Professional Documents
Culture Documents
Enabling Secure Wireless Communications Via Intelligent Reflecting Surfaces
Enabling Secure Wireless Communications Via Intelligent Reflecting Surfaces
Abstract—In this paper, we propose to utilize intelligent reflect- There are several studies on the design of IRS-assisted wire-
ing surfaces (IRSs) for enhancing the physical layer security of less systems [6]–[9]. A point-to-point multiple-input single-
wireless communications systems. In particular, an IRS-assisted output (MISO) system was investigated in [6], [7], where
secure wireless system is considered, where a multi-antenna
transmitter communicates with a single-antenna receiver in the the IRS is implemented with continuous and discrete phase
presence of an eavesdropper. To maximize the secrecy rate, both shifters, respectively. Based on the semidefinite relaxation
the beamformer at the transmitter and the IRS phase shifts (SDR) method, approximate solutions for the beamformer
are jointly optimized. Based on the block coordinate descent at the access point and the phase shifts at the IRS were
(BCD) and minorization maximization (MM) techniques, two developed. The authors of [8] considered a downlink multiuser
efficient algorithms are developed to solve the resulting non-
convex optimization problem for small- and large-scale IRSs, communication system, where the signal-to-interference-plus-
respectively. Simulation results show that IRSs can significantly noise ratio (SINR) was maximized for given phase shifts.
improve physical layer security if the proposed algorithms are Energy efficiency maximization was tackled in [9], where sub-
employed. Furthermore, we reveal that deploying large-scale IRSs optimum zero-forcing beamforming was assumed at the access
is more efficient than enlarging the antenna array size of the point. We note that none of these existing works considers
transmitter for both boosting the secrecy rate and enhancing the
energy efficiency. physical layer security, despite its great importance for modern
wireless systems.
I. I NTRODUCTION To fill this gap, this paper investigates physical layer se-
curity provisioning for IRS-assisted wireless systems. Assume
Physical layer security has received considerable attention a transmitter equipped with multiple antennas communicates
from both academia and industry in recent years [1]. Vari- with one legitimate receiver in the presence of an eavesdropper.
ous approaches to improve physical layer security have been Both the legitimate receiver and the eavesdropper are assumed
proposed in the literature, e.g., cooperative relaying schemes, to use a single antenna, and the IRS is implemented via
artificial noise-aided beamforming [2], and cooperative jam- programmable phase shifters. Our goal is to maximize the
ming [3]. However, deploying a large number of relays or other secrecy rate of the considered system by optimizing both
helpers in secure wireless systems inevitably incurs an exces- the beamformer at the transmitter and the phase shifts at
sive cost. Furthermore, cooperative jamming and transmitting the IRS, which leads to a non-convex optimization problem.
artificial noise consume additional power for security provi- Based on the block coordinate descent (BCD) and minorization
sioning. These shortcomings of existing approaches urgently maximization (MM) techniques, two efficient algorithms are
call for a new paradigm for secure wireless systems, which is proposed for solving the problem. The first algorithm is more
both cost-effective and energy-efficient. suitable for small-scale IRSs, while the second algorithm is
With the rapid development of radio frequency (RF) micro- advantageous for large-scale IRSs. Unlike existing works [6],
electro-mechanical systems (MEMS), programmable and re- [8], [9], we obtain locally optimal solutions for both the
configurable meta-surfaces have recently found abundant ap- beamformer and the phase shifts. To the best of the authors’
plications in the public and civil domains, among which intel- knowledge, this is the first work that studies the design of
ligent reflecting surfaces (IRSs) have drawn special attention secure IRS-assisted wireless systems.
for their applications in wireless communications [4]. The Notations: The√imaginary unit of a complex number is
artificial thin films of IRSs can be easily coated on existing denoted by ȷ = −1. Matrices and vectors are denoted by
infrastructures such as walls of buildings, which reduces imple- boldface capital and lower-case letters, respectively. Cm×n
mentation cost and complexity. In addition, unlike conventional denotes the set of all m × n complex-valued matrices. Im is
communication transceivers, IRSs consume no power as they the m-dimensional identity matrix. The i-th element of vector
are passive devices. Furthermore, different from conventional a is denoted as ai . A∗ and AH stand for the conjugate and
transceivers, IRSs help transmit information without generating conjugate transpose of matrix A. diag(a1 , · · · , an ) denotes a
new signals. Instead, they smartly transform or recycle existing diagonal matrix whose diagonal entries are a1 , · · · , an . The
signals. To sum up, IRSs are passive cost-effective devices with largest eigenvalue of matrix A and the corresponding eigen-
the ability to control the radio propagation environment [5]. vector are denoted by λmax (A) and λmax (A), respectively.
These characteristics make IRSs promising key enablers for , means “defined as”. Expectation and the real part of a
improving the physical layer security of wireless communica- complex number are denoted by E[·] and ℜ(·), respectively.
tions in an economical and energy-efficient manner. The operation ∠(A) constructs a matrix by extracting the
l
l and the optimal solution is provided in the following lemma.
maximize 2
f ,Φ 1 + σ12 |hH e ΦGf |
P1 : e (3) Lemma 1. Given the phase shift matrix Φ of the IRS, the
2
subject to ∥f ∥ ≤ P optimal solution for beamforming vector f is given by
( ) √ ( )
Φ = diag eȷθ1 , eȷθ2 , · · · , eȷθM , f ⋆ = P λmax X−1 e Xl , (5)
where P ≥ 0 is the given total transmit power. where
P H H
1 While the CSI of the eavesdropper is generally difficult to acquire, the Xi = INt + G Φ hi hH
i ΦG, i ∈ {l, e}. (6)
results in this paper serve as theoretical performance upper bounds for the σi2
considered system. These bounds and the insights gained from them can be Proof: It was shown in [13] that allocating all transmit
used to guide the system design for the case when the CSI of the eavesdropper 2
is not perfectly known. power for beamforming is optimal, i.e., ∥f ⋆ ∥ = P . Then, the
2 The secrecy rate is zero if the transmission is turned off. Hence, the term
( ) (
numerator and denominator of the objective function of P2 can
2 2 )
log 1 + 12 hH Φf − log 1 + 12 hH e Φf
will always be non- be rewritten as
P 2
σl l σe
negative if the beamformer and the phase shifts are optimized to maximize P H( H H )
the secrecy rate. 1 + 2 hH i ΦGf̃ = f̃ f̃ + 2 f̃
H
G Φ hi hH i ΦG f̃
σi σi
Authorized licensed use limited to: Cisco. Downloaded on December 12,2021 at 11:32:30 UTC from IEEE Xplore. Restrictions apply.
ce,k dl,k cos (pl,k ) − cl,k de,k cos (pe,k ) dl,k de,k sin (pe,k − pl,k )
θ̃k = arctan − arccos √ (12)
ce,k dl,k sin (pl,k ) − cl,k de,k sin (pe,k ) ce,k dl,k + cl,k d2e,k − 2cl,k ce,k dl,k de,k cos (pl,k − pe,k )
2 2 2
di,k = 2 h∗i,k gkH f hi,m e−ȷθm f H gm , (11) v v = M . Therefore, optimization problem P3 can be recast
H
σi as
m̸=k
∑ v H Yl v
maximize g(v) =
pi,k = ∠ h∗i,k gkH f hi,m e−ȷθm f H gm , i ∈ {l, e}, P4 : v v H Ye v (15)
m̸=k subject to |vk | = 1, k ∈ {1, 2, · · · , M }.
Authorized licensed use limited to: Cisco. Downloaded on December 12,2021 at 11:32:30 UTC from IEEE Xplore. Restrictions apply.
Algorithm 2 Alternating Optimization With Minorization
Maximization (AO-MM)
g (v)
vz(t+1) f v|vz(t+1) 1:
(0)
Construct an initial vz and set t = 0;
2: repeat
(t)
g vz(t+1) ≥ g vz(t)
3: Fix vz and optimize f (t) according to (5);
vz(t) 4: Optimize vz
(t+1)
according to (20) and (21);
5: t ← t + 1;
f v|vz(t)
6: until convergence.
Fig. 2. The procedure of minorization maximization.
P
R ff H RH
σl2 l l is positive definite. Because of the convexity, we
Remark 3: The main difficulty in solving P4 is the unit have the following inequality
modulus constraint, which is element-wise and highly non- ( )
v H Yl v ℜ vzH Yl v vzH Yl vz H
convex. By introducing V = vvH as the optimization variable ≥ 2 − 2 v Ye v
and relaxing the rank-one constraint of V, P4 can be solved vH Ye v vzH Ye vz (vzH Ye vz ) (18)
via semidefinite relaxation (SDR). The resulting problem is a (b)
≥ f (v|vz ) + [g(vz ) − f (vz |vz )] ,
quasi-convex problem, whose optimal solution can be obtained
by solving a series of semidefinite programming (SDP) prob- where (b) applies [15, Lemma 1].
lems. However, there is no guarantee that the obtained solution
Proposition 1. The phase shift optimization problem in each
V is a rank-one matrix, and therefore the SDR approach
iteration of the AO is equivalent to
can only provide an approximate solution for v. As a result, [(
the objective function does not necessarily increase in each )H ]
P5 : vz(t+1)
= arg max ℜ w (t)
v , (19)
iteration of the AO, and the convergence of the SDR-based |vi |=1
algorithm cannot be guaranteed. In addition, the computational
where
complexity of solving a large number of SDP problems in each ( )H
iteration of the AO is prohibitively high. (t)
(t)
vz
(t)
Yl v z
In this paper, we propose to solve P4 by the MM technique Yl v z
w (t)
= ( )H − [( )H ]2
[14], whose main idea is illustrated in Fig. 2. In particular, (t) (t) (t) (t) (20)
vz Ye v z vz Ye v z
assuming the value of v in the t-th iteration of the AO is
(t)
denoted as vz , we construct a lower bound on the objective × [Ye − λmax (Ye )IM ] vz(t) .
function g(v) that (touches) the objective function at point
(t) (t)
vz , denoted as f v|vz . We adopt this lower bound as The optimal solution of P5 is given by3
( ) ( )
a surrogate objective function, and the maximizer of this ∠ vz(t+1) = ∠ w(t) . (21)
surrogate objective function is then taken as the value of v
(t+1)
in the next iteration of the AO, i.e., vz . In this way, the Proof: With Lemma 3, the optimization problem that
)
objective value is (monotonically )
(increasing from one iteration needs to be solved is the maximization of f (v|vz ) in (17)
(t+1)
to the next, i.e., g vz
(t)
≥ g vz . The key to the success with respect to v. Since vH v is a constant, the proposition
of( MM lies can be proved with basic algebraic manipulations.
) in constructing a surrogate objective function The AO-MM algorithm is presented in Algorithm 2. With
(t) (t+1)
f v|vz for which the maximizer vz is easy to find.
the closed-form solutions in (5) and (21), the objective function
For phase shift matrix optimization problem P4 , a surrogate is guaranteed to monotonically increase and to converge to a
objective function is composed in the following lemma. local optimum. There are some open issues that require further
Lemma 3. The objective function g(v) is lower bounded by remarks.
(1) Initial point: In both the element-wise BCD and AO-MM
v H Yl v algorithms, we require an initialization for the phase shifts of
g(v) = ≥ f (v|vz ) + [g(vz ) − f (vz |vz )] , (16)
v H Ye v the IRS. As P1 is a non-convex problem, the quality of the sub-
where optimal solution depends to some extent on the initialization.
( ) Here, we propose an effective way to construct the initial values
ℜ vzH Yl v vzH Yl vz { H of the phase shifts. In particular, we set
f (v|vz ) = 2 H − 2 v λmax (Ye )v
v z Ye v z (vzH Ye vz ) ( ) ( )
( )} ∠ vz(0) = ∠ (u) , Φ(0) = diag vz(0) , (22)
+2ℜ vzH [Ye − λmax (Ye )IM ] v , (17)
where u is the dominant left singular vector of Rl in (13).
and g(vz ) − f (vz |vz ) is a constant term that is irrelevant for This initialization is a heuristic obtained by ignoring the
optimization. denominator of the objective function.
Proof: Defining y = vH Ye v, the objective function 3 Since v is a unit modulus vector, it is sufficient to determine the phases
z
H
g(v) , v yYl v is jointly convex in {v, y} since Yl = M
1
IM + of the elements of the vector.
Authorized licensed use limited to: Cisco. Downloaded on December 12,2021 at 11:32:30 UTC from IEEE Xplore. Restrictions apply.
22
Secrecy capacity
(bits/s/Hz)
Element-wise BCD
5.6 20 AO-MM
Element-wise BCD Without IRS
14
(bits/s/Hz)
11
gain with the IRS
6
10.8 AO-MM
M=40 Element-wise BCD 4
10.6 5 6 7 8 9 10 11 12 13 14 15
0 0.5 1 1.5 2 2.5 3 3.5
Number of block updates 104 Transmit power (dBm)
Fig. 3. Convergence of the proposed algorithms for different values of M Fig. 4. Average secrecy rate achieved by different algorithms when M = 10
when Nt = 5, α = 4, P = 5 dBm, rTR = 250 m, rRl = rRe = 160 m. and Nt = 8.
(2) Complexity comparison: As closed-form solutions are M gradually increases, i.e., M = 40, a large number of blocks
obtained for all the blocks of both algorithms, the compu- (41 blocks) have to be updated in each iteration of the element-
tational complexity of the proposed algorithms is critically wise BCD algorithm, which slows down the convergence. In
determined by the total number of block updates until con- contrast, only two blocks per iteration have to be updated in
vergence. In particular, the number of blocks in each iteration the AO-MM algorithm. This offsets the drawback that the AO-
of the element-wise BCD algorithm is M2+1 times higher than MM requires more iterations due to the sub-optimal solutions
that of the AO-MM algorithm. On the other hand, as globally for each phase shift update block. In summary, the element-
optimal solutions (11) are obtained for all the blocks of the wise BCD algorithm is suitable for small-scale IRS systems
element-wise BCD algorithm, the number of iterations needed while the AO-MM algorithm is preferable for large-scale IRS
for convergence is less than for the AO-MM algorithm, in systems.
which the phase shifts are updated in parallel but sub-optimally B. Average Secrecy Rate Evaluation
by (21). Therefore, there is a trade-off between the number of
blocks per iteration and the number of iterations required for In Fig. 4, the average achievable secrecy rate is plotted for
convergence, which shall be investigated in the next section. different algorithms. First, we observe that the average secrecy
rate achieved by both proposed algorithms is the same. We
IV. S IMULATION R ESULTS also compare our approach with a benchmark system which
does not employ an IRS for security provisioning. In this case,
In this section, we numerically evaluate the performance
the distance between the transmitter and the legitimate receiver
of the proposed algorithms. The channels are assumed to be
is denoted by rTl while the distance between the transmitter
independent Rayleigh fading, and the path loss exponent is
and the eavesdropper is denoted by rTe 4 . To maximize the
denoted by α with reference distance 10 meters. The noise
secrecy rate, optimal transmit beamforming according to (5)
power at both the legitimate receiver and the eavesdropper
is adopted. As can be observed from Fig. 4, the system with
is set to σl2 = σe2 = −80 dBm. The distance between the
IRS provides a significant performance gain in terms of the
transmitter and the IRS is denoted as rTR , while rRl and rRe
secrecy rate, which indicates that deploying IRSs is a promising
are the distances from the IRS to the legitimate receiver and
approach for improving the physical layer security of wireless
the eavesdropper, respectively. The simulation results in Figs.
communications systems.
4 and 5 are averaged over 1000 channel realizations.
C. Massive MIMO or Massive IRS?
A. Comparison of the Proposed Algorithms
For conventional wireless communications systems, deploy-
The convergence of the proposed algorithms is investigated ing large-scale antenna arrays at the transceivers is an ef-
for three typical examples in Fig. 3. The stopping criterion for fective way to boost communication performance, including
convergence is that the increment of the normalized objective the network capacity and physical layer security. Therefore,
function is less than ϵ = 10−6 . We first investigate the scenario it is intriguing to investigate how much performance gain we
where the thin film employed to implement the IRS has a can obtain from large-scale IRSs. In Fig. 5, we first increase
small area, e.g., M = 5. Although the number of blocks the number of reflecting elements of the IRS while keeping
per iteration of the element-wise BCD algorithm is slightly the number of antenna elements as Nt = 10 (red curve).
larger for the AO-MM algorithm (6 versus 2), the element-wise In addition, to illustrate the effectiveness of the IRS, we
BCD algorithm needs much fewer iterations for convergence
4 The values of r
since globally optimal solutions are obtained for all the blocks. Tl and rTe and the performance gains achievable with IRSs
depend on the geometry of the network. In Fig. 4, we investigate different
Therefore, the element-wise BCD algorithm converges faster geometries by providing one example with rTl < rRl and rTe < rRe , and
when M is relatively small. On the other hand, as the value of one example with rTl > rRl and rTe > rRe .
Authorized licensed use limited to: Cisco. Downloaded on December 12,2021 at 11:32:30 UTC from IEEE Xplore. Restrictions apply.
13 This equation can be further simplified by some basic trigono-
Nt=10
metric manipulations as follows,
Average secrecy rate (bits/s/Hz) 12
11
M=10
Ak sin θk + Bk cos θk = dl,k de,k sin(pe,k − pl,k ), (25)
M=N t=10
10 where
9 Ak , ce,k dl,k cos pl,k − cl,k de,k cos pe,k ,
(26)
8 Secrecy capacity
gain with 10 IRS
Bk , ce,k dl,k sin pl,k − cl,k de,k sin pe,k .
Nt=10, Increasing M, Element-wise BCD
reflecting elements
7 Nt=10, Increasing M, AO-MM When Ak ≥ 0, by introducing an auxiliary angle, the equation
M=10, Increasing N t, Element-wise BCD
6 can be recast as
M=10, Increasing N t, AO-MM
( )
Without IRS, Increasing N t Ak dl,k de,k sin(pe,k − pl,k )
5 no IRS
cos θk − arctan = √ . (27)
10 20 30 40 50 60 Bk A2 + B 2
Number of elements
It can be readily shown by checking the second derivative that
Fig. 5. Average secrecy rate achieved for different values of M and Nt , when the objective function is maximized when
P = 5 dBm, α = 4, rTR = 200 m, rRl = 150 m, rRe = 100 m, rTl = 300
m, and rTe = 110 m. Ak dl,k de,k sin(pe,k − pl,k )
θk = arctan − arccos √ . (28)
Bk A2 + B 2
also evaluate the average secrecy rate achieved for different The optimal solution when Ak < 0 can be obtained in a similar
numbers of transmit antenna elements when using an IRS manner, which completes the proof.
with M = 10 (blue curve). As can be observed from Fig. R EFERENCES
5, increasing the number of IRS reflecting elements is more
[1] Y. Shiu, S. Y. Chang, H.-C. Wu, S. C.-H. Huang, and H.-H. Chen,
beneficial for improving the secrecy rate than increasing the “Physical layer security in wireless networks: A tutorial,” IEEE Wireless
number of antenna elements. Moreover, as the IRS is a passive Commun., vol. 18, no. 2, pp. 66–74, Apr. 2011.
device, deploying large-scale IRSs is more energy-efficient [2] Y. Sun, D. W. K. Ng, J. Zhu, and R. Schober, “Robust and secure
resource allocation for full-duplex MISO multicarrier NOMA systems,”
than installing more energy-consuming RF chains and power IEEE Trans. Commun., vol. 66, no. 9, pp. 4119–4137, Sep. 2018.
amplifiers as is needed for increasing the number of antenna [3] L. Dong, Z. Han, A. P. Petropulu, and H. V. Poor, “Improving wireless
elements at the transmitter. These results clearly demonstrate physical layer security via cooperating relays,” IEEE Trans. Signal
Process., vol. 58, no. 3, pp. 1875–1888, Mar. 2010.
the superiority of IRSs compared with conventional system [4] M. Di Renzo et al., “Smart radio environments empowered by AI
designs in terms of both communication performance and reconfigurable meta-surfaces: An idea whose time has come,” arX-
energy consumption. iv:1903.08925, Mar. 2019.
[5] C. Liaskos, S. Nie, A. Tsioliaridou, A. Pitsillides, S. Ioannidis, and
V. C ONCLUSIONS I. Akyildiz, “A new wireless communication paradigm through software-
controlled metasurfaces,” IEEE Commun. Mag., vol. 56, no. 9, pp. 162–
In this paper, we proposed to improve the physical layer 169, Sep. 2018.
security of wireless communications networks by deploying [6] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless
network: Joint active and passive beamforming design,” in Proc. IEEE
IRSs. Two efficient algorithms, i.e., the element-wise BCD Global Commun. Conf. (GLOBECOM), Abu Dhabi, United Arab Emi-
and AO-MM algorithms, were developed for joint optimization rates, Dec. 2018, pp. 1–6.
of the beamformer at the transmitter and the phase shifts [7] ——, “Towards smart and reconfigurable environment: Intelligent reflect-
ing surface aided wireless network,” arXiv:1905.00152, May 2019.
at the IRS. The element-wise BCD algorithm was shown to [8] Q.-U.-A. Nadeem, A. Kammoun, A. Chaaban, M. Debbah, and M.-
be preferable for small-scale IRS-assisted systems, while the S. Alouini, “Large intelligent surface assisted MIMO communications,”
AO-MM algorithm is advantageous for wireless systems with arXiv:1903.08127, Mar. 2019.
[9] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and C. Yuen,
large-scale IRSs. Simulation results have confirmed the huge “Reconfigurable intelligent surfaces for energy efficiency in wireless
potential of IRSs to improve the security and energy efficiency communication,” IEEE Trans. Wireless Commun., to appear.
of future communications systems. [10] F. Oggier and B. Hassibi, “The secrecy capacity of the MIMO wiretap
channel,” IEEE Trans. Inf. Theory, vol. 57, no. 8, pp. 4961–4972, Aug.
A PPENDIX 2011.
[11] P. Netrapalli, P. Jain, and S. Sanghavi, “Phase retrieval using alternating
As the phase shift matrix Φ is a diagonal matrix, the minimization,” in Proc. Adv. in Neural Inf. Process. Syst. (NIPS), Lake
objective function of P3 can be rewritten as a function of the Tahoe, NV, USA, Dec. 2013, pp. 2796–2804.
[12] X. Yu, J.-C. Shen, J. Zhang, and K. B. Letaief, “Alternating minimization
k-th reflecting element as follows: algorithms for hybrid precoding in millimeter wave MIMO systems,”
2 IEEE J. Sel. Topics Signal Process., vol. 10, no. 3, pp. 485–500, Apr.
1 + σ12 hH
l ΦGf cl,k + dl,k cos(θk + pl,k ) 2016.
l
2 = c , (23) [13] S. Shafiee and S. Ulukus, “Achievable rates in gaussian MISO channels
1 + 2 |he ΦGf |
1 H e,k + de,k cos(θk + pe,k )
σe with secrecy constraints,” in Proc. IEEE Int. Symp. Inf. Theory, Nice,
France, Jun. 2007, pp. 2466–2470.
where ci,k , di,k , and pi,k are given in (11). By taking the [14] Y. Sun, P. Babu, and D. P. Palomar, “Majorization-minimization algo-
derivative of the objective function with respect to θk , and rithms in signal processing, communications, and machine learning,”
IEEE Trans. Signal Process., vol. 65, no. 3, pp. 794–816, Feb. 2017.
setting the derivative to zero, we obtain the following equation [15] J. Song, P. Babu, and D. P. Palomar, “Optimization methods for design-
ing sequences with low autocorrelation sidelobes,” IEEE Trans. Signal
dl,k sin(θk + pl,k ) [ce,k + de,k cos(θk + pe,k )] Process., vol. 63, no. 15, pp. 3998–4009, Aug. 2015.
(24)
= de,k sin(θk + pe,k ) [cl,k + dl,k cos(θk + pl,k )] .
Authorized licensed use limited to: Cisco. Downloaded on December 12,2021 at 11:32:30 UTC from IEEE Xplore. Restrictions apply.