460 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 16, NO.

3, APRIL 2022

Distributed Machine Learning for Multiuser

Mobile Edge Computing Systems
Yinghao Guo , Rui Zhao, Shiwei Lai , Lisheng Fan , Member, IEEE, Xianfu Lei ,
and George K. Karagiannidis , Fellow, IEEE

Abstract—In this paper, we investigate a distributed machine and industry, since it can fulfill the intelligent identification,
learning approach for a multiuser mobile edge computing (MEC) positioning, tracking, etc. To support various applications of IoT,
network in a cognitive eavesdropping environment, where multiple mobile edge computing (MEC) has been proposed to overcome
secondary devices (SDs) have some tasks with different priorities to
be computed. The SDs can be allowed to use the wireless spectrum the limitations of traditional cloud computing [1]–[3]. In the
as long as the interference to the primary user is tolerated, and MEC networks, mobile users can offload the computational
an eavesdropper in the network can overhear the confidential tasks to the nearby computational access points (CAPs), in order
message from the SDs, which threatens the data offloading. For the to meet the requirement of latency and energy consumption.
considered system, we firstly present three optimization criteria, Hence, offloading strategy has become a key design in MEC
whereas criterion I aims to minimize the linear combination of
latency and energy consumption, criterion II tries to minimize the networks [4]–[6]. In general, computational offloading can be di-
latency under a constraint on the energy consumption, and criterion vided into two types of binary offloading and partial offloading.
III is to minimize the energy consumption under a constraint on the In binary offloading, the tasks from the users should be computed
latency. We then exploit a federated learning framework to solve either locally or by the CAPs through offloading [7]. In [8], the
these optimization problems, by optimizing the offloading ratio, authors studied a binary offloading and user attack model, and
bandwidth and computational capability allocation ratio. Simu-
lation results are finally demonstrated to show that the proposed then designed an offloading strategy based on the Lyapunov
method can effectively reduce the system cost in terms of latency optimization framework to protect the user privacy and ensure
and energy consumption, and meanwhile ensure more bandwidth the user experience. In the partial offloading, some parts can be
and computational capability allocated to the user with a higher computed locally, while the rest parts are computed by the CAPs
task priority. through offloading [9]. For massive multiple-input multiple-
Index Terms—Computational offloading, federated learning, output (MIMO) MEC networks, a partial offloading strategy
resource allocation, task priority. can be designed by taking into account the latency constraint,
through a nested algorithm based on convex optimization [10].
I. INTRODUCTION
Moreover, the technique of intelligent reflecting surfaces (IRS)
A. Literature Review technique can be integrated into the MEC networks to enhance
N RECENT years, the emerging paradigm of Internet of the system transmission performance significantly, where the
I Things (IoT) has attracted much attention from academy key system design on the IRS was given in the works [11]–[13].
The development of machine learning algorithms such as deep
Manuscript received July 21, 2021; revised November 23, 2021 and December learning (DL) and deep reinforcement learning (DRL) provides
21, 2021; accepted December 26, 2021. Date of publication January 5, 2022; a novel way to solve the problems in wireless communication
date of current version May 17, 2022. This work was supported in part by NSFC
under Grant 61871139 and Grant 62101145, in part by the Natural Science systems, which has attracted a lot of attention [14]–[20]. For
Foundation of Guangdong Province under Grant 2021A1515011392, and in part example, the authors in [14] proposed a novel framework based
by the Research Program of Guangzhou University under Grant YJ2021003. on deep neural network (DNN) to address the problem of antenna
The work of Xianfu Lei was supported in part by the National Key R&D
Program of China under Grant 2019YFB1803400, in part by the National Natural selection and power allocation in the wireless communication
Science Foundation of China under Grant 61971360, in part by the Fundamental systems. By mapping the hybrid beamforming architecture to
Research Funds for the Central Universities under Grant XJ2021KJZK007, a neural network (NN), [15] presented a novel design scheme
and in part by the open research fund of National Mobile Communications
Research Laboratory, Southeast University under Grant 2021D05. The guest based on the machine learning for multiple beamforming ar-
editor coordinating the review of this manuscript and approving it for publication chitectures, which provides critical guidance in the design of
was Prof. Ping Yang. (Corresponding authors: Lisheng Fan; Xianfu Lei.) intelligent radio networks. The authors in [16] developed two
Yinghao Guo, Rui Zhao, Shiwei Lai, and Lisheng Fan are with the School
of Computer Science, Guangzhou University, Guangzhou 510006, China novel antenna selection methods based on the support vector
(e-mail: [email protected]; [email protected]; shiweilai@ machine and DNN, in order to solve the antenna selection prob-
e.gzhu.edu.cn; [email protected]). lem in full-duplex spatial modulation system. In addition, for
Xianfu Lei is with the Provincial Key Lab of Information Coding and Trans-
mission, Southwest Jiaotong University, Chengdu 610031, China, and also with vehicular fog computing networks, a DRL-based task offloading
National Mobile Communications Research Laboratory, Southeast University, method was proposed by jointly exploiting the task priority,
Nanjing 210096, China (e-mail: [email protected]). mobility of vehicles and service availability [17]. Although the
George K. Karagiannidis is with the Aristotle University of Thessaloniki,
54636 Thessaloniki, Greece (e-mail: [email protected]). aforementioned researches can find a proper offloading strategy
Digital Object Identifier 10.1109/JSTSP.2022.3140660 for the MEC networks, they are centralized optimization in

1932-4553 © 2022 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: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
GUO et al.: DISTRIBUTED MACHINE LEARNING FOR MULTIUSER MOBILE EDGE COMPUTING SYSTEMS 461

essence, which requires to know the full information of all the convergence of the FL. To optimize the system performance
users. This is however harmful to the privacy protection, and constrained by a latency threshold, the authors in [31] proposed
meanwhile causes a high communication overhead. two bandwidth allocation strategies to maximize the number
Some distributed learning algorithms have been proposed of users participating into FL. Therefore, FL provides a new
to solve the optimization problems in wireless networks [21]– inspiration to the computation offloading optimization in the
[23], by taking into account the following advantages of the MEC networks, which becomes the motivation of the work in
distributed learning framework. Firstly, the distributed learning this paper.
allows mobile devices not to upload their full information to It is notable that there still exist some challenges in the
the CAPs, so as to protect the privacy of the mobile devices. federated learning, from the aspects of implementation and com-
Secondly, the distributed learning can help reduce the large munication overhead. One challenge is the limited battery supply
communication overhead in the process of centralized learning. at the mobile devices. Once these mobile devices run out of their
Thirdly, unlike the centralized learning, the distributed learning batteries, they will be dropped from the system operation, which
provides a flexible way to the system design, thanks to its char- may cause a severe performance degradation in the federated
acteristic of multi-module cooperation. Due to these reasons, learning. Another challenge is the communication overhead, as
distributed optimization methods are more attractive in solving mobile devices need to upload their local models for multiple
the optimization problems of the MEC networks. For example, a rounds, in order to reach a high learning performance. However,
fully distributed computation offloading algorithm was devised this may cause a large communication overhead and result in a
to reduce the average system-wide execution cost in [21], and a training bottleneck, particularly when some training networks
distributed branch-and-bound method could be applied to min- consist of millions of parameters, such as convolutional neural
imize the long-term average delay subject to the constraints of networks.
computation resources and power consumption [22]. However,
it is notable that there still exist some challenges in the con-
ventional distributed learning. One challenge is the distribution
characteristics of data, where the data is often assumed to follow B. Contribution
the independent and identical distribution (IID). This however In this paper, we investigate distributed machine learning for a
may not hold in the practical MEC networks, where the task char- multiuser mobile edge computing system, where multiple users
acteristics of mobile devices are varying and non-independently have some computational tasks of different priorities, which can
and identically distributed (nIID). Another challenge is the lack be computed through offloading to multiple CAPs. We consider
of interaction among users, which may cause the whole system a cognitive eavesdropping environment where the users can
to lose robustness and result in a high exploration cost. use the wireless spectrum as long as the interference to the
To solve the problems in the conventional distributed learning, primary user is tolerated, and an eavesdropper in the network
Google researchers proposed a pioneering concept of federated can overhear the confidential message from the users. For the
learning (FL) to break the data island in the machine learn- considered system, we firstly present three optimization criteria,
ing [24]–[27], which puts the research of machine learning whereas criterion I aims to minimize the linear combination of
to a new altitude. In general, the process of FL contains four latency and energy consumption, criterion II tries to minimize
steps. Firstly, each user trains its own data at local, and then the latency with a given threshold of energy consumption, while
uploads the trained network parameters to a central node. Sec- criterion III is to minimize the energy consumption with a given
ondly, according to the federated learning framework, the central latency threshold. We then solve the optimization problem in a
node aggregates parameters of users participating into the FL. distributed machine learning way by optimizing the offloading
After the aggregation, the central node sends the aggregated ratio as well as the allocation of bandwidth and computational
network parameters to each user. The above steps are repeated capability, by taking into account the task priority among users.
until the system performance becomes convergent. Different Specifically, a FL optimization framework is used, where each
from the conventional distributed learning, FL is a multi-user user employs DRL to solve the optimization. Simulation results
cooperative learning without data share, and it allows users are finally demonstrated to show that the proposed method
to interact and have nIID data [28], [29]. Compared with the can effectively reduce the system cost in terms of latency and
centralized learning, the federated learning has been shown to energy consumption, and meanwhile ensure more bandwidth
achieve some advantages in the following three folds. Firstly, and computational capability allocated to the user with a higher
as a typical distributed learning, the federated learning allows task priority.
mobile devices not to upload their full information to the CAPs, The main contributions of this work can be summarized as
so as to protect the privacy of the mobile devices. Secondly, follows,
the federated learning can help reduce the large communica- r This paper studies distributed machine learning for a mul-
tion overhead in the process of centralized learning. Thirdly, tiuser multi-CAP MEC network in cognitive and secure
unlike the centralized learning, the federated learning provides environments, where the task priority of users is taken into
a flexible way to the system design, thanks to its characteristic account.
of multi-module cooperation. Due to these advantages, a lot r We present three optimization criteria for the MEC net-
of researches have been put into the study of FL [30]–[32]. In works, which provide a flexible choice for the system
particular, a new resource allocation framework was proposed design and optimization, by taking into account the re-
for the wireless federated learning in [30], in order to optimize quirement of latency and energy consumption.

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
462 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 16, NO. 3, APRIL 2022

Let Dm = (sm , dm ) denote the characteristics of the task of the

user SDm , where sm is the task length, and dm is the task priority.
In practice, the task characteristics may be dynamically varying
due to many factors, which can be characterized by the uniform
distribution, without loss of generality. Specifically, sm is sub-
ject to the uniform distribution of U (smin , smax )2 , where smin
and smax are the minimum and maximum lengths, respectively.
As the users have limited computational capabilities, the tasks
can not be computed at local completely, given the require-
ment from the perspectives of latency and energy consumption.
Hence, the tasks of users should be partially or fully offloaded to
the CAPs. Specifically, let αm,k ∈ [0, 1] denote the offloading
ratio from the user SDm to CAPk through wireless transmission.
Accordingly, αm,0 part of the task will be computed at local,
while the rest part will be computed by the CAPs. As the users
have different task characteristics, such as task length or priority,
Fig. 1. System model of a multiuser and multi-CAP MEC network in cognitive
the communication and computational resources should be allo-
eavesdropping environments. cated dynamically, in order to enhance the system performance
by reducing the latency and energy consumption.
In the following, we will firstly describe the task priority, and
r We solve the optimization problem through optimizing the
then present the offloading and computational model for the
offloading ratio as well as the allocation of bandwidth and considered secure MEC network.
computational capability in the FL framework, where each
user employs DRL to solve the optimization.
r Simulation results are demonstrated to show the effec- A. Task Priority
tiveness of the proposed method in reducing the system In practice, different tasks may have different priorities in the
latency and energy consumption. In particular, the user latency and energy consumption. For example, the signal control
with a higher task priority is ensured to be allocated more tasks such as car navigation and autopilot, have a higher priority
bandwidth and computational capability. in the latency. For the IoT nodes with limited energy, the energy
resources are quite limited, and computational tasks have a high
C. Structure priority in the energy consumption. Overall, the task priority
plays an important role in the task implementation and system
The rest of this paper is organized as follows. After the performance.
introduction, Section II describes the system communication To support the tasks with a higher priority, the system commu-
and computation models, and Section III provides three criteria nication and computational resources should be allocated more
for the system optimization in the MEC networks. After that, to the tasks, in order to meet the task requirement. In particular,
Section IV presents the FL based distributed algorithm to solve we can allocate the transmission bandwidth and computational
the offloading optimization problem in the MEC networks. capability, by taking into account the task priority. As to the user
Simulation results are demonstrated in Section V to verify the SDm with priority dm , we need to allocate the bandwidth from
effectiveness of the proposed scheme, and finally we conclude the k-th edge node CAPk for the user SDm . A feasible solution
this work in Section VI. of the bandwidth allocation is,
βm,k dm
II. SYSTEM MODEL βm,k = M , (1)
Fig. 1 depicts the system model of a distributed wireless mo- m1 =1 βm1 ,k dm1

bile edge computing network in eavesdropping environments, where βm,k denotes the expected bandwidth allocation ratio
where there are M secondary devices (SDs) within the coverage from CAPk to the user SDm , while βm,k is a weighted version
of K CAPs. The users work in an underlay spectrum-sharing of βm,k by taking into account the task priority. As the training
mode, and they can use the spectrum of the primary user, as of users operates in a distributed way, each user can adjust its
long as their interference to the primary user is below a given own bandwidth only, which may cause the sum of bandwidth
threshold IP . We consider the eavesdropping environments, allocated to all users be larger than the system total bandwidth.
where there is one eavesdropper such as a unmanned aerial
vehicle (UAV), which can overhear the confidential commu-
2 In practice, the task size s may vary due to many factors, such as dynamic
nication from the users1 to the CAPs. The users have some m
application requirement and processes in the user terminal. Without loss of
computational tasks to be computed, which can be assisted by generality, we assume that sm follows the uniform distribution within the
the CAPs with much more powerful computational capabilities. interval of [smin , smax ]. Moreover, when the task size follows some other
distributions, such as normal distribution or exponential distribution, the results
in this work will be changed accordingly, whereas the optimization methods in
1 In this paper, the SDs and users are used interchangeably. this work can be easily extended.

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
GUO et al.: DISTRIBUTED MACHINE LEARNING FOR MULTIUSER MOBILE EDGE COMPUTING SYSTEMS 463

To solve this issue, we use (1) to determine the allocation ratio where hm,U AV ∼ CN (0, ζU AV ) denotes the channel parameter
of the bandwidth for each user. Specifically, for the right-hand 2
of the link from SDm to the UAV, and σU AV is the variance of
side (RHS) of (1), the numerator represents the bandwidth share the AWGN at the UAV. From (5) and (7), the secrecy data rate
which the user SDm with task priority dm expects to be allocated under eavesdropping can be obtained as3 ,
from CAP k, while the denominator is the total bandwidth s
Rm,k = [Rm,k − Rm,e ]+ , (8)
share which M SDs expect to be allocated from CAP k. This
normalized processing is exploited to ensure the total allocation where [X]+ returns X if X is positive, or zero otherwise. From
ratio of the bandwidth equal to unity. (8), we can write the transmission latency of the offloading as,
Similar to the process of allocating the bandwidth ratio, we tran sm αm,k
can allocate the computational capability to the users by taking lm,k = s . (9)
Rm,k
into account the task priority as,
From (3) and (9), we can write the transmission energy con-
γm,k dm
m,k = M
γ , (2) sumption of the offloading as,
m1 =1 γm1 ,k dm1
etran tran
m,k = lm,k Pm ,
where γm,k denotes the expected allocation ratio of the com-
putational capability from CAPk to the user SDm , while γ m,k tran IP
= lm,k . (10)
is a weighted version of γm,k by taking into account the task |gm |2
priority. In a word, by allocating more resources such as the
bandwidth and computational capability, the tasks with a higher C. Computation Model
priority can be implemented with a lower latency and energy Note that the computational task can be executed either at the
consumption, which can help enhance the system performance local device or at the edge nodes. When the αm,0 part of the task
of the considered secure MEC network. Dm is executed locally, the local latency is expressed as,
local sm αm,0
B. Offloading Model lm = , (11)
fm
When the local users cannot calculate the computational tasks where fm denotes the computational capability of the user SDm .
completely by themselves, some or full parts of the task should Accordingly, the local energy consumption is,
be offloaded to the CAPs through wireless channels from the
users to the CAPs. In this paper, the secondary users work in the elocal
m
comp local
= Pm lm , (12)
cognitive environments, and the transmit power of the secondary comp
where Pm is the computational power of the user SDm .
users are limited by the peak interference from the primary user. When the αm,k part of the task Dm is executed by CAPk , the
Let IP denote the peak interference power to the primary user computational latency is,
imposed by the secondary users, and the transmit power of the
comp sm αm,k
user SDm is given by, lm,k = total , (13)
fk γ m,k
IP
Pm = , (3) where fktotal denotes the total computational capability of CAPk .
|gm |2
Moreover, the energy consumption at the CAPk is,
where gm ∼ CN (0, ζp ) denotes the channel parameter of the
link from the user SDm to the primary user. From Pm , we can em,k = Pkcomp lm,k
comp
, (14)
write the data rate of the link from SDm to CAPk as,
  where Pkcomp is the computational power at CAPk .
total  Pm |hm,k |2 Since the task can be computed by the CAPs after the offload-
Rm,k = Wk βm,k log2 1 + , (4)
σk2 ing, the offloading and computational latency from the user SDm
  to CAPk is,
total  IP |hm,k |2
= Wk βm,k log2 1 + 2 , (5) tran comp
σk |gm |2 lm,k = lm,k + lm,k . (15)

where Wktotal is the total wireless bandwidth of CAPk , hm,k ∼ As the processing of offloading and local computation can be
CN (0, ζk ) denotes the channel parameter of the link from the carried in parallel, the latency of computing the task of the user
user SDm to CAPk , and σk2 is the variance of the additive white SDm is,
Gaussian noise (AWGN) at CAPk . When the UAV attacker total
lm local
= max{lm , max{lm,1 , 1m,2 , . . . , lm,K }}. (16)
overhears the confidential communication from the users to the
CAPs, the eavesdropping data rate of the link from the user SDm Accordingly, the total latency of the whole system is,
to the UAV is, Ltotal = max{l1total , l2total , . . . , lM
total
}. (17)
 
total  Pm |hm,U AV |2
Rm,e = Wk βm,k log2 1 + 2 , (6)
σU AV
3 Note that if there is no eavesdropper to overhear the confidential message
  from the users, the data transmission rate from the user SDm to the CAPk will
total  IP |hm,U AV |2 become into Rm,k . As Rm,k s ≤ Rm,k holds, we can find that the existence of
= Wk βm,k log2 1 + 2 , (7)
σU AV |gm |2 eavesdroppers will be harmful to the data transmission rate.

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
464 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 16, NO. 3, APRIL 2022

As shown in the above equation, the system total latency is the equal to 0, minimizing the linear combination is equivalent to
maximal latency among M users, as the users perform the of- minimizing the energy consumption, while when λ is equal to 1,
floading and computing in parallel. In particular, the computing minimizing the linear combination is equivalent to minimizing
is performed among M users in parallel, as each user uses its the latency. Secondly, the linear combination can provide a flex-
individual computational resource, either local computational ible optimization form of the latency and energy consumption
capability or the allocated computational capability from the for the MEC networks. We can increase the value of λ, when
CAPs, to calculate the tasks. Moreover, the offloading is im- the latency becomes dominant in some scenarios, while we can
plemented in parallel among M users, as each user employs an decrease the value of λ, when the energy consumption becomes
orthogonal spectrum frequency for the wireless transmission. dominant. Due to these two reasons, most of existing works,
In further, the energy consumption at the user SDm is given such as [33], use the linear combination of the latency and energy
by, consumption to measure the system cost of the MEC networks.
K
In addition to criterion I, we provide criterion II in the fol-
   lowing to minimize the system latency with a given constraint
etotal
m = elocal
m + em,k + etran
m,k . (18)
on the energy consumption, given by,
k=1

Hence, the total energy consumption of the whole system is, min ΦII = Ltotal , (21a)
m,k ,
{αm,k ,β γm,k }
M
 s.t. C1 : Etotal ≤ Eth , (21b)
Etotal = etotal
m . (19)
K

m=1
C2 : αm,0 + αm,k = 1, (21c)
III. PROBLEM FORMULATION k=1

∀m ∈ M,
As different offloading and bandwidth allocation strategies
yield different latency and energy consumption for the con- M

sidered secure MEC network, we need to optimize the system C3 : βm,k = 1, ∀k ∈ K, (21d)
performance by allocating the system resources. To this end, m=1

we need to firstly formulate the optimization problem. In this M


paper, we provide three optimization criteria for the considered C4 : m,k = 1, ∀k ∈ K,
γ (21e)
secure MEC network. Specifically, criterion I is to minimize a m=1

linear combination of latency and energy consumption, through where Eth is the energy consumption threshold predetermined
adjusting the offloading ratio, bandwidth allocation, and com- by the system, and constraint C1 ensures that the system energy
putational capability allocation, given by, consumption should not exceed the threshold. Criterion II can
be used in the latency-sensitive scenarios, such as the automatic
min ΦI = λLtotal + (1 − λ)Etotal , (20a) driving in Internet of Vehicles and control systems.
m,k ,
{αm,k ,β γm,k }
K
In further, we provide criterion III in the following to minimize
 the system energy consumption with a given constraint on the
s.t. C1 : αm,0 + αm,k = 1, ∀m ∈ M,
k=1
latency, given by,
(20b) min ΦIII = Etotal , (22a)
M m,k ,
{αm,k ,β γm,k }

C2 : βm,k = 1, ∀k ∈ K, (20c) s.t. C1 : Ltotal ≤ Lth , (22b)
m=1
K

M
C3 : m,k = 1, ∀k ∈ K,
γ (20d) C2 : αm,0 + αm,k = 1, (22c)
k=1
m=1
∀m ∈ M,
where λ ∈ [0, 1] is a weight factor, and constraint C1 indicates
that the computational task is fully computed by the local com- M

puting and the CAPs through offloading. Constraint C2 ensures C3 : βm,k = 1, ∀k ∈ K, (22d)
that the bandwidth allocated to the uses should not exceed the m=1

total bandwidth of the CAPs, and constraint C3 guarantees that M


the computational capability allocated to the users should not C4 : m,k = 1, ∀k ∈ K,
γ (22e)
exceed the total computational capability of the CAPs. Note m=1

that the linear combination form in criterion I is feasible to where Lth is the latency threshold predetermined by the system,
measure the system cost in some application scenarios, where the and constraint C1 ensures that the system latency should not
system needs to adjust the importance of the latency and energy exceed the threshold. Criterion III can be used in the energy-
consumption dynamically, such as smart home, UAV networks sensitive scenarios, such as the IoT networks where the nodes
and electric autopilot, from the following two aspects. Firstly, have limited energy.
the latency and energy consumption can be jointly optimized Note that the above three criteria can be applicable to different
by minimizing the linear combination. In particular, when λ is application scenarios with different requirements on the system

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
GUO et al.: DISTRIBUTED MACHINE LEARNING FOR MULTIUSER MOBILE EDGE COMPUTING SYSTEMS 465

performance. Specifically, for the application scenarios such as where j is the index of Am (t) and ξ denotes the step size.
smart home, electric autopilot and UAV networks, criterion I According to the range of j, we can update the offloading
tends to be used, since both the metrics of latency and energy ratio, expected bandwidth and computational capability allo-
consumption are important. On the other hand, for the applica- cation ratio. Specifically, when 0 ≤ j ≤ 2K + 1 holds, the task
tion scenarios such as video transmission and navigation, the offloading ratio is updated, which can be expressed as,
latency plays a much more important role in the system perfor-
mance, and we should use criterion II to design the MEC system. αm,v (t) = αm,v (t − 1) + ajm (t − 1), (26)
In further, for the application scenarios such as energy-aware
where v = [0, 1, 2, . . . , K] indicates the local user or the CAPs.
IoT networks, criterion III should be used for the system design.
When 2K + 2 ≤ j ≤ 4K + 1 holds, the expected bandwidth
Overall, this paper provides a flexible choice for the system
allocation ratio is updated as,
design and optimization of the MEC networks.
βm,k (t) = βm,k (t − 1) + ajm (t − 1). (27)
IV. FL BASED DISTRIBUTED OPTIMIZATION
When 4K + 2 ≤ j ≤ 6K + 1 holds, the expected computa-
In this part, we propose a FL based distributed optimization tional capability allocation ratio is updated as,
scheme, which aims to minimize the system cost by adjusting
the offloading ratio, bandwidth allocation ratio and computa- γm,k (t) = γm,k (t − 1) + ajm (t − 1). (28)
tional capability ratio. Specifically, we will firstly formulate the
task offloading and resource allocation into a Markov decision After defining the action command, we then describe the way
process (MDP) problem, and then design the state and action these actions execute. In general, we expect that the agent can
spaces in the DQN. In further, we employ the FL framework not only exploit the prior experience, but also explore some
to solve the optimization problems in (20)-(22) in a distributed unknown state actions. Therefore, we use the -greedy algorithm
way, which can help reduce the communication overhead and to explore the action Am (t). In particular, when the probability
protect the data privacy. The details of the FL based distributed is , the action is selected randomly; while when the probability
optimization is given as follows. is 1 − , the action with the maximum Q function is selected.
The action selection is modeled by,
A. Markov Decision Process
random, ,
Since the characteristics of computational tasks and the wire- Am (t) = arg max Qm (Sm (t), Am (t); ωm ), 1 − , (29)
less channels from the users to the CAPs are time-varying, we Am

use the MDP to characterize the tasks offloading and resource where 0 ≤  ≤ 1. After selecting the action Am (t), the state
allocation problem. In particular, the MDP can be characterized Sm (t) is changed accordingly.
by a three-tuple {S, A, R}, where S is the state space, A is the 3) Reward Function: After the state space is updated, we
action space, and R is the reward function. The MDP model is compare the cost of the user SDm at the current time slot with
elaborated in detail as follows. that at the previous time slot, in order to obtain a reward function
1) State Space: Let Sm (t) = {C(t), B(t), F(t)} denote the for the agent SDm . In particular, when the cost of the user SDm
state space of the user SDm at time slot t (t = 1, 2, . . . , tmax ), at time slot t − 1 is equal to that at the current time slot t, the
where tmax is the maximum time slot. The state space Sm (t) is agent SDm gets a zero feedback. When the cost of SDm at the
expanded as, current moment is lower than that at the previous time slot, the
⎧
⎨ C(t) = {αm,0 (t), αm,1 (t), . . . , αm,K (t)}, agent SDm gets a positive feedback. Otherwise, the agent SDm
Sm (t) = B(t) = {βm,1 (t), βm,2 (t), . . . , βm,K (t)}, (23) gets a negative feedback.
⎩
F(t) = {γm,1 (t), γm,2 (t), . . . , γm,K (t)}, When the system chooses criterion I as the optimization
objective, the reward function of the user SDm is given by,
where αm,0 (t) denotes the local computational ratio of the user ⎧
SDm at time slot t, αm,k (t) is the task offloading ratio from ⎨ 0, If ΦI (t) = ΦI (t − 1),
I
SDm to CAPk at time slot t, and βm,k (t) and γm,k (t) represent Rm (t) = 1, If ΦI (t) < ΦI (t − 1), (30)
⎩
the expected bandwidth and computational capability allocation −1, If ΦI (t) > ΦI (t − 1),
ratio from CAPk to SDm , respectively.
2) Action Space: A discrete vector Am (t) is used to denote where ΦI (t) is the cost of the user SDm under criterion I at time
the action space of the user SDm at time slot t, and we use the slot t, while ΦI (t − 1) is the cost at time slot t − 1.
index of the action space Am (t) to denote the action command. Similarly, when the system chooses criterion II as the opti-
In general, the state space will change after performing an action, mization objective, the reward function of the user SDm is,
and the action can update only one element of the state space ⎧
⎨ 0, If ΦII (t) = ΦII (t − 1),
once. The action space Am (t) is expressed by, Rm II
(t) = 1, If ΦII (t) < ΦII (t − 1), (31)
⎩
Am (t) = {ajm (t)|0 ≤ j ≤ 6K + 1}, (24) −1, If ΦII (t) > ΦII (t − 1),

j ξ, If j%2 = 0, where ΦII (t) is the cost of the user SDm under criterion II at
am (t) = (25)
−ξ, If j%2 = 1, time slot t, while ΦII (t − 1) is the cost at time slot t − 1.

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
466 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 16, NO. 3, APRIL 2022

Algorithm 1: Proposed FL Based Distributed Optimization

Scheme.
Input: Dm , Wktotal , fktotal
Output: αm,k , αm,0 , βm,k , γ
m,k
1: Initialize experience relay pool G
−
2: Initialize weighted parameters ωm and ωm
3: for Epoch=1 : R do
4: for Episode=1 : Z do
5: for m=1 : M do
6: Initialize state Sm and environment of agent m
Fig. 2. Structure of DQN of the agent SDm . 7: end for
8: for t=1 : tmax do
9: for m=1 : M do
When the system chooses criterion III as the optimization
10: Choose an action Am (t) according to (29)
objective, the reward function of the user SDm is expressed as,
11: Execute the action Am (t) and get reward
⎧
⎨ 0, If ΦIII (t) = ΦIII (t − 1), Rm (t), then observe the next state Sm (t + 1)
III 12: Store transition
Rm (t) = 1, If ΦIII (t) < ΦIII (t − 1), (32)
⎩ (Sm (t), Am (t), Rm (t), Sm (t + 1)) in
−1, If ΦIII (t) > ΦIII (t − 1),
experience relay pool G
where ΦIII (t) is the cost of the user SDm under criterion III at 13: Choose random mini-batch of transitions
time slot t, while ΦIII (t − 1) is the cost at time slot t − 1. (Sm (t), Am (t), Rm (t), Sm (t + 1)) from G
14: Calculate loss function L̂m (t) according to
B. Local Training (33)
15: Perform a gradient descent step with respect to
Since the state space and the action space in a dynamic ωm
system are enormous and complicated, most of reinforcement 16: Update Q(t + 1) = Q(t) every c steps
algorithms such as Q-learning cannot work well. Therefore, we 17: Update the weighted parameter of the user
turn to employ DQN-based algorithm to train the local models. SDm according to (35)
Different from Q-learning, DQN uses a DNN to approximate the 18: end for
Q function and leverages an experience relay to break the corre- 19: if t% ==0 then
lation between the data. Fig. 2 depicts the structure of the DQN 20: Aggregate the weighted parameters of all users
used in this paper, where the user SDm puts the current state at the central node according to (36)
and observation from experience pool into evaluation network, 21: The central node distributes the global
and then obtains the action-state value Qm (Sm (t), Am (t); ωm ), parameter to all users according to (37)
in which ωm denotes the weighted parameter in the evaluation 22: end if
network. Specifically, the loss function is the difference between 23: end for
the action-state value in the evaluation network and the next state 24: end for
value in the target network, which is given by, 25: end for
L̂m (t) = (Ym (t) − Qm (Sm (t), Am (t); ωm ))2 , (33)

where Ym (t) is the target value function, represented as, C. Model Aggregation and Model Distribution
−
Ym (t) = Rm (t) + δ arg max Qm (Sm (t + 1), Am (t + 1); ωm ), After elaborating the process that each user finds its own allo-
Am (t+1) cation strategy through DQN, we further describe how the users
(34) cooperate in the FL framework to optimize the whole system
performance. Fig. 3 shows the process of the FL framework,
−
in which δ is a discount factor and ωm denotes the weighted which consists of several key steps: local training, model upload,
parameter in the target network. Note that the target network has model aggregation and model distribution. Specifically, the users
the same structure and initial weights as the evaluation network. firstly initialize and train their own DQN models locally for
However, the evaluation network is updated at each iteration, rounds, and then upload their own model parameters to a central
while the target network is updated every c steps. In particular, node which is responsible for aggregation. According to the
the update process of the evaluation network of SDm is, federated average algorithm, the model aggregation at time slot
i is expressed as,
∂ L̂m (t)
ωm (t + 1) = ωm (t) − η̂ , (35)
∂ωm (t) 1 
M
ω(i + 1) = ωm (i ), (36)
where η̂ is the learning rate of the users. M m=1

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
GUO et al.: DISTRIBUTED MACHINE LEARNING FOR MULTIUSER MOBILE EDGE COMPUTING SYSTEMS 467

Fig. 3. The process of the federated learning.

where i (i = 1, 2, . . . , imax ) represents the number of the ag- Fig. 4. Convergence of the proposed method with K = 2 and λ = 0.5:
gregation, and ω(i + 1) denotes the global model parameter Criterion I.
of the evaluation network at time slot i + 1. Afterwards, the
central node distributes the global model parameter to all users.
The model distribution from the central node to the users can be
Fig. 4 depicts the convergence of the proposed method under
formulated as,
criterion I versus the epoch, where the number of users varies
ωm (t + 1) = ω(t + 1). (37) in {5, 6, 8}, the number of CAPs is 2, λ is 0.5, and the number
of epochs varies from 1 to 20. As observed from Fig. 4, we can
After the users participating into the FL have updated their local find that for various numbers of users, the weighted cost declines
model parameters of the evaluation network, one iteration of the along with the increasing number of epochs. In particular, the
FL process is completed. This iterative process repeats until the cost with 5 users declines roughly from 10 to 4 and it converges
system performance becomes convergent. The whole procedure at epoch 2, the cost with 6 users declines roughly from 18 to 6
of the FL based distribution resource allocation is summarized and it converges at epoch 7, and the cost with 8 users declines
in Algorithm 1. roughly from 32 to 11 and it converges at epoch 9. Moreover,
the weighted cost with M = 6 fluctuates at epoch 5 and 6, as the
system may perform a bad exploration in some poor channels.
V. SIMULATION RESULTS In further, we can see that the weighted cost increases along
In this section, we perform some simulations to evaluate with a larger M , since more users increase the burden on the
the performance of the proposed schemes. The channels in the communication and computation and result in a larger cost.
network follow Rayleigh flat fading [34], [35], and the average Therefore, Fig. 4 verifies that the proposed method can help
channel gains of the links from the users to the UAV, primary reduce the system cost effectively for various numbers of users
user and CAPs are set to 0.01, 1 and 1, respectively. The variance under criterion I.
of the AWGN at the UAV and CAPs is set to 0.1. Moreover, Fig. 5 depicts the convergence of the proposed method under
the power of the peak interference to the primary user is set to criterion II versus the epoch, where the number of users varies
2 W, while the maximum transmit power of users is limited to in {5, 6, 8}, the number of CAPs is 2, and the number of epochs
5 W. The computational power of the local users and CAPs are varies from 1 to 20. As observed from Fig. 5, we can find that
1 W. The step size ξ for adjusting the offload ratio, bandwidth for various numbers of users, the latency declines along with
allocation ratio, and computing capability allocation ratio is set the increasing number of epochs. In particular, the latency with
to 0.1. Without loss of generality, the wireless bandwidth of 5 users declines roughly from 5.5 to 1.3 and it converges at
CAP k is set to (50 + k) MHz, and the computational capability epoch 6, the latency with 6 users declines roughly from 5.5 to
of CAP k is set to (8 + 0.1 k) 106 cycles/s, where k varies in 1.7 and it converges at epoch 6, and the latency with 8 users
{1, . . . , K}. The local computational capability of the user SDm declines roughly from 7.3 to 2.5 and it converges at epoch 17.
is (5 + 0.1 m) 104 cycles/s, where m varies in {1, . . . , M }. In Moreover, the latency with M = 8 fluctuates at epoch 10 and
further, the task size of the user SDm is uniformly distributed 15, as the system may perform a bad exploration in some poor
in the range of [10 + 5 m, 10 + 5(m + 1)] MB, and we assume channels. In further, we can see that the latency increases along
that the priority of the last user is 3, while that of the other users with a larger M , since more users increase the burden on the
is set to 1. Considering the requirements from the application communication and computation and result in a larger latency.
scenarios in criterion II and criterion III, we set the latency Therefore, Fig. 5 verifies that the proposed method can help
threshold Lth and the energy consumption Eth to 10 and 20, reduce the system latency effectively for various numbers of
respectively, when M = 5 and K = 2. users under criterion II.

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
468 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 16, NO. 3, APRIL 2022

Fig. 7. Weighted cost of criterion I versus λ with K = 2.

Fig. 5. Convergence of the proposed method under with K = 2: Criterion II.

criterion III. Fig. 7 demonstrates the linearly weighted cost of

criterion I versus the weight factor λ, where λ varies from 0.1
to 0.9, the number of users varies in {5, 6, 8} and the number
of CAPs is 2. As observed from this figure, we can find that
the weighted cost decreases along with a larger λ, as the energy
consumption dominates in the weighted cost of the considered
system. In particular, the weighted cost with M = 5 declines
from 6 to 2, the weighted cost with M = 6 declines from 9 to
3, and the weighted cost with M = 8 declines from 18 to 5.
Moreover, we can see that the weighted cost increases with a
larger M , since more users will increase the system overhead in
the communication and computation. Hence, the results in Fig. 7
further verify that the proposed method can provide an effective
strategy for various values of λ under criterion I. Fig. 8 shows the
performance comparison of the three criteria in terms of system
cost, where the number of users varies from 5 to 10, the number
of CAPs is 2 or 4, and λ = 0.5. Specifically, Fig. 8(a), 8(b)
Fig. 6. Convergence of the proposed method with K = 2: Criterion III. and 8(c) correspond to criterion I, II and III, respectively. For
comparison, we also plot the system cost of “All-Local” which
Fig. 6 depicts the convergence of the proposed method under computes all tasks at local, and “All-MEC” which computes
criterion III versus the epoch, where the number of users varies all tasks at the CAPs. As observed from these three figures,
in {5, 6, 8}, the number of CAPs is 2, and the number of epochs we can find that the system cost of the three criteria increases
varies from 1 to 20. As observed from Fig. 6, we can find that along with a larger M , as more users will cause more burden
for various numbers of users, the energy consumption declines on the communication and computation, which leads to a higher
along with the increasing number of epochs. In particular, the system cost. Moreover, the system cost in these figures decreases
energy consumption with 5 users declines roughly from 21 to with a larger K, as more CAPs can help reduce the burden on
7 and it converges at epoch 6, the energy consumption with the communication and computation, which leads to a lower
6 users declines roughly from 28 to 11 and it converges at system cost. In further, the proposed method outperforms the
epoch 5, and the energy consumption with 8 users declines conventional “All-Local” and “All-MEC” methods under the
roughly from 45 to 18 and it converges at epoch 9. Moreover, three criteria. For example, when K = 2, the system cost of
the energy consumption with M = 8 fluctuates at epoch 8 and “All-Local” in Fig. 8(a) roughly increases from 40 to 109, and
18, as the system may perform a bad explorations in some poor that of “All-MEC” roughly increases from 9 to 21. In contrast, the
channels. In further, we can see that the energy consumption system cost of the proposed method slightly increases from 4 to
increases along with a larger M , since more users will in- 19. Therefore, the results in these figures verify the effectiveness
crease the burden on the communication and computation and of the proposed method for the considered system furthermore.
result in a larger energy consumption. Therefore, Fig. 6 verifies To show the effect of task priority on the bandwidth allocation
that the proposed method can help reduce the system energy under the three criteria, Fig. 9(a) and 9(b) are presented to show
consumption effectively for various numbers of users under the allocated bandwidth of each user from the CAPs, where

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
GUO et al.: DISTRIBUTED MACHINE LEARNING FOR MULTIUSER MOBILE EDGE COMPUTING SYSTEMS 469

Fig. 9. Allocated bandwidth from CAPs to the users under the three criteria
with M = 5 and K = 2.

allocated to SD1 is lower than SD2−4 , due to the wireless link

from SD1 to CAP1 is better than those from the other users to
CAP1 . Similarly, as observed from Fig. 9(b), we can see that
for the three criteria, the bandwidth allocated from CAP2 to the
user SD5 roughly reaches 20.3 MHz, 19.5 MHz and 21.1 MHz,
respectively, while that from CAP2 to other users does not exceed
10.2 MHz. Therefore, Fig. 9 validates the effectiveness of the
proposed priority based bandwidth allocation policy.
To show the impact of task priority on the computational
Fig. 8. Performance comparison of the three criteria versus the number of capability allocation under the three criteria, Fig. 10(a) and
users. 10(b) are demonstrated to show the allocated computational
capability of each user from the CAPs, where M = 5, K = 2,
λ = 0.5, and the task priority of the five users is {1, 1, 1, 1, 3}.
M = 5, K = 2, λ = 0.5, and the task priority of the five users As observed from Fig. 10(a), we can find that for the three
is {1, 1, 1, 1, 3}. As observed from Fig. 9(a), we can find that criteria, the computational capability allocated from CAP1 to the
for the three criteria, the bandwidth allocated from CAP1 to the user SD5 roughly reaches 3.4 × 106 cycles/s, 4.3 × 106 cycles/s
user SD5 roughly reaches 15.7 MHz, 17.1 MHz and 17.7 MHz and 3.5 × 106 cycles/s, respectively, while that allocated from
under the three criteria, respectively, while that from CAP1 to CAP1 to other users does not exceed 1.6 × 106 cycles/s. Sim-
other users does not exceed 10.7 MHz. Moreover, the bandwidth ilarly, as observed from Fig. 10(b), we can see that for the

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
470 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 16, NO. 3, APRIL 2022

Fig. 11. Allocated resources from CAPs to the user SD5 with equal and non-
Fig. 10. Allocated computational capability from CAPs to the users under the
equal priorities under the three criteria.
three criteria with M = 5 and K = 2.

priority is about 11.3 MHz, 9.1 MHz and 8.7 MHz, respectively.
three criteria, the bandwidth allocated from CAP2 to the user By comparing the allocated bandwidth of SD5 with equal and
SD5 roughly reaches 3.5 × 106 cycles/s, 3.7 × 106 cycles/s and non-equal priorities under three criteria, we can find that SD5
3.6 × 106 cycles/s, respectively, while that from CAP2 to other with a higher priority can be allocated more bandwidth than SD5
users does not exceed 1.6 × 106 cycles/s. Therefore, Fig. 10 with an equal priority. Similarly, as observed from Fig. 11(b), we
verifies the effectiveness of the proposed priority based com- can see that for the three criteria, the computational capability al-
putational capability allocation policy. Fig. 11 compares the located from CAP1 to SD5 with a higher priority roughly reaches
allocated bandwidth and computational capability from CAPs 3.4 × 106 cycles/s, 4.3 × 106 cycles/s and 3.5 × 106 cycles/s, re-
to the user SD5 with equal and non-equal priorities under the spectively, while that allocated from CAP1 to SD5 with an
three criteria, where M = 5, K = 2, and λ = 0.5. The equal equal priority is about 1.6 × 106 cycles/s, 1.5 × 106 cycles/s and
task priority of the users is {1, 1, 1, 1, 1}, while the non-equal 1.6 × 106 cycles/s, respectively. In further, the computational
priority of the users is set to {1, 1, 1, 1, 3}. As observed from capability allocated from CAP2 to SD5 with a higher prior-
Fig. 11(a), we can find that for the three criteria, the bandwidth ity roughly reaches 3.5 × 106 cycles/s, 3.7 × 106 cycles/s and
allocated from CAP1 to SD5 with a higher priority roughly 3.6 × 106 cycles/s, respectively, while that allocated from CAP2
reaches 15.7 MHz, 17.1 MHz and 17.7 MHz, respectively, while to SD5 with an equal priority is about 1.7 × 106 cycles/s, 1.3 ×
that allocated from CAP1 to SD5 with an equal task priority is 106 cycles/s and 1.6 × 106 cycles/s, respectively. By comparing
about 1.0 MHz, 0.7 MHz and 4.5 MHz, respectively. Moreover, the allocated computational capability of SD5 with equal and
the bandwidth allocated from CAP2 to SD5 with a higher priority non-equal priorities under three criteria, we can find that SD5
roughly reaches 20.3 MHz, 19.5 MHz and 21.1 MHz, respec- with non-equal priority can be allocated more computational ca-
tively, while that allocated from CAP2 to SD5 with an equal pability than SD5 with equal priority. Therefore, Fig. 11 further

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
GUO et al.: DISTRIBUTED MACHINE LEARNING FOR MULTIUSER MOBILE EDGE COMPUTING SYSTEMS 471

TABLE I by multiple CAPs. We considered a cognitive eavesdropping

NOTATION LIST
environment where the users could use the wireless spectrum as
long as the interference to the primary user was tolerated, and
an eavesdropper in the network could overhear the confidential
message from the users. For the considered system, we firstly
presented three optimization criteria and then solved the opti-
mization problem in a distributed machine learning way through
optimizing the offloading ratio as well as the allocation of
bandwidth and computational capability, by taking into account
the task priority among users. In particular, a FL optimization
framework was used, and each user employed DRL to solve the
optimization. Simulation results were finally demonstrated to
show that the proposed method can effectively reduce the system
cost in terms of latency and energy consumption, and meanwhile
ensure that more bandwidth and computational capability are
allocated to the user with a higher task priority.
Regarding future works extending from this paper, we will
take into account some discrepancy in the fairness among users,
due to the parameters such as the channel condition, task size
and computational capability. We can alleviate the unfair issue
by some methods such as imposing some different weights on
the SDs’ latency and energy consumption. Moreover, as most
of mobile devices are powered by a limited battery, it is hard
to keep all mobile devices online. Therefore, how to provide
an efficient and flexible energy supply to the mobile devices
becomes a critical issue in the MEC networks. Motivated by
this, we tend to exploit the energy harvesting model to solve the
TABLE II above issue in the future.
PARAMETER SETTING

REFERENCES
[1] J. Gong, Q. Kuang, and X. Chen, “Joint transmission and computing
scheduling for status update with mobile edge computing,” in Proc. IEEE
Int. Conf. Commun., 2020, pp. 1–6.
[2] X. Lai and J. Xia, “Secure mobile edge computing networks in the presence
of multiple eavesdroppers,” IEEE Trans. Commun., early access, 2022,
doi: 10.1109/TCOMM.2021.3119075.
[3] C. Park and J. Lee, “Mobile edge computing-enabled heterogeneous
networks,” IEEE Trans. Wireless Commun., vol. 20, no. 2, pp. 1038–1051,
Feb. 2021.
[4] C. Liu, M. Bennis, M. Debbah, and H. V. Poor, “Dynamic task offloading
and resource allocation for ultra-reliable low-latency edge computing,”
IEEE Trans. Commun., vol. 67, no. 6, pp. 4132–4150, Jun. 2019.
[5] X. Lai and J. Xia, “Outdated access point selection for mobile edge
computing with cochannel interference,” IEEE Trans. Veh. Technol., early
access, 2022.
[6] S. Tang and L. Chen, “Computational intelligence and deep learning for
next-generation edge-enabled industrial IoT,” IEEE Trans. Netw. Sci. Eng.,
early access, 2022.
[7] S. Bi and Y. J. Zhang, “Computation rate maximization for wireless
powered mobile-edge computing with binary computation offloading,”
validates that the proposed allocation strategy can dynamically IEEE Trans. Wireless Commun., vol. 17, no. 6, pp. 4177–4190, Jun. 2018.
allocate the bandwidth and computational capability according [8] X. He, R. Jin, and H. Dai, “Peace: Privacy-preserving and cost-efficient
task offloading for mobile-edge computing,” IEEE Trans. Wireless Com-
to the task priority. mun., vol. 19, no. 3, pp. 1814–1824, Mar. 2020.
[9] M. Sheng, Y. Wang, X. Wang, and J. Li, “Energy-efficient multiuser
partial computation offloading with collaboration of terminals, radio ac-
VI. CONCLUSION cess network, and edge server,” IEEE Trans. Commun., vol. 68, no. 3,
pp. 1524–1537, Mar. 2020.
In this work, we studied the FL framework for the system [10] R. Malik and M. Vu, “Energy-efficient computation offloading in delay-
design of a multiuser MEC system, where multiple users had constrained massive MIMO enabled edge network using data partitioning,”
some computational tasks of different priorities, to be computed IEEE Trans. Wireless Commun., vol. 19, no. 10, pp. 6977–6991, Oct. 2020.

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
472 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 16, NO. 3, APRIL 2022

[11] X. Hu, J. Wang, and C. Zhong, “Statistical CSI based design for intelligent Yinghao Guo received the bachelor’s degree in soft-
reflecting surface assisted MISO systems,” Sci. China, Inf. Sci., vol. 63, ware engineering from the Zhengzhou University of
no. 12, 2020, Art. no. 222303. Education, Zhengzhou, China, in June 2019. He is
[12] J. Zhang, Y. Zhang, C. Zhong, and Z. Zhang, “Robust design for intelligent currently a Graduate Student with the School of Com-
reflecting surfaces assisted MISO systems,” IEEE Commun. Lett., vol. 24, puter Science and Cyber Engineering, Guangzhou
no. 10, pp. 2353–2357, Oct. 2020. University, Guangzhou, China. His current research
[13] Q. Tao, J. Wang, and C. Zhong, “Performance analysis of intelligent interests include Machine Learning and mobile edge
reflecting surface aided communication systems,” IEEE Commun. Lett., computing resource scheduling algorithms.
vol. 24, no. 11, pp. 2464–2468, Nov. 2020.
[14] P. Yang, Y. Xiao, M. Xiao, Y. L. Guan, S. Li, and W. Xiang, “Adaptive
spatial modulation MIMO based on machine learning,” IEEE J. Sel. Areas
Commun., vol. 37, no. 9, pp. 2117–2131, Sep. 2019.
[15] J. Chen et al., “Hybrid beamforming/combining for millimeter wave
MIMO: A machine learning approach,” IEEE Trans. Veh. Technol., vol. 69,
no. 10, pp. 11353–11368, Oct. 2020.
[16] H. Liu, Y. Xiao, P. Yang, J. Fu, S. Li, and W. Xiang, “Transmit antenna
selection for full-duplex spatial modulation based on machine learning,” Rui Zhao received the B.E. degree in computer sci-
IEEE Trans. Veh. Technol., vol. 70, no. 10, pp. 10695–10708, Oct. 2021. ence and technology from Bohai University, Jinzhou,
[17] J. Shi, J. Du, J. Wang, J. Wang, and J. Yuan, “Priority-aware task offloading China, in 2018, and the master’s degree from the
in vehicular fog computing based on deep reinforcement learning,” IEEE School of Computer Science and Cyber Engineering,
Trans. Veh. Technol., vol. 69, no. 12, pp. 16067–16081, Dec. 2020. Guangzhou University, Guangzhou, China, in 2022.
[18] S. Tang, “Dilated convolution based CSI feedback compression for mas- His current research interests include Machine Learn-
sive MIMO systems,” IEEE Trans. Veh. Technol., early access, 2022. ing and mobile edge computing resource scheduling
[19] L. He and K. He, “Towards optimally efficient search with deep learning for algorithms.
large-scale MIMO systems,” IEEE Trans. Commun., early access, 2022.
[20] K. He and Y. Deng, “Efficient memory-bounded optimal detection for
GSM-MIMO systems,” IEEE Trans. Commun., early access, 2022.
[21] H. Cao and J. Cai, “Distributed multiuser computation offloading for
cloudlet-based mobile cloud computing: A game-theoretic machine learn-
ing approach,” IEEE Trans. Veh. Technol., vol. 67, no. 1, pp. 752–764,
Jan. 2018.
[22] R. Lin et al., “Distributed optimization for computation offloading in
edge computing,” IEEE Trans. Wireless Commun., vol. 19, no. 12,
pp. 8179–8194, Dec. 2020.
[23] X. Cai, X. Mo, J. Chen, and J. Xu, “D2D-enabled data sharing for Shiwei Lai received the bachelor’s degree in com-
distributed machine learning at wireless network edge,” IEEE Wireless puter science and technology from the Guangdong
Commun. Lett., vol. 9, no. 9, pp. 1457–1461, Sep. 2020. University of Education, Guangzhou, China, in June,
[24] M. Wu, D. Ye, J. Ding, Y. Guo, R. Yu, and M. Pan, “Incentivizing differen- 2019. She is currently a Graduate Student with the
tially private federated learning: A multidimensional contract approach,” School of Computer Science and Cyber Engineer-
IEEE Internet Things J., vol. 8, no. 13, pp. 10639–10651, Jul. 2021. ing, Guangzhou University, Guangzhou, China. Her
[25] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. Y Arcas, current research interests include Machine Learning
“Communication-efficient learning of deep networks from decentralized and mobile edge computing resource scheduling al-
data,” in Proc. Artif. Intell. Statist., 2017, pp. 1273–1282. gorithms.
[26] S. Zheng, C. Shen, and X. Chen, “Design and analysis of uplink and
downlink communications for federated learning,” IEEE J. Sel. Areas
Commun., vol. 39, no. 7, pp. 2150–2167, 2021.
[27] Y. Liu et al., “A communication efficient vertical federated learning
framework,” 2020, arXiv:1912.11187.
[28] W. Y. B. Lim et al., “Federated learning in mobile edge networks:
A comprehensive survey,” IEEE Commun. Surv. Tut., vol. 22, no. 3,
pp. 2031–2063, Mar. 2020.
[29] C. Shen, J. Xu, S. Zheng, and X. Chen, “Resource rationing for wireless Lisheng Fan (Member IEEE) received the bachelor’s
federated learning: Concept, benefits, and challenges,” IEEE Commun. degree from the Department of Electronic Engineer-
Mag., vol. 59, no. 5, pp. 82–87, May 2021. ing, Fudan University, Shanghai, China, in 2002, the
[30] R. Yu and P. Li, “Toward resource-efficient federated learning in mobile master’s degree from the Department of Electronic
edge computing,” IEEE Netw., vol. 35, no. 1, pp. 148–155, Jan. 2021. Engineering, Tsinghua University, Beijing, China, in
[31] Z. Zhao, J. Xia, L. Fan, X. Lei, and G. K. Karagiannidis, “System opti- 2005, and the Ph.D. degree from the Department
mization of federated learning networks with a constrained latency,” IEEE of Communications and Integrated Systems, Tokyo
Trans. Veh. Technol., early access, 2022, doi: 10.1109/TVT.2021.3128559. Institute of Technology, Tokyo, Japan, in 2008. He
[32] X. Huang, P. Li, R. Yu, Y. Wu, K. Xie, and S. Xie, “Fedparking: A federated is currently a Professor with the School of Computer
learning based parking space estimation with parked vehicle assisted edge Science, GuangZhou University, Guangzhou, China.
computing,” IEEE Trans. Veh. Technol., vol. 70, no. 9, pp. 9355–9368, His research interests include wireless cooperative
Sep. 2021. communications, physical-layer secure communications, intelligent communi-
[33] R. Yadav, W. Zhang, O. Kaiwartya, H. Song, and S. Yu, “Energy-latency cations, and system performance evaluation. He has authored or coauthored
tradeoff for dynamic computation offloading in vehicular fog computing,” many papers in international journals such as IEEE TRANSACTIONS ON WIRE-
IEEE Trans. Veh. Technol., vol. 69, no. 12, pp. 14198–14211, Dec. 2020. LESS COMMUNICATIONS, IEEE TRANSACTIONS ON COMMUNICATIONS, IEEE
[34] N. Yang, L. Wang, G. Geraci, M. Elkashlan, J. Yuan, and M. D. Renzo, TRANSACTIONS ON INFORMATION THEORY, and also papers in conferences such
“Safeguarding 5G wireless communication networks using physical layer as IEEE ICC, IEEE Globecom, and IEEE WCNC. He is the Editor of the China
security,” IEEE Commun. Mag., vol. 53, no. 4, pp. 20–27, Apr. 2015. Communications, and was the Guest Editor of many journals such as Physical
[35] C. Liu, N. Yang, R. A. Malaney, and J. Yuan, “Artificial-noise-aided Communication, EURASIP Journal on Wireless Communications and Network-
transmission in multi-antenna relay wiretap channels with spatially ran- ing, and Wireless Communications and Mobile Computing. He has been awarded
dom eavesdroppers,” IEEE Trans. Wireless Commun., vol. 15, no. 11, as Exemplary Reviewer by IEEE TRANSACTIONS ON COMMUNICATIONS and
pp. 7444–7456, Nov. 2016. IEEE COMMUNICATIONS LETTERS.

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.
GUO et al.: DISTRIBUTED MACHINE LEARNING FOR MULTIUSER MOBILE EDGE COMPUTING SYSTEMS 473

Xianfu Lei received the Ph.D. degree from South- George K. Karagiannidis (Fellow, IEEE) was born
west Jiaotong University, Chengdu, China, in 2012. in Pithagorion, Samos Island, Greece. He received
Since 2015, he has been an Associate Professor with the University Diploma (five years) and the Ph.D.
the School of Information Science and Technology, degree in electrical and computer engineering from
Southwest Jiaotong University. From 2012 to 2014, the University of Patras, Patras, Greece, in 1987 and
he was a Research Fellow with the Department of 1999, respectively. From 2000 to 2004, he was a
Electrical and Computer Engineering, Utah State Senior Researcher with the Institute for Space Appli-
University, Logan, UT, USA. His research interests cations and Remote Sensing, National Observatory of
include 5G/6G networks, cooperative and energy har- Athens, Athens, Greece. In June 2004, he joined the
vesting networks, and physical-layer security. He is Faculty of the Aristotle University of Thessaloniki,
currently the Area Editor of IEEE COMMUNICATIONS Thessaloniki, Greece, where he is currently a Profes-
LETTERS and an Associate Editor for IEEE WIRELESS COMMUNICATIONS LET- sor with the Electrical and Computer Engineering Department and the Head of
TERS and IEEE TRANSACTIONS ON COMMUNICATIONS. He was the Senior Editor the Wireless Communications and Information Processing (WCIP) Group. He
and an Associate Editor for IEEE COMMUNICATIONS LETTERS from 2014 to is also an Honorary Professor with South West Jiaotong University, Chengdu,
2019. He was the recipient of the Best Paper Award in IEEE/CIC ICCC2020, China. His research interests include the broad area of digital communica-
the Best Paper Award in WCSP2018, the WCSP Ten-Year Anniversary Excel- tions systems and signal processing, with emphasis on wireless communica-
lent Paper Award, IEEE Communications Letters Exemplary Editor 2019, and tions, optical wireless communications, wireless power transfer and applica-
Natural Science Award of China Institute of Communications (2019). tions, and communications and signal processing for biomedical engineering.
Dr. Karagiannidis has been involved as the General Chair, the Technical Program
Chair, and a member of the Technical Program Committees in several IEEE and
non-IEEE conferences. In the past, he was an Editor of several IEEE journals.
From 2012 to 2015, he was the Editor-in-Chief of IEEE COMMUNICATIONS
LETTERS. He is currently the Associate Editor-in-Chief for the IEEE OPEN
JOURNAL OF COMMUNICATIONS SOCIETY. He is one of the highly-cited authors
across all areas of Electrical Engineering, recognized from Clarivate Analytics
as Web-of-Science Highly-Cited Researcher in the six consecutive years 2015–
2020.

Authorized licensed use limited to: NASATI. Downloaded on July 05,2022 at 07:23:36 UTC from IEEE Xplore. Restrictions apply.