Covert Federated Learning Via Intelligent Reflecting Surfaces

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 71, NO.
8, AUGUST 2023 4591
Covert Federated Learning via Intelligent

Reflecting Surfaces
Jie Zheng , Member, IEEE, Haijun Zhang , Fellow, IEEE, Jiawen Kang ,
Ling Gao, Member, IEEE, Jie Ren , and Dusit Niyato , Fellow, IEEE
Abstract— Over-the-air computation (OAC) is a promising preserving, is regarded as a significant way to realize
technology that can achieve rapid model aggregation by utilizing ubiquitous AI for 6th generation mobile networks (6G) [1],
the wireless waveform superposition feature to harness the [2]. FL is also an emerging technique for edge intelligence
interference of multiple-access channel for wireless federated
learning (FL). However, OAC-based aggregation for OAC faces in wireless access networks because it enables collaboratively
critical security challenges due to unfavorable and wireless training of an AI model for mobile edge computing [3].
broadcast properties, such as privacy leaks and eavesdropping However, since the mobile devices typically access the base
attacks. In this paper, we propose to utilize an intelligent station (BS) equipped with an edge server through an uncertain
reflecting surface (IRS) to support covert OAC-based FL. We first wireless environment, the model parameters aggregated by
derive the optimal condition for covertness in OAC with IRS and
formulate a joint optimization problem to select the maximum BS will be inevitably distorted by fading and additive noise
covert devices participating in the model aggregation while of wireless channel. The limited radio resources also restrict
satisfying the mean squared error (MSE) requirement. We then the number of mobile devices that can access the BS and
design a covert difference-of-convex-functions program (CDC) participate in FL. As the number of mobile devices increases,
to efficiently determine the transmission power of the device, it gives rise to unacceptable latency in the model aggregation
aggregation beamforming of base station (BS), phase shifts, and
reflection amplitudes at the IRS. Simulation results demonstrate process, which is a critical performance bottleneck of FL.
that our proposed approach can achieve significant performance To overcome these challenges, over-the-air computation
gain compared to the baseline algorithms by deploying IRS into (OAC) combines the concurrent data transmission from a
covert OAC-based FL. large number of devices and performs the model aggregation
Index Terms— Over-the-air computation, federated learning, by leveraging the air-wave superposition nature of wireless
covert communication, intelligent reflecting surfaces. broadcast channel [4], [5], [6]. This enables wireless federated
learning (WFL) to achieve the accurate learning performance
I. I NTRODUCTION while meeting lower latency requirements despite the limited
frequency bandwidth [7]. Furthermore, since the BS in
F EDERATED learning (FL), as a burgeoning distributed

artificial intelligence (AI) scheme with local data privacy-
FL mainly focuses on the global model aggregation rather
than the local models of devices, OAC is recognized as
the non-orthogonal multiple access (NOMA) to accomplish
Manuscript received 26 April 2022; revised 29 October 2022 and 7 March
2023; accepted 22 May 2023. Date of publication 1 June 2023; date of spectrum effectiveness and low latency for WFL. However,
current version 16 August 2023. This research is supported in part by the complex and uncertainty wireless propagation channels
National Natural Science Foundation of China (Grants nos. 62072362 and
62072373), MOE Tier 1 (RG87/22), the National Research Foundation
of OAC-based FL inevitably result in the secure problem,
(NRF), Singapore and Infocomm Media Development Authority under such as the model parameter privacy leaks or eavesdropping
the Future Communications Research Development Programme (FCP), and attacks [8].
DSO National Laboratories under the AI Singapore Programme (AISG
Award No: AISG2-RP-2020-019), under Energy Research Test-Bed and
The physical layer security technologies concentrate on
Industry Partnership Funding Initiative, part of the Energy Grid (EG) securing the content of the transmitted message from the
2.0 programme, and under DesCartes and the Campus for Research Excellence eavesdropper by using secrecy rate (SR) from information
and Technological Enterprise (CREATE) programme. The associate editor
coordinating the review of this article and approving it for publication was
theoretic [9]. In cognitive radio, secondary users are required
R. F. Schaefer. (Corresponding author: Jie Zheng.) to sense the spectrum to determine its availability, which
Jie Zheng and Ling Gao are with the School of Information Science and is usually done without any attempt to conceal the primary
Technology, Northwest University, Xi’an, Shaanxi 710127, China (e-mail:
jzheng@nwu.edu.cn; gl@nwu.edu.cn). user’s wireless transmissions [10]. However, emerging covert
Haijun Zhang is with the Institute of Artificial Intelligence, University communication technology aims to hide the presence of
of Science and Technology Beijing, Beijing 100083, China (e-mail: wireless transmission, offering a high level of security and
haijunzhang@ieee.org).
Jiawen Kang is with the Automation of School, Guangdong University of privacy [11]. Recently IRS-aided wireless transmission has
Technology, Guangzhou 510006, China (e-mail: kjwx886@163.com). been regarded as a revolutionizing technique and has received
Jie Ren is with the School of Computer Science, Shaanxi Normal University, intensive investigation [12], [13], [14], [15], [16]. Utilizing
Xi’an 710062, China (e-mail: renjie@snnu.edu.cn).
Dusit Niyato is with the School of Computer Science and Engi- IRS has the potential to enhance signals for various devices
neering, Nanyang Technological University, Singapore 639798 (e-mail: with different magnitudes, enabling signal alignment in the
dniyato@ntu.edu.sg). aggregation node and resulting in lower losses and higher
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TCOMM.2023.3281880. accuracy in FL systems [17]. In addition, IRS can not only
Digital Object Identifier 10.1109/TCOMM.2023.3281880 strengthen signals to improve the quality of legitimate wireless
0090-6778 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: SASI INSTITUTE OF TECHNOLOGY & ENGINEERING. Downloaded on November 21,2023 at 09:37:11 UTC from IEEE Xplore. Restrictions apply.
4592 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 71, NO. 8, AUGUST 2023
transmission but also suppress signals to prevent an attacker better accuracy in simulations compared to other baseline
from detecting malicious signals [18]. The performance gain of algorithms.
IRS-aided covert communications had been examined in [19]. • We formulate an optimization problem to jointly optimize
For a single-user system, the authors in [20] jointly optimized the transmission power of devices, device selection,
the transmitting beamforming of Alice and passive phase shift aggregation beamforming of the BS, IRS reflection
of IRS for maximization of covert rate. Furthermore, in multi- amplitudes, and phase shifts. This is necessary because
user systems, the transmitter beamforming and phase shift at optimizing only the IRS phase shift is suboptimal for
the IRS were jointly designed to maximize the rate of covert covert communications, and setting the IRS amplitudes
communication [21]. to one for covert transmission is not always ideal. The
Meanwhile, the security challenges associated with OAC- formulated problem involves mixed integer programming,
based FL have garnered increasing attention [8], [22], which is hardly tractable due to the integer objective
[23]. One approach to address this issue is to deploy a function and the nonconvex constraints with coupled
friendly jammer to make the signal at the legitimate receiver integral variable and continuous variables that must
stronger than that at the eavesdropper, thereby protecting satisfy both the MSE requirement and covertness
the analog OAC from passive eavesdropping [8]. Another constraint simultaneously.
method involves using receiver noise as a disturbance injected • We design a covert difference-of-convex-functions (CDC)
into the global model aggregated process for OAC to scheme to solve the mixed integer programming problem.
protect model parameters from various privacy attacks [22]. The integer objective function is transformed into a
Additionally, anonymizing the transmitting mobile devices has sparse optimization problem by inducing sparsity in the
been proposed as a means of enhancing the privacy of the selected devices. To handle the covertness constraint and
OAC system in the WFL scenario [23]. Yet, research on IRS- MSE requirement using matrix lifting, we transform them
assisted secure OAC for FL systems is limited, particularly for into rank-one constraints belonging to the semidefinite
IRS-assisted covert OAC. This paper aims to address this gap program. Then, the semidefinite programming problem is
by exploring the benefits of using an IRS to boost covert OAC converted into CDC programming problems, which can
for FL. In an IRS-aided covert OAC for FL system, the IRS be effectively tackled with the alternating optimization
can reflect the desired signals in phase at the BS to enhance method. Simulation results illustrate that our proposed
the signals to boost the model aggregated accuracy, referred CDC approach can obtain significant performance gains
to as wireless model aggregation strengthens. Additionally, compared to other algorithms in an IRS-aided covert
unwanted signals received by attackers can be suppressed by OAC-based FL system.
reflecting them in the opposite phase, regarded as wireless The remainder of the article is organized as follows.
model aggregation leaks. However, it is vitally challenging Section II describes the system model. Section III presents
to achieve the trade-off between enhancing the signal of the problem formulation and optimal covertness OAC condition.
legitimate receiver and deteriorating the detection signal of Section IV outlines a two-step framework for transforming the
the attacker. It has been proved that the IRS’s amplitudes formulated problem. In section V, we design an alternating
should be designed optimally together with the phase shifts; covert DC algorithm to solve the transformed problem.
otherwise, it can result in performance loss for covert Section VI presents our analysis and simulation results.
communication [24]. Our work fills this gap and provides Finally, we conclude the article in Section VII.
new guides on IRS optimization for covert OAC-based FL Notations: Rk×n and Ck×n represent the k × n complex
systems. We jointly optimize the device’s transmission power, space. R[·] expresses the real component of a complex number.
the number of devices selected, the aggregation beamformer E[·] represents the statistical expectation. The operators (·)T ,
of BS, the phase shift, and the reflecting amplitude of IRS (·)H , T r(·), and diag(·) represent the transpose, Hermitian
to improve the performance of the covert OAC-based FL transpose, trace, and diagonal matrix, respectively. The
system. In the future, our work can be extended to NOMA operators | · |, ∥ · ∥, ∥ · ∥2 , and ∥ · ∥F express the absolute
network [25], imperfect channel state information (CSI) [26] value, Euclidean norm, spectral norm, and Frobenius norm,
and multi-antenna case for the user and Willie [27] in an IRS- respectively.
aided covert FL system. The main contributions of our work
II. S YSTEM M ODEL
are summarized as follows:
• We propose an IRS-aided covert OAC framework that With the IRS-assisted, we investigate a joint computation
can efficiently and securely aggregate models in WFL and communication design for a rapid and secure model
systems. We first derive the optimal covertness OAC aggregated in an OAC-based covert FL system. The system
condition for IRS-assisted WFL systems with available comprises a BS with M -antennas, I single-antenna users, and
CSI. By deploying an IRS, more users can participate an IRS with N passive reflection elements. A warden Willie,
in covert OAC-based FL under the MSE requirement equipped with a single antenna, continuously tries to detect
and covertness level, preventing model parameters from the existence of the aggregation, as illustrated in Figure 1.
leaking information in the aggregation process and
achieving higher accuracy and lower loss in the FL A. IRS-Assisted OAC-Based FL
system. Our proposed approach ensures OAC-based FL In this paper, we take into account IRS-aided OAC as the
secure aggregation resulting in faster convergence and model aggregating method to train a global model for FL.
ZHENG et al.: COVERT FEDERATED LEARNING VIA INTELLIGENT REFLECTING SURFACES 4593
postprocessing function ẑ = ψ(x̂). To verify the aggregated

performance of the global model for OAC, we utilize MSE
to compute the error of the aggregation model between the
estimated value x̂ and the target value x, which are as follows:
M SE(x̂, x) = E[|x̂ − x|2 ]
X 1
= ∈ I| √ aH (hH
rb Θhir + hib )pi − 1|
2
i
β
σ 2 ∥a∥2
+ √ . (5)
β
Fig. 1. The IRS-assisted OAC-based covert FL system. Here, the transmitting symbols follow an independent and
normal distribution with unit variance, i.e., E(|xi |2 ) = 1.
OAC [16], regarded as a function of uplink NOMA technology, Thus, we have the transmit power of device i being limited
is used to effectively promote the model aggregated from by E(|pi xi |2 ) = |pi |2 ≤ P0 .
distributed devices for the FL system. Suppose that the device i B. IRS-Assisted Covert OAC
pre-processes function ϕi (z) = z, and then the post-processes
1 For the covert OAC-based FL system, we consider a finite
function at the BS ψi (z) = |I| z. Afterward, the BS aggregates
number of channel L as delay-constrained covert, which means
the target function of the model, which is shown as
X that in some low delay applications the block length should
z = ψ( ϕi (z i )), (1) be limited with delay constraints. OAC can be regarded as
i∈I NOMA [16], and we assume that the CSI is available for
where I denotes the set of devices participating in FL. Before willie so that willie can identify one of the signals of all
the target function ϕi (·) is calculated at the BS, target-function users, which is considered due to the fact that the willie
variable s should be collected as can detect the existence of transmission of OAC. Specifically,
X the channel gain between i-th user-to-Willie is denoted as
x= xi and xi = ϕi (z i ), i ∈ I, (2) hiw , and the channel gain between IRS-to-Willie is denoted
i∈I as hH rw , where Willie has a single antenna. In our paper,
where xi ∈ C denotes the transmit symbol pre-processed by we assume that Willie is a legitimate user in considered
the device i. system so that the Willie can obtain the accuracy CSI, but
We use hib ∈ CM , hir ∈ CN , and hH rb ∈ C
M ×N Willie in the system is poisoned into malicious user as
to represent the channel gain for i-th device-to-BS, i- eavesdroppers. Thus, we can characterize the fundamental
th device-to-IRS, and the IRS-to-BS. We assume that performance gain that can be achieved by IRS in the context
the channel conditions are flat and slowly varying within of covert communications. The Willie detects an existence of
one transmission slot. Furthermore, since various channel the transmission by distinguishing the following hypotheses:
estimation schemes have been proposed for IRS-aided wireless (
nw [k], H0 ,
transmission [28], [29], we can assume that the perfect yw [k] = (6)
CSI can be available, as in [7], [13], and [16]. We use (hH
rw Θhir + hib )pi xi [k] + nw [k], H1 ,
Θ = diag(ρ1 ejθ1 , ρ2 ejθ2 , . . . , ρN ejθN ) to represent the IRS’s where yw [k] denotes the signal received by Willie for k-
diagonal reflecting matrix, where ρn ∈ [0, 1] denotes reflecting th channel use, H0 represents the null hypothesis that the
amplitude, and θn ∈ [0, 2π) denotes the phase shift by the n-th device cannot send information, H1 represents the alternative
element of IRS, n = 1, 2, . . . , N , respectively. hypothesis that the device sends information to aggregate at
With reflecting signals of the IRS, the signal received by the BS, pi denotes the i-th user transmission power and nw
the BS can be obtained from denotes the GWN with mean 0 and variance δw 2
. Thus, we can
obtain the false alarm rate P r{D1 |H0 } and the miss detection
X
y= (hHrb Θhir + hib )pi xi + n, (3)
i∈I
rate P r{D0 |H1 } at Willie, where D1 and D0 denote the events
whether the device is transmitting or not. So, the detection
where pi ∈ C denotes the i-th device’s transmit power and error probability at Willie is computed by
n denotes the Gaussian white noise (GWN), which is n ∼
CN (0, σ 2 I). ξ = P r{D1 |H0 } + P r{D0 |H1 }. (7)
With the aggregation beamforming a ∈ CM at the BS, the Willie aims to minimize the detected error probability ξ to
target function ahead of postprocessing is estimated by identify the existing transmission of OAC. The optimal ξ is
1 the test of likelihood-ratio, which can be represented by
x̂ = √ aH y
β L
Y
1 HX H 1 P1 = f (yw [k]|H1 ) ≥ 1, ∀k ∈ D1 ,
=√ a (hrb Θhir + hib )pi xi + √ aH n, (4)
β i∈I
β k=1
L
where β represents the parameter used for denoise. Thus, the
Y
P0 = f (yw [k]|H0 ) ≤ 1, ∀k ∈ D0 , (8)
global aggregation model at the BS can be obtained using the k=1
where P0 and P1 denote the likelihood detection at Willie Before we address the optimization problem of the passive
with L wireless channel usage with H0 and H1 , respectively. beamforming of IRS, we analyze the optimal covertness OAC,
2
We use f (yw [k]|H0 ) = CN (0, δw ) and f (yw [k]|H1 ) = which can be referred to D(P0 ∥P1 ) = 0 and the minimum
2 H 2 2
CN (0, pi |hrw Θhir + hiw | + δw ) to express the likelihood error of the aggregation signal min M SE, especially for non-
detection of yw with H0 and H1 , respectively. Under H0 , the zero transmit power of the device. However, achieving the
receive signal at Willie consists of AWGN only and then the perfect covertness OAC is impossible without an IRS with
likelihood function of y w can be obtained by finite blocklength, since the channel between each device and
L Willie without IRS is SISO, making zero forcing impossible.
Y
f (y w |H0 ) = f (yw [k]|H0 )
k=1 A. Optimal Covertness OAC With Perfect CSI
PL 2
1 k=1 |yw [k]| Proposition 1: If the transmission power of the device is
= 2 )L
exp(− 2
) (9)
(πδw δw non zero in the IRS-assisted OAC-based FLP system, if and only
N
Under H1 , the likelihood function of yw can be obtained if the channel condition should satisfy as n=1 |gn | ≥ |hiw |
by to obtain the optimal covert OAC by designing properly IRS
L
reflect elements, where gn denotes the n-th element of g =
diag(hH rw )hir . The optimal covert OAC can be guaranteed
Y
f (y w |H1 ) = f (yw [k]|H1 ) 2 2
k=1 by both the minimizing MSE σ ∥a∥ β and the KL divergence

1 D(P0 ∥P1 ) = 0.
= Proof: Refer to Appendix A.
(π(p2i |hH
rw Θh 2 2 L
ir + hiw | + δw ))
PL 2
Proposition 1 shows that the perfect covert OAC can be
|yw [k]| achieved when the wireless link quality of the reflection is
× exp(− 2 H k=1 )
pi |hrw Θhir + hiw |2 + δw2
larger than the direction wireless link quality from device to
(10) Willie. The proposition also points out that the reflection link
∗ can effectively suppress the ability of Willie’s detection for
We can derive the optimal probed threshold ξ at Willie OAC aggregation when the reflect amplitudes and phase shifts
by (11). However, the formal formulation of ξ ∗ ≥ 1 − ϵ are jointly designed.
contains the gamma function, which can make optimization It is hard to check the feasibility of the optimal solution
and solution challenging. To address this, we utilize a lower for all users and it is usually not feasible for all users to
bound for ξ ∗ as proposed in [30] take part in FL. The aggregation error, convergence rate and
r
1 covertness lie with the number of participant devices and
ξ∗ ≥ 1 − D(P0 ∥P1 ), (11) wireless resource allocation [7], [13]. Thus, we aim to select
2
as more covert participants as possible with wireless allocation
where D(P0 ∥P1 ) denotes the Kullback-Leibler (KL) diver-
to ensure fast convergence, while meeting MSE guarantee for
gence between P0 and P1 derived from f (yw |H0 ) to f (yw |H1 )
prediction accuracy and covertness requirements.
for each user device i as follows, which is based on the total
variation distance bound on the hypothesis testing performance
and Pinskers inequality [31]. B. Optimization Formulation
p2 |hH Θhir + hiw |2 In this paper, the objective function is to maximize the
D(P0 ∥P1 ) = L[ln(1 + i rw 2
) number of selection users problem and we jointly design the
σw
transmit power of the user, receive beamforming of BS, phase
p2 |hH Θhir + hiw |2
− 2 iH rw ]. (12) shift and reflecting amplitude of IRS with MSE requirement
pi |hrw Θhir + hiw |2 + σw
2
and covertness constraint.
In this paper, we adopt a more stringent covertness
max |I| (13a)
constraint D(P0 ∥P1 ) ≤ 2ϵ2 than that typical covertness I,a,Θ,pi
constraint ξ ∗ ≥ 1 − ϵ, where ϵ denotes the covertness level, ∥a∥2

which should be a small value depending on the specific s.t. max ≤ γ, ∀i ∈ I (13b)
i∈I |aH (hHrb Θhir + hib )|
2
requirements of a practical system. The model aggregation
of OAC may be affected by unfavorable wireless links, ln(1 + i rw 2 )
which results in misalignment error and information leak of σw
transmitted signals. Therefore, we propose to leverage the IRS p2 |hH Θhir + hiw |2 2ϵ2
− 2 iH rw ≤ , ∀i ∈ I
to guarantee the accuracy and security of OAC during the pi |hrw Θhir + hiw |2 + σw 2 L
aggregation process. (13c)
|Θn,n | ≤ 1, ∀n = 1, . . . , N, (13d)
III. F ORMULATION OF THE O PTIMIZATION P ROBLEM 2
|pi | ≤ P0 , ∀i ∈ I, (13e)
The passive beamforming of IRS needs to be designed
appropriately so that the reflected signals can reduce the where γ > 0 denotes the MSE requirement for the OAC
aggregation error at the BS and avoid detection by Willie. aggregated model in (13b), and |I| represents the set of
selected devices to upload the parameters of local models,

where(13c) is the covertness constraint and the constraint in
(13d) represents that the reflection amplitude is limited in [0, 1]
and the phase shift is limited in [0, 2π] for each reflecting
element of IRS, and (13e) denotes the maximum transmission
power of the device. The optimization problem (13) is mixed
integer programming, which is highly intractable due to
the integer objective function |I| and the nonconvex MSE
constraint with coupled ingetral variable I and continuous
variable a, pi and Θ. Moreover, the covertness constraint (13c)
is non-convex since the tight coupling of pi and Θ in KL
divergence [19], [32]. Fig. 2. The transform and decomposition of problem for solving.
To solve the problem, the MSE requirement (13b)
can be transformed as nonconvex constraints ∥a∥2 −
A. Two-Step Covert Framework
γ|aH (hH 2
rb Θhir + hib )| ≤ 0 and ∥a∥ ≥ 1, which has been
proved in [13]. We introduce an auxiliary variable s [33] to The two problems P1 and P2 are described as follows:
transform the problem of maximizing the number of users into 1) The covert priority of devices selected: To solve problem
that of minimizing the number of nonzero si . Therefore, the P in (14), we replace the nonconvex sparse objective function
following sparse optimization problem becomes l0 -norm with the l1 -norm as a convex substitution [32], which
can be solved by CVX. The problem P1 with the objective of
(P ) : min ∥s∥0 (14a) l1 -norm can be expressed as follows
s,a,Θ,pi
s.t. ∥a∥ − γ|aH (hH

2 2
rb Θhir + hib )| ≤ si , ∀i ∈ J (P1 ) : min ∥s∥1
s,a,Θ,pi
(14b)
s.t. (14b), (14c), (14d), (14e), (14f ). (15)
ln(1 + i rw 2
)
σw By solving problem P1 , we can obtain the value si , which
p2 |hH Θhir + hiw |2 2ϵ2 represents the priority of the i-th device selection, and then we
− 2 iH rw ≤ , ∀i ∈ J to identify the feasibility of the covertness users selection
turn
pi |hrw Θhir + hiw |2 + σw2 L
for OAC as the second subproblem.
(14c)
2) Maximum covert device set for minimized MSE: The
∥a∥2 ≥ 1, (14d) results si of the P1 subproblem characterizes the MSE of
|Θn,n | ≤ 1, ∀n = 1, . . . , N, (14e) device i, but it cannot achieve the requirement of MSE and
2
|pi | ≤ P0 , ∀i ∈ J , (14f) covertness. The larger si represents the lower priority device to
be selected. We sort {si }Ji=1 as sπ(1) ≤ sπ(2) ≤ · · · ≤ sπ(K)
where J ⊆ I. The sparsity structure of s implies that the in the ascending order to ensure the priority of device from
selection of devices should satisfy the MSE requirement. P1 , where sπ(i) represents the i-th largest element for {si }Ji=1 .
The problem P is still nontrivial to solve because of the We can employ bisection method to obtain the maximum set of
sparse objective function and the nonconvex (14b), (14c) and j that assures devices in the set to be feasibly selected for P2 .
(14d) constraints. Therefore, we design an alternating covert For a given set I [j] , we check the feasibility set to minimize
framework to handle the issue in the next section. maximal MSE of selected devices in I [j] . Thus, with the
results from P1 , we select the maximum covertness devices to
IV. P ROBLEM D ECOMPOSITION minimize MSE of selected devices with the bisection method
To solve problem P , the proposed covert DC framework as follows:
is first divided into two problems P1 and P2 , and then the ∥a∥2
problems P1 and P2 are solved by the covert DC algorithm, (P2 ) : min max
pi ,a,Θ i∈I |aH (hH Θhir + hib )|2
rb
which is shown in Figure 2. The proposed alternating
s.t. (14c), (14e), (14f ). (16)
covert optimization algorithm (P1,1 , P1,2 and P2,1 , P2,2 ) can
be carried out in a central manner, which can be easily If the objective of solving problem P2 for the set of I k is
implemented in BS. The BS then sends the power allocation less than the requirement of MSE, then we can obtain the
parameters together with the aggregated gradient results to the optimal covert set of devices with the transmission power
user in the downlink, and sends the phase parameters to the of the device, aggregation beamforming of BS, reflection
IRS with the backhaul connection. In the first subproblem, amplitudes and phase shifts of IRS.
the sparsity of x is introduced to ensure the covert priority
of devices selected. In the second subproblem, based on
the priority of devices, we aim to minimize MSE with the B. Alternating Low-Rank Covert Optimization
MSE requirement and covertness constraint to obtain the Due to the nonconvex of the objective functions of P1 and
maximum feasible device set to participate in the covert P2 and their constraints, we propose to employ alternative
OAC-based FL. optimization with matrix lifting to solve this problem.
1) The covert priority of devices selected: In this Proof : The detailed proof is provided in Appendix B.
subproblem, we alternately optimize variables (s, a, pi ) and Θ With (22) and (23), we can divide problems P1 into
for problem P1 . Firstly, we fix the reflect beamforming matrix problems P1,1 and P1,2 by the convex rearranged and the
Θ, and then the problem P1 is represented as matrix lifting method [35]. By defining A = aaH , problem
min ∥s∥1 (17) can be recast P1,1 as follows:
s,a,pi
s.t. (14b), (14c), (14d), (14f ). (17) P1,1 min ∥s∥1 (24a)
s,A,p,U
We first note that the term |hH 2 s.t. T r(A) − γT r(AH i ) ≤ si , ∀i ∈ J , (24b)
rw Θhir + hiw | for
the nonconvex covertness constraint (14c) is equivalently T r(Qi U i ) T r(Qi U i )
(1 + 2
)ln(1 + 2
)
transformed to σw σw
v H gg H v + 2Re(v H gh∗iw t) + |hiw |2 |t|2 , ∀i ∈ J , (18) 2ϵ2 T r(Qi U i ) 2ϵ2
−(1 + ) 2
≤ , ∀i ∈ J , (24c)
L σw L
where g = diag(h)H
rw hir
and v = [v1 , v2 , . . . , vN ], vn = T r(A) ≥ 1, A ⪰ 0, rank(A) = 1, (24d)
ρn e−jθn . The auxiliary variable t satisfies |t|2 = 1. hiw 2
is scalar, which means that the device and Willie are both |pi | ≤ P0 , U n,n ≤ p2i , U N +1,N +1 = p2i ,
configured with one antenna. Furthermore, if we set di = hiw , U ⪰ 0, rank(U ) = 1, ∀i ∈ J , (24e)
then (18) is rewritten as a quadratic form eH Ge, where where H i = hi hH and hi = hH
i rb Θhir
+ hib .
gi gH H

e = [v, t]T , Qi = i g i di For variables s and a, problem P1 can be simplified into a
H . (19)
di g H
i di di feasible identified problem of reflect beamforming matrix Θ.
In fact, (14c) and (14e) can be further rewritten as follows: By defining v = [ρejθ1 , . . . , ρejθN ]T , bH H H
i = a hrb diag(hir )
H
and ci = a hib , the problem P1 is recast as
p2i eH Qi e p2i eH Qi e 2ϵ2
ln(1 + ) − ≤ , ∀i ∈ J , Solve v (25a)
2
σw p2i eH Qi e + σw2 L
2 H 2
|en | ≤ 1, ∀n = 1, 2, . . . , N, s.t. ∥a∥ − γ|b v + ci | ≤ si , ∀i ∈ J , (25b)
|eN +1 | = 1. (20) uH Q ui uH Q ui 2ϵ2
ln(1 + i 2 i ) − H i i 2
≤ , ∀i ∈ J ,
σw ui Q i ui + σw L
It is obvious that vt is also the optimal result for the original
P1 when e is the optimal result for P1 . However, it is still hard (25c)
to solve P1 because the power pi of the device is coupled with where solve v represent the feasible set of reflection
the reflect beamforming e of IRS in the covertness constraint amplitudes and phase shifts at the IRS as follows
(20). Especially, pi is a scalar variable with |pi |2 ≤ P0 in
(14f), which allows us to simplify (20) as |vn | ≤ 1, ∀n ∈ 1, . . . , N . (25d)
T
uH uH 2ϵ2 Since we have e = [v, t] in (19). The constraint (25b) is
i Qi ui i Qi ui
ln(1 + ) − ≤ , ∀i ∈ J , further recast as
2
σw uH 2
i Q i ui + σw L
|pi |2 ≤ P0 , ∀i ∈ J , ∥a∥2 − γ(eH S i e) ≤ si , ∀i ∈ J
|un | ≤ pi , ∀n = 1, 2, . . . N, (26)
|uN +1 | = pi .∀i ∈ J (21) where
bi bH

where ui = pi e. With the non-convex constraint (21), it is i b i ci
Si = H 2 . (27)
generally difficult to handle P1 directly [34]. Nevertheless, cH
i bi |ci |
we set U i = ui uH i , the constraint (21) is represented as To facilitate the development of the problem, we set V = eeH
2 by lifting e with rank(V ) = 1 by a positive semidefinite matrix
T r(Qi U i ) T r(Qi U i ) 2ϵ
ln(1 + 2
)− 2
≤ , lifting. Problem (25) is transformed into the optimization
σw T r(Qi U i ) + σw L
|pi |2 ≤ P0 , ∀i ∈ J , problem with a low rank matrix:
U n,n ≤ p2i , ∀n = 1, 2, . . . , N, P1,2 : Solve V (28a)
U N +1,N +1 = p2i , ∀i ∈ J 2
s.t.∥a∥ − γT r(S i V ) ≤ si , ∀i ∈ J (28b)
U ⪰ 0, 2 2
p T r(Qi V ) p T r(Qi V )
rank(U ) = 1. (22) (1 + i 2
)ln(1 + i 2
)
σw σw
Proposition 2: The covertness constraint in (22) is rearranged 2ϵ2 p2i T r(Qi V ) 2ϵ2
into a convex constraint with U , which is described as − (1 + ) 2
≤ , ∀i ∈ J (28c)
L σw L
T r(Qi U i ) T r(Qi U i ) V n,n ≤ 1, ∀n ∈ 1, . . . , N + 1; V ⪰ 0; rank(V ) = 1. (28d)
(1 + 2
)ln(1 + 2
)
σw σw
2) Maximum covert device set for minimized MSE: On
2ϵ2 T r(Qi U i ) 2ϵ2 the other hand, we can reformulated problem P2 with
− (1 + ) 2
≤ , ∀i ∈ J . (23)
L σw L Proposition 3.
Proposition 3: Problem P2 can be rewritten equivalently as wi2 T r(Gi V ) wi2 T r(Gi V )

(1 + 2
)ln(1 + 2
)
follows: σw σw
2ϵ2 wi2 T r(Gi V ) 2ϵ2
min ∥a∥2 (29a) − (1 + ) 2
≤ , ∀i ∈ I [j] (33c)
p,a,Θ L σw L
|aH (hH 2 |V n,n | ≤ 1, ∀n ∈ 1, . . . , N + 1; V ⪰ 0; rank(V ) = 1.
s.t. rb Θhir + hib )| ≥ 1, ∀i ∈ J (29b)
(33d)
ln(1 + i rw 2 ) Our proposed alternating covert DC algorithm described in
σw
Algorithm V-B obtains the priority of devices and ensures the
p2 |hH Θhir + hiw |2 2ϵ2
− 2 iH rw ≤ , ∀i ∈ J feasible covertness and MSE of devices to solve the problem
pi |hrw Θhir + hiw |2 + σw2 L P . Firstly, problem P is divided into two problems: the covert
(29c) priority of devices selected priority as P1 and maximum covert
|Θn,n | ≤ 1, ∀n ∈ 1, . . . , N . (29d) device set for minimized MSE P2 . By using matrix lifting and
|pi |2 ≤ P0 , ∀i ∈ J , (29e) convex covertness rearranged, P1 can be optimized alternately
between P1,1 and P1,2 , and P2 can be optimized between
Proof: Refer to Appendix C. P2,1 and P2,2 alternately. However, problems P1,1 , P1,2 , P2,1 ,
For fixed Θ, we optimize the aggregation beamforming a and P2,2 are still nonconvex due to the rank one constraint,
of BS and transmission power of device pi from the following even with the convex transforming covertness constraint.
formulation By dropping the rank one constraints, the nonconvex P1,1 ,
P1,2 , P2,1 , and P2,2 problems can be solved by semidefinite
min ∥a∥2 (30a) relaxation (SDR) [35], which can be referred to covert SDR
p, a
s.t. |aH hi |2 ≥ 1, (30b) method. The solution obtained from SDR method fails to
satisfy the rank one constraint, so a suboptimal solution can be
|pi g i |2 |pi g i |2 2ϵ2 obtained by the Gaussian randomization. As the scale of the
ln(1 + 2
)− 2 2
≤ . (30c)
σw |pi g i | + σw L problem grows, especially for a large number of antennas and
|pi |2 ≤ P0 , ∀i ∈ J , (30d) number of reflection elements in our problem, the SDR with
Gaussian randomization fails to meet the rank one constraint
where hi = hH H
rb Θhir + hib , g i = diag(h)rw hir . in high probability. Therefore, we incorporate the covertness
H
With pi = pi g i and P i = pp , we transform (30) into the constraint into the DC programming and propose a novel
following problems P2,1 and P2,2 with low-rank matrix: covert DC scheme for handling the rank one constraint and
covertness constraint in the next section.
P2.1 : min T r(A) (31a)
A,P
V. S OLVING THE C ONVERTED O PTIMIZATION P ROBLEM
s.t. T r(AHi ) ≥ 1, ∀i ∈ I [j] , (31b)
In this section, we design a covert alternating DC scheme
T r(P i ) T r(P i ) that incorporates the rank one constraint and covertness
(1 + 2
)ln(1 + 2
)
σw σw constraint to solve P1,1 , P1,2 , P2,1 , and P2,2 and obtain the
2ϵ2 T r(P i ) 2ϵ2 solution of problem P for the IRS-aided covert OAC-based
− (1 + ) 2
≤ , ∀i ∈ I [j] (31c)
L σw L FL system.
A ⪰ 0, rank(A) = 1, (31d)
A. DC Forumulation and Scheme
P i ≤ P0 Gi , P ⪰ 0, rank(P ) = 1, ∀i ∈ I [j] . It is very important to detect the rank-one constraint in
(31e) nonconvex quadratic constraint for our covert OAC-based FL
On the other hand, given the transmission power of device framework with device selection. The rank one constraint
I
pi and the aggregation beamforming vector a, we have for matrix A ∈ Ca×a is rewritten as ||σi (A)i=1 ||0 = 1,
where σi (A) represents the largest singular value in matrix
Solve v (32a) A. In addition, if we
PAdenote the trace norm and the spectral
H 2
s.t.|b v + ci | ≥ 1, ∀i ∈ I , [j]
(32b) norm as T r(A) = i=1 σi (A) and ||A||2 = σ(A), then we
H 2 H 2 obtain rank(A) = 1 ⇔ T r(A)−||A||2 = 0 with T r(A) ≥ 0.
|g v + di | |g v + di |
(1 + )ln(1 + ) Thus, we handle rank one constraint with DC penalty. The
2
σw σw2 ′
DC expression of problem P1,1 can be described as
2ϵ2 |g H v + di |2 2ϵ2
− (1 + ) 2
≤ , ∀i ∈ I [j] (32c) ′
P1,1 : min ||s||1 + α(T r(A) − ||A||2 )
L σw L s,A,p,U
|vn | ≤ 1, ∀n ∈ 1, . . . , N . (32d) + β(T r(U ) − ||U ||2 ) (34a)
Furthermore, the low-rank matrix problem of (32) can be s.t. T r(A) − γT r(AH i ) ≤ si , ∀i ∈ I, (34b)
obtained by T r(U Gi ) T r(U Gi )
(1 + 2
)ln(1 + 2
)
σw σw
P2.2 : Solve V (33a) 2ϵ2 T r(U Gi ) 2ϵ2
[j] − (1 + ) ≤ , ∀i ∈ I (34c)
s.t.T r(S i V ) ≥ 1, ∀i ∈ I , (33b) L σw2 L
M ⪰ 0, T r(M ) ≥ 1, U ⪰ 0, (34d) design an efficient algorithm by linearizing successively with

′
2
|pi | ≤ P0 , U n,n ≤ pi , subgradient [36]. Thus, the objective functions of problem P1,1
′
and problem P1,2 are represented as f1 − h1 and f2 − h2 ,
∀n = 1, 2, . . . , N, ∀i ∈ I (34e)
respectively, where
where α ≥ 0 and β ≥ 0 represent the penalty factors.
Therefore, we can satisfy the rank one matrix constraint if f1 = ||s||1 + λT r(A) + µT r(U ), h1 = λ||A||2 + µ||U ||2 ,
the DC penalty part is equal to zero. Furthermore, we can (38)
compute the feasible aggregation beamforming a by Cholesky f2 = T r(V ), h2 = ||V ||2 . (39)
decomposition with a matrix A∗ = aaH . Then, we ensure the
′ The concave part −h1 of the objective function in the problem
feasibility of problem P1,2 by minimizing the DC expression, ′
P1,1 in (34) is linearized as follows
which is described as follows
′′
′
P1.2 : min T r(V ) − ||V ||2 (35a) P1,1 : min f1 − < ∂A[k−1] h1 , A > − < ∂U [k−1] h1 , U >
s,A,p,U
V
(40a)
s.t. ∥a∥2 − γT r(S i V ) ≤ si , ∀i ∈ I, (35b)
2
p T r(Gi V ) 2
p T r(Gi V ) s.t. T r(A) − γT r(AH i ) ≤ si , ∀i ∈ I, (40b)
(1 + i 2
)ln(1 + i 2
) T r(U Gi ) T r(U Gi )
σw σw (1 + )ln(1 + )
2
σw 2
σw
2ϵ2 p2i T r(Gi V ) 2ϵ2
−(1 + ) 2
≤ , ∀i ∈ I, 2ϵ2 T r(U Gi ) 2ϵ2
L σw L − (1 + ) ≤ , ∀i ∈ I (40c)
L σw2 L
(35c)
V n,n ≤ 1, ∀n ∈ 1, . . . , N + 1; V ⪰ 0. A ⪰ 0, T r(A) ≥ 1, (40d)
2
(35d) U ⪰ 0, |pi | ≤ P0 , U n,n ≤ pi ,
∀n = 1, 2, . . . N + 1, ∀i ∈ I (40e)
If the objective value of P1,2 turns into zero, we can also
compute a feasible solution for rank one extract v = [v0 , t0 ]T where < X, Y >= ℜ[T r(X H Y )] represents the inner
by Cholesky decomposition V ∗ = vv H . With v = v 0 /t0 , product between one matrix and another matrix, and ∂M [k−1] h
we further obtain the reflecting beamforming matrix of IRS represents the subgradient of function h in X for iteration
according to Θ = diag(v). ′
k − 1. Likewise, we can solve problem P1,2 by iterative
′ ′
Similarly, we formulate the problems P2,1 and P2,2 for optimization
problem P2 with CDC framework to ensure the feasible rank
′′
′
one constraint and covertness constraint. Problem P2,1 can be P1.2 : min f2 − < ∂V [k−1] h2 , V > (41a)
V
expressed as s.t. ∥a∥2 − γT r(S i V ) ≤ xi , ∀i ∈ I, (41b)
′ 2 2
P2.1 min T r(A) + α(T r(A) − ||A||2 ) p T r(Gi V ) p T r(Gi V )
A,U ,p (1 + i 2
)ln(1 + i 2
)
σw σw
+ β(T r(P ) − ||P ||2 ) (36a)
2ϵ2 p2i T r(Gi V ) 2ϵ2
s.t. T r(AHi ) ≥ 1, ∀i ∈ I , [j]
(36b) − (1 + ) 2
≤ , ∀i ∈ I (41c)
L σw L
T r(P i ) T r(P i ) V n,n ≤ 1, ∀n ∈ 1, . . . , N + 1; V ⪰ 0. (41d)
(1 + 2
)ln(1 + 2
)
σw σw
2ϵ2 T r(P i ) 2ϵ2 Similarly, we can also use the covert DC programs to address
− (1 + ) 2
≤ , ∀i ∈ I [j] , (36c) ′
problems P2,1 ′
and P2,2 by linearizing successively. Specially,
L σw L ′ ′
we represent the objective functions of P2,1 and P2,2 with
A ⪰ 0, P i ≤ P0 Gi , P ⪰ 0, ∀i ∈ I [j] (36d) f3 − h3 and f4 − h4 as follows
′
and problem P2,2 is described as
f3 = (1 + λ)T r(A) + µT r(P ), h3 = λ||A||2 + µ||P ||2 ,
′
P1.2 : min T r(V ) − ||V ||2 (37a) (42)
V
s.t. ∥a∥2 − γT r(S i V ) ≤ si , ∀i ∈ I, (37b) f4 = T r(V ), h4 = ||V ||2 . (43)
p2i T r(Gi V ) p2i T r(Gi V ) We can obtain A[t] for P2,1
′
by solving
(1 + 2
)ln(1 + 2
)
σw σw
′′
2ϵ2 p2i T r(Gi V ) 2ϵ2 P2.1 : min f3 − < ∂A[k−1] h3 , A > − < ∂P [k−1] h3 , P >
A,P ,w
−(1 + ) 2
≤ , ∀i ∈ I,
L σw L (44a)
(37c) s.t. T r(AHi ) ≥ 1, ∀i ∈ I , [j]
(44b)
V n,n ≤ 1, ∀n ∈ 1, . . . , N + 1; V ⪰ 0. (37d) T r(P i ) T r(P i )
(1 + 2
)ln(1 + 2
)
σw σw
B. Our Proposed Algorithm 2ϵ2 T r(P i )
− (1 + )
The CDC programming is still non-convex, but we can L σw2
utilize the different structures of the convex function to (44c)
A ⪰ 0, P i ≤ P0 Gi , Algorithm 1 Alternating Covert DC Approach to Solve

Problem P for FL System
P ⪰ 0.∀i ∈ I [j] (44d)
1: Initialize:
We can obtain V [t] for P2,2
′
by solving Initialize channel gains hib , hir , hrb , hrw , hiw ;
Initialize point Θ0 and predefined threshold ϵ > 0;
′′
P2.2 : min f4 − < ∂V [k−1] h4 , V > (45a) 2: The covert priority of devices selected:
V
[j] Repeat Update sk , ak , Θk , pki ;
s.t. T r(S i V ) ≥ 1, ∀i ∈ I , (45b) Given Θk−1 , obtain solution sk , ak , pi by solving
2 2
w T r(Gi V ) w T r(Gi V ) problem P1,1 ;
(1 + i )ln(1 + i )
2
σw 2
σw Given sk , ak , pi , obtain solution Θk by solving problem
2ϵ2 wi2 T r(Gi V ) 2ϵ2 P1,2 ;
− (1 + ) 2
≤ , ∀i ∈ I [j] Until the objective result of problem P1 is below ϵ, we can
L σw L
(45c) obtain s∗ = sk .
3: Maximum covert device set for minimized MSE:
|V n,n | ≤ 1, ∀n ∈ 1, . . . , N + 1; V ⪰ 0. (45d)
Users set I [J] = {π(1), π(2), . . . , π(J)} can be computed
Therefore, we have ∂A h1 = ∂A h3 = λ∂||A||2 , ∂U h1 = by sorting s∗ in an ascending sequence with sπ(1) ≤
µ||U ||2 , ∂P h3 = µ||P ||2 and ∂V h2 = ∂V h4 = ∂||V ||2 , sπ(2) ≤ . . . ≤ sπ(J) , and update feasible set of devices
where ∂||A||2 is efficiently calculated by a1 aH 1 [7] and j = J with bisection method L = 0, H = J;
a1 denotes the eigenvector in line with the largest eigenvalue Repeat Update ak , Θk , pi ;
of the matrix A. Therefore, we can solve efficiently the above Given Θk−1 , compute result ak , pi by solving problem
convex problems (40), (41), (44) and (45) by utilizing CVX P2,1 ;
tool [37]. If max M SE ≤ γ in P2 , then L = j, j = ⌊ L+H 2 ⌋, â =
The convergence to the critical value of the DC program- ak , pˆi = pki , break; else if max M SE ≤ ϵ, then H = j,
ming has been proved in [36]. The computational complexity j = ⌊ L+H2 ⌋, where ⌊·⌋ rounds down the integer number,
of our proposed covert DC algorithm is larger than that of the break;
covert SDR method which can be considered as the lower Given ak , pi , compute result Θk by solving problem P2,2 ;
bound [38]. By simply removing the rank one constraints, Until the objective value of problem P2 is below ϵ
we consider the complexity of the covert SDR method as and M SE ≤ γ, we can obtain a∗ = âk , Θ∗ =
shown in i.e, problems P1,1 ′′ ′′
and P2,1 have the complexity of Θk , p∗i = p̂ki and the set of selected users I [j] =
2 3.5 ′′
O((A +I) ) and problem P1,2 and P2,2 ′′
have the complexity {I(1), I(2), . . . , I(j ∗ )}.
2 3.5
of O((N + I) ). With the more number of iterations K,
′′ ′′
P1,1 and P2,1 have the complexity of O(K(A2 + I)3.5 ) at the
′′ ′′
lowest, P1,2 and P2,2 have the complexity of O(K(N 2 +I)3.5 )
at lowest. Yet, the high cost of the computation complexity
Baseline schemes: the following schemes are compared with
leads to prominent performance gains of system, which is
our approach:
shown by the numerical results presented in Section VI.
• Alternating DC covert with IRS: Our proposed approach
VI. N UMERICAL R ESULTS is to jointly optimize the transmit power of the device,
device selection, the amplitudes and phases of the IRS
A. Evaluation Scenario and Methodology and the beamforming of the BS for the IRS-assisted
Simulation setup: In the section, we present simulation covert OAC problem of the FL system.
results to evaluate the performance of our proposed covert DC • Alternating SDR Covert with IRS: In the method, the two-
scheme for IRS-assisted covert OAC-based FL with device step covert SDR method is exploited to tackle the covert
selection. To demonstrate the effectiveness of IRS-assisted OAC problem of the FL system.
covert OAC in the FL system, we consider the 3D coordinate • Maximizing SR on MSE with IRS: The iterative SDR
system, where the BS, the IRS and the Willie are located algorithm by approximating secrecy rate formulation with
at (0, 25, 0), (50, 60, 0) and (0,0,0) m, respectively. The Taylor polynomial [40] is to maximize the secrecy rate
users are randomly located in the square area of ([40, 120], while satisfying MSE requirement with the IRS in OAC-
[-40, 40], 0) m. Similar to [13] and [39], we consider path based FL system.
loss to be the large-scale fading for the wireless channel • Minimizing MSE on SR with IRS: The iterative SDR
model. The path loss factors of the user-to-IRS, the user-to- algorithms by approximating secrecy rate formulation
BS, the IRS-to-BS, the user-to-Willie and the IRS-to-Willie with Taylor polynomial is to minimize the MSE to
are αur = 2.2, αub = 3.5, αrb = 2.3, αuw = 3.5 and achieve a predefined one-bit [41] secrecy rate requirement
αrw = 2.3, respectively. We assume that the small-scale fading at the BS in OAC-based FL system.
follows a Rayleigh distribution the wireless channel model. • Minimizing MSE at BS with IRS: The heuristic IRS
Other system parameters are as follows: P0 = 30 dBm and passive beamforming with SDR method is designed to
σ 2 = σw 2
= −60 dBm, ϵ = 10−3 , N = 64, M = 20, and minimize the MSE at the BS for a feasible number of
K = 20. users satisfying covert constraints.
Fig. 5. The average selected devices varying with MSE requirment under
the two covertness levels (a) ϵ = 0.1 and (b) ϵ = 0.01.
Fig. 3. Convergence behaviour of the covert priority of devices selected

problem under different ϵ. levels with and without IRS. We observe that the number
of selected users reduces as the covertness level ϵ decreases
for all algorithms. This is because decreasing ϵ leads to a
stricter covertness constraint. As expected, the schemes with
IRS can select more devices than those without IRS, clearly
demonstrating the benefit of integrating IRS into the covert FL
system. We also observe that the number of devices selected
by all algorithms increases with an increase in the MSE
threshold γ. The reason is that a lower requirement for the
OAC aggregation error allows more devices to participate
in the FL training. Compared to schemes without IRS, the
IRS-aided covert OAC in FL can accommodate more devices
Fig. 4. Convergence behaviour of maximum covert device set for minimized while meeting the covertness level for parameters aggregation
MSE problem under different ϵ. of OAC under the requirement of MSE. In addition, it is
essential to jointly optimally design the transmission power
• Minimizing RSP at Willie with IRS: The heuristic IRS of the device, the selected devices, the amplitudes and phase
passive beamforming with SDR method is designed to of IRS and the beamforming of BS, while, without these
minimize the received signal power (RSP) at Willie for optimizations, the random phase of IRS is less effective
a feasible number of users satisfying MSE requirment. than the covert DC and covert SDR schemes. The number
• Random phase shifts: In the method, the phase of of devices selected by maximizing SR on MSE with IRS
IRS with the proposed covert DC scheme is generated is lower than the proposed covert DC scheme under the
following the uniform and independent distribution. covertness levels in Figure 5a. However, we can observe that
• Without IRS: In the method without IRS, the proposed maximizing SR on MSE with IRS can achieve the largest
covert DC scheme is used to solve the covert OAC amount of devices selected because covert communications
problem by setting Θ = 0. adopts a more restrictive requirement than the secrecy rate
Evaluation criteria: We use three metrics to evaluate under the covertness levels ϵ = 0.1 in Figure 5b. The reason
our proposed method: (1) Convergence behavior; (2) device is that the method by approximating this secrecy rate with
selection under different levels of covertness; and (3) Taylor series expansion leads to performance loss than the
performance comparison of the covert federated learning. DC scheme under the covertness levels ϵ = 0.1, while the
number of devices selected by maximizing SR on MSE is
B. Convergence Behaviour With Covertness Constraint larger than the alternating DC Covert with IRS due to the more
restrictive constraint for covert communications under the
To identify the validity of our proposed covert DC scheme,
covertness levels ϵ = 0.01. Minimizing MSE on SR with IRS
we first demonstrate the convergence of the first step of the
ensures a higher requirement with one-bit SR so the devices
covert priority of devices selected for equations (40) and (41)
selected are fewer than maximizing SR on MSE with IRS.
and the second step of selecting the maximum covert device
However, minimizing MSE on SR with IRS still results in the
set while minimizing MSE for equations (44) and (45) under
performance loss compared to alternating DC covert with IRS.
the different level of covertness in Figure 3 and Figure 4,
Nonetheless, minimizing MSE on SR with IRS selects more
respectively. It can be observed that both steps of the covert
users than the alternating SDR Covert with IRS due to the
DC scheme converge by solving effective solutions of rank-
more stringent covert aggregation scheme. Based on the above
one condition with the effective convex approximation of
observation, the relationship between secrecy rate requirement
covertness constraint.
and the covertness level constraint for OAC-based FL system
needs further investigation. Furthermore, minimizing MSE at
C. Device Selection With Covertness Constraint BS with IRS scheme can ensure the covert constraint but it
In Figure 5, we evaluate the devices selected by different minimizes the MSE for feasible users so that users can not
methods versus the MSE requirement γ for various covertness be selected properly. And minimizing RSP at Willie with IRS
Fig. 8. (a) The channel power received by Willie varying with the number of
Fig. 6. The average selected devices varying with the number of BS’s reflection elements of IRS with ϵ = 0.1 and (b) Average number of selected
antennas under two covertness levels (a) ϵ = 0.1 and (b) ϵ = 0.01. devices varying with IRS horizontal locations with ϵ = 0.01.
level becomes strict, the more reflecting elements can enable

more devices to participate in the training process than the
scheme without IRS while satisfying the MSE and covertness
requirements. Since the random phase schemes cannot make
full use of the advantage of IRS and the covert SDR scheme
cannot deal with effectively the rank-one constraint of our
problem, these methods lead to the larger gap between the
random scheme, the covert SDR scheme and the proposed
covert DC scheme with more reflecting elements of IRS.
Fig. 7. The average selected devices varying with the number of reflection
elements of IRS under two covertness levels (a) ϵ = 0.1 and (b) ϵ = 0.01.
And the gap between maximizing SR on MSE with IRS and
minimizing MSE on SR with IRS is also larger than other
schemes.
scheme can ensure MSE requirement but it minimizes RSP In Figure 8a, we plot the channel power gain received for
at Willie which is more stringent than the KL divergence Willie i.e.,|hH 2
rw Θhir + hiw | , versus the number of reflection
so that few users can be selected than the minimizing MSE elements, from which we explain the basic reason that the IRS
at BS scheme. Our proposed covert DC scheme outperforms promotes the performance of covert OAC. Since we employ
the covert SDR scheme because of the transformation of the D(P0 ∥P1 ) ≤ 2ϵ2 to model the covert limitation, the D(P0 ∥P1 )
rank-one constraint together with the convex approximation of increases with the received power of Willie wi2 |hH rw Θhir +
covertness constraint. hiw |2 . From Figure 8a, we solved that the schemes with IRS
In Figure 6, we plot the number of devices selected by can achieve lower received power for Willie than that of the
different methods by varying the number of antennas for BS method without IRS. The lower transmission power of the
for different covertness levels. We also observe that fewer device can enhance the performance of covert communication
devices are selected with a smaller covertness level ϵ for all with the aid of the IRS. This observation demonstrates that the
algorithms, but this is alleviated by increasing the number proposed method with IRS not only improves the aggregate
of antennas at the BS. This is due to the fact that more performance at the BS, but also reduces the probability of
antennas at the BS can strengthen the aggregation signal and interception at the Willie. The proposed DC covert scheme can
reduce the gathering error for alternating DC covert with IRS, achieve a lower received power at Willie than the maximizing
which allows the devices to reduce their transmission power SR on MSE with IRS scheme and the Minimizing MSE on
and satisfy the covertness constraint. Therefore, for a certain SR with IRS. In addition, the minimizing RSP at Willie with
MSE requirement and covertness level, we can schedule more IRS can be lower than the other schemes due to the direct
devices to participate in covert OAC for the FL system. improvement of covert communication performance.
In addition, this figure shows that the scheme without IRS In Figure 8b, we plot the number of selected devices
still has a more significant gap with the schemes of IRS even versus the horizontal location of IRS under covertness level
when the antennas of BS are more than three times larger. ϵ = 0.01 in the worse condition that Willie is near the BS
This observation indicates that IRS-assisted covert FL can while locating at the other side of BS and IRS. Similarly, our
improve system performance at a lower cost, for example, proposed covert DC scheme with IRS outperforms the other
by using fewer antennas for the BS. We also observed similar schemes except for maximizing SR on MSE with IRS for
performance variations for minimizing MSE on SR with IRS different IRS locations due to the more restrictive constraint
and minimizing MSE on SR with IRS under corresponding for covert communications than secrecy rate, which illustrates
covertness levels, as shown in Figure 5, Figure 6 and Figure 7. the advantage of the proposed covert DC scheme. More
In Figure 7, we plot the devices selected by different importantly, we observe that more users are selected with IRS
methods varying with the reflecting elements under two deploying near BS and near users, but the better horizontal
covertness levels. With more reflecting elements, the IRS can location of IRS should be located near the BS in our simulation
perform more effective passive beamforming to weaken the setup. Due to the fact that Willie usually keeps close to the
signal at Willie and to mitigate the aggregated model error of BS to detect the OAC aggregation, it is challenging to satisfy
OAC at the BS in the FL system. Even when the covertness the covertness constraint as IRS moves closer to Willie [19].
to participate in covert OAC-based FL under the MSE

requirement, we have proposed a covert DC algorithm that
jointly optimizes the phase shifts and reflecting amplitudes at
the IRS, the beamforming of the BS, and the transmission
power of device while satisfying low-rank optimization and
covertness constraint. Simulation results illustrated that our
proposed approach outperforms other methods and methods
without IRS by achieving higher prediction accuracy and
faster convergence rate with secure model aggregation. In the
future, we can investigate robust covert OAC-based WFL with
Fig. 9. Federated learning performance with (a) Training loss and (b) Test
accuracy under covertness level ϵ = 0.1.
imperfect CSI or statistical CSI to address more practical
situations.
However, when the Willie is located in the other side of BS and A PPENDIX A
IRS, we can observe that deploying the IRS closer to the BS is P ROOF OF P ROPOSITION 1
better than to the users. Therefore, we find that the deployment
of IRS should be properly designed, and the greater passive The aggregated model performance of OAC is mea-
beamforming gain can be obtained by deploying IRS near sured by the MSE min M SE(x̂, x) = E[|x̂ − x|2 ] =
H σ 2 ∥a∥2
√1 H 2
P
the BS, especially for the worse condition that the Willie i∈I | β a (hrb Θhir + hib )pi − 1| + β . The trans-
is near the BS for covert OAC-based FL. We can conclude mitter scalar p i has the zero-forcing structure to enforce
H
√1 aH (hrb Θhir + hib )pi − 1|2
P
that the optimal location of the IRS should strike a tradeoff i∈I | η = 0. We also
between the MSE of OAC from users to BS and the covertness have the form of zero-forcing transmitter scalar p∗i =
√ |aH (hH Θhir +hib )|H
constraint depending on the location of Willie, which would be β |aH (hrbH Θh +h )|2 .
ir ib
Since |pi |2 ≤ P0 , then β =
rb
a noteworthy topic for future. Figure 8b shows the maximizing P0 mini∈I |aH (hH 2
rb Θhir + hib )| .
SR on MSE with IRS and minimizing MSE on SR with IRS The optimal covertness OAC condition is given as
have similar performance varying with the location of IRS H √ |aH (hH Θhir +hib )|H
p∗2 2 ∗
i |hrw Θhir +hiw | = 0. since pi = β |aH (hrb
H Θh +h )|2
with Figure 6-8. However, the comparison between Figure 8b ir ib rb
and Figure 6-8 shows the secrecy rate OAC is more sensitivity and β = P0 mini∈I |aH (hH 2
rb Θhir + hib )| , we have the
σ 2 ∥a∥2
than the covert OAC scheme with respect to the location of minimum M SE(x̂, x) = β , which can be written as
IRS. σ2 ∥a∥2
M SE(x̂, x) = P0 maxi∈I |aH (hH 2. Since the p∗i is
rb Θhir +hib )|
D. Federated Learning Performance of Our Proposed not equal to zero, we only focus on the equation |hH rw Θhir +
Approach hiw |2 = 0, which is rewritten as |v H g + hiw |2 = 0, where
v = [v1 , v2 , . . . , vN ]T and vn P
= ρn ejθn , ∀n. Furthermore,
In order to illustrate the effectiveness of our proposed covert N
H 2
it follows that |v g + hiw | = | n=1 ρn |gn |ej(arg(gn +θn )) +
DC scheme with devices selection in distributed OAC-based
|hiw |ejarg(hiw ) |2 = 0. The sum of real and imaginary parts
FL, the CIFAR-10 dataset is employed to train a classifier
for the first term and second termPshould be equal to zero.
using support vector machine (SVM) [37]. In this simulation, N
θ can ensure the opposite of the n=1 ρn |gn |ej(arg(gn +θn ))
the size of training sets is 50000 and the size of the test sets is N
and |hiw |ejarg(hiw ) , and when
P
10000. Furthermore, the local data sets of devices are followed PN n=1 |gn | ≥ |hiw | and
uniform and non-independent and identically distributed (non- ρn ∈ [0, 1], we can obtain n=1 ρn |gn | = |haw |. Thus the
IID). After running 20 times realization, the average results reflecting amplitude ρ and phase shift θn always exist to enable
of training loss and test accuracy for different schemes are D(P0 ∥P1 ) = 0.
demonstrated in Figure 9. The training loss obtained from loss
function with the input-output data pair shows the convergence A PPENDIX B
of various user selection schemes. The prediction accuracy P ROOF OF P ROPOSITION 2
is given by the ratio of the number of correct predictions The first term of (23) is reexpressed with
to the size of test set. We observe that the proposed covert
T r(QU ) T r(QU )
DC scheme can achieve the lowest loss and highest accuracy f (U ) = (1 + )ln(1 + ). (46)
2
σw 2
σw
compared to the other methods since it selects much more
devices, enabling larger data sets to be used in FL while The affine transformation xln(x) can be a convex function
keeping the covert OAC in fewer communication rounds. of x when x ≥ 0 [32], so f (U ) is convex for U with
VII. C ONCLUSION Q ⪰ 0 and U ⪰ 0. In addition, the second term of (23)
is a linear function of U .
In this paper, we have presented a novel IRS-aided covert
OAC framework for achieving fast and highly secure model
A PPENDIX C
aggregation in FL systems. Our approach relies on the
P ROOF OF P ROPOSITION 3
condition that the channel gain of the reflected wireless
link should be greater than that of the direct wireless link, We assume that θ = min |aH (hH 2
rb Θhir + hib )| , then
i
which we have proven. To enable a larger number of devices the problem P2 can represent equivalently the following
expressions: [18] X. Lu, E. Hossain, T. Shafique, S. Feng, H. Jiang, and D. Niyato,

“Intelligent reflecting surface enabled covert communications in wireless
∥a∥2 networks,” IEEE Netw., vol. 34, no. 5, pp. 148–155, Sep. 2020.
min [19] X. Zhou, S. Yan, Q. Wu, F. Shu, and D. W. K. Ng, “Intelligent
p, a,Θ θ reflecting surface (IRS)-aided covert wireless communications with
s.t. (14c), (14e), (14f ). (47) delay constraint,” IEEE Trans. Wireless Commun., vol. 21, no. 1,
pp. 532–547, Jan. 2022.
Furthermore, we can introduce a new optimization variable [20] C. Wang, Z. Li, J. Shi, and D. W. K. Ng, “Intelligent reflecting surface-
assisted multi-antenna covert communications: Joint active and passive
â = √aθ , so the problem (47) can be rewritten as beamforming optimization,” IEEE Trans. Commun., vol. 69, no. 6,
pp. 3984–4000, Jun. 2021.
min ∥â∥2 [21] J. Si et al., “Covert transmission assisted by intelligent reflecting
p, â,Θ
surface,” IEEE Trans. Commun., vol. 69, no. 8, pp. 5394–5408,
s.t. |aH (hH 2
rb Θhir + hib )| ≥ θ, ∀i ∈ J
Aug. 2021.
[22] Y. Koda, K. Yamamoto, T. Nishio, and M. Morikura, “Differentially
s.t. (14c), (14e), (14f ). (48) private aircomp federated learning with power adaptation harnessing
receiver noise,” in Proc. GLOBECOM IEEE Global Commun. Conf.,
The above problem (48) is the same as problem (28). Dec. 2020, pp. 1–6.
[23] B. Hasircioglu and D. Gunduz, “Private wireless federated learn-
ing with anonymous over-the-air computation,” in Proc. IEEE
R EFERENCES Int. Conf. Acoust., Speech Signal Process. (ICASSP), Jun. 2021,
pp. 5195–5199.
[1] S. Niknam, H. S. Dhillon, and J. H. Reed, “Federated learning for [24] S. Yan, X. Zhou, D. W. K. Ng, J. Yuan, and N. Al-Dhahir, “Intelligent
wireless communications: Motivation, opportunities, and challenges,” reflecting surface for wireless communication security and privacy,”
IEEE Commun. Mag., vol. 58, no. 6, pp. 46–51, Jun. 2020. 2021, arXiv:2103.16696.
[2] P. S. Bouzinis, P. D. Diamantoulakis, and G. K. Karagiannidis, “Wireless [25] L. Lv, Q. Wu, Z. Li, Z. Ding, N. Al-Dhahir, and J. Chen, “Covert
federated learning (WFL) for 6G networks—Part I: Research challenges communication in intelligent reflecting surface-assisted NOMA systems:
and future trends,” IEEE Commun. Lett., vol. 26, no. 1, pp. 3–7, Design, analysis, and optimization,” IEEE Trans. Wireless Commun.,
Jan. 2022. vol. 21, no. 3, pp. 1735–1750, Mar. 2022.
[3] W. Y. B. Lim et al., “Federated learning in mobile edge networks: [26] P. Zeng, D. Qiao, H. Qian, and Q. Wu, “Joint beamforming design
A comprehensive survey,” IEEE Commun. Surveys Tuts., vol. 22, no. 3, for IRS aided multiuser MIMO with imperfect CSI,” IEEE Trans. Veh.
pp. 2031–2063, 3rd Quart., 2020. Technol., vol. 71, no. 10, pp. 10729–10743, Oct. 2022.
[4] W. Liu, X. Zang, Y. Li, and B. Vucetic, “Over-the-air computation [27] X. Chen, T.-X. Zheng, L. Dong, M. Lin, and J. Yuan, “Enhancing MIMO
systems: Optimization, analysis and scaling laws,” IEEE Trans. Wireless covert communications via intelligent reflecting surface,” IEEE Wireless
Commun., vol. 19, no. 8, pp. 5488–5502, Aug. 2020. Commun. Lett., vol. 11, no. 1, pp. 33–37, Jan. 2022.
[5] X. Cao, G. Zhu, J. Xu, and S. Cui, “Transmission power control for [28] Z. Wang, L. Liu, and S. Cui, “Channel estimation for intelligent
over-the-air federated averaging at network edge,” IEEE J. Sel. Areas reflecting surface assisted multiuser communications: Framework,
Commun., vol. 40, no. 5, pp. 1571–1586, May 2022. algorithms, and analysis,” IEEE Trans. Wireless Commun., vol. 19,
[6] G. Zhu and K. Huang, “MIMO over-the-air computation for high- no. 10, pp. 6607–6620, Oct. 2020.
mobility multimodal sensing,” IEEE Internet Things J., vol. 6, no. 4, [29] A. L. Swindlehurst, G. Zhou, R. Liu, C. Pan, and M. Li,
pp. 6089–6103, Aug. 2019. “Channel estimation with reconfigurable intelligent surfaces—A gen-
[7] K. Yang, T. Jiang, Y. Shi, and Z. Ding, “Federated learning via over- eral framework,” Proc. IEEE, vol. 110, no. 9, pp. 1312–1338,
the-air computation,” IEEE Trans. Wireless Commun., vol. 19, no. 3, Sep. 2022.
pp. 2022–2035, Mar. 2020. [30] B. A. Bash, D. Goeckel, and D. Towsley, “Limits of reliable
[8] M. Frey, I. Bjelakovic, and S. Stanczak, “Towards secure over-the-air communication with low probability of detection on AWGN channels,”
computation,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Jul. 2021, IEEE J. Sel. Areas Commun., vol. 31, no. 9, pp. 1921–1930,
pp. 700–705. Sep. 2013.
[9] D. Wang, B. Bai, W. Zhao, and Z. Han, “A survey of optimization [31] S. Yan, B. He, X. Zhou, Y. Cong, and A. L. Swindlehurst, “Delay-
approaches for wireless physical layer security,” IEEE Commun. Surveys intolerant covert communications with either fixed or random transmit
Tuts., vol. 21, no. 2, pp. 1878–1911, 2nd Quart., 2019. power,” IEEE Trans. Inf. Forensics Security, vol. 14, no. 1, pp. 129–140,
[10] A. Ali and W. Hamouda, “Advances on spectrum sensing for cognitive Jan. 2019.
radio networks: Theory and applications,” IEEE Commun. Surveys Tuts., [32] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
vol. 19, no. 2, pp. 1277–1304, 2nd Quart., 2017. Cambridge Univ. Press, 2004.
[11] K. S. K. Arumugam and M. R. Bloch, “Covert communication over [33] Y. Shi, J. Cheng, J. Zhang, B. Bai, W. Chen, and K. B. Letaief,
a K-user multiple-access channel,” IEEE Trans. Inf. Theory, vol. 65, “Smoothed Lp -minimization for green cloud-RAN with user admission
no. 11, pp. 7020–7044, Nov. 2019. control,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 1022–1036,
[12] Q. Wu and R. Zhang, “Towards smart and reconfigurable environment: Apr. 2016.
Intelligent reflecting surface aided wireless network,” IEEE Commun. [34] Y. Huang and S. Zhang, “Complex matrix decomposition and quadratic
Mag., vol. 58, no. 1, pp. 106–112, Jan. 2020. programming,” Math. Oper. Res., vol. 32, no. 3, pp. 758–768, Aug. 2007.
[13] Z. Wang et al., “Federated learning via intelligent reflecting surface,” [35] Z.-Q. Luo, W.-K. Ma, A. So, Y. Ye, and S. Zhang, “Semidefinite
IEEE Trans. Wireless Commun., vol. 21, no. 2, pp. 808–822, Feb. 2022. relaxation of quadratic optimization problems,” IEEE Signal Process.
[14] X. Zhai, G. Han, Y. Cai, and L. Hanzo, “Beamforming design Mag., vol. 27, no. 3, pp. 20–34, May 2010.
based on two-stage stochastic optimization for RIS-assisted over-the- [36] P. D. Tao and L. T. An, “Convex analysis approach to DC programming:
air computation systems,” IEEE Internet Things J., vol. 9, no. 7, Theory, algorithms and applications,” Acta Mathematica, vol. 22, no. 1,
pp. 5474–5488, Apr. 2022. pp. 287–367, 1997.
[15] W. Ni, Y. Liu, Z. Yang, and H. Tian, “Over-the-air federated learning [37] B. Grant and S. Boyd. (2014). CVX: MATLAB Software for Disci-
and non-orthogonal multiple access unified by reconfigurable intelligent plined Convex Programming, Version 2.1. [Online]. Available: http://
surface,” in Proc. IEEE INFOCOM Conf. Comput. Commun. Workshops cvxr.com/cvx/
(INFOCOM WKSHPS), May 2021, pp. 1–6. [38] M. Cui, G. Zhang, and R. Zhang, “Secure wireless communication via
[16] W. Ni, Y. Liu, Z. Yang, H. Tian, and X. Shen, “Federated learning intelligent reflecting surface,” IEEE Wireless Commun. Lett., vol. 8,
in multi-RIS-aided systems,” IEEE Internet Things J., vol. 9, no. 12, no. 5, pp. 1410–1414, Oct. 2019.
pp. 9608–9624, Jun. 2022. [39] Z. Wang, Y. Shi, Y. Zhou, H. Zhou, and N. Zhang, “Wireless-
[17] K. Yang, Y. Shi, Y. Zhou, Z. Yang, L. Fu, and W. Chen, “Federated powered over-the-air computation in intelligent reflecting surface-aided
machine learning for intelligent IoT via reconfigurable intelligent IoT networks,” IEEE Internet Things J., vol. 8, no. 3, pp. 1585–1598,
surface,” IEEE Netw., vol. 34, no. 5, pp. 16–22, Sep. 2020. Feb. 2021.
[40] K. Cumanan, Z. Ding, B. Sharif, G. Y. Tian, and K. K. Leung, “Secrecy Jiawen Kang received the Ph.D. degree from
rate optimizations for a MIMO secrecy channel with a multiple-antenna the Guangdong University of Technology, China,
eavesdropper,” IEEE Trans. Veh. Technol., vol. 63, no. 4, pp. 1678–1690, in 2018. He was a Post-Doctoral Researcher
May 2014. with Nanyang Technological University, Singapore,
[41] G. Zhu, Y. Du, D. Gunduz, and K. Huang, “One-bit over-the-air from 2018 to 2021. He is currently a Full Professor
aggregation for communication-efficient federated edge learning: Design with the Guangdong University of Technology. His
and convergence analysis,” IEEE Trans. Wireless Commun., vol. 20, research interests include blockchain, security, and
no. 3, pp. 2120–2135, Mar. 2021. privacy protection in wireless communications and
networking.
Ling Gao (Member, IEEE) received the B.S.

degree in computer science from Hunan University
Jie Zheng (Member, IEEE) received the B.Sc. in 1985, the M.S. degree in computer science
degree in communications engineering from Nan- from Northwest University in 1988, and the Ph.D.
chang University in 2008 and the Ph.D. degree from degree in computer science from Xi’an Jiaotong
the Department of Telecommunications Engineering, University, Xi’an, China, in 2005. His research
Xidian University, China, in 2014. He is currently interests include network security and management,
an Associate Professor with the Information Science and embedded internet service. He was the Director
and Technology Institute, Northwest University, of the China Higher Educational Information
Xi’an, China. His research interests include energy- Academy, the Chairperson of the China Computer
efficient transmission, wireless resource allocation, Federation Network and Data Communications
and edge intelligence. Technical Committee, and the CAET Director.
Jie Ren is currently an Assistant Professor with

Shaanxi Normal University, Xi’an, China. He serves
as a PC Member for IJCAI 2021 and SocialCom
2020. He servers as the Session Chair for UIC 2018.
Haijun Zhang (Fellow, IEEE) was a Post-Doctoral
Research Fellow with the Department of Electrical
and Computer Engineering, The University of
British Columbia (UBC), Canada. He is currently
a Full Professor and an Associate Dean of the
University of Science and Technology Beijing,
China. He received the IEEE CSIM Technical
Committee Best Journal Paper Award in 2018, the Dusit Niyato (Fellow, IEEE) received the B.Eng.
IEEE ComSoc Young Author Best Paper Award degree from the King Mongkut’s Institute of Tech-
in 2017, and the IEEE ComSoc Asia-Pacific Best nology Ladkrabang (KMITL), Thailand, in 1999,
Young Researcher Award in 2019. He serves/served and the Ph.D. degree in electrical and computer
as the Track Co-Chair for WCNC 2020, the Symposium Chair for engineering from the University of Manitoba,
GLOBECOM 2019, the TPC Co-Chair for INFOCOM 2018 Workshop on Canada, in 2008. He is currently a Professor with
Integrating Edge Computing, Caching, and Offloading in Next Generation the School of Computer Science and Engineering,
Networks, and the General Co-Chair for GameNets 2016. He serves Nanyang Technological University, Singapore. His
as an Editor for IEEE T RANSACTIONS ON C OMMUNICATIONS, IEEE research interests include the Internet of Things
T RANSACTIONS ON N ETWORK S CIENCE AND E NGINEERING, and IEEE (IoT), machine learning, and incentive mechanism
T RANSACTIONS ON V EHICULAR T ECHNOLOGY. design.

Covert Federated Learning Via Intelligent Reflecting Surfaces

Uploaded by

Copyright:

Available Formats

Covert Federated Learning Via Intelligent Reflecting Surfaces

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Covert Federated Learning Via Intelligent Reflecting Surfaces

Uploaded by

Copyright:

Available Formats

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 71, NO.

8, AUGUST 2023 4591

Covert Federated Learning via Intelligent

F EDERATED learning (FL), as a burgeoning distributed

postprocessing function ẑ = ψ(x̂). To verify the aggregated

k=1 by both the minimizing MSE σ ∥a∥ β and the KL divergence

selected devices to upload the parameters of local models,

s.t. ∥a∥ − γ|aH (hH

Proposition 3: Problem P2 can be rewritten equivalently as wi2 T r(Gi V ) wi2 T r(Gi V )

M ⪰ 0, T r(M ) ≥ 1, U ⪰ 0, (34d) design an efficient algorithm by linearizing successively with

A ⪰ 0, P i ≤ P0 Gi , Algorithm 1 Alternating Covert DC Approach to Solve

Fig. 3. Convergence behaviour of the covert priority of devices selected

level becomes strict, the more reflecting elements can enable

to participate in covert OAC-based FL under the MSE

expressions: [18] X. Lu, E. Hossain, T. Shafique, S. Feng, H. Jiang, and D. Niyato,

Ling Gao (Member, IEEE) received the B.S.

Jie Ren is currently an Assistant Professor with

You might also like