient Desc’. Note that
of 2.7 [40]. All simulation results are averaged over 1000 the conventional approach requires the highest time while the
independent channel realizations, unless explicitly mentioned. gradient descent approach takes the lowest time. However,
In the following simulations, we compare the performance the ‘Const Prec’ and the ‘Const-Dest Prec’ algorithms require
of the proposed approaches with that of the conventional almost identical time to execute. Next, for a fairer comparison
AN-aided precoding scheme in [31] as the benchmark. and noting that the proposed optimizations need to be solved
We start the performance evaluation of the proposed con- on a symbol basis, we analyze the average execution time
structive interference based secure AN precoding schemes per frame considering the LTE Type 2 downlink TDD frame
with varying extent of CSI of the eavesdropping nodes avail- structure defined in [41]. In a Type 2 downlink TDD frame,
able at the transmitter. Fig. 3 shows the average transmit 5 out of the 10 sub-frames are designated for downlink
2266 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 9, SEPTEMBER 2018

Fig. 6. Average execution time (s) versus K for slow/fast fading channels
with NT = 6, d = 10 (dB), and e = −5 (dB).
Fig. 4. Histogram of the average Eves’ SINR normalized by the threshold e
with NT = 6, K = 4. (a) Conventional precoding. (b) Constructive precoding.

Fig. 7. Transmit power PT versus required SINR e with NT = 6, K = 4,

and d = 10 (dB).
Fig. 5. Average execution time (seconds) versus K with NT = 6,
d = 10 (dB), and e = −5 (dB).
transmit power PT versus e for NT = 6, K = 4 and
d = 10(d B). The results in Fig. 7 are consistent with those
transmission, each containing 14 symbols. Therefore, in Fig. 3 in the sense that the proposed constructive inter-
the downlink adopts a block size of 70. We consider two cases; ference precoding schemes yield the best performance. Note
a slow fading case where the channel remains constant for the that the required transmit power for the isotropic beamforming
whole duration of the frame, and a fast fading case where the schemes (No CSI) are invariant of the Eves’ SINR thresh-
channel is constant only for a signle sub-frame. In a typical old since they do not consider blocking the eavesdroppers.
slow fading environment, channel coefficients are assumed to However, for the other schemes, with the increase in the
be constant over one frame duration and hence updated only allowable SINR threshold at the Eves, the required transmit
once. Thus the conventional precoding scheme executes only power gradually decreases due to the relaxed eavesdropping
once over a frame duration. However, the proposed symbol- constraints. In any case, the constructive interference based
by-symbol precoding schemes need to execute 70 times over precoding schemes outperform the conventional AN-aided
one frame period. For the fast fading case, the CSI and secure precoding scheme.
hence the conventional optimization is updated 5 times per Next, we demonstrate the effects of the constructive and
frame. It can be seen from Fig. 6 that, while higher than that destructive AN on the IR’s as well as the Eves’ SINR con-
of the conventional schemes, the per-frame complexity of straints. The histograms of the instantaneously obtained SINRs
the proposed approaches is still comparable. The significant at the IR and Eves normalized by the corresponding thresholds
performance gains offered by our approaches therefore make (i.e., d for the IR and e for the Eves) with NT = 6, K = 4,
their performance-complexity trade-off favourable. for different schemes have been plotted in Fig. 8. The red
In the next example, we examine the transmit power require- lines at position 1 indicate the normalized threshold value of
ment against the maximum allowable eavesdropping SINR e the corresponding constraints. It can be observed that in many
assuming perfect CSI of all nodes. Fig. 7 plots the average cases the instantaneous SINR thresholds are not satisfied under
KHANDAKER et al.: CONSTRUCTIVE INTERFERENCE BASED SECURE PRECODING: A NEW DIMENSION IN PLS 2267

Fig. 8. Normalized histogram of the average SINR with NT = 6, K = 4.

Fig. 9. Transmit power PT versus required SINR d with NT = 6, K = 3, Fig. 10. Transmit power PT versus Eves’ CSI error bound εe with NT =
e = −5 (dB), and εd = 0.1, εe = 0.3 (dB). 6, K = 3, εd = 0.1, e = −5 (dB), and d = 5 (dB), 10 (dB).

for εd = 0.1, εe = 0.3. On the other hand, the ‘Non-robust’

scheme is designed treating the imperfect channel estimates
the conventional average SINR constraints for both the IR and available at the BS as the perfect CSI, hence yields noticeable
the Eves. Indeed, the IR has instantaneous SINRs that are performance degradation. However, the proposed construc-
below the threshold requirements, which would lead to SINR tive interference based robust secure beamforming schemes
outages. More importantly, the Eves’ receive instantaneous demonstrate significant transmit power gains. Fig. 10 shows
SINR above the secrecy threshold jeopardizes the information the required transmit power of the robust algorithms across a
secrecy. However, the SINR threshold is always satisfied for wide range of Eves’ CSI uncertainty with NT = 6, K = 3,
the IR under the constructive AN schemes, although the Eves’ εd = 0.1, e = −5 (dB), and d = 5 (dB), 10 (dB). It can
SINR is perfectly constrained only under the ‘Const-Dest Prec’ be observed that as the CSI error bound increases (i.e., with
scheme. These results demonstrate significant gain in terms of lower extent of CSI available at the transmitter), the required
information secrecy by the proposed schemes. transmit power significantly increases in order to satisfy the
Finally, we turn our attention to the imperfect CSI case SINR requirements.
(Section VIII), where we analyze the performance of the
proposed robust beamforming designs in Figs. 9 and 10 with
X. C ONCLUSIONS
NT = 6, K = 3, e = −5 (dB). In Fig. 9, the robust schemes
indicate the solutions to the problems (36) and (47), respec- We proposed the novel idea of designing the AN-aided
tively, for conventional and constructive precoding schemes secure precoding schemes as constructive to the IR and
2268 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 13, NO. 9, SEPTEMBER 2018

destructive to the Eves. This introduces a major breakthrough [21] Q. Li and W.-K. Ma, “Spatially selective artificial-noise aided transmit
in the conventional approach of transmitting AN for improving optimization for MISO multi-eves secrecy rate maximization,” IEEE
Trans. Signal Process., vol. 61, no. 10, pp. 2704–2717, May 2013.
PLS. The concept opens up new opportunities for expanding [22] M. R. A. Khandaker and K.-K. Wong, “Masked beamforming in
the secrecy rate regions. We studied the downlink transmit the presence of energy-harvesting eavesdroppers,” IEEE Trans. Inf.
power minimization problem considering both perfect and Forensics Security, vol. 10, no. 1, pp. 40–54, Jan. 2015.
[23] M. R. A. Khandaker and K.-K. Wong, “Robust secrecy beamforming
imperfect CSI at the BS. Simulation results demonstrated that with energy-harvesting eavesdroppers,” IEEE Wireless Commun. Lett.,
significant performance gain is achievable by the proposed vol. 4, no. 1, pp. 10–13, Feb. 2015.
constructive AN precoding schemes compared to the conven- [24] A. Mukherjee, S. A. A. Fakoorian, J. Huang, and A. L. Swindlehurst,
“Principles of physical layer security in multiuser wireless networks:
tional schemes and have established the proposed approach as A survey,” IEEE Commun. Surveys Tuts., vol. 16, no. 3, pp. 1550–1573,
a new dimension in the design of PLS. 3rd Quart., 2014.
[25] M. R. A. Khandaker, K.-K. Wong, Y. Zhang, and Z. Zheng, “Probabilis-
tically robust SWIPT for secrecy MISOME systems,” IEEE Trans. Inf.
R EFERENCES Forensics Security, vol. 12, no. 1, pp. 211–226, Jan. 2017.
[1] M. R. A. Khandaker and Y. Rong, “Interference MIMO relay channel: [26] M. P. Daly and J. T. Bernhard, “Directional modulation technique
Joint power control and transceiver-relay beamforming,” IEEE Trans. for phased arrays,” IEEE Trans. Antennas Propag., vol. 57, no. 9,
Signal Process., vol. 60, no. 12, pp. 6509–6518, Dec. 2012. pp. 2633–2640, Sep. 2009.
[2] M. R. A. Khandaker and K.-K. Wong, “Joint source and relay optimiza- [27] M. P. Daly, E. L. Daly, and J. T. Bernhard, “Demonstration of directional
tion for interference MIMO relay networks,” EURASIP J. Adv. Signal modulation using a phased array,” IEEE Trans. Antennas Propag.,
Process., vol. 24, p. 24, Dec. 2017. vol. 58, no. 5, pp. 1545–1550, May 2010.
[3] C. Masouros, T. Ratnarajah, M. Sellathurai, C. B. Papadias, and [28] A. Kalantari, M. Soltanalian, S. Maleki, S. Chatzinotas, and B. Ottersten,
A. K. Shukla, “Known interference in the cellular downlink: A per- “Secure M-PSK communication via directional modulation,” in Proc.
formance limiting factor or a source of green signal power?” IEEE IEEE ICASSP, Shanghai, China, Mar. 2016, pp. 3481–3485.
Commun. Mag., vol. 51, no. 10, pp. 162–171, Oct. 2013. [29] A. Kalantari, M. Soltanalian, S. Maleki, S. Chatzinotas, and B. Ottersten,
[4] G. Zheng, I. Krikidis, C. Masouros, S. Timotheou, D.-A. Toumpakaris, “Directional modulation via symbol-level precoding: A way to enhance
and Z. Ding, “Rethinking the role of interference in wireless networks,” security,” IEEE J. Sel. Topics Signal Process., vol. 10, no. 8,
IEEE Commun. Mag., vol. 52, no. 11, pp. 152–158, Nov. 2014. pp. 1478–1493, Dec. 2016. [Online]. Available: http://arXiv:1606.
[5] C. Masouros and E. Alsusa, “Dynamic linear precoding for the exploita- 04488v2
tion of known interference in MIMO broadcast systems,” IEEE Trans. [30] M. R. A. Khandaker, C. Masouros, and K.-K. Wong, “Constructive
Wireless Commun., vol. 8, no. 3, pp. 1396–1404, Mar. 2009. interference based secure precoding,” in Proc. IEEE ISIT, Aachen,
[6] C. Masouros, “Correlation rotation linear precoding for MIMO broad- Germany, Jun. 2017, pp. 2875–2879.
cast communications,” IEEE Trans. Signal Process., vol. 59, no. 1, [31] W.-C. Liao, T.-H. Chang, W.-K. Ma, and C.-Y. Chi, “QoS-based transmit
pp. 252–262, Jan. 2011. beamforming in the presence of eavesdroppers: An optimized artificial-
[7] C. Masouros, M. Sellathurai, and T. Ratnarajah, “Vector perturbation noise-aided approach,” IEEE Trans. Signal Process., vol. 59, no. 3,
based on symbol scaling for limited feedback MISO downlinks,” IEEE pp. 1202–1216, Mar. 2011.
Trans. Signal Process., vol. 62, no. 3, pp. 562–571, Feb. 2014. [32] C. Masouros and E. Alsusa, “Two-stage transmitter precoding based on
[8] C. Masouros and G. Zheng, “Exploiting known interference as green data-driven code-hopping and partial zero forcing beamforming for Mc-
signal power for downlink beamforming optimization,” IEEE Trans. CDMA communications,” IEEE Trans. Wireless Commun., vol. 8, no. 7,
Signal Process., vol. 63, no. 14, pp. 3628–3640, Jul. 2015. pp. 3634–3645, Jul. 2009.
[9] E. Alsusa and C. Masouros, “Adaptive code allocation for interference [33] C. Masouros and E. Alsusa, “Interference exploitation using adaptive
management on the downlink of DS-CDMA systems,” IEEE Trans. code allocation for the downlink of precoded multiple carrier code
Wireless Commun., vol. 7, no. 7, pp. 2420–2424, Jul. 2008. division multiple access systems,” IET J. Commun., vol. 9, no. 9,
[10] M. Alodeh, S. Chatzinotas, and B. Ottersten, “Constructive multiuser pp. 1118–1130, Oct. 2008.
interference in symbol level precoding for the MISO downlink channel,” [34] C. Masouros and E. Alsusa, “A novel transmitter-based selective-
IEEE Trans. Signal Process., vol. 63, no. 9, pp. 2239–2252, May 2015. precoding technique for DS/CDMA systems,” IEEE Signal Process.
[11] S. Timotheou, G. Zheng, C. Masouros, and I. Krikidis, “Exploiting Lett., vol. 14, no. 9, pp. 637–640, Sep. 2007.
constructive interference for simultaneous wireless information and [35] A. L. Swindlehurst, “Fixed SINR solutions for the MIMO wire-
power transfer in multiuser downlink systems,” IEEE J. Sel. Areas tap channel,” in Proc. IEEE ICASSP, Taipei, Taiwan, Apr. 2009,
Commun., vol. 34, no. 5, pp. 1772–1784, May 2016. pp. 2437–2440.
[12] A. Li and C. Masouros, “Exploiting constructive mutual coupling in [36] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
P2P MIMO by analog-digital phase alignment,” IEEE Trans. Wireless Cambridge Univ. Press, 2004.
Commun., vol. 16, no. 3, pp. 1948–1962, Mar. 2017. [37] M. Grant and S. Boyd. (Apr. 2010). CVX: MATLAB Software for
[13] K. L. Law, C. Masouros, and M. Pesavento, “Transmit precoding for Disciplined Convex Programming (Web Page and Software). [Online].
interference exploitation in the underlay cognitive radio Z-channel,” Available: http://cvxr.com/cvx
IEEE Trans. Signal Process., vol. 65, no. 14, pp. 3617–3631, Jul. 2017. [38] A. Ben-Tal and A. Nemirovski, Lectures on Modern Convex Optimiza-
[14] P. V. Amadori and C. Masouros, “Interference-driven antenna selection tion: Analysis, Algorithms, and Engineering Applications (MPS SIAM
for massive multiuser MIMO,” IEEE Trans. Veh. Technol., vol. 65, no. 8, Series on Optimization). Philadelphia, PA, USA: SIAM, 2001.
pp. 5944–5958, Aug. 2016. [39] M. Alodeh, S. Chatzinotas, and B. Ottersten, “Symbol-level multiuser
[15] P. V. Amadori and C. Masouros, “Constant envelope precoding MISO precoding for multi-level adaptive modulation,” IEEE Trans.
by interference exploitation in phase shift keying-modulated multi- Signal Process., vol. 16, no. 8, pp. 5511–5524, Aug. 2017. [Online].
user transmission,” IEEE Trans. Wireless Commun., vol. 16, no. 1, Available: http://arXiv:1601.02788
pp. 538–550, Jan. 2017. [40] IEEE P802.11 Wireless LANs, TGn Channel Models, IEEE Standard
[16] P. V. Amadori and C. Masouros, “Large scale antenna selection and 802.11-03/940r4, May 2004.
precoding for interference exploitation,” IEEE Trans. Commun., vol. 65, [41] Evolved Universal Terrestrial Radio Access (E-UTRA); Physical Layer;
no. 10, pp. 4529–4542, Oct. 2017. General Description, Release 11, document 3GPP TS 36.201, V11.1.0,
[17] A. Li, C. Masouros, F. Liu, and L. Swindlehurst, “Massive MIMO 1-bit 2008.
DAC transmission: A low-complexity symbol scaling approach,” IEEE
Trans. Wireless Commun., to be published.
[18] A. D. Wyner, “The wire-tap channel,” Bell Syst. Tech. J., vol. 54, no. 8, Muhammad R. A. Khandaker, photograph and biography not available at
pp. 1355–1387, 1975. the time of publication.
[19] I. Csiszár and J. Körner, “Broadcast channels with confidential mes- Christos Masouros, photograph and biography not available at the time of
sages,” IEEE Trans. Inf. Theory, vol. IT-24, no. 3, pp. 339–348, publication.
May 1978.
[20] S. Goel and R. Negi, “Guaranteeing secrecy using artificial noise,” IEEE Kai-Kit Wong, photograph and biography not available at the time of
Trans. Wireless Commun., vol. 7, no. 6, pp. 2180–2189, Jun. 2008. publication.