See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/358401090

Energy-Efficient Federated Learning over UAV-Enabled Wireless Powered

Communications

Article  in  IEEE Transactions on Vehicular Technology · February 2022

CITATIONS READ

0 1

5 authors, including:

Quoc-Viet Pham Mai Le

Pusan National University Inje University
130 PUBLICATIONS   1,878 CITATIONS    3 PUBLICATIONS   238 CITATIONS   

SEE PROFILE SEE PROFILE

Thien Huynh-The won-Joo Hwang

Kumoh National Institute of Technology Pusan National University
106 PUBLICATIONS   1,064 CITATIONS    182 PUBLICATIONS   1,538 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Detecting malicious domain names View project

Swarm Intelligence for Next-Generation Wireless Systems View project

All content following this page was uploaded by Quoc-Viet Pham on 07 February 2022.

The user has requested enhancement of the downloaded file.

 1

Energy-Efficient Federated Learning over

UAV-Enabled Wireless Powered Communications
Quoc-Viet Pham, Member, IEEE, Mai Le, Thien Huynh-The, Member, IEEE,
Zhu Han, Fellow, IEEE, and Won-Joo Hwang, Senior Member, IEEE

Abstract—Since the invention in 2016, federated learning (FL) require central storage and training of the learning models.
has been a key concept of artificial intelligence, in which the Such AI methods cause serious privacy issues since the user’s
data of FL users needs not to be uploaded to the central personal data should be uploaded to the learning server, which
server. However, performing FL tasks may not be feasible
due to the unavailability of terrestrial communications and the may include private information, such as bank account, gender,
battery limitation of FL users. To address these issues, we make and blood type [3], [4]. For example, the Self-Diagnosis
use of unmanned aerial vehicles (UAVs) and wireless powered App1 helps manage the COVID-19 self-quarantine situation
communications (WPC) for FL networks. In order to enable in South Korea. However, this application requires users to
sustainable FL solutions, the UAV equipped with edge computing report personal information, such as location, temperature, and
and WPC capabilities is deployed as an aerial energy source as
well as an aerial server to perform FL tasks. We propose a symptoms, to the central server, which may cause privacy
joint algorithm of UAV placement, power control, transmission issues. Centralized learning also raises critical privacy issues in
time, model accuracy, bandwidth allocation, and computing re- smart healthcare, in which highly sensitive and private health-
sources, namely energy-efficient FL (E2FL), aiming at minimizing related information must be regulated according to health
the total energy consumption of the aerial server and users. bodies, such as United States Health Insurance Portability and
The E2FL overcomes the original nonconvex problem by an
efficient algorithm. We show that sustainable FL solutions can Accountability Act (HIPPA)2 . Consequently, there is an urgent
be provided via UAV-enabled WPC through various simulation need to go towards distributed AI approaches for scalable and
results. Moreover, the outperformance of E2FL in terms of energy privacy-preserving IoT applications at mobile edge networks.
efficiency over several benchmarks emphasizes the need for a Federated learning (FL) is a new AI concept that was
joint resource allocation framework rather than optimizing a
originated by Google in 2016 [5]–[7]. In FL, multiple users
subset of optimization factors.
collaboratively build a global learning model without the
Index Terms—Energy Harvesting, Federated Learning (FL), need for data collection and sharing with the central server.
Mobile Edge Computing (MEC), UAV Communications, Wireless
Powered Communication (WPC).
Considered as a promising privacy-preserving solution, FL
has found its applications in different engineering areas and
problems, such as healthcare [8], intelligent radio access net-
I. I NTRODUCTION works [9], IoT intrusion detection [10], and industrial IoT [11].

T HE exponential increase of the Internet-of-Things (IoT)

and emerging applications (e.g., face recognition, zero-
touch network, autonomous driving, and virtual reality) have
However, the conventional FL method requires the existence
of a central server, e.g., a multi-access edge computing (MEC)
server located at the network edge [6]. Hence, performing
generated a huge amount of data [1]. Different artificial FL tasks may not be feasible, especially when the terrestrial
intelligence (AI) approaches, such as deep learning, deep rein- communication infrastructure is unavailable. Moreover, since
forcement learning, and transfer learning, have been developed FL users, such as IoT devices and tiny sensors, may not have
to leverage various features of the data [2]. Conventionally, sufficient power budgets to participate actively in the FL pro-
AI methods rely on the existence of centralized servers and cedure, providing sustainable and energy-efficient solutions is
necessary to enhance the performance of FL-enabled networks.
Copyright (c) 2015 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be For example, the work in [12] develops an iterative algorithm
obtained from the IEEE by sending a request to [email protected]. for energy efficiency in FL, where closed-form solutions are
Quoc-Viet Pham is with the Korean Southeast Center for the 4th Industrial obtained at every step. In [13], intelligent reflecting surface
Revolution Leader Education, Pusan National University, Busan 46241, Korea
(e-mail: [email protected]). (IRS), an emerging technology in 6G wireless systems, is
Mai Le is with the Department of Information Convergence En- leveraged to further reduce the energy consumption of FL
gineering, Pusan National University, Busan 46241, Korea (e-mail: users compared with that of without IRS.
[email protected]).
Thien Huynh-The is with the ICT Convergence Research Center, Kumoh Aerial access networks are considered as a key-enabling
National Institute of Technology, Gumi, Gyeongbuk 39177, Korea (email: concept of the future sixth-generation (6G) wireless sys-
[email protected]).
Zhu Han is with the Department of Electrical and Computer Engineering
tems [14]–[16]. Fundamentally, an aerial access network is
in the University of Houston, Houston, TX 77004 USA, and also with the composed of different platforms, which include low-altitude
Department of Computer Science and Engineering, Kyung Hee University, platforms, high-altitude platforms, and satellite communica-
Yongin 446-701, Korea (e-mail: [email protected]).
Won-Joo Hwang is with the Department of Biomedical Convergence
1 http://ncov.mohw.go.kr/selfcheck/
Engineering, Pusan National University, Yangsan 50612, Korea (e-mail:
[email protected]). 2 https://www.cdc.gov/phlp/publications/topic/hipaa.html
 2

tions. In addition, there are several features of aerial access been investigated in our recent work [29]. However, different
networks, such as ubiquity, mobility, availability, simultaneity, from [29], the current work considers the whole FL training
and scalability, which make aerial access networks distinguish- process, in which the local training accuracy and multiple
able [16]. These key features will carry a full complement global communication rounds are taken into consideration.
of conventional terrestrial communications, thus improving Moreover, rather than a simple linear model in [29], we design
communication services and user satisfaction. For example, sustainable FL solutions considering a more realistic nonlinear
positioned at the low-altitude tier of future 6G aerial access EH model3 .
networks, unmanned aerial vehicles (UAVs) can be dispatched In spite of promising results in the available literature, no
as access points and radio towers that provide wireless and prior work focuses on UAV communications for sustainable
computing services to ground users from the air, when the ter- FL networks with a nonlinear EH model. Motivated by this
restrial communication network is unreachable [17]. Another consideration, we introduce a resource optimization framework
real example of UAV is data collection from network sites in this work for designing sustainable solutions of FL-enabled
[18]. In this project, Ericsson and Vodafone UK collaborate networks under the nonlinearity induced by the nonlinear
to deploy UAVs to collect data from 30 network sites and EH model. In summary, our paper presents the following
then use the collected data for network upgrading purposes. contributions and features.
The application of aerial access networks for FL networks • Problem Formulation of UAV-FL Networks: To over-
can be found in recent studies. In [19], multiple UAVs are come the battery limitation of FL users and the unavail-
dispatched to capture and then locally process environmental ability of terrestrial communications, we leverage UAV
images before sending the model updates to an aggregation communications to enable sustainable FL networks. We
server located at the ground fusion center. A holistic air quality design the sustainable FL solutions under a more practical
sensing framework with two main segments, including air EH model, in which a nonlinear and exponential function
sensing and ground sensing, is investigated [20], in which of the input power is used to calculate the harvested
multiple UAVs are deployed as FL users, and a swarm leader power. We formulate a problem of minimizing the overall
is regarded as the aggregation server. However, these prior energy consumption of FL users and the aerial server
studies consider UAVs as learners, which is different from while taking into account various constraints of UAV
our work where the UAV acts as the FL model aggregator. placement, power control, transmission time, FL model
The integration of aerial communications and wireless accuracy, bandwidth allocation, and computing resources.
powered communication (WPC) has the potential to provide • Problem Transformation via Alternating Approach:
sustainable solutions for terrestrial networks via exploiting Solving the formulated problem optimally is very chal-
distinctive features of aerial energy sources [21]–[23]. For lenging because of the coupling among optimization
example, battery-limited sensors can wirelessly harvest energy variables and the nonlinearity of underlying models, e.g.,
from aerial sources before sending their data to the point of channel model and nonlinear EH model. We develop an
interest for further processing. The work in [24] demonstrates efficient iterative algorithm in this work. In particular,
that the mobility of aerial energy transmitters can be exploited each update step divides the entire set of the optimization
to enhance the performance and operation management of variables into three blocks, which are solved alternately.
WPC networks, when compared to ones with fixed locations of • Proposed Solution via Successive Convex Approxima-
the energy transmitters as well as provide a fairness guarantee tion: For each subproblem, we propose a path-following
among energy receivers. In [25], a UAV is deployed to serve procedure, namely energy-efficient FL (E2FL), to address
as both the aerial MEC server and energy transmitter for its nonconvexity difficulty. By solving a sequence of con-
ground users. The trajectory of the UAV is optimized in [25] vex approximation problems, the solution is eventually
to enhance the performance in terms of computation bits. convergent after a sufficient number of iterations.
However, the existing works on UAV-aided WPC [21]–[25] • Performance Evaluation: In order to show the effective-
normally focus on the communication aspects and ignore the ness of aerial FL networks and the performance of the
potential of UAV-aided WPC in FL-enabled networks. We E2FL algorithm, we conduct various performance results.
note that the consideration of energy harvesting (EH) from We observe that employing UAV communications with
stochastic sources for FL is investigated in [26]. However, how MEC and WPC capabilities is promising for FL networks
to leverage the potential of UAV-WPC for the implementation with battery-limited FL users. Moreover, we observe that
of FL is not addressed yet. In this regard, a recent work in [27] the E2FL algorithm is significantly superior to several
utilizes deep reinforcement learning to solve the long-term benchmarks, such as alternative schemes with preset local
energy problem of a UAV-FL WPC network. Nevertheless, accuracy, location of the aerial FL server, local computing
multiple communication rounds and the nonlinearity of the resources, and transmission time.
EH model are not considered in [27].
B. Paper Organization
A. Motivations and Contributions
Motivated by the potential of UAV-aided WPC over conven- The remainder of this paper is composed of the following
tional systems [28], we leverage the deployment of the UAV parts. The joint problem is formulated to minimize the overall
to overcome the energy limitation of FL users. The outper- 3 The study of this work with the linear EH model and preliminary analysis
formance of UAV-FL over conventional FL approaches has has been accepted for presentation at IEEE ICC 2022 [30].
 3

Alg. 1 A Seminal FL Algorithm

Local model updates
1: Initialization: Set the global index n = 0 and the initial
Global model broadcating
Data
model parameter ω (0) .
2: repeat
User 1 Local model 3: Increase the global index n = n + 1.
4: for All users k ∈ K do
5: Update the local difference ∇0k .
6: User k uses the gradient method to update ∇k until
Data the condition (3) holds.
7: end for
User 2 Local model
Global 8: The FL server does model aggregation based on (4).
model
...

Edge server 9: The FL server broadcasts the update to all users.

10: until The global model accuracy 0 is achieved.

Data
specifically, the local FL problem is optimized to minimize the
User K Local model objective function Gk (ω (n) + ∇k ) and to find the difference
∇k between the local and global models. Denote by ξ a
Fig. 1. Illustrative example of the FL architecture. hyper-learning parameter that is used to train the local models.
Minimizing the local FL problem of each user k
energy consumption of the aerial server and FL users in Gk (ω (n) + ∇k ) =Fk (ω (n) + ∇k )
Section II. The solution approach is detailed in Section III.
− (∇Fk (ω (n) ) − ξ∇Fk (ω (n) ))T ∇k (2)
Section IV presents simulation results to evaluate the perfor-
mance of the E2FL algorithm. Finally, the conclusion is drawn requires multiple local iterations to achieve an accuracy level
in Section V. of η such that
Gk (ω (n) + ∇k ) − Gk (ω (n) + ∇0k )
II. P RELIMINARY AND P ROBLEM F ORMULATION  
This section firstly presents an essential preliminary to the ≤ η Gk (ω (n) + ∇k ) − Gk (ω (n) + ∇∗k ) , (3)
FL concept in Subsection II-A, then describes the network
where ∇0k and ∇∗k denote the difference of user k at the first
model, EH model, computing model, and communication
local iterations and the optimal solution, respectively. Then,
model in Subsections II-B to II-E, and finally formulates the
all users report the updates ω (n) + ∇k to the FL server who
problem of sustainable UAV-FL networks in Subsection II-F.
will aggregate the received updates to create a new global
parameter as follows:
A. Preliminary to Federated Learning
1 X
FL is an up-to-minute concept of AI that allows different ω (n+1) = ω (n) + ∇k . (4)
K
users to build a collaborative learning model, and thus FL k∈K

makes the privacy attacks more difficult to launch. Fig. 1 The new parameter ω (n+1)
is distributed to all scheduled users
shows an illustrative example of the FL architecture, where K in the downlink to begin a new communication round. The
FL users share their local model updates with an edge server above procedure is repeated until the global model achieves
for model aggregation and the global model is broadcast to FL the desired accuracy 0 as follows:
users for the next training round. Denote by Fk (ω) the loss  
function of each user k, which can be different for different F (ω (n) ) − F (ω ∗ ) ≤ 0 F (ω (n) ) − Fk (ω (0) ) . (5)
learning tasks and applications. The training problem at the
Based on the above seminal algorithm, the last few years
FL server can be formulated as
X Dk have witnessed numerous improvements to the FL design, such
min F (ω) , Fk (ω), (1) as the convergence of FL on non-independent and identically
ω D distributed data, personalized FL, multiple FL services, secure
k∈K
model aggregation, serverless FL, and privacy-preserving so-
where K FL users are indexed in the set K = {1, . . . , K}.
lutions for FL. Thanks to its distinctive features and recent
In (1), F (ω) represents the global loss function, Dk represents
P advancements, FL has shown great potential for many appli-
the number of samples owned by user k, and D = k Dk
cations and services in wireless and mobile networks since the
represents the total number of samples.
invention in 2016. For more details, we invite the readers to
A seminal approach to solve problem (1) is developed in
check recent reviews in [2], [6], [7], [9].
[5]. The details of this approach are summarized in Alg. 1.
The global model is updated iteratively via multiple global
communication rounds. At a specific global update step n, B. Network Model
each user receives the global model parameter ω (n) and runs An FL network setup, as depicted in Fig. 2, is considered.
the gradient method to solve the local FL problem. More The network comprises a UAV, which acts as the aerial
 4

Wireless powered flow

where the reference channel power β̂0 is specified at a distance
of d0 = 1 m [29], [31], and α is the pathloss exponent. The
Local model update horizontal positions of the aerial server (i.e., UAV) and user
Model
Aerial FL aggregation Model broadcasting k are denoted as q = {x, y} and qk = {xk , yk }, respectively.
server The distance dk between the aerial server and user k is
computed as
Malfunctioning
q
dk = H 2 + kq − qk k22 . (7)
/Overloaded
! C. Energy Harvesting Model
There are two popular EH models that are typically lever-
aged in the literature on WPC. When the linear model is
Local
model employed, the amount of energy of user k harvested from the
UAV5 is calculated as Êklin = η0 T P gk . Here, T , P , and η0
Local data Local represent the time frame, the transmit power of the energy
Terrestrial
FL users with model source, and the energy conversion efficiency at FL users,
BS
limited battery respectively [29], [33]. This model is widely considered in the
literature such as [25], [29], [31] for the sake of optimization
and insightful lessons. However, the nonlinear model has been
Fig. 2. An example of aerial FL WPC networks, where the UAV is dispatched
as the aerial sever and aerial energy source.
shown to be more realistic than the linear one. According to
[34], the energy harvested by user k, when the nonlinear model
is applied, can be computed as follows:
server and aerial energy source, and K FL users (e.g., smart
π̂k − Mk νk
sensors, wearable devices, and industrial IoT devices). An Êknln = T , (8)
MEC server is mounted on the UAV to provide computing 1 − νk
services and broadcast energy to the users. The deployment where
of the UAV as an aerial aggregation server is helpful in areas, Mk 1
where the computing infrastructure is currently unavailable π̂k = and νk = .
1 + exp(−σk (Pk − ιk )) 1 + exp(σk ιk )
or the terrestrial base station is malfunctioning/overloaded (as (9)
illustrated in Fig. 2) [20], [25], [29]. There are other different
use cases of UAVs in FL networks, such as aerial FL users in In this nonlinear model, Mk is the maximum amount of energy
the sky and aerial relaying for FL. Model aggregation in the that user k can harness when its EH circuit approaches the
air is a basic task that does not necessitate a lot of computing saturation point, σk and ιk are the two constant coefficients
capabilities compared with the task computing and learning that are determined by the circuit characteristics, such as
[5], making our study differ from most of the existing studies resistance and capacitance, and Pk = P gk denotes the input
such as [19], [20], where UAVs are deployed as the learners. power of user k. As the sigmoid function is used in (8), the
As a rotary-wing UAV, the aerial server can hold up in the harvested energy of user k increases when the input power
air and does not need to move continuously. The deployment increases, but it becomes saturated when the input power is
of fixed-wing UAVs is also promising for several application sufficiently large. Thus, from (8), the harvested energy Êknln
scenarios and will be investigated in future work. FL users is nonlinear in Pk , which in turn depend on the transmit
are equipped with on-board computing resources to train the power and placement of the aerial server. As a result, in
local models and energy circuits to wirelessly garner energy comparison to the linear EH model, the EH nonlinearity
from the aerial energy source (i.e., UAV). As assumed in many further complicates the optimization problem.
studies such as [25], [29], we suppose that the aerial server has
simultaneous energy transmitting and local model receiving
D. Local Computing Model
operations. Similarly, FL users can harvest energy, execute
local computation, upload the model updates simultaneously. The data available at FL users is used to update the local
The aerial server hovers at a preset altitude H 4 , but the models and does not need to be uploaded to the aerial server
horizontal coordinate of the aerial server is taken into account because of privacy concerns. The learning model of each user
to maximize the EH and network performance. The channel k is denoted by a tuple (fk , Ck , Dk ), where fk is the local
gain of user k is computed as follows: computing capability (usually measured in CPU cycles per
−α second), Ck is the amount of computing resource required by
gk = β̂0 (dk /d0 ) , (6) user k to process one data bit, and Dk is the data size. Note
4 When FL users are sparsely distributed, the multidimensional path plan- that Ck and Dk are also called the computation workload to
ning of the aerial server can be taken into account to improve the learning and
charging efficiencies. Moreover, three-dimensional (3D) placement of multiple 5 To provide a more stable WPC capability for the implementation of FL
UAVs is necessary to improve the network performance and extend the service over longer time, the deployment of tethered UAVs is promising [32]. In that
coverage of FL-enabled networks. The multidimensional path planning and case, additional constraints, such as tether length and inclination angle need
3D placement are promising directions for future work. to be considered in future research.
 5

process one sample and the number of samples. The local Assumption 1. Fk (·) is L-Lipschitz continuous and γ-strongly
computing time for each iteration is calculated as convex
Ck Dk Fk (w) ≤ Fk (w0 ) + h∇Fk (w0 ), w − w0 i
tcp
k = , (10)
fk L
+ kw − w0 k2 , (14)
which corresponds to the energy consumption Ekcp = 2
ζk Ck Dk fk2 , where ζk is the coefficient of user k’s chip [35]. Fk (w) ≥ Fk (w0 ) + h∇Fk (w0 ), w − w0 i
γ
Denote by Nloc the number of location iterations, the total + kw − w0 k2 , (15)
time and energy consumption for local computing are given 2
as Nloc tcp cp
k and Nloc Ek , respectively. The transmission time
where w refers to as the model parameter, h·, ·i is the inner
and CPU frequency vectors are denoted as f = [f1 , . . . , fK ]T product operation, k·k denotes Euclidean norm, and ∇Fk (w0 )
and t = [tcm cm T
1 , . . . , tK ] , respectively.
is the gradient of Fk (·) with respect to w0 .
With Assumption 1, the minimum number of local iterations
E. Communication Model for each global round required to achieve the accuracy η is
derived as follows [12]:
After local training is completed, FL users transmit their  
model updates to the aerial server in the uplink for ag- 2 1
Nloc = log , (16)
gregation. In this work, frequency division multiple access (2 − Lδ)δγ η
(FDMA) is adopted since FL systems using FDMA can be where the step δ < 2/L. Similarly, to reach the global
implemented in an asynchronous fashion [12]. Other strategies, accuracy 0 , the global model should be updated at least Nglo
such as massive multi-input multi-output (MIMO) [36] and times as the following
non-orthogonal multiple access (NOMA) [37] are promising
2L2
 
provided that the synchronization among FL users is taken into 1 1
Nglo = 2 log , (17)
consideration. Another promising direction is to employ hybrid γ ξ 0 1 − η
NOMA (e.g., NOMA + FDMA) for the uplink transmission, where the hyper-learning parameter ξ < γ/L. More details
which would enable more FL users to transmit their model about the proofs of these bounds can be found in [12]. We
updates to the aggregation server. do note that [12] focuses on energy-efficient FL designs over
The transmission rate (in bits per second) of user k is wireless networks without considering the UAV deployment

gk pk
 and WPC.
Rk = bk log2 1 + , (11) From the computing and communication models, the total
bk N0
completion time of user k is as
where pk , bk , and N0 represent transmit power, bandwidth,  
Nloc Ck Dk
and noise power density of user k. The transmit power and Tk = Nglo + tcm
k , (18)
bandwidth allocation vectors are denoted as p = {p1 , . . . , pK } fk
and b = {b1 , . . . , bK }, respectively. For consistency, the where the minimum numbers of local and global iterations
model size is denoted as s, which is considered the same for all in (16) and (16) are leveraged. Similarly, the energy consump-
model updates. To guarantee that the model updates from FL tion of user k to complete the FL training is given as
users can be uploaded successfully to the aerial server within
Ek = Nglo Nloc ζk Ck Dk fk2 + tcm


the transmission time tcm k pk . (19)
k , the following constraint is imposed
Accordingly, the overall energy consumption is computed as
s ≤ Rk tcm
k , ∀k ∈ K. (12)
follows:
The energy consumed by user k for model uploading is given E=
X
Ek + ρP T
as follows: k∈K
Ekcm = tcm
k pk . (13) X
Nglo Nloc ζk Ck Dk fk2 + tcm


= k pk + ρP T, (20)
k∈K

F. Problem Formulation where ρ is a scaling coefficient that is used to balance the

In general, the energy consumption and completion time energy cost of FL users and the aerial FL server. The objective
of the FL processes (i.e., local learning and model update) function in (20) is to minimize the overall energy of both the
depend on the desirable accuracies, loss functions, and hyper- aerial server and FL users as well as optimize the learning
parameters of the learning models. More specifically, those parameters in FL. Moreover, we note that since we concern
factors directly affect the number of global iterations the global with the communication and computation aspects of FL by
model should be updated and the number of local iterations means of optimizing the UAV horizontal placement, we do
FL users should execute for each global round. To characterize not include the energy consumption in hover of the aerial
these bounds, we make the following two assumptions on the server. According to [16, Eq. (5)], the amount of power
smoothness and convexity of the loss function Fk (·). consumed by the UAV in hover can be modeled as a function
of the frame and battery weight, payload weights, gravitational
 6

acceleration, air density, and rotor disk area. Given a set of and bandwidth partitioning factors are constrained in (21b).
those parameters, the power consumption in hover of the UAV Optimizing the UAV placement also arises the nonlinearity
can be preset. This inspires future works to take account of of problem (21), which derives from the channel model,
the energy consumption in hover when the path planning is EH model, and logarithmic transmission rates. Moreover, the
considered. Moreover, the energy consumption of the UAV energy harvested by user k is nonlinear in the UAV transmit
needs to be considered when the UAV has computation tasks power, which is caused by the EH model nonlinearity. The
to process. For example, if resource-poor IoT devices cannot optimization of the UAV placement further exacerbates the
process all the tasks locally, they can offload a part of the nonlinearity. In what follows, we first transform the nonconvex
total workload to the aerial server for remote computing and problem in (21) into a more tractable form by introducing a
learning, i.e., semi-federated learning is considered. We will new slack variable. Then, an efficient iterative algorithm is
consider this promising extension in our future work. devised in this section to obtain the solution to (21).
In order to enable sustainable solutions of aerial FL commu- To solve problem (21), we start by introducing a slack vector
nication networks, we investigate a resource allocation frame- u = {uk }k∈K to overcome the nonlinearity caused by the
work that minimizes the overall energy consumption under UAV placement problem. We represent (21) as the following
various resource constraints. In particular, the optimization problem
problem of interest is formulated as X a
ν log(1/η)ζk Ck Dk fk2 + tk pk


min
min E (21a) P,p,f ,b,t,q,η,u 1−η
k∈K
P,p,f ,b,t,q,η
tcm ≥ s, ∀k ∈ K, + ρP T (22a)
s.t. k Rk (21b)  
pk g0
Ek ≤ Êknln , ∀k ∈ K, (21c) s.t. tk bk log2 1 + ≥ s, ∀k ∈ K, (22b)
bk uk
Tk ≤ T, ∀k ∈ K, (21d)  
a ν log(1/η)Ck Dk
0≤P ≤P max
, (21e) + tk
1−η fk
0 ≤ pk ≤ pmax
k , ∀k
∈ K, (21f) ≤ T, ∀k ∈ K, (22c)
X    
bk ≤ B, bk ≤ 0, ∀k ∈ K, (21g) a 1
k∈K ν log ζk Ck Dk fk2 + tk pk
fkmin ≤ fk ≤ fkmax , ∀k ∈ K, (21h) 1−η η
≤ Eknln , ∀k ∈ K, (22d)
kqk22 ≤ C 2 , (21i)
α
2 2 2
0 < η ≤ 1. (21j) uk ≥ H + kq − qk k2 , ∀k ∈ K, (22e)
(21e), (21f), (21g), (21h), (21i), (21j), (22f)
where B refers to as the system bandwidth, fkmin and fkmax
denote the minimum and maximum CPU frequencies of user where g0 = β̂0 /N0 , a = 2L2 log (1/0 ) /γ 2 ξ, ν =
k, respectively, and pmax
k and P max denote the maximum 2/(2 − Lδ)δγ, and tcm is replaced by by tk for ease of
k
transmit power of the UAV and user k, respectively. presentation. We note that a slack vector u has been leveraged
In problem (21), (21b) denotes transmission rate constraints in well-known literature, such as joint UAV trajectory and
and (21c) expresses that each user k cannot use more than resource optimization in [38] and secure MEC systems via
the harvested energy for local computing and transmission, UAV placement optimization in [39]. The harvested energy
which is calculated using the nonlinear EH model in (8). Eknln of user k is now computed as
Constraint (21d) guarantees that the completion time of users
is less than the time frame T . While the transmit power of πk − Mk νk
Eknln = T (23)
the UAV is imposed in constraint (21e), the feasible regions 1 − νk
of the transmit power of users and the bandwidth partitioning where
factors are defined in (21f) and (21g), respectively. The CPU Mk
πk = . (24)
frequency of users is imposed in (21h). Constraint (21i) defines 1 + exp(−σk (P β̂0 /uk − ιk ))
the horizontal coordinate constraint; that is, the aerial server
should position within a circle with a radius of C. The final However, finding the solution to problem (22) is still a
constraint in (21j) captures the feasible accuracy range of challenge because of its nonlinearity and nonconvexity. In
local models. Constraints (21f) and (21h) also indicates that what follows, we propose to decompose problem (22) into
users are heterogeneous with different transmit power and smaller subproblems and solve them by an iterative algorithm.
computing capabilities. In particular, the entire set of optimization variables is parti-
tioned into three block variables. In what follows, we present
how to update these block variables in more detail. We denote
III. P ROPOSED S OLUTION
the iteration index of the proposed algorithm as κ.
It is not difficult to see that the objective function (21a)
and three constraints, including (21b), (21c), and (21d), are
nonconvex. Moreover, it is challenging to solve problem (21) A. First Block Optimization
directly because there exist multiple coupled optimization The first block variables include the transmission time t,
variables. For example, the transmission time, transmit power, CPU frequencies f , and the UAV transmit power P . Given
 7

the other two blocks, the optimization problem of this block As a result, the RHS of (28) can be approximated as
(κ)
can be written as follows: ψ0 θk (ik ) − ψ0 νk . Accordingly, (28) can be represented as
X a
ν log(1/η)ζk Ck Dk fk2 + tk pk


min    
P,f ,t 1−η a 1
k∈K ν log ζk Ck Dk fk2 + tk pk
1−η η
+ ρP T (25a) (κ)

pk g0
 ≤ ψ0 θk (ik ) − ψ0 νk , (32)
s.t. tk bk log2 1 + ≥ s, ∀k ∈ K, (25b)
bk u k
  which is a convex expression since the LHS is jointly convex
a ν log(1/η)Ck Dk
+ tk ≤ T, ∀k ∈ K, (25c) in (fk , tk) and the RHS is linear in ik . Moreover, (29) is
1−η f
  k  convex since the LHS is linear in ik and the RHS is convex
a 1 in P . In sum, the following convex problem is solved at the
ν log ζk Ck Dk fk2 + tk pk
1−η η step κ in order to update the first block (t, f , P )
nln
≤ Ek , ∀k ∈ K, (25d) X a
ν log(1/η)ζk Ck Dk fk2 + tk pk


(21e), (21h). (25e) min
P,f ,t,i 1−η
k∈K
When the other blocks are given, the objective function (25a) + ρP T (33a)
is jointly convex in (fk , tk ). From (25b), the constraint of the    
a 1 2
transmission time of user k can be represented as s.t. ν log ζk Ck Dk fk + tk pk
1−η η
s
tk ≥ . (26) (κ)
≤ ψ0 θk (ik ) − ψ0 νk , ∀k ∈ K, (33b)
bk log2 (1 + pk g0 /bk uk ) !!
For given (bk , pk , uk ), (26) is a half space, and thus con- P β̂0
ik ≥ 1 + exp −σk − ιk , ∀k ∈ K, (33c)
straint (25b) is convex. Similarly, for given η, the left-hand- uk
side (LHS) of (25c) is jointly convex in (fk , tk ), whereas the (21e), (21h), (25b), (25c), (33d)
right-hand-side (RHS), i.e., the maximum completion time T ,
is constant. Accordingly, constraint (25c) is convex. where i = [i1 , . . . , iK ]T .
Constraint (25d) for user k is rewritten as
   
a 1 2
ν log ζk Ck Dk fk + tk pk
1−η η B. Second Block Optimization
πk − Mk νk
≤T . (27) For given transmission time, CPU frequencies, UAV trans-
1 − νk
mit power, and model accuracy (t, f , P, η), the optimization
This constraint is not convex because of the nonlinearity of P
problem of the second block variables (p, b, q, u) boils down
caused by the nonlinear EH model. The nonconvexity of (27)
into the following optimization problem
is handled by introducing auxiliary variables ik and rewriting
it as follows: X a
ν log(1/η)ζk Ck Dk fk2 + tk pk


min
1−η
   
a 1 ψ0 p,b,q,u
ν log ζk Ck Dk fk2 + tk pk ≤ − ψ0 νk , k∈K
1−η η ik + ρP T (34a)
(28)  
pk g0
ik ≥ 1 + exp(−σk (P β̂0 /uk − ιk )), (29) s.t. tk bk log2 1 + ≥ s, ∀k ∈ K, (34b)
b u
  k k 
where ψ0 = T Mk /(1 − νk ). However, constraint (28) is still a 1
ν log ζk Ck Dk fk2 + tk pk
not convex since both sides are convex. 1−η η
Lemma 1. Over the domain x > 0 and x̄ > 0, we have the ≤ Eknln , ∀k ∈ K, (34c)
α
following inequality uk ≥ H 2 + kq − qk k22 2 , ∀k ∈ K,


(34d)
1 2 1 (21f), (21g), (21i). (34e)
f (x) , ≥ − 2 x. (30)
x x̄ x̄
Proof. Over the domain x > 0, f (x) is a convex function Problem (34) of the block (p, b, q, u) is nonconvex because
and thus the inequality (30) can be derived from the standard of the nonlinearity caused by the coupled variables and log-
condition of the convex function f (x) [40]. arithmic operations. It is noted that the second term ρP T
can be dropped from (34) without affecting the block solution
Based on Lemma 1, the RHS of (28) can be handled as optimality.
follows: We first observe that the objective (34a) is convex as
1 2 1 (κ) it involves only transmit power variables of this block. To
≥ (κ−1) −  2 ik , θk (ik ). (31)
ik ik (κ−1) solve the nonconvex problem in (34), we first introduce the
ik
following lemma.
 8

Lemma 2. Over the domain x > 0, y > 0, τ > 0, x̄ > 0, problem (34), we need to address the nonconvexity of (34c).
ȳ > 0, τ̄ > 0, we have Constraint (34c) is equivalently represented as
 
1
f (τ, x, y) , τ ln 1 +
   
a 1 2
xy ν log ζk Ck Dk fk + tk pk
  1−η η
1
≥ 2τ̄ ln 1 +
 
x̄ȳ T Mk  1
≤  − νk  . (42)
τ̄ 2 ln(1 + 1/x̄ȳ)
   
τ̄ x y 1 − νk 1 + exp −σ P β̂ /u − ι
+ 2− − − . k 0 k k
1 + x̄ȳ x̄ ȳ τ
(35)
It is difficult to handle this constraint due to the fractional
Proof. Over the domain x > 0, y > 0, and τ > 0, f (τ, x, y) exponential expression in the RHS. To overcome this difficulty,
is a convex function [41]. Therefore, inequality (35) can be we introduce auxiliary variables zk and wk and equivalently
obtained by applying the first-order optimality condition, i.e.,, rewrite (42) as the following
f (τ̄ , x̄, ȳ) + h∇f (τ̄ , x̄, ȳ),(τ, x, y) − (τ̄ , x̄, ȳ)i.
We apply inequality (35) for constraint (34b) with zk ≥ 1 + exp (wk ) , (43)
(κ−1) ψ5
τ = bk , x = uk /pk g0 , y = bk , τ̄ = bk , x̄ = ψ3 + ψ4 pk ≤ , (44)
(κ−1) (κ−1) (κ−1)
uk /pk g0 , ȳ = bk , and thus obtain the following  zk 
approximation wk ≥ −σk P β̂0 /uk − ιk , (45)
 
pk g0 1
bk log2 1 + ≥
bk uk ln(2) where the coefficients are defined as
(κ−1)
! !
uk pk bk υk  
× λk + µk 2 − − (κ−1) − , (36) a 1 T Mk νk
pk u(κ−1) b bk ψ3 = ν log ζk Ck Dk fk2 + , (46)
k k 1−η η 1 − νk
with a
ψ4 = tk , (47)
(κ−1)
! 1−η
(κ−1) pk g0 T Mk
λk = 2bk ln 1 + (κ−1) (κ−1)
, (37) ψ5 = . (48)
bk uk 1 − νk
(κ−1) (κ−1) −1
!
(κ−1) bk uk
µk = bk 1+ , (38) Since the RHS of (43) is convex in wk and the LHS is affine in
(κ−1)
pk g0 zk , constraint (43) is convex. Fortunately, based on Lemma 1,
(κ−1)
!
(κ−1) 2 pk g0 we can approximate constraint (44) as follows:
υk = (bk ) ln 1 + (κ−1) (κ−1) . (39)
bk uk (κ)
ψ3 + ψ4 pk ≤ ψ5 θk (zk ). (49)
However, the RHS of (36) is still nonconvex because of the
existence of a fractional term. To solve this issue, we introduce
Similarly, constraint (45) can be approximated as
a slightly tricky inequality as follows:
(κ−1)  
uk pk (κ)
wk + σk P β̂0 θk (uk ) − ιk ≥ 0. (50)
pk u(κ−1)
k 
(κ−1) 2
! ! 
(κ−1) 2
1 uk pk uk pk In summary, the step κ, the block (p, b, q, u) is updated by
= (κ−1)
+ − (κ−1)
−  solving the following convex problem
4 uk pk uk pk
(κ−1) 2
! X a
ν log(1/η)ζk Ck Dk fk2 + tk pk


1 uk pk (κ) min
≤ + , Λk (pk , uk ). (40) p,b,q,u,z,w 1−η
4 u(κ−1) pk k∈K
k (51a)
Therefore, inequality (36) can be represented as follows: s.t.
(κ)
Rk (bk , pk , uk ) ≥ s/tk , ∀k ∈ K, (51b)
 
pk g0 zk ≥ 1 + exp (wk ) , ∀k ∈ K, (51c)
bk log2 1 +
bk uk (κ)
! ! ψ3 + ψ4 pk ≤ ψ5 θk (zk ), ∀k ∈ K, (51d)
1 (κ) bk υk 
(κ)

≥ λk + µk 2 − Λk (pk , uk ) − (κ−1) − wk + σk P β̂0 θk (uk ) − ιk ≥ 0, ∀k ∈ K,
ln(2) b bk
k (51e)
(κ)
, Rk (bk , pk , uk ). (41) (21f), (21g), (21i), (34d). (51f)
We note that the feasible solution set composed of three
constraints (21f), (21g), and (21i) is convex. Hence, to solve where z = [z1 , . . . , zK ]T and w = [w1 , . . . , wK ]T .
 9

Alg. 2 Bisection Method for Local Model Accuracy Alg. 3 Iterative Algorithm to Solve Problem (21) with Non-
1: Initialization: Initialize Emin , Emax , and the tolerance ε. linear Energy Harvesting (E2FL)
2: repeat 1: Set the step index κ = 0 and initialize a feasible solution
3: Set Eloa = (Emin + Emax )/2. (P (κ) , p(κ) , f (κ) , b(κ) , t(κ) , q(κ) , u(κ) , η)
4: Check the feasibility optimization problem (58). 2: repeat
5: if problem (58) is infeasible then 3: Set κ ← κ + 1.
6: Set Emin = Eloa . 4: Solve (25) to update the block (P (κ) , f (κ) , t(κ) ).
7: else 5: Solve (51) to update the block (p(κ) , b(κ) , q(κ) ).
8: Set Emax = Eloa . 6: Run Alg. 2 to update the FL model accuracy η.
9: end if 7: until The relative tolerance is less than  for two consec-
10: until Emax − Emin ≤ ε utive times.
11: Output: the FL model accuracy η.

Constraint (52b) can be equally expressed as

C. Third Block Optimization ν log(1/η)Ck Dk T
+ tk ≤ (1 − η), (56)
The third block involves the FL model accuracy η for fk a
a given solution obtained from the first and second block which is a convex constraint as the LHS is convex and the RHS
variables. is linear in η. Similarly, constraint (52c) can be rewritten as
X a
E nln (1 − η)
 
ν log(1/η)ζk Ck Dk fk2 + tk pk 1


min (52a) ν log ζk Ck Dk fk2 + tk pk ≤ k , (57)
η 1−η η a
k∈K
 
a ν log(1/η)Ck Dk which is also a convex constraint. As the objective func-
s.t. + tk ≤ T, ∀k ∈ K, (52b)
1−η f tion (52a) is quasiconvex and constraints (52b)-(52d) are all
  k 
a 1 2 convex, problem (52) is quasiconvex.
ν log ζk Ck Dk fk + tk pk
1−η η A well-known method to solve the quasiconvex problem
≤ Eknln , ∀k ∈ K, (52c) in (52) is the bisection algorithm, as shown in Alg. 2. In partic-
0 < η ≤ 1. (52d) ular, this algorithm solve the following feasibility optimization
problem at each step
To solve problem (52), we first show its quasi-convexity in
the following theorem. min Eloa (58a)
η
 
Theorem 1. Problem (52) is strictly quasiconvex. 1
s.t. ν log Ck Dk /fk + tk
η
Proof. We denote the objective function (52a) as f (η) and
T (1 − η)
rewrite it as follows: ≤ , ∀k ∈ K, (58b)

  a
a ν log(1/η)ζk Ck Dk fk2 + tk pk
P
1
k∈K ν log ζk Ck Dk fk2 + tk pk
f (η) = . (53) η
1−η
E nln (1 − η)
≤ k , ∀k ∈ K, (58c)
We observe that the numerator and denominator of f (η) are a a
convex function and a linear function, respectively. We denote 0 < η ≤ 1, (58d)
its sublevel sets as Se = {η ∈ (0 1] | f (η) ≤ e}, ∀e ∈ R+ . X   
1

Se can be equivalently represented as a ν log ζk Ck Dk fk2 + tk pk
η
k∈K
Se = {η ∈ (0 1] | g(η) ≤ 0}, ∀e ∈ R+ , (54) ≤ Eloa (1 − η) , (58e)

where where Eloa is an auxiliary parameter. Alg. 2 starts by initial-

X izing the interval [Emin Emax ] that is guaranteed to contain
g(η) = a(ν log(1/η)ζk Ck Dk fk2 + tk pk ) − e (1 − η) . the optimal solution Eloa . At each step, the feasibility of prob-
k∈K lem (58) is solved at the midpoint Eloa = (Emin + Emax )/2
(55) to check whether the optimal value lies in the upper half or
If Se is a strictly convex set if for any η1 , η2 ∈ Se and any 0 < the lower half. This procedure is repeated until the interval
θ < 1, θη1 + (1 − θ) η2 ∈ Se [40]. This condition holds when Emax − Emin is less than a tolerance ε.
g (θη1 + (1 − θ) η2 ) ≤ 0. Here, it is not difficult to show that
g(η) is a strictly convex function with respect to η. Therefore,
g (θη1 + (1 − θ) η2 ) ≤ θg(η1 ) + (1 − θ) g(η2 ) for any 0 < D. Proposed Algorithm of Sustainable FL
θ < 1. Since η1 , η2 ∈ Se , we have g(η1 ) ≤ 0, g(η2 ) ≤ 0, and The proposed iterative algorithm, namely E2FL, to solve
the condition θη1 + (1 − θ) η2 ∈ Se holds. Consequently, Se the optimization problem in (22) is presented in Alg. 3. Each
is convex, and thus f (η) is quasiconvex. iteration involves three subproblems of three block variables,
 10

which are solved in a cyclical fashion. Alg. 3 ends when 8

K = 80
the relative tolerance falls below the threshold  for two K = 70
consecutive times. For the results presented below, we set the 7 K = 60

Total Energy Consumption (J)

K = 50
threshold  to be 10−3 . As Alg. 2 is adopted, the bisection K = 40
6
method eventually reaches a unique optimal solution of the K = 30
local model accuracy at each step due to strict quasiconvexity
5 Increasing the
of problem (52). Moreover, since the objective function is number of FL users
bounded and increases after each block optimization, Alg. 3
4
generates a sequence of improved feasible solutions and is
finally convergent. According to [42], Alg. 3 has a sublinear
3
convergence rate and O(log(1/)) update steps are needed for
Alg. 3 to reach the -optimal solution. By simulation, we will
2
show that Alg. 3 takes only a few iterations to converge.
Since problem (33) involves (3K + 1) variables and 1
5K + 1 constraints, the computational complexity is O (O1 ) 1 2 3 4 5 6 7 8 9 10
2 2.5 3.5 Iteration Index
with O1 = (3K + 1) (5K + 1) + (5K + 1) . The
complexity to solve problem (51) is O (O2 ) with O2 = Fig. 3. Convergence curve of the E2FL algorithm.
2 2.5 3.5
(5K + 2) (6K + 2) + (6K + 2) as it involves (5K + 2)
variables and (6K +2) constraints. Á a result, the total compu-
tational complexity of Alg. 3 is O (I (O1 + O2 + O3 )). Here, that the relative difference between the two consecutive update
O3 is the complexity of solving the quasi-convex problem steps falls below a threshold  two consecutive times. In other
in (52), which is O3 = log2 (1/ε) according to [40]. words, the algorithm stops when the following occurs
E (τ ) − E (τ −1)
IV. P ERFORMANCE E VALUATION ≤  for two consecutive times, (59)
E (τ −1)
This section provides numerical results to evaluate the E2FL where
algorithm and compare it with four benchmarks. X a
E (τ ) =ρP (τ ) T + ×
k∈K
1 − η (τ )
A. Simulation Settings    2 
1 
(τ ) (τ ) (τ )
We simulate an FL network with an aerial server and ν log ζ C
k k kD fk + t k pk . (60)
η (τ )
K = 50 users. FL users are randomly positioned within
a circular area with a radius of 20 m, whereas the UAV We also note that iteration is a fully-cycle update, i.e., three
placement is optimized via our algorithm to improve the blocks are updated cyclically. We present the convergence
learning efficiency. The altitude H = 5 m, the bandwidth curve of the E2FL algorithm in Fig. 3 when varying the
B = 20 MHz, the maximum transmit powers P max = 30 dBm number of FL users from 30 to 80. As shown in the figure, the
and pmax = 10 dBm, the noise power N0 = −174 dBm/Hz, E2FL algorithm takes only a few iterations to reach the optimal
k
and the pathloss exponent α = 2.01. The parameters of the solution, demonstrating the efficiency of the E2FL algorithm,
nonlinear EH model are identical for all users, which are as presented in Alg. 3.
M = 24 mW (i.e., the maximum energy each user can harvest In the following, the proposed E2FL algorithm is compared
is 24 mW), σk = 1500, and ιk = 0.0014 [43]. The stopping with the four benchmarks, including FCF with fixed CPU
measure used to stop Alg. 3 is set as  = 10−3 . For the frequencies of FL users, FTT with fixed transmission, FUP
local computing model, we set the maximum and minimum with fixed placement of the aerial server, and FLA with fixed
CPU frequencies of FL users as fkmax = 2.0 GHz and local model accuracy. Here we add a note that the scheme
fkmin = 0.1 GHz, respectively, the model size is s = 50 Kb, proposed in [26] is similar to FUP which considers a fixed
and the coefficient ζk = 10−28 for all users. Further, the deployment of the FL server for the implementation of FL
data size Dk (in Mb) and the computation workload Ck (in over EH networks. Similar to the FLA scheme, the work [29]
cycles per bit) are random values in [5, 10] Mb and[10, 20], considers one communication round and does not take account
respectively. For the smoothness and convexify of the loss of the local model accuracy (i.e., the local accuracy is preset).
functions, we set γ = 2 and L = 4, and further set the hyper- 2) Performance versus the maximum completion time:
learning parameters δ = 1/4 < 2/L and ξ = 1/3 < γ/L [12]. Fig. 4 shows the overall energy consumption when the maxi-
Finally, unless stated otherwise, the global accuracy 0 are set mum completion time T increases from 50 to 90. We notice
to be 10−3 . that the energy increases with the increase in the maximum
completion time. It is since the local accuracy η, CPU fre-
quency fk , and transmission time tk are relaxed when the
B. Simulation Results maximum completion time T is higher, thus reducing the
1) Convergence Analysis: The convergence rate of the overall energy consumption. Because the energy consumption
proposed algorithm, namely E2FL, is firstly investigated. We of local computing increases quadratically with the CPU
note that the stopping criteria used to stop the algorithm is frequency, the FCF with fixed CPU frequencies does not
 11

4.5 4.5

4
4
Total Energy Consumption (J)

Total Energy Consumption (J)

3.5

E2FL 3.5
3 FCF
FLA
FTT
2.5 FUP 3

2
E2FL
2.5 FCF
1.5 FLA
FTT
FUP
1 2
50 60 70 80 90 30 40 50 60 70 80
Maximum Completion Time (s) Model Size (Kb)

Fig. 4. Overall energy consumption vs. the maximum completion time. Fig. 6. Overall energy consumption vs. the model size.

4.5
from 30 Kb to 80 Kb. From the figure, the overall energy
consumption increases when the model size becomes larger.
4
Total Energy Consumption (J)

The explanation for this is that in order to transmit a larger

model to the aerial server, each FL user needs to use a higher
3.5
transmit power or increase the transmission time, all of which
lead to more energy.
We also plot the overall energy consumption of FL users
3 and the transmit power of the aerial server when the model
size varies. As shown in Fig. 7, the FUP scheme with a preset
E2FL location of the aerial server results in similar overall energy
2.5 FCF
FLA
consumption of FL users; however, the transmit power of the
FTT UAV is very high. It is reasonable since the UAV placement
FUP
2 is not optimized, the channel conditions between the aerial
2 3 4 5 6 server and some FL users are not good, and thus the UAV as
Aerial Server Altitude (m)
the energy source should increase its transmit power so that
Fig. 5. Overall energy consumption vs. the aerial server altitude. these FL users can harvest sufficient energy for performing the
local training and transmitting the model updates. Moreover,
the CPU frequencies of FL users are not optimized in the
exploit the increase in the maximum completion time. Hence, FCF scheme, and the energy consumption for local computing
the overall energy consumption of the FCF scheme reduces at a increases quadratically with the CPU frequency. Accordingly,
slower speed when compared with the other approaches. From the overall energy consumption of FL users is the highest
the figure, the E2FL algorithm achieves better performance in the FCF scheme among the benchmarks, as illustrated in
than all the benchmarks for different maximum completion Fig. 7a.
times. 5) Performance versus the global model accuracy: The
3) Performance versus the UAV altitude: We plot the over- performance evaluation of the proposed E2FL and the bench-
all energy consumption of all the approaches when varying marks is compared in Fig. 8 for different global model
the aerial server (i.e., UAV) altitude. As shown in Fig. 5, the accuracies. From the figure, the overall energy consumption
energy increases as the server altitude increases. It is since the of FL users and the aerial server decreases as the global
UAV should transmit higher transmit powers to compensate for model accuracy increases. It is reasonable since the higher
the worst channel conditions, which is implicitly indicated in the global accuracy is, the larger the number of global
the nonlinear EH model (8) and constraints (22d). Moreover, rounds is. For example, with the defaults parameters and
because of worse channel gains, the energy consumed by FL η = 0.5, {332, 255, 222, 144, 111} global rounds are required
users is higher to transmit their model updates to the server, to achieve the accuracies 0 = {10−3 , 5 × 10−3 , 10−2 , 5 ×
as shown in (22b). The figure also suggests the importance 10−2 , 10−1 }, respectively. Accordingly, FL users consume
of a joint resource optimization framework, as devised by our more energy to finish the training within the maximum
E2FL approach in Alg. 3. completion time. For a relatively low global accuracy (e.g.,
4) Performance versus the model size: The performance η0 = 10−1 ), it is observed that the performance of the
in terms of the overall energy consumption of different ap- alternative approaches becomes closer to that of the proposed
proaches is plotted in Fig. 6 when the model size varies E2FL algorithm. However, the E2FL algorithm is significantly
 12

Overall Energy Consumption of FL Users (J) 2.5 4.5

4 E2FL
FCF

Total Energy Consumption (J)

3.5 FLA
FTT
FUP
2 3

2.5

1.5 1.5

E2FL 1
FCF
FLA
0.5
FTT
FUP
1 0
30 40 50 60 70 80 1e-3 5e-3 1e-2 5e-2 1e-1
Model Size (Kb) Global Model Accuracy
(a)
Fig. 8. Overall energy consumption vs. the global model accuracy.
0.5
E2FL
FCF
FLA V. C ONCLUSION
0.45
Transmit Power of the UAV (W)

FTT
FUP The application of aerial communications for FL-enabled
0.4 wireless networks has been investigated in this work. Par-
ticularly, we have deployed a UAV as the FL server for
aerial model aggregation and the energy source to wirelessly
0.35
power energy-limited FL users via a practical nonlinear EH
model. An energy-efficient FL framework has been designed
0.3
to minimize the overall energy consumption of the aerial
server and FL users under various resource constraints. The
0.25 nonlinearity and nonconvexity of the formulated problem have
been addressed by an efficient iterative algorithm. Further, we
0.2 have provided various simulations to show the application
30 40 50 60 70 80
Model Size (Kb) of UAV communications in enhancing the performance of
(b) FL-enabled wireless networks. The simulation results have
also justified the importance of jointly optimizing all the
Fig. 7. Performance evaluation vs. the model size: (a) energy consumption
of FL users vs. model size; (b) transmit power of the UAV vs. model size. optimization variables, as investigated in the proposed E2FL
algorithm when it is compared with ones that only optimize
a subset of optimization variables, such as the FL server
location, computing resources, and learning model accuracy.
superior to the benchmarks when the global accuracy is large
(e.g., η0 = 10−3 ). This emphasizes the need to jointly optimize R EFERENCES
UAV placement, computing resources, local accuracy, and [1] K. Lin, Y. Li, Q. Zhang, and G. Fortino, “AI-driven collaborative
transmission time to lower the energy consumption of the resource allocation for task execution in 6G enabled massive IoT,” IEEE
Internet of Things Journal, vol. 8, no. 7, pp. 5264–5273, Apr. 2021.
aerial server and FL users. [2] H. Yang, A. Alphones, Z. Xiong, D. Niyato, J. Zhao, and K. Wu,
“Artificial-intelligence-enabled intelligent 6G networks,” IEEE Network,
Finally, we show the overall energy consumption of FL vol. 34, no. 6, pp. 272–280, Nov.-Dec. 2020.
users and the transmit power of the aerial server for different [3] C. T. Dinh, N. H. Tran, M. N. Nguyen, C. S. Hong, W. Bao, A. Y.
global accuracies. Similar to the observation from Fig. 8, the Zomaya, and V. Gramoli, “Federated learning over wireless networks:
Convergence analysis and resource allocation,” IEEE/ACM Transactions
performance discrepancies among the approaches reduce when on Networking, vol. 9, no. 1, pp. 398–409, Feb. 2021.
the FL model does not need to be learned accurately (i.e., [4] W. Y. B. Lim, J. S. Ng, Z. Xiong, J. Jin, Y. Zhang, D. Niyato,
the global model accuracy is low or the 0 value is high). C. Leung, and C. Miao, “Decentralized edge intelligence: A dynamic
resource allocation framework for hierarchical federated learning,” IEEE
Same as before, Fig. 9a emphasizes the need to optimize Transactions on Parallel and Distributed Systems, vol. 33, no. 3, pp.
computing resources because the energy consumption of the 536–550, Mar. 2022.
FCF approach is the highest out of the benchmarks. From [5] J. Konečný, H. B. McMahan, F. X. Yu, P. Richtarik, A. T. Suresh, and
D. Bacon, “Federated learning: Strategies for improving communication
Fig. 9b, the FUP scheme results in the highest transmit power efficiency,” in NIPS Workshop on Private Multi-Party Machine Learning,
of the UAV when the global model accuracy is sufficiently Barcelona, Spain, Dec. 2016.
large (i.e., the 0 value is low). This result again confirms [6] W. Y. B. Lim, N. C. Luong, D. T. Hoang, Y. Jiao, Y.-C. Liang, Q. Yang,
D. Niyato, and C. Miao, “Federated learning in mobile edge networks:
the advantage of deploying UAVs as the aerial FL server over A comprehensive survey,” IEEE Communications Surveys & Tutorials,
conventional FL systems and FL-enabled WPC networks. vol. 22, no. 3, pp. 2031–2063, Third Quarter 2020.
 13

2.5 0.45
E2FL
Overall Energy Consumption of FL Users (J)
E2FL
FCF FCF
FLA 0.4
FLA

Transmit Power of the UAV (W)

2 FTT FTT
FUP 0.35 FUP

0.3
1.5
0.25

0.2
1
0.15

0.5 0.1

0.05

0 0
1e-3 5e-3 1e-2 5e-2 1e-1 1e-3 5e-3 1e-2 5e-2 1e-1
Global Model Accuracy Global Model Accuracy
(a) (b)
Fig. 9. Performance evaluation vs. the global model accuracy: (a) energy consumption of FL users vs. the aerial server altitude; (b) transmit power of the
UAV vs. the aerial server altitude.

[7] D. C. Nguyen, M. Ding, P. N. Pathirana, A. Seneviratne, J. Li, and erated learning in the sky: Aerial-ground air quality sensing framework
H. V. Poor, “Federated learning for internet of things: A comprehensive with UAV swarms,” IEEE Internet of Things Journal, vol. 8, no. 12, pp.
survey,” IEEE Communications Surveys & Tutorials, vol. 23, no. 3, pp. 9827–9837, Jun. 2021.
1622–1658, Third Quarter 2021. [21] L. Xie, J. Xu, and Y. Zeng, “Common throughput maximization for
[8] D. C. Nguyen, Q.-V. Pham, P. N. Pathirana, M. Ding, A. Seneviratne, UAV-enabled interference channel with wireless powered communica-
Z. Lin, O. A. Dobre, and W.-J. Hwang, “Federated learning for smart tions,” IEEE Transactions on Communications, vol. 68, no. 5, pp. 3197–
healthcare: A survey,” ACM Computing Surveys (CSUR), vol. 55, no. 3, 3212, May 2020.
pp. 1–37, Apr. 2023. [22] W. Feng, N. Zhao, S. Ao, J. Tang, X. Zhang, Y. Fu, D. K. So, and K.-K.
[9] Z. Zhao, C. Feng, H. H. Yang, and X. Luo, “Federated-learning- Wong, “Joint 3D trajectory design and time allocation for UAV-enabled
enabled intelligent fog radio access networks: Fundamental theory, key wireless power transfer networks,” IEEE Transactions on Vehicular
techniques, and future trends,” IEEE wireless communications, vol. 27, Technology, vol. 69, no. 9, pp. 9265–9278, Sep. 2020.
no. 2, pp. 22–28, Apr. 2020. [23] X. Yuan, T. Yang, Y. Hu, J. Xu, and A. Schmeink, “Trajectory design for
[10] S. A. Rahman, H. Tout, C. Talhi, and A. Mourad, “Internet of things UAV-enabled multiuser wireless power transfer with nonlinear energy
intrusion detection: Centralized, on-device, or federated learning?” IEEE harvesting,” IEEE Transactions on Wireless Communications, vol. 20,
Network, vol. 34, no. 6, pp. 310–317, Nov.-Dec. 2020. no. 2, pp. 1105–1121, Feb. 2021.
[11] Y. Liu, S. Garg, J. Nie, Y. Zhang, Z. Xiong, J. Kang, and M. S. [24] J. Xu, Y. Zeng, and R. Zhang, “UAV-enabled wireless power transfer:
Hossain, “Deep anomaly detection for time-series data in industrial IoT: Trajectory design and energy optimization,” IEEE Transactions on
a communication-efficient on-device federated learning approach,” IEEE Wireless Communications, vol. 17, no. 8, pp. 5092–5106, Aug. 2018.
Internet of Things Journal, vol. 8, no. 8, pp. 6348–6358, Apr. 2021. [25] F. Zhou, Y. Wu, R. Q. Hu, and Y. Qian, “Computation rate maximization
[12] Z. Yang, M. Chen, W. Saad, C. S. Hong, and M. Shikh-Bahaei, “Energy in UAV-enabled wireless-powered mobile-edge computing systems,”
efficient federated learning over wireless communication networks,” IEEE Journal on Selected Areas in Communications, vol. 36, no. 9,
IEEE Transactions on Wireless Communications, vol. 20, no. 3, pp. pp. 1927–1941, Sep. 2018.
1935–1949, Mar. 2021. [26] R. Hamdi, M. Chen, A. B. Said, M. Qaraqe, and H. V. Poor, “Federated
[13] T. Zhang and S. Mao, “Energy-efficient federated learning with intelli- learning over energy harvesting wireless networks,” IEEE Internet of
gent reflecting surface,” IEEE Transactions on Green Communications Things Journal, vol. 9, no. 1, pp. 92–103, Jan. 2022.
and Networking, 2021, in press. [27] Q. V. Do, Q.-V. Pham, and W.-J. Hwang, “Deep reinforcement learning
[14] C. De Alwis, A. Kalla, Q.-V. Pham, P. Kumar, K. Dev, W.-J. Hwang, for energy-efficient federated learning in UAV-enabled wireless powered
and M. Liyanage, “Survey on 6G frontiers: Trends, applications, re- networks,” IEEE Communications Letters, vol. 26, no. 1, pp. 99–103,
quirements, technologies and future research,” IEEE Open Journal of Jan. 2022.
the Communications Society, vol. 2, pp. 836–886, Apr. 2021. [28] Q.-V. Pham, M. Zeng, T. Huynh-The, Z. Han, and W.-J. Hwang, “Aerial
[15] X. You et al., “Towards 6G wireless communication networks: Vision, access networks for federated learning: Applications and challenges,”
enabling technologies, and new paradigm shifts,” Science China Infor- IEEE Network, 2022, in press.
mation Sciences, vol. 64, no. 1, pp. 1–74, Jan. 2021. [29] Q.-V. Pham, M. Zeng, R. Ruby, T. Huynh-The, and W.-J. Hwang, “UAV
[16] N.-N. Dao, Q.-V. Pham, N. H. Tu, T. T. Thanh, V. N. Q. Bao, D. S. communications for sustainable federated learning,” IEEE Transactions
Lakew, and S. Cho, “Survey on aerial radio access networks: Toward a on Vehicular Technology, vol. 70, no. 4, pp. 3944–3948, Apr. 2021.
comprehensive 6G access infrastructure,” IEEE Communications Surveys [30] Q.-V. Pham, M. Le, T. Huynh-The, Z. Han, and W.-J. Hwang, “UAV-
and Tutorials, vol. 23, no. 2, pp. 1193–1225, Second Quarter 2021. enabled wireless powered communication for energy-efficient federated
[17] L. Xu, M. Chen, M. Chen, Z. Yang, C. Chaccour, W. Saad, and C. S. learning,” in IEEE International Conference on Communications (ICC),
Hong, “Joint location, bandwidth and power optimization for THz- Seoul, Korea, May 2022.
enabled UAV communications,” IEEE Communications Letters, vol. 25, [31] Z. Yang, W. Xu, and M. Shikh-Bahaei, “Energy efficient UAV com-
no. 6, pp. 1984–1988, Jun. 2021. munication with energy harvesting,” IEEE Transactions on Vehicular
[18] “Ericsson partners with Vodafone UK to speed up network Technology, vol. 69, no. 2, pp. 1913–1927, Feb. 2020.
upgrades with drone surveys,” Apr. 15, 2021. [Online]. Available: [32] V. U. Pai and B. Sainath, “UAV selection and link switching policy for
https://www.ericsson.com/en/press-releases/3/2021/4/ericsson-partners- hybrid tethered UAV-assisted communication,” IEEE Communications
with-vodafone-uk-to-speed-up-network-upgrades-with-drone-surveys Letters, vol. 25, no. 7, pp. 2410–2414, Jul. 2021.
[19] H. Zhang and L. Hanzo, “Federated learning assisted multi-UAV net- [33] T.-T. Nguyen, Q.-V. Pham, V.-D. Nguyen, J.-H. Lee, and Y.-H. Kim,
works,” IEEE Transactions on Vehicular Technology, vol. 69, no. 11, “Resource allocation for energy efficiency in OFDMA-enabled WPCN,”
pp. 14 104–14 109, Nov. 2020. IEEE Wireless Communications Letters, vol. 9, no. 12, pp. 2049–2053,
[20] Y. Liu, J. Nie, X. Li, S. H. Ahmed, W. Y. B. Lim, and C. Miao, “Fed- Dec. 2020.
 14

[34] E. Boshkovska, D. W. K. Ng, N. Zlatanov, and R. Schober, “Practical [39] Y. Zhou, C. Pan, P. L. Yeoh, K. Wang, M. Elkashlan, B. Vucetic, and
non-linear energy harvesting model and resource allocation for SWIPT Y. Li, “Secure communications for UAV-enabled mobile edge computing
systems,” IEEE Communications Letters, vol. 19, no. 12, pp. 2082–2085, systems,” IEEE Transactions on Communications, vol. 68, no. 1, pp.
Dec. 2015. 376–388, Jan. 2020.
[35] Q.-V. Pham, H. T. Nguyen, Z. Han, and W.-J. Hwang, “Coalitional [40] S. Boyd, S. P. Boyd, and L. Vandenberghe, Convex optimization.
games for computation offloading in NOMA-enabled multi-access edge Cambridge university press, 2004.
computing,” IEEE Transactions on Vehicular Technology, vol. 69, no. 2, [41] A. A. Nasir, H. D. Tuan, T. Q. Duong, and H. V. Poor, “UAV-enabled
pp. 1982–1993, Feb. 2020. communication using NOMA,” IEEE Transactions on Communications,
[36] T. T. Vu, D. T. Ngo, N. H. Tran, H. Q. Ngo, M. N. Dao, and R. H. vol. 67, no. 7, pp. 5126–5138, Jul. 2019.
Middleton, “Cell-free massive MIMO for wireless federated learning,” [42] M. Hong, M. Razaviyayn, Z.-Q. Luo, and J.-S. Pang, “A unified
IEEE Transactions on Wireless Communications, vol. 19, no. 10, pp. algorithmic framework for block-structured optimization involving big
6377–6392, Oct. 2020. data: With applications in machine learning and signal processing,” IEEE
[37] H. Sun, X. Ma, and R. Q. Hu, “Adaptive federated learning with gra- Signal Processing Magazine, vol. 33, no. 1, pp. 57–77, Jan. 2016.
dient compression in uplink NOMA,” IEEE Transactions on Vehicular [43] T.-T. Nguyen, V.-D. Nguyen, Q.-V. Pham, J.-H. Lee, and Y.-H. Kim,
Technology, vol. 69, no. 12, pp. 16 325–16 329, Dec. 2020. “Resource allocation for AF relaying wireless-powered networks with
[38] Z. Li, M. Chen, C. Pan, N. Huang, Z. Yang, and A. Nallanathan, “Joint nonlinear energy harvester,” IEEE Communications Letters, vol. 25,
trajectory and communication design for secure UAV networks,” IEEE no. 1, pp. 229–233, Jan. 2021.
Communications Letters, vol. 23, no. 4, pp. 636–639, Apr. 2019.

View publication stats