You are on page 1of 8

ACCEPTED FROM OPEN CALL

UAV-assisted Communication Efficient Federated Learning in the Era of the Artificial


Intelligence of Things
Wei Yang Bryan Lim, Sahil Garg, Zehui Xiong, Yang Zhang, Dusit Niyato, Cyril Leung, and Chunyan Miao

Abstract server for processing. However, this is no longer


sustainable due to the following reasons. First,
Artificial Intelligence (AI) based models are privacy laws, for example, the General Data Pro-
increasingly deployed in the Internet of Things tection Regulation (GDPR) [1], are increasingly
(IoT), paving the evolution of the IoT into the stringent. In addition, the increasingly privacy-sen-
AI of things (AIoT). Currently, the predominant sitive data owners can opt out of data sharing with
approach for AI model training is cloud-centric third-parties. Second, the transfer of massive quan-
and involves the sharing of data with external par- tities of data to the distant cloud burdens the net-
ties. To preserve privacy while enabling collabora- works and incurs unacceptable latency especially
tive model training across distributed IoT devices, for time-sensitive tasks. This potentially impedes the
the machine learning paradigm called Federated further development of the AIoT.
Learning (FL) has been proposed. The future FL In the face of these challenges, Federated
network is envisioned to involve up to millions Learning (FL) has first been proposed in [2]. In
of distributed IoT devices involved in collabora- each iteration of FL, a model owner transmits a
tive learning. However, communication failures set of model parameters to each data owner, for
and dropouts by nodes can lead to inefficient FL. example, IoT devices. Then, the model parame-
Inspired by the UAV-assisted communications in ters are updated with the locally stored training
5G heterogeneous networks (HetNet), we pro- data. Only the updated model parameters, rather
pose the UAV-assisted FL in this article. The FL than the raw data, need to be transmitted back to
model owner may employ UAVs to provide the the model owner for global aggregation. The FL
intermediate model aggregation in the sky and training process thus enables privacy preserving
mobile relay of the updated model parameters collaborative machine learning.
from data owners to the model owner. This there- However, the communication inefficiency
fore increases the reach of FL to data owners that remains a key bottleneck in FL. The FL network
face uncertain network conditions and improves can involve thousands of distributed IoT devices.
the communication efficiency. To incentivize the The communication cost thus often dominates the
UAV service providers, we adopt the multi-di- computation cost. In addition, node failures and
mensional contract incentive design as a case device dropouts due to communication failures
study. The incentive compatibility of the contract can lead to inefficient FL, for example, in terms of
ensures that the UAVs only choose an incentive lower model accuracy or slower convergence. As
package corresponding to its type, for example, such, it is of paramount importance to improve
traveling cost. The simulation results show that the communication efficiency of FL. Few solutions
the UAV-assisted FL achieves significant improve- exist to improve the communication efficiency in
ment in communication efficiency and validates FL, for example, worker selection protocols and
the incentive compatibility of our contract design. compression techniques [3]. However, they do
not well-address the issue. For example, data own-
Introduction ers with severely limited network access are still
The recent advances of the Internet of Things (IoT), unable to participate in the FL process.
Artificial Intelligence (AI), and communication tech- Inspired by recent advancements in the
nologies are gradually paving the evolution of the Unmanned Aerial Vehicle (UAV) assisted commu-
IoT into the Artificial Intelligence of Things (AIoT). In nications as part of the 5G heterogeneous networks
particular, the enhanced perception capabilities of (HetNets) [4], we propose UAV-assisted commu-
IoT devices enable a wealth of data to be collected. nication for FL in this article. To improve the com-
Coupled with the advent of 5G communications, munication efficiency, the FL model owners can
a seamlessly connected network of AI-empowered employ the third-party UAV service providers, for
IoT devices is enabled to bring intelligence to both example, Drones-as-a-Service (DaaS), to provide
consumers and industrial end users. network coverage and relay support from the sky,
Currently, the predominant approach for AI thereby increasing the reach to more devices and
model training is cloud-centric, that is, the IoT reducing the device dropout rate during training.
devices transmit the training data to a distant cloud However, there exist various UAV types in the mar-

Wei Yang Bryan Lim, Dusit Niyato, and Chunyan Miao are with Nanyang Technological University (NTU); Sahil Garg is with the Université du Québec;
Digital Object Identifier: Zehui Xiong is with Singapore University of Technology and Design, Yang Zhang (corresponding author) is with Nanjing University of
10.1109/MNET.002.2000334 Aeronautics and Astronautics; Cyril Leung is with The University of British Columbia.

188 0890-8044/21/$25.00 © 2021 IEEE IEEE Network • September/October 2021

XIONG_LAYOUT.indd 188 10/22/21 10:18 PM


Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:54:20 UTC from IEEE Xplore. Restrictions apply.
IoT IoT

2. Transmission 3. Aggregation 2. Transmission


1. Local model training 1. Local model training

2. Transmission 2. Transmission

1. Local model training 1. Local model training

IoT IoT

FIGURE 1. The FL working process. The model training is conducted locally. Then, only the model parameters
are subsequently exchanged for aggregation.

ket, for example, different in terms of mobility and aggregated through IoT sensing, for example, tem-
communication capabilities. The types are hidden perature readings or images. Note that these private
from the model owner amid information asymmetry, data may not be willingly shared with the model
that is, the DaaS providers may misreport their UAV owner or among other workers. On one hand, the
types to achieve greater rewards for their services. other workers may be competitors, for example, in
As such, to appropriately incentivize the DaaS pro- the case of industrial applications. On the other hand,
viders, we leverage the multi-dimensional contract the model owner may be a third-party developer of
theoretic incentive mechanism design. The incentive AI applications with no affiliations to the data owners.
compatibility of the contract ensures that each DaaS In the FL process, the model owner first trans-
provider only selects the contract corresponding mits a set of model parameters to each worker.
to its actual type, thus ensuring profit maximization This set of parameters may be randomized, or
amid information asymmetry. The contributions of may comprise parameters from a previous train-
this article is as follows: ing round. The FL process consists of a number
• We provide a tutorial of the fundamentals of of iterations, each of which consists of three main
FL. Then, we review the communication relat- steps as shown in Fig. 1:
ed issues and existing solutions proposed to 1. Local Model Training: The worker trains the
improve the communication efficiency. received model using the local sensing data,
• We provide an overview of the issues and for example, through the Stochastic Gradient
applications related to UAV-assisted commu- Descent to minimize a local loss function.
nication in 5G HetNet. Then, we propose 2. Wireless Transmission: The worker transmits
the UAV-assisted communication efficient FL. the updated local parameters to the model
• We develop the multi-dimensional con- owner.
tract-matching incentive mechanism design 3. Global Parameter Update: The local parame-
to reward the DaaS providers amid informa- ter updates from all workers are aggregated
tion asymmetry. and averaged in the case of the conventional
Federated Averaging algorithm [2].
fundAmentAls of federAted leArnIng After Step 3, the model parameters are trans-
mitted back to the worker for the next iteration of
federAted leArnIng training. The process continues until a desired test
The two main entities in the FL system are the data accuracy is achieved.
owners and model owners (Fig. 1). The data own- In fact, Step 3 of each iteration is usually con-
ers, hereinafter workers, each own a private dataset ducted in a synchronous manner, that is, the glob-

IEEE Network • September/October 2021 189

XIONG_LAYOUT.indd 189 10/22/21 10:18 PM


Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:54:20 UTC from IEEE Xplore. Restrictions apply.
al parameter update is conducted only after all devices exceeds that of the uplink speed. How-
workers have completed the local model training. ever, the model parameter updates have to be
In contrast, an asynchronous aggregation implies transmitted to the model owner for aggregation
that the global parameter is updated whenever after local training. The delay in wireless transmis-
the model owner receives a set of local param- sion of complex, high-dimensional model updates
eter updates from a worker. Though the asyn- during the FL training process can lead to the
chronous scheme enables workers to join while straggler effect. The synchronous FL global aggre-
a training round is in progress, there have been gation only occurs after every worker’s parameter
convergence issues. As such, most studies of FL update has been transmitted to the model owner.
presently adopt the synchronous scheme. This implies that each iteration of model training
only takes place as quickly as the slowest worker.
Applications of Federated Learning Besides the inefficiency in transmission of local
The main advantages of FL are as follows: model updates, there exists a significant num-
• FL enables privacy preserving collaborative ber of worker dropouts while training is taking
machine learning. place due to network and communication fail-
• FL leverages the computation capabilities of end ures. Moreover, the dropout frequencies are
IoT devices for local model training, thus reduc- time-correlated, that is, dropouts are more likely
ing the computation workload of the cloud. to occur in the daytime, and be geographically
• Model parameters are usually smaller in size correlated, for example, workers from a poorly
than raw data, thus alleviating the burden on connected region may have more limited band-
backbone communication networks. width and thus collectively have higher dropout
This has enabled several practical applications of FL. rates. The worker dropouts can lead to reduced
Mobile Keyboard Prediction: The next-word- model accuracy given a decline in the number of
prediction model of Gboard (gboard.app.goo.gl/ available training samples. The FL model may also
get), developed by Google, has been successfully under-represent the data distribution from regions
trained using FL. This ensures that sensitive data, with limited connectivity.
for example, user text messages, do not need to
be shared with a third-party server. Review of Potential Solutions
Healthcare: Nvidia Clara FL (https://blogs.nvid- In the following, we review the potential solutions
ia.com/blog/2019/12/01/clara-federated-learn- toward improving communication efficiency in FL.
ing/) has been exploring the use of FL to enable Edge Computing: The edge computing par-
collaborative learning across multiple hospitals. adigm utilizes the enhanced computation capa-
This guarantees that the sensitive data of patients bilities of end devices and edge servers to bring
remain kept in the local servers of hospitals. computation closer to where data is located, that
Finance: WeBank FedAI (https://www.fedai. is, at the network edge. Solutions from edge com-
org/) has explored the use of FL to enable collab- puting have been incorporated to solve the com-
oration across financial institutions while adher- munication issues in FL. The study in [2] proposes
ing to user privacy protection regulations in the that each worker can implement multiple rounds of
finance industry. training passes over the local dataset before com-
Recommender System: To serve the interests munication for aggregation. This implies that more
of users while meeting privacy laws, several stud- local updates are computed before each round of
ies have also explored the use of FL to build rec- global communication, which reduces the number
ommender systems. The comprehensive interest of communication rounds required, at the expense
profile of a user group can be derived across plat- of increased computation expense on end devices.
forms for e-commerce, such as Taobao, to enter- Alternatively, the hierarchical FL [7] can be adopt-
tainment applications such as YouTube. ed, in which the workers do not communicate
Internet of Flying Things: UAVs are increas- directly with the model owner, that is, a distant
ingly utilized in smart city applications (https:// cloud server. Instead, the local parameters are first
spectrum.ieee.org/tech-talk/sensors/remote-sens- communicated to edge servers for intermediate
ing/uavs-prove-usefuldelivering-remote-pow- aggregation. Then, communication with the model
er-charging-services) [5], for example, to enable owner is further established for global aggregation.
data collection or for package delivery. Recent Besides reducing the instances of communication
studies have proposed the use of FL for the coor- with the distant servers of the model owner, this
dination of independently owned UAV swarms or relay reduces the dropout rate of devices.
for FL model training using UAV sensing data [6]. Worker Selection Protocols: With the expecta-
The FL network for the above applications can tion of worker dropouts, several studies propose
potentially involve up to millions of distributed IoT the selection of an excess of the required number
devices, each with heterogeneous network condi- of workers to participate in training. Moreover,
tions and communication capabilities. As such, it resource-aware worker selection algorithms are
is of paramount importance to improve the com- proposed, for example, to select only workers
munication efficiency in FL. with reliable network connections or sufficient
energy capacity to complete the FL training. How-
Communication Efficiency in ever, the study in [8] finds that the selection algo-
rithms that leave workers out, for example, as a
Federated Learning result of being geographically situated in an area
with limited bandwidth, can lead to “unfair” mod-
Communication Issues in FL els that generalize poorly, that is, the models have
In distributed learning schemes such as FL, the low accuracy when applied on test samples from
communication cost often dominates the compu- workers that are left out during training. As such,
tation cost. In addition, the downlink speed of IoT a fairness-aware FL is proposed to assign higher

190 IEEE Network • September/October 2021

XIONG_LAYOUT.indd 190 10/22/21 10:18 PM


Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:54:20 UTC from IEEE Xplore. Restrictions apply.
weights to the parameter updates from underrep- of-sight (LOS) communication links with ground
resented workers. devices, for example, mobile sensors and vehicles.
Model Compression: To reduce communica- As such, the UAV-assisted wireless communica-
tion cost, the study in [3] proposes the structured tions are used in several scenarios.
and sketched updates that encode the model Network Coverage and Relay Support From
updates from workers in a compressed form for the Sky: To enable ubiquitous network coverage,
more efficient communication with the server, UAVs are envisioned to be part of 5G HetNets,
which subsequently decodes the update. In addi- for example, to facilitate enhanced Mobile Broad-
tion, the quantization and subsampling approaches band (eMBB) [11] or Ultra reliable low-latency
are also proposed. In particular, the quantization communication (URLLC) for time-sensitive tasks.
approach reduces the number of bits per param- In addition, UAV-enabled mobile relaying can sig-
eter communicated whereas the subsampling nificantly reduce link distances between ground
approach involves the selective communication IoT devices and the distant cloud servers [4].
of only a random subset of the model parameters. Caching and Computation Offloading in the
This reduces the size of model parameters required Sky: UAVs can serve as flying edge computing
to be uploaded to the model owner. platforms to provide caching and computation
Multi-Access Bandwidth Allocation: The study offloading support for ground IoT devices with
in [9] proposes the multi-access Broadband Ana- limited computing capabilities. Dynamic cach-
log Aggregation (BAA) with over-the-air compu- ing can be provided to mobile devices at a lower
tation to reduce the communication latency in transmission latency [11].
FL through exploiting the signal superposition Sensing from the Sky: In the “data ferry-
property of a multi-access channel. The communi- ing model,” the UAV travels to the data source,
cation latency of the BAA transmission is indepen- receives the data, and subsequently travels clos-
dent of the number of devices in the FL network, er to the destination for data transmission. UAVs
and thus it enhances the scalability of the FL work- can be deployed to conduct sensing itself, for
ing process. example, through their equipped sensors. How-
However, these existing solutions do not solve ever, due to the limited resources of UAVs and
the issue of communication inefficiency in FL suf- their heterogeneity, the group of studies propose
ficiently. The edge computing inspired solutions, to address different resource allocation issues in
for example, hierarchical FL, still require an edge UAV communications. Some approaches that are
server, for example, base station, to be in close related to FL are as follows.
proximity to the workers for intermediate aggre- Trajectory Optimization: In general, UAVs
gation. Such a communication infrastructure can expend the most energy resource on traveling,
be infeasible to establish, especially when the FL that is, flying. As such, trajectory optimization
training is not recurrent. Moreover, the cover- is of paramount importance to maximize route
age of the infrastructure may be reduced due to coverage given onboard energy constraints. For
shadowing effects. Even with model compression, example, the study in [12] proposes the dynam-
workers with severely limited connectivity are still ic programming (DP) and convex optimization
unable to participate in the FL training. solution.
Naturally, these workers will be omitted by Efficient Deployment: Different UAVs have
worker selection algorithms due to their high different capabilities and characteristics, requiring
probability of dropout. The omission of such train- the matching of the most suitable UAV to per-
ing samples during the FL training can cause a form varying tasks [6, 12]. Moreover, machine
deterioration in model performance, that is, the learning techniques have been used for the pre-
model’s ability to generalize is affected. Though dictive deployment of UAVs.
the fairness-aware selection protocols may assign Spectrum Management: To perform commu-
higher weights to incorporate the parameter nication functions efficiently, UAVs are required
updates from these underrepresented workers, to manage the spectrum efficiency well, that is, to
the communication inefficiency and high dropout maximize throughput. However, this comes at the
rates of workers with limited connectivity is still cost of greater energy expense required to main-
unsolved. Therefore, we propose the UAV-assist- tain shorter link distances with ground nodes .
ed FL as a solution in the following section. Incentive Mechanism: It has been envisioned
that the DaaS providers will be increasingly
UAV-Assisted Communication in employed by network operators, for example, to
provide network coverage support during large-
Federated Learning scale events. To incentivize participation, an
incentive scheme has to be designed [13].
Overview of UAV-Assisted Communication Inspired by the above studies, we propose the
In recent years, the studies regarding the subject UAV-assisted FL.
of UAV-assisted communication are increasingly
popular. UAV-Assisted
UAVs have two key advantages that increase Communication Efficient Federated Learning
their appeal for utilization in wireless communica- We consider a system model (Fig. 2) in which
tions. First, UAVs are cheaper and more flexible a model owner initiates an FL task with a group
to deploy [10] as compared to terrestrial commu- of workers, for example, independent owners of
nication infrastructure or other high-altitude plat- IoT devices. However, the workers may either
forms, for example, balloons [4]. UAVs are more face unreliable network conditions and thus have
cost effective for ad-hoc, non-recurrent events higher dropout rates, or the workers may have
such as FL training. Second, UAVs are highly insufficient transmit power to complete the wire-
mobile and can thus maneuver to establish line- less transmission of model updates effectively,

IEEE Network • September/October 2021 191

XIONG_LAYOUT.indd 191 10/22/21 10:18 PM


Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:54:20 UTC from IEEE Xplore. Restrictions apply.
Model Transmission
UAV-assi sted FL wi thin subr egi on i nvolving th ree n odes
Aggregation

UAV Trajectory to collect node updates

Contract
ct Contract Contract

UAV-assisted edge aggregation


Transmission Node 1 Node 2 Node 3
Point

2(a): The UAV assisted FL with mobile relay in the sky 2(b): Multi-dimensional contract matching

FIGURE 2. The system model.

thus delaying the FL training process. As such, the Given the involvement of third-party DaaS
third-party DaaS providers, hereinafter UAVs, are providers, an incentive scheme has to be
employed by the model owner to provide net- designed to compensate the UAVs for task
work coverage and relay support from the sky. completion. However, there exist different UAV
The area in which the workers are situated is types with different traveling, computation, and
divided into subregions (Fig. 2b). Each of the sub- communication capabilities. These heteroge-
regions consists of a number of nodes in which a neous types are often the private information
subset of workers are situated. A UAV is deployed of the DaaS providers that are unknown to the
to serve each subregion. Each iteration of the model owner. In addition, the DaaS providers
UAV-assisted FL consists of the following steps may have an incentive to misreport their types
(Fig. 2a): to seek higher rewards. As such, we leverage the
Local Computation: At the beginning of each self-revealing mechanism of the contract theoret-
training round, the model owner transmits a set of ic incentive mechanism design.
parameters to the workers, that is, data owners.
Then, each worker conducts model training using
its local data. tHe multI-dImensIonAl
Edge Aggregation in the Sky: The UAV vis-
its selected workers to collect the updated local
contrAct-mAtcHIng desIgn
model parameters for intermediate aggregation. uAv And model owner utIlItIes
Mobile Relay: The aggregated parameters are UAV types related to the FL task completion are
transmitted from the UAV to the model owner as follows.
for global aggregation across all subregions. To Traveling Cost: UAVs are required to trav-
enable efficient wireless transmission to the model el to the subregion during the task, and visit
owner, the UAV travels to a transmission point the worker nodes for reception of parameter
which has favorable channel conditions for com- updates and then travel to a transmission point
munication with the model owner (Fig. 2a). for wireless relay to the model owner. The ener-
The benefits of the proposed UAV-assisted gy required for this process varies according to
FL are as follows. First, the flexible deployment the propulsion power expended for traveling.
of UAVs implies that model owners are able to The propulsion power [12] is determined by the
obtain the training inputs of workers that are geo- required power to balance the parasitic drag
graphically distant from terrestrial communication caused by skin friction, and drag force of air redi-
infrastructure. Second, UAVs have higher transmit rection. These factors are variable according to
power than individual IoT devices. This increases the reference area, wing aspect ratio, and weight
the speed of model updating. Third, UAVs are able of the UAV.
to conduct mobile relaying, that is, the link distance Computation Cost: The UAVs are required
of reception and transmission of the model param- to perform the intermediate aggregation of local
eters is reduced. As such, UAVs enable a more reli- model parameters. The energy required depends
able connection, thus reducing the worker dropout on the UAV computation capacity and cycles per
rates during the training. Moreover, the privacy bit for computing.
of the workers is preserved since the UAVs and Communication Cost: The varying communi-
model owner only receive the model updates, rath- cation cost incurred by the UAV depends on the
er than the raw data of the workers. bandwidth allocation and transmission power.

192 IEEE Network • September/October 2021

XIONG_LAYOUT.indd 192 10/22/21 10:18 PM


Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:54:20 UTC from IEEE Xplore. Restrictions apply.
Note that each UAV type is characterized by the the proportion of the nodes it is able to cover
minimum energy required to cover a proportion of based on the contract bundle it selected. Besides
nodes, in terms of the propulsion power, compu- contract rewards for covering the proportion of
tation energy, and transmission power required by nodes, the UAVs are also granted an addition-
the UAV to complete the task within the stipulated al fixed reward for traveling to the node, which
time constraint. The UAV types are often the private is separate from the contractual reward since
knowledge unknown to the model owner. it does not vary with node coverage. Note that
On the other hand, the profit of the model there may be more than one UAV which prefers
owner is the revenue gains from FL model accu- to serve a subregion which is situated closer to
racy minus incentive expense. FL model accuracy the UAVs’ bases. Likewise, there may be more
is assumed to be a concave function [14] of the than one subregion which prefers a UAV that can
aggregate training data collected, that is, to reflect cover a greater proportion of nodes.
that the profit increases at a diminishing rate As such, the Gale-Shapley (GS) matching algo-
when more worker nodes are covered. As such, rithm is applied to assign the UAV to each subre-
to maximize profit, a model owner should select gion based on preferences. Before matching, we
UAVs with the lowest marginal cost of node cov- first derive the list of preferences for each subre-
erage to cover the most number of nodes. gion (decided by the model owner) and the UAV.
The matching procedure is as follows.
Multi-Dimensional Contract Design Initialization: The model owner announc-
We utilize the contract theory to design node es the contract bundles to the UAVs. Then, the
coverage-reward contract pairs that correspond UAVs pick the contract bundles that best reflect
to different UAV types in the market (Fig. 2b). A their types. A preference list is constructed for
contract is feasible if it fulfills two conditions: both the UAVs and subregions, in which the
• Individual rationality (IR): Each UAV type UAVs rank the subregions and vice versa. Then, a
achieves non-negative utility if it chooses the list of unmatched subregions is initialized.
contract item designed for its type. Iterative Matching: In each iteration, a subre-
• Incentive compatibility (IC): Each UAV type gion that is still currently unmatched proposes to
achieves maximum utility if it chooses the its most preferred UAV.
contract item designed for its type. If the UAV has not already been assigned
In particular, the IR constraint guarantees participa- to a subregion, the proposal is accepted. If the
tion of the UAVs, whereas the IC constraint ensures UAV has already been assigned to a subregion,
self-revelation, that is, the UAVs have no incentive to the prevailing subregion and the new proposal is
misreport their types amid information asymmetry. compared. There exist two separate scenarios for
In conventional contract theory studies, it is consideration:
usually assumed that the worker type is one-di- • If the prevailing subregion is preferred, the
mensional. In the case of UAVs, there are multiple new proposal is rejected. Then, the UAV is
dimensions of variations in types to be consid- removed from the preference list of the pro-
ered. To solve this challenge, we first sort the UAV posing subregion.
types according to an auxiliary, one-dimensional • If the proposing subregion is preferred, the
variable to collectively reflect the marginal cost prevailing subregion is rejected and added
of node coverage. Then, we derive the necessary into the pool of unmatched subregion. Fur-
conditions for the contract feasibility as follows: thermore, the UAV is removed from the
• For the optimal contract, the node coverage preference list of the newly rejected subre-
is higher if and only if the corresponding gion.
rewards are higher. Note that the removal of rejected UAVs from
• The UAV with the lower marginal cost of the preference lists of subregions ensures that no
node coverage will cover a greater propor- subregion proposes to the same UAV twice. The
tion of the worker nodes. GS algorithm has been proven to achieve stability
For the first condition, the UAVs should only and optimality in matching outcomes [12].
cover a greater proportion of nodes if and only if it
can receive a higher reward. The second condition Performance Evaluation
is the monotonic condition of the contract, which Following the empirical findings in [14], we use
is in line with the IC constraints of the contract. the logarithmic function to model the relation-
In particular, a UAV with lower marginal cost of ship between number of workers participating in
node coverage will not choose the contract bundle FL training, probability of device dropouts, and
designed for high cost UAVs even though the bun- model owner profit derived from the model accu-
dle involves a lower proportion of nodes to be cov- racy. In particular, we consider a range between
ered, given that it is able to derive higher rewards if 500 and 1000 workers each assumed to own the
it chooses the contract designed for its type. same data quantities, and a dropout rate of 0 per-
However, the IC and IR constraints involved cent to 50 percent. Figure 3 shows that as the
are still numerous and non-convex in nature. As number of workers increases and dropout rates
such, procedures [6] to relax the constraints have fall, the model owner profit increases. The rea-
to be implemented to derive the set of sufficient son is that there is an increase in training samples
conditions. With that, we are able to solve for the available, which contributes to an improvement in
optimal contract pairs. model accuracy. As such, the UAV-assisted FL can
increase the model owner profit since it is able
UAV-Subregion Assignment to reach out to more workers without a reliable
With the optimal contract bundles derived, the network connection. In addition, the network and
model owner has the knowledge of the hidden relay support that the UAV provides can reduce
UAV types. For each subregion, the UAV declares the device dropout probability.

IEEE Network • September/October 2021 193

XIONG_LAYOUT.indd 193 10/22/21 10:18 PM


Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:54:20 UTC from IEEE Xplore. Restrictions apply.
FIGURE 4. Wireless transmission latency for UAV-as-
FIGURE 3. Relationship between model owner profit, number of workers, and sisted FL vs. non UAV-assisted.
dropout rate.

Besides, the UAV assisted FL aids in commu-


nication latency reduction. As an illustration, we
consider five UAV types with the descending
order of transmission power, bandwidth, and
channel gain, that is, type 1 UAV has the most
favorable conditions for transmission to the
model owner. An IoT device with a fixed set of
values of transmission power, bandwidth, and
channel gain is also considered for comparison.
The simulations are then conducted with refer-
ence from commonly used parameters. In Fig. 4,
the communication time required by the UAVs
and IoT device to transmit a 100MB model
update to the model owner is plotted. Even with-
out considering dropout probability, it is clear
that the UAV transmission is much faster than
that of the IoT device.
To illustrate the optimality of our contract FIGURE 5. Utility for each UAV type vs. contract
design, we consider the case of six UAVs and items.
six subregions. The UAVs are arranged based
on the marginal cost of node coverage in
ascending order, with type 1 UAV, hereinafter denoted j. The UAVs naturally prefer the subre-
“type 1,” incurring the lowest marginal cost per gions that are closer to their bases due to greater
proportion of nodes covered. Moreover, the gains from the fixed rewards portion. For exam-
UAV bases and subregions are situated at differ- ple, type 1 prefers subregion 6 to subregion 1.
ent coordinates. Similarly, type 2 prefers subregion 6 to subregion
In Fig. 5, each UAV type is assumed to take 1. Given this conflict in preferences, we adopt the
on the contract items designed for each of the GS algorithm discussed earlier for the UAV-subre-
six types, that is, marginal cost of node coverage. gion assignment.
Then, the utility of the UAV is computed. We The match between the UAV and subregions
observe that type 6 achieves non-negative utility is derived and represented by the dotted lines
when it chooses the contract item designed for its in Fig. 6, for example, subregion 6 is matched to
type, that is, type 6 contract. In contrast, if type 6 type 1. Though type 2 also prefers subregion 6, it
chooses contract items designed for other types, is unable to be matched to the subregion given
it achieves negative utility. This validates that our that type 1 is higher up on the list of preferences
contract satisfies the IC constraints. Moreover, the of subregion 6. As such, type 2 is matched to the
satisfaction of the IR constraint is also validated, second closest subregion 1. Similarly, type 6 is
given that all UAVs achieve non-negative utilities matched to one of its least favored preferences,
when they choose the contract designed for their that is, subregion 4, given that the other subre-
types. gions have been matched to the more favored
After each UAV type is known to the model UAVs.
owner, the UAV has to be matched to the sub- The illustration in this section validates the con-
region. For ease of exposition, we assume that vergence and stability of the proposed matching
each subregion has the same number of work- algorithm. Together with the IC constraints of the
er nodes and data quantity. In Fig. 6, we show contract, we can ensure that the UAVs are effi-
the coordinates of the UAVs and subregions. The ciently matched to the subregion based on the
subregions are denoted n whereas the UAVs are truthful reporting of their types.

194 IEEE Network • September/October 2021

XIONG_LAYOUT.indd 194 10/22/21 10:18 PM


Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:54:20 UTC from IEEE Xplore. Restrictions apply.
conclusIon
In this article, we discuss the issues of communi-
cation efficiency in FL and present potential solu-
tions. The UAV-assisted communication efficient
FL is then proposed. To reward the DaaS provid-
ers, we propose the multi-dimensional contract
matching incentive scheme. The simulation results
validate the communication efficiency of the
UAV-assisted FL, and the incentive compatibility
of our contract incentive design.
For future work, we aim to incorporate the
mobility of workers into the system model where
workers may move from one subregion to the
next, thus leading to dynamic variations of data
quantity, communication, and computation work-
loads in the subregion. Moreover, we aim to
study the wireless charging techniques with ener-
gy harvesting [15] for UAVs, as well as scheduling
for the efficient FL in the sky.
FIGURE 6. UAV-subregion assignment.
AcKnowledgment [11] B. Li, Z. Fei, and Y. Zhang, “UAV Communications for 5G
This research is supported in part by Alibaba and Beyond: Recent Advances and Future Trends,” IEEE
Group through the Alibaba Innovative Research Internet of Things J., vol. 6, no. 2, 2018, pp. 2241–63.
[12] Z. Zhou et al., “When Mobile Crowd Sensing Meets UAV:
(AIR) Program and Alibaba-NTU Singapore Energy-Efficient Task Assignment and Route Planning,” IEEE
Joint Research Institute (JRI); by the National Trans. Commun., vol. 66, no. 11, 2018, pp. 5526–38.
Research Foundation, Singapore, under its AI [13] W. Y. B. Lim et al., “Incentive Mechanism Design for
Singapore Programme (AISG Award No: AISG2- Resource Sharing in Collaborative Edge Learning,” arXiv
preprint arXiv:2006.00511, 2020.
RP-2020-019 and AISG-GC-2019-003); by WASP/ [14] Y. Zhan et al., “A Learning-Based Incentive Mechanism for
NTU grant M4082187 (4080); by the Singapore Federated Learning,” IEEE Internet of Things J., 2020.
Ministry of Education (MOE) Tier 1 (RG16/20); [15] C. Wang, S. Guo, and Y. Yang, “An Optimization Frame-
by the National Natural Science Foundation of work for Mobile Data Collection in Energy-Harvesting Wire-
less Sensor Networks,” IEEE Trans. Mobile Computing, vol.
China (Grant No. 62071343); and by SUTD 15, no. 12, 2016, pp. 2969–86.
SRG-ISTD-2021-165. Any opinions, findings and
conclusions or recommendations expressed in BIOGRAPHIES
this material are those of the authors and do not WEI YANG BRYAN LIM is currently an Alibaba Ph.D. candidate
reflect the views of the National Research Foun- with the Alibaba Group and Alibaba-NTU Joint Research Insti-
dation, Singapore. tute, Nanyang Technological University, Singapore.

S AHIL G ARG is currently a research fellow with the Electrical


REFERENCES Engineering Department, École de Technologie Supérieure, Uni-
[1] W. Y. B. Lim et al., “Federated Learning in Mobile Edge Net- versité du Québec, Montréal, Canada.
works: A Comprehensive Survey,” IEEE Commun. Surveys &
Tutorials, 2020, pp. 1–1. ZEHUI XIONG is currently an assistant professor in the Pillar of
[2] B. McMahan et al., “y arcas, ba 2017. Communication-Effi- Information Systems Technology and Design, Singapore Univer-
cient Learning of Deep Networks from Decentralized Data,” sity of Technology and Design. Prior to that, he was a researcher
Proc. 20th Int’l. Conf. Artificial Intelligence and Statistics with the Alibaba-NTU Joint Research Institute, Singapore. He
(AISTATS). received the Ph.D. degree from Nanyang Technological Univer-
[3] J. Konečnỳ et al., “Federated Learning: Strategies for sity, Singapore.
Improving Communication Efficiency,” arXiv preprint
arXiv:1610.05492, 2016. YANG ZHANG is currently an associate professor with the Col-
[4] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless Communica- lege of Computer Science and Technology, Nanjing University
tions with Unmanned Aerial Vehicles: Opportunities and of Aeronautics and Astronautics, China. He obtained a Ph.D.
Challenges,” IEEE Commun. Mag., vol. 54, no. 5, 2016, pp. degree in computer engineering from Nanyang Technological
36–42. University (NTU), Singapore, in 2015.
[5] S. Garg et al., “Tree-Based Attack-Defense Model for Risk
Assessment in Multi-UAV Networks,” IEEE Consumer Elec- DUSIT NIYATO [F] is currently a professor in the School of Com-
tronics Mag., vol. 8, no. 6, 2019, pp. 35–41. puter Science and Engineering and, by courtesy, the School of
[6] W. Y. B. Lim et al., “Towards Federated Learning in Physical & Mathematical Sciences, at the Nanyang Technological
UAV-Enabled Internet of Vehicles: A Multi-Dimensional University, Singapore. He has published more than 380 technical
Contract-Matching Approach,” IEEE Trans. Intelligent Trans- papers in the area of wireless and mobile networking, and is an
portation Systems, to appear. inventor of four U.S. and German patents. He was named a 2017,
[7] W. Y. B. Lim et al., “Hierarchical Incentive Mechanism 2018, and 2019 highly cited researcher in computer science.
Design for Federated Machine Learning in Mobile Net-
works,” IEEE Internet of Things J., 2020, pp. 1–1. CYRIL LEUNG has been with the Department of Electrical and
[8] T. Li, M. Sanjabi, and V. Smith, “Fair Resource Allocation in Computer Engineering, University of British Columbia (UBC),
Federated Learning,” Proc. Int’l. Conf. Learning Representa- Vancouver, Canada, since 1980, where he is a professor and
tions, 2019. currently holds the PMC-Sierra Professorship in Networking and
[9] K. Yang et al., “Federated Learning via Over-the-Air Compu- Communications.
tation,” IEEE Trans. Wireless Commun., vol. 19, no. 3, 2020,
pp. 2022–35. CHUNYAN MIAO is currently a professor in the School of Com-
[10] J. Wang et al., “Joint UAV Hovering Altitude and Power puter Science and Engineering, Nanyang Technological Uni-
Control for Space-Air-Ground IoT Networks,” IEEE Internet versity (NTU), and the director of the Joint NTU-UBC Research
of Things J., vol. 6, no. 2, 2018, pp. 1741–53. Centre of Excellence in Active Living for the Elderly (LILY).

IEEE Network • September/October 2021 195

XIONG_LAYOUT.indd 195 10/22/21 10:18 PM


Authorized licensed use limited to: OREGON STATE UNIV. Downloaded on June 25,2022 at 05:54:20 UTC from IEEE Xplore. Restrictions apply.

You might also like