Professional Documents
Culture Documents
Cooperative Federated Learning and Model Update Verification in Blockchain Empowered Digital Twin Edge Networks
Cooperative Federated Learning and Model Update Verification in Blockchain Empowered Digital Twin Edge Networks
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
data of the physical devices by considering about privacy. (iii) and Proof of Stake, suffers from limitations of throughput
It is insecure for users to transmit the enormous amount of data and scalability. Therefore, a generic blockchain may not
in untrusted and nontransparent wireless network environment. be capable of processing a large number of model updates
To this end, state of the art demands new approaches for micro-transactions in digital twin edge networks. Different
a more flexible digital twin construction while satisfying from generic blockchain, Directed Acyclic Graph (DAG)-
security and privacy protection requirements. based IOTA ledger [29] has been designed to process amounts
Federated learning is emerging as a distributed learning of micro-transactions simultaneously but asynchronously. It
method to preserve privacy by enabling users train machine fulfills the facilities of security and transparency provided by
learning models locally and requiring the users to upload only the traditional blockchain, but also provides high through-
model parameters instead of sending original data to server put and less resource consumption in verifying the micro-
[12]. Due to the characteristics of distribution and privacy transactions. However, with the vast and growing scale of
preservation, federated learning has been envisioned as one of micro-transactions, the more wireless resource (e.g., band-
the key technologies in IoT [13], [14] and 6G networks [15]. width, energy) are needed to achieve the consensus of the
In [11], the authors leveraged federated learning to alleviate entire network. This causes the DAG being compromised to
privacy leakage in digital twin modeling. Specifically, each the scalability of resource constrained physical devices with
client device trains local model based on their local dataset. multiple uses cases. Recently, hybrid blockchain has been
The BSs collect the trained models from the client devices and developed to accommodate the different use cases in IoT by
construct the digital twin models of client devices. Moreover, integrating multiple blockchains [30]. In hybrid blockchain,
the BSs continuously interact with the client devices to keep multiple sub-blockchains, which are similar with side chains,
consistency, so that the established digital twin models can are generated according to the different uses cases, and an
give reaction to the states based on the rules and states of inter-connector framework is exploited to manage the sub-
physical devices. However, in digital twin modeling, the BS blockchains, and also enhance the interoperability among the
which owns all of the physical devices digital twins may different sub-blockchains.
mislead the behavior model of the users. Besides, malicious In this paper, we leverage hybrid blockchain of DAG and
physical devices may disseminate false data or low-quality consortium blockchain to propose a new digital twin edge
models to the BSs to adversely affect the digital twin mod- networks framework for coordinating the wireless resource and
eling. Additionally, the heterogeneous communication and enabling secure digital twin construction at the network edges.
computing resources of physical devices and the stochastic The main contributions are as follows:
wireless transmission environment may hinder the digital twin
• We exploit DAG and consortium blockchain to propose
modeling process.
a new digital twin edge networks framework, where
Blockchain is a promising solution to provide transparent
federated learning is leveraged to construct the digital
network environment and guarantee secure interactions among
twin of smart devices at the BSs belonging to different
untrusted users by maintaining a tamper-proof ledger [19]–
MNOs. Besides, to secure both local model updates and
[23]. The integration of federated learning and blockchain
global model updates, we design a model update chain in
in digital twin edge networks has been studied in some
which the APs and the smart devices verify local model
recent works such as in [27], where a blockchain empowered
updates and reputation to form local model update tangle.
federated learning scheme was proposed to strengthen com-
Then, the BSs confirm the global model updates to form
munication security and data privacy protection in digital twin
a main chain.
edge networks. Furthermore, users scheduling for local model
• To improve the performance of federated learning and
relaying and bandwidth allocation was designed to improve the
local model updates verification, we propose cooperative
communication efficiency of the proposed scheme. Neverthe-
federated learning through the APs to help resource
less, the proposed blockchain empowered federated learning
limited smart devices in local model training. Moreover,
scheme is designed without considering a global network with
APs cluster based multicast communication is developed
multiple mobile network operators (MNOs), which may not
to establish links in local model update tangle so as to re-
adapt to the diverse Quality of Service (QoS) requirements
duce wireless resource consumption in micro-transactions
from a global view. In addition, in local model relaying policy,
consensus.
it is difficult to establish a stable relaying link since the neigh-
• To incentivize the APs to help in local model training
bors may not be willing to help relaying the local model due
of resource limited smart devices, and also encourage
to extra overhead, such as bandwidth and energy consumption.
them to contribute in local model update verification, we
Some incentive schemes were designed in federated learning
design an iterative double auction based joint cooperative
[7], [17], [18], where an aggregator pays rewards to clients to
federated learning and local model update verification
compensate for their resource consumption in model training.
scheme, where the controllers are responsible for solving
However, the designed incentive schemes neglect the security
the optimal unified time for cooperative federated learn-
issues in federated learning, which cannot be directly applied
ing and local model update verification, and forwards the
in our developed digital twin edge networks.
decisions to the APs and MNOs.
Moreover, the security and trustworthiness of blockchain
mainly depends on consensus process. It is noted that tradi- The remainder of this paper is organized as follows. In
tional proof-based consensus process, such as Proof-of-Work Section II, we discuss the related works. In Section III, we
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
'LJLWDO
WZLQ å å
1HWZRUN
%6 %6
%6 ,QIRUPDWLRQ
012 012
Controllers
$3 $3
$ $3
$
5HVRXUFH $XFWLRQ
$OORFDWLRQ %URNHU
'HFLVLRQ
/RFDOGDWD
FROOHFWLRQ
6' 6' 6PDUW 6' 6' 6PDUW
6' 'HYLFHV 'HYLFHV 6' 6'
/RFDOPRGHO
6' 6'
6' XSORDGLQJ
present the proposed blockchain empowered digital twin edge these works didn’t leverage blockchain to secure both global
networks framework and introduce the main components. In model update and local model update in federated learning,
Section IV, we formulate an iterative double auction based thus they cannot be directly applied in our developed digital
joint cooperative federated learning and local model update twin edge networks.
verification scheme. In Section V, we illustrate the perfor-
mance of the proposed scheme through extensive simulations.
Section VI concludes this paper.
B. DAG blockchain for digital twin
II. R ELATED W ORKS
Blockchain has been widely utilized to address the security
A. Federated learning for digital twin and privacy issues associated with resources sharing and
Federated learning is one of the most promising distributed trading among untrusted individuals [24], [25]. In [24], O.
machine learning framework, in which geo-distributed devices Novo et al. leveraged blockchain to propose a new and
train machine learning models locally and upload only model distributed access control mechanism for IoT. In [25], L. Jiang
parameters instead of sending original data to the server. et al. utilized blockchain of DAG and consortium blockchain
Therefore, federated learning mitigates the privacy concerns to develop a new distributed UAVs-assisted wireless power
of users and provides a machine learning framework for transfer framework. When leveraging the federated learning to
distributed users [12]. However, the resource limited edge construct digital twin models, the users might not be willing
devices and the stochastic wireless transmission environment to participate in learning process due to the untrusted and
lead to signal distortion and aggregation error, which makes nontransparent federated learning environment. In [28], Y. Lu
the federated learning encounter the challenges of degraded et al. proposed a blockchain empowered federated learning
convergence rate and prediction accuracy. In [16], M. Chen framework running in the digital twin wireless networks to
et al. derived optimal solutions of user selection and wireless improve the reliability and security of the system and enhance
resource allocation to minimize the loss function of federated data privacy. Nonetheless, the existing works of blockchain
learning. In [17], T. Le et al. formulated an auction game and federated learning in digital twin edge networks are
based incentive mechanism to motivate users to contribute designed without considering a global network with multiple
their resources to join federated learning. In [7], W. Sun et al. MNOs, which may not adapt to the diverse QoS requirements
studied dynamic digital twin and federated learning for air- from a global view. Moreover, the challenges associated with
ground networks, where an incentive scheme was designed incentivizing the clients to help in federated learning and
for federated learning based on Stackelberg game, in order designing an efficient blockchain verification scheme have
to motivate clients to collaboratively train the model. In [18], not been well studied. Motivated by such considerations, in
J. Kang et al. proposed an incentive mechanism combining this paper, we design blockchain empowered digital twin
reputation with contract theory to motivate high-reputation edge networks framework and jointly optimize the cooperative
mobile devices to participate in federated learning. However, federated learning and model update confirmation procedure.
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
III. B LOCKCHAIN E MPOWERED D IGITAL T WIN E DGE cooperative federated learning will consume APs’ limited
N ETWORKS F RAMEWORK resource for their own traffic demand.
The proposed blockchain empowered digital twin edge Moreover, we leverage DAG blockchain to develop a mod-
networks framework is shown in Fig.1, which captures various el update chain so as to enhance the security and privacy
unique characteristics of the digital twin construction problem. preservation in digital twin construction. The BSs, the APs
Accordingly, the main components of cooperative federated and the smart devices act as nodes in the model update
learning and DAG blockchain are introduced in detail. chain. Each node has a unique digital identity consisting of its
public/privacy keys to ensure trust in digital twin construction.
The model update chain is maintained by all the nodes.
A. System Model Specifically, the APs and the smart devices leverage the DAG
∆
There are a set P = {1, 2, ..., P } of mobile network to verify micro-transactions, which include the identities of
operators (MNOs). The operator p ∈ P has a set of Qp BSs. both the APs and the smart devices 1 , and event records (such
Each BS is equipped with mobile edge computing server, thus as smart devices’ data sharing, local model training results
∆ and the quality of local model updates), so as to form local
has enough computing resource. Denote Q = {1, 2, ..., Q}
model update tangle and record a long-term reputation of the
as the set of all BSs, where Q = ∪p∈P Qp . Each MNO
APs. In order to reduce the wireless resources consumption
serves a group of smart devices that have different traffic
in achieving consensus of the local model update tangle,
types and are randomly distributed in the coverage areas of
AP cluster based multicast communication is introduced to
their associated BSs. The BS q ∈ Q has set of Nq smart
∆ establish the links among APs and smart devices, which will
devices, and N = {1, 2, ..., N } is the set of all smart devices, be illustrated in detail in part C of Section III. While the BSs
∆
where N = ∪q∈Q Nq . Besides, a set J = {1, 2, ..., J} of different MNOs utilize the consortium blockchain and act
of APs are placed at the different locations of the BSs. as verifiers to verify the global model update results to form
Each AP can be a Wi-Fi or a femtocell access point that a main chain. This hybrid blockchain of DAG and consortium
is also equipped with edge server and operates in a sepa- blockchain provides both trust and efficiency for construction
rate channel with the macro cellular network. The data of of digital twin edge networks.
smart device n is Dn = {(xn1, , yn1 ), ..., (xnDn , , ynDn )}, To adapt to the distribution of resources in digital twin edge
and D¦n is the data size. Additionally, © the data of AP j is networks, many controller instances are placed in proximity
Dj = (xj1, , yj1 ), ..., (xjDj , , yjDj ) , and Dj is the data size. to edge nodes (i.e., BSs) to cooperatively manage the digital
Both the smart devices and the APs synchronize their data twin construction in edge nodes. The controller is responsible
with the corresponding digital twins that are maintained by for designing the resource allocation strategy for digital twin
the associated BSs. Meanwhile, the BSs also have their own construction. We consider a digital twin construction mar-
twin mapping. The digital twins of smart device n and AP ket, where the controller needs to design a proper incentive
j in the BS are represented as DTn = (Mn , Dn , sn (t)) and mechanism that is essential for the MNOs to determine the
DTj = (Mj , Dj , sj (t)), respectively, where Mn and Mj are optimal reward for the APs to motivate them to contribute with
behavior model of the smart device n and the AP j, Dn and cooperative federated learning and coordination of local model
Dj are static running data, and sn (t) and sj (t) are realtime update tangle from their wireless resource, and also determine
dynamic state, respectively. In order to reduce communication the optimal reward for the verifiers to incentivize them to
cost and data leakage risk, we use federated learning to train contribute their backhaul resource in verifying the transactions
digital twin models instead of uploading all the raw data to of global model update. Since in our previous works [20],
the BSs. [25], we have studied the mechanism to incentivize the ver-
In federated learning, we consider each BS acts as server ifiers to participate in transactions verification in consortium
and trains the task-specific global model individually. The blockchain, here we mainly focus on the incentive mechanism
smart devices and the APs act as clients of their associated BS design to motivate the APs to cooperate in federated learning
and obtain a shared global model to train the local models over and coordinate the consensus of local model update tangle.
their local data. Then the smart devices and the APs upload In the designed incentive mechanism for cooperative feder-
the new local models to the associated BS for updating the ated learning and local model update verification, the AP own-
global model. However, the smart devices always have limited ers offer their services (i.e., participating time for cooperative
computation and energy resource, which may be insufficient federated learning and for coordinating the local model update
for local models training and uploading. As a result, the tangle) in exchange of compensation. The MNOs declare how
constructed digital twins may be inaccurate and cannot give much rewards they are willing to pay each AP for its service.
an efficient reaction as the real smart devices. In this case, the The controller collects the MNOs’ demands and the APs’
APs can cooperatively train and aggregate local models for offers, and determines the service time of each AP for each
the resource limited smart devices. We consider a general case MNO and the corresponding reward so as to maximize the
where each AP can serve more than one MNO, and each MNO social welfare of digital twin construction market. Then, the
owns several BSs and may employ multiple APs at different controller forwards the decisions to the MNOs and the APs.
locations to help the local models training and aggregating
for the resource limited smart devices. However, AP owners 1 The smart device changes its identity (i.e., public key) after each micro-
are expected to ask for compensation since participating in transaction to ensure anonymity.
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
However, in realistic scenario with asymmetric information, one sample of data for smart device. Hence, the computation
the controller may be not aware of the actual needs of the time of a local iteration in training data for Kj,q smart devices
MNOs and the APs due to the concern of their own interests. is denoted as
Therefore, we develop an iterative double auction mechanism
PK
to induce the MNOs and the APs to truthfully reveal their cmp
j,q
bn Dn cj
needs, which will be illustrated in detail in Section IV. Tj,q = n=1
, n ∈ Kj,q , q ∈ Q, (5)
fjC
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
update micro-transaction dissemination benefit for each MNO should satisfy the maximum cooperative federated learning
p ∈ P can be respectively expressed by tolerate time of BS q, T gt,max
q
(−1)
, q ∈ Q. Constraint C2
Qp
ensures that the admitted cooperative federated learning time
X cannot exceed the minimum cooperative federated learning
Fp Tgt(−1) = Fq Tgt(−1) , (17)
p q time provided by AP j ∈ J . Constraint C3 indicates that
q=1
the cooperative federated learning time that each AP admits
Qp should satisfy the respective requested learning time of the
X
Fp′ Tde(−1)
p = Fq′ Tde(−1)
q , (18) MNOs. Constraint C4 indicates that the local model update
q=1 micro-transaction dissemination time that each MNO decides
We consider the utility function Fq (·) and Fq′ (·) are pos- to bid for its BS q should satisfy the maximum micro-
(−1)
itive, increasing and strictly concave function related with transaction dissemination tolerate time of BS q, T de,max
q ,
vector Tq
gt(−1)
and vector Tq
de(−1)
, q ∈ Q, respectively. Each q ∈ Q. Constraint C5 ensures that the admitted local model
AP j ∈ J responds to cooperative federated learning requests, update micro-transaction dissemination time cannot exceed
gt(−1) the minimum micro-transaction dissemination time provided
and admits Γj,q time for one global iteration for each
BS q ∈ Qp of every MNO p ∈ P. Furthermore, each AP by AP j ∈ J . Constraint C6 indicates that the local model
j ∈ J also responds to local model update micro-transaction update micro-transaction dissemination time that each AP ad-
de(−1) mits should satisfy the respective requested micro-transaction
dissemination requests, and admits Γj,q time for local
model update micro-transaction dissemination for each BS dissemination time of the MNOs.
q ∈ Qp of every MNO p ∈ P. We define the admitted coop- In SWO problem (19), the controller tries to find an e-
erative federated learning vector and the admitted local mod- quilibrium where the time of supply equals to the demand,
gt(−1) gt(−1) de(−1) de(−1)
el update micro-transaction dissemination vector Γj
gt(−1) ∆
= i.e., Γj,q ≥ Tq,j and Γj,q ≥ Tq,j . Since the
objective function of social welfare is strictly concave, and
gt(−1) de(−1) ∆ de(−1)
Γj,q , ∀q ∈ Q and Γj = Γj,q , ∀q ∈ Q , re- the feasible region defined by constraints C1-C7 is convex,
gt(−1)
spectively, where Γj,q depends on the cooperative feder- optimization problem (19) has a unique optimization solution
◦ ◦ ◦ ◦
ated learning request Tq,j
gt(−1) de(−1)
, and Γj,q depends on the (Tgt(−1) , Γgt(−1) ) and (Tde(−1) , Γde(−1) ) meeting con-
local model update micro-transaction
dissemination
request straints C1-C7. We first define Lagrange function of SWO
and Cj′ Γj
de(−1) gt(−1) de(−1) problem in (20), where λ, δ, τ , α, β, µ are Lagrange multipli-
Tq,j . We use Cj Γj to denote
the cost incurred by AP j for providing cooperative federated er matrixes corresponding to constraints C1-C6, respectively.
learning service and local model update micro-transaction The Karush-Kuhn-Tucker (KKT) conditions can be derived
dissemination, respectively, which are positive, increasing and by
gt(−1) de(−1)
strictly convex function about vector Γj and Γj . gt(−1)◦
It can be observed that the objective of the MNOs and ∂Fq (Tq )
gt(−1)
= λq,j + τq,j , (21a)
APs are conflicting with each other. If they independently ∂Tq,j
to decide the request time or the admit time, it will be gt(−1)◦
hard to reach an agreement. In order to maintain fairness ∂Cj (Γj )
gt(−1)
= τq,j − σj,q , (21b)
and stimulate the cooperative federated learning and local ∂Γj,q
model update micro-transaction dissemination,
the controller
de(−1)◦
finds optimal Tq
gt(−1)
, Γj
gt(−1)
and Tq
de(−1) de(−1)
, Γj by ∂Fq′ (Tq )
= αq,j + µq,j , (21c)
solving a Social Welfare Optimization (SWO) problem de(−1)
∂Tq,j
P P de(−1)◦
max Fq (Tq
gt(−1)
)− Cj (Γj
gt(−1)
) ∂Cj′ (Γj )
Tgt(−1) ,Γgt(−1) q∈Q j∈J de(−1)
= µq,j − βj,q , (21d)
Tde(−1) ,Γde(−1)
∂Γj,q
P de(−1) P de(−1)
+ Fq′ (Tq )− Cj ′ (Γj ) However, due to hidden information of utility function
q∈Q
gt(−1)
j∈J
gt,max(−1)
Fq (·), Cj (·), F ′ q (·) and C ′ j (·) to the controller, it is infeasible
s.t.C1 : Tq,j ≥ Tq , ∀q ∈ Q, ∀j ∈ J , for the controller to obtain the optimal solutions by directly
gt(−1) gt,min(−1)
C2 : Γj,q ≤ Γj , ∀q ∈ Q, ∀j ∈ J , solving (21a)-(21d). Therefore, the controller needs to design
gt(−1) gt(−1) a proper pricing scheme to induce both the MNOs and APs
C3 : Γj,q = Tq,j , ∀q ∈ Q, ∀j ∈ J ,
bid truthfully, so as to elicit their hidden information, and
de(−1) de,max(−1)
C4 : Tq,j ≥ Tq , ∀q ∈ Q, ∀j ∈ J , eventually achieve equilibrium.
de(−1) de,min(−1)
C5 : Γj,q ≤ Γj , ∀q ∈ Q, ∀j ∈ J ,
de(−1) de,min(−1) A. SWO Problem Transformation
C6 : Γj,q = Tq,j , ∀q ∈ Q, ∀j ∈ J ,
Each MNO p ∈ P submits bids xgt q,j ≥ 0 and xq,j ≥ 0 for
de
gt(−1) gt(−1) de(−1) de(−1)
C7 : Γj,q ≥ 0, Tq,j ≥ 0, Γj,q ≥ 0, Tq,j
≥ 0, each one of its BSs q ∈ Qp , and for each AP j ∈ J . Similarly,
gt
(19) each AP j ∈ J submits bids yj,p ≥ 0 and yj,p
de
≥ 0 for each
where constraint C1 indicates that the cooperative federated BS q ∈ Qp of each MNO p ∈ P. We define the bid vector
learning time that each MNO decides to bid for its BS q of each MNO p ∈ P and of each AP j ∈ J , respectively, as
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
P P P P
Fq′ (Tq Cj′ (Γj
gt(−1) gt(−1) de(−1) de(−1)
LSW O (λ, δ, τ , α, β, µ) = Fq (Tq )− Cj (Γj )+ )− )−
q∈Q j∈J q∈Q j∈J
Q
J P
P P Q
J P Q
J P
P
gt,max(−1) gt(−1) gt(−1) gt,min(−1) gt(−1) gt(−1)
λq,j Tq − Tq,j − δj,q Γj,q − Γj − τq,j Tq,j − Γj,q − (20)
j=1 q=1 j=1 q=1 j=1 q=1
PJ P Q J P
P Q PJ P Q
de,max(−1) de(−1) de(−1) de,min(−1) de(−1) de(−1)
αq,j Tq − Tq,j − βj,q Γj,q − Γj − µq,j Tq,j − Γj,q ,
j=1 q=1 j=1 q=1 j=1 q=1
∆
gt ∆
p = (xq,j : ∀q ∈ Q, j ∈ J ), xp = (xq,j : ∀q ∈ Q, j ∈
xgt de de
gt(−1)◦
∆ ∆ ∂Fq Tq
J ), yjgt = gt
: ∀q ∈ Q), and
(yj,q : ∀q ∈ Q).
yjde = de
(yj,q xgt
gt(−1)∗
q,j = Tq,j gt(−1)
, (24a)
These bids signal the cooperative federated learning and local ∂Tq,j
model update micro-transaction dissemination requirements of
de(−1)◦
MNOs, and also signal the serving cost of the APs. Then, de(−1)∗
∂Fq Tq
the controller determines the time allocation for cooperative xde
q,j = Tq,j de(−1)
, (24b)
∂Tq,j
federated learning and local model update micro-transaction
dissemination based on the bids by solving a new allocation gt(−1)◦
gt 1 ∂Cj Γj
problem (NAP) as follows yj,q = gt(−1)∗ gt(−1)
, (24c)
Γj,q ∂Γj,q
de(−1)◦
de 1 ∂Cj Γj
PJ P Q gt yj,q = de(−1)∗ de(−1)
, (24d)
gt(−1) yj,q gt(−1) 2 Γj,q ∂Γj,q
max xgt
q,j log Tq,j − 2 Γ j,q
Tgt(−1) ,Γgt(−1) j=1 q=1
Tde(−1) ,Γde(−1) Then the optimal solution of NAP problem is identical to the
Q ∆
PJ P
de(−1) y de de(−1) 2 unique optimal solution of the SWO problem, i.e., Tgt(−1)◦ =
+ xde
q,j log Tq,j − j,q2 Γ j,q gt(−1)∗ de(−1)◦ ∆ de(−1)∗ gt(−1)◦ ∆ gt(−1)∗
j=1 q=1 T , T = T , Γ = Γ , and
gt(−1) gt,max(−1) de(−1)◦ ∆ de(−1)∗
s.t.C1 : Tq,j ≥ Tq , ∀q ∈ Q, ∀j ∈ J , Γ = Γ . Thus, a proper pricing scheme needs
gt(−1) gt,min(−1) to be derived to induce both the MNOs and the APs to bid
C2 : Γj,q ≤ Γj , ∀q ∈ Q, ∀j ∈ J , according to (24).
gt(−1) gt(−1)
C3 : Γj,q = Tq,j , ∀q ∈ Q, ∀j ∈ J ,
de(−1) de,max(−1)
C4 : Tq,j ≥ Tq , ∀q ∈ Q, ∀j ∈ J , B. Pricing Scheme
de(−1) de,min(−1)
C5 : Γj,q ≤ Γj , ∀q ∈ Q, ∀j ∈ J , According to the NAP problem in (22), the MNOs and the
de(−1)
C6 : Γj,q
de,min(−1)
= Tq,j , ∀q ∈ Q, ∀j ∈ J , APs find their optimal bids
by solving
their payoff optimiza-
and Z ′ p Tp
gt(−1) de(−1)
gt(−1) gt(−1) de(−1) de(−1)
tion problem. Denote Zp Tp as the
C7 : Γj,q ≥ 0, Tq,j ≥ 0, Γj,q ≥ 0, Tq,j
≥ 0, MNO p payment to the controller for cooperative federated
(22) learning and for local model update micro-transaction dissem-
Notice that the NAP problem has the same constraints as ination services obtained from the APs. The MNO p ∈ P
the SWO problem, but has a different strictly concave ob- finds the optimal bids xgt∗
p and xde∗
p by solving the payoff
jective function which is designed according to the alloca- maximization problem as follows
tion
rule in [35]. Thus,
it has a unique optimal
solution
gt(−1) gt(−1)
Tgt(−1)∗ , Γgt(−1)∗ and Tde(−1)∗ , Γde(−1)∗ . According max
gt
Fp Tp − Zp Tp +
xp ,xde
to KKT conditions, we can obtain a set of equations as follows p
Fp′ Tp
de(−1)
− Zp′ Tp
de(−1) (25)
s.t. xgt
q,j ≥ 0, xde
q,j ≥ 0, ∀j ∈ J, q ∈ Q,
gt(−1)∗ xgt
q,j
Tq,j = , (23a) The optimal solution of problem (25) satisfies the following
λq,j + τq,j conditions
gt(−1)∗ τq,j − σj,q
Γj,q = gt , (23b)
yj,q gt(−1) gt(−1)
∂Fp Tp ∂Zp Tp
= (τq,j + λq,j ) , (26)
de(−1)∗ xde
q,j ∂Tq,j
gt(−1)
∂xgt
q,j
Tq,j = , (23c)
αq,j + µq,j
de(−1) de(−1)
∂Fp Tp ∂Zp Tp
de(−1)∗ µq,j − βj,q = (αq,j + µq,j ) , (27)
Γj,q = de
, (23d) ∂Tq,j
de(−1) ∂xde
q,j
yj,q
and H ′ j Γj
gt(−1) de(−1)
It can be noted that (23a)-(23d) are different from (21a)-(21d). Similarly, denote Hj Γj as com-
If the MNOs and the APs submit the following bids pensation from the controller for the AP j ∈ J . The APs
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
10
find their optimal bids yjgt and yjde by solving the payoff Specifically, both the twin APs and twin BSs belonging to
maximization problem as follows different MNOs act as bidders, and iteratively interact with the
controller to solve the desirable solutions. Then, the controller
gt(−1) gt(−1)
max
gt
Hj Γj − Cj Γj + forwards the solutions to the physical BSs and the physical
yj ,yjde
APs. Therefore, compared to the iterative double auction in
Hj′ Γj
de(−1)
− Cj′ Γj
de(−1) (28)
traditional edge networks, the iterative double auction in our
gt
s.t. yq,j ≥ 0, yq,j
de
≥ 0, , ∀j ∈ J, q ∈ Q. developed digital twin edge networks can significantly reduce
signaling cost among the controller, the BSs and the APs.
The optimal solution of problem (28) satisfies the following
conditions
Algorithm 1: Iterative double auction based joint coopera-
2 tive federated learning and local model update verification
gt(−1) gt(−1) gt
∂Cj Γj ∂Hj Γj yj,q gt(−1)(0) gt(−1)(0) de(−1)(0) de(−1)(0) (0)
= , (29) 1 Input: Tq,j , Γj,q , Tq,j , Γj,q , λq,j ,
∂Γj,q
gt(−1) gt
∂yj,q δj,q − τq,j (0) (0) (0) (0) (0)
δq,j , τq,j , αq,j , βj,q and µq,j , ∀j ∈ J ,q ∈ Q;
de(−1)
de(−1)
de
2 2 Output: Tgt(−1) , Γgt(−1) , Tde(−1) and Γde(−1) ;
∂Cj Γj ∂Hj Γj yj,q 3 t ← 0, and conv ← 0;
= , (30)
∂Γj,q
de(−1) de
∂yj,q βj,q − µq,j 4 while conv = 0 do
gt(−1)(t) gt(−1)(t)
5 The controller announces Tq,j , Γj,q ,
In order to induce the MNOs and the APs to bid according de(−1)(t) de(−1)(t) (t) (t) (t) (t) (t)
to (24), the pricing rules are defined as follows Tq,j , Γj,q , λq,j , δq,j , τq,j , αq,j , βj,q and
(t)
X X
µq,j , ∀j ∈ J ,q ∈ Q;
Zp xgt
p = xgt
q,j , (31) 6 t ← t + 1;
gt(t)
j∈J q∈Q 7 Each MNO computes the optimal bids xp and
X X de(t)
Zp′ xde = xde xp by (25);
p q,j , (32) gt(t) de(t)
j∈J q∈Q
8 Each AP computes the optimal bids yj and yj
by (28);
2
X X (δj,q − τq,j ) 9 Each MNO and AP submit their bids to the
Hj yjgt = gt , (33)
j∈J q∈Q
yj,q controller;
gt(−1)(t)
10 The controller computes the new Tp ,
2
X X (βj,q − µq,j ) Γj
gt(−1)(t) de(−1)(t)
, Tp and Γj
de(−1)(t)
by (22);
Hj′ yjde = de
, (34)
j∈J q∈Q
yj,q 11 The controller checks the convergence;
12 if convergence then
The NAP problem (22) and the pricing rules (31)-(34) 13 conv ← 1;
guarantee the social welfare in (19) maximization. It can be 14 end
observed that the optimal requested time and the admitted 15 else
gt(−1)∗ gt(−1)∗ de(−1)∗ de(−1)∗
time of Tq,j , Γj,q , Tq,j and Γj,q are related 16 Enter a new iteration and repeat lines 4-12;
with the bids of the MNOs and the APs in (25) and (28), 17 end
respectively. Similarly, the optimal bids of the MNOs and 18 end
the APs are also related with the the optimal time allocation
19 The controller computes the payment Zp xgt p ,
and the Lagrange multipliers in (22). Therefore, problems
(22), (25) and (28) need to be solved iteratively to achieve Z ′ p xde gt
p , the complementary Hj yj and H ′ j yjde
the desirable solutions. The controller initializes the time by (31)- (34).
variables Tgt(−1) , Γgt(−1) , Tde(−1) , Γde(−1) and the Lagrange
multiplier matrixes λ, δ, τ , α, β, µ. Then, the controller
broadcasts the initialization variables to both the MNOs and V. N UMERICAL R ESULTS
the APs. Each MNO and AP find their optimal bids by
solving their payoff optimization problem (25) and (28), and A. Experiment Setting
submit their new bids to the controller. Then, the controller We evaluate the performance of our proposed iterative
computes the new time variables and updates the Lagrange double auction based joint cooperative federated learning and
multiplier matrixes by utilizing a gradient descent method. The local model update verification scheme. We consider a small
iteration is executed until the convergence of the solutions. digital twin construction market settings of P = 2 operators,
The implementation process of iterative double auction based each one with one BS. J = 5 APs are placed at the different
joint cooperative federated learning and local model update locations of the BSs. Each BS has Np = 50 smart devices.
verification scheme can be illustrated in Algorithm 1. It The smart devices are randomly distributed in the coverage of
is worth noting that after the construction of digital twin BSs and APs. We conduct our cooperative federated learning
edge networks, the developed iterative double auction based experiment on a real-world data set CIFAR10 [36], which
joint cooperative federated learning and local model update consists of 60,000 images in 10 classes, including 40,000
verification scheme can be executed by the digital twin models. training images, 10,000 valid images and 10,000 test images.
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
11
2.5
The CIFAR10 data set is shuffled and randomly assigned to The proposed scheme
the smart devices. The Convolutional Neural Network (CNN) The IDA-woDAG scheme
Random selection scheme
is utilized as machine learning model for cooperative federated 2
Loss
For the purpose of comparing with our proposed iterative
1
double auction based joint cooperative federated learning and
local model update verification scheme, we consider two other
optimization schemes: (i) iterative double auction without local 0.5
0.8
0.7
TABLE 1
PARAMETERS 0.6
Accuracy
0.5
Parameter Value The proposed scheme
0.4
Tolerate time for learning Tqgt,max [200, 300]s The IDA-woDAG scheme
Random selection scheme
Minimum supply learning time Γgt,min
0.3
j [60, 200]s
Tolerate time for dissemination Tqde,max [5, 10]s 0.2
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
12
2500 8
Tde , f de =0.02
11 11
The proposed scheme 7 Tde, f de =0.04
12 21
2000 The optimal solution
Tde, f de =0.02
21 12
6 Tde, f de =0.04
de
de de
1500 G11, f 11 =0.02
5
Social welfare
de de
G12, f 21 =0.04
de de
G21, f 12 =0.02
de
1000 4
Gde , f de =0.04
22 22
3
500
0
1
-500 0
2 4 6 8 10 12 14 16 18 0 5 10 15 20 25
Iteration number Iteration number
22 22
j,q
(provided Ggt
140 gt gt
G11, f 11 = 0.01
120 gt gt
G12, f 21 = 0.03
p to AP j is equal to the provided cooperative learning time
gt gt
G21, f 12 = 0.01
Γgt
j,p of AP j for MNO p, which proves the effectiveness of the
q,j
100
The required Tgt
Ggt
22
, f gt
22
= 0.03 proposed iterative double auction based cooperative federated
gt
80
learning scheme (Tp,j = Γgtj,p ). Furthermore, it can be noted
60 gt
that both the required cooperative learning time Tp,j of MNO
40
p and the provided cooperative learning time Γgt j,p of AP j
20
decreases with the increase of cooperative learning cost ϕgt j,p .
0
0 5 10 15 20 25
The reason is that with increasing cooperative learning cost
Iteration number
ϕgt
j,p , the MNO p needs to pay more to compensate for the
gt resource consumption of AP j. In order to maximize the payoff
Fig. 6: The required cooperative learning time Tp,j and the
gt
provided cooperative learning time Γp,j . in (25), the MNO p decreases its required cooperative learning
gt
time Tp,j .
Fig. 7 shows the required local model update micro-
de
transaction dissemination time Tp,j of MNO p and the provid-
proposed scheme has the highest accuracy, since the local de
ed dissemination time Γj,p of AP j under different dissemi-
model updates provided by the APs are verified by the model
nation cost ϕdep,j . Similarly, we take p = 1, p = 2 and j = 1
update chain in each global round, the low-quality model
and j = 2 for example, it can be seen that the trends of the
updates will be deleted. In the IDA-woDAG scheme, both the de
required dissemination time Tp,j of MNO p and the provided
local model training results and the quality of local model
dissemination time Γde p,j of AP j are consistent with Fig. 6.
updates are not verified by the DAG, thus the accuracy of de
Specifically, the required dissemination time Tp,j of MNO p
the IDA-woDAG scheme is lower than that of the proposed
to AP j is equal to the provided dissemination time Γde j,p of
scheme. The random selection scheme randomly selects the
AP j for MNO p. Besides, the required dissemination time
APs without considering the learning time requirement and de
Tp,j of MNO p and the provided dissemination time Γde j,p of
the model quality, thus its accuracy is the lowest.
AP j decreases with the increase of dissemination cost ϕde j,p ,
Fig. 5 proves the convergence of the proposed iterative
since the increase of dissemination cost ϕde j,p drives the MNO
double auction based joint cooperative federated learning and
p to pay more for compensating AP j. The MNO p decreases
local model update verification scheme. It can be seen that de
Tp,j to maximize the payoff in (28).
the proposed iterative double auction based joint cooperative
federated learning and local model update verification scheme
gradually converges to the optimal solution of social welfare VI. C ONCLUSION
after 8 iterations, which implies that the proposed scheme can This paper proposed a new blockchain empowered digital
induce both the MNOs and the APs to bid truthfully to elicit twin edge networks framework, where cooperative federated
their hidden information (i.e., utility function of the MNOs learning was developed to construct digital twin of resource
and cost function of the APs), and eventually converge to the limited smart devices at the network edges belonging to
optimal solution. different MNOs. In order to provide a secure digital twin
gt
Fig. 6 shows the required cooperative learning time Tp,j of construction environment, we designed a model update chain
gt
MNO P and the provided cooperative learning time Γj,p of by exploiting DAG blockchain, in which APs cluster based
AP j under different cooperative learning cost ϕgtp,j . Here, we multicast communication was developed to establish links
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
13
in local model update tangle to reduce wireless resource [18] J. Kang, Z. Xiong, D. Niyato, S. Xie, and J. Zhang, “Incentive
consumption during the consensus. Moreover, we designed Mechanism for Reliable Federated Learning: A Joint Optimization
Approach to Combining Reputation and Contract Theory,” IEEE Internet
an iterative double auction based joint cooperative federated Things J., vol. 6, no. 6, pp. 10700-10714, Dec. 2019.
learning and local model update verification scheme to incen- [19] M. Conti, E. Kumar, C. Lal, and S. Ruj, “A Survey on Security and
tivize the APs to help in local models training and also to con- Privacy Issues of Bitcoin,” IEEE Commun. Surv. Tutor., vol. 20, no. 4,
pp. 3416-3452, May 2018.
tribute their resource in local model update verification. The [20] L. Jiang, S. Xie, S. Maharjan, and Y. Zhang, “Joint Transaction Relaying
optimal unified time for cooperative federated learning and and Block Verification Optimization for Blockchain Empowered D2D
local model update verification were solved to maximize the Communication,” IEEE Trans. Veh. Tech., vol. 69, no. 1, pp. 828-841,
Oct. 2019.
social welfare. Numerical results illustrated that the proposed [21] L. Jiang, S. Xie, S. Maharjan, and Y. Zhang, “Blockchain Empowered
scheme can improve the performance of global model update Wireless Power Transfer for Green and Secure Internet of Things,” IEEE
in cooperative federated learning and achieve the optimal Netw., vol. 33, no. 6, pp. 164-171, Dec. 2019.
[22] Z. Xiong, Y. Zhang, N. Luong, D. Niyato, P. Wang, and N. Guizani,
social welfare by eliciting the hidden information of MNOs “The Best of Both Worlds: A General Architecture for Data Management
and APs. in Blockchain-Enabled Internet-of-Things,” IEEE Netw., vol. 34, no. 1,
pp. 166-173, Jan. 2020.
[23] Z. Xiong, S. Feng, W. Wang, D. Niyato, P. Wang, and Z.
Han, “Cloud/Fog Computing Resource Management and Pricing for
R EFERENCES Blockchain Networks,” IEEE Internet Things J., vol. 6, no. 3, pp. 4585-
4600, Jun. 2019.
[1] S. Haag and R. Anderl, “Digital Twin – Proof of Concept,” Manuf. [24] O. Novo, “Blockchain Meets IOT: An Architecture for Scalable Access
Lett., vol. 15, pp. 64-66, Jan. 2018. Management in IoT,” IEEE Internet Things J., vol. 5, no. 5, pp. 1184-
[2] B. Minerva, G. Lee, and N. Crespi, “Digital Twin in the IoT Context: 1195, Apr. 2018.
A Survey on Technical Features, Scenarios, and Architectural Models,” [25] L. Jiang, B. Chen, S. Xie, S. Maharjan, and Y. Zhang, “Incentiviz-
P. IEEE, vol. 108, no. 10, pp. 1785-1824, Oct. 2020. ing Resource Cooperation for Blockchain Empowered Wireless Power
[3] I. Yaqoob, K. Salah, M. Uddin, R. Jayaraman, M. Omar, and M. Imran, Transfer in UAV Networks,” IEEE Trans. Veh. Tech., vol. 69, no. 12,
“Blockchain for Digital Twins: Recent Advances and Future Research pp. 15828-15841, Dec. 2020.
Challenges,” IEEE Netw., vol. 34, no. 5, pp. 290-298, Apr. 2020. [26] N. Tran, W. Bao, A. Zomaya, M. Nguyen, and C. Hong, “Federated
[4] Gartner, [Online]. Available: https://www.gartner.com/en/newsroom/ Learning over Wireless Networks: Optimization Model Design and
pressreleases/2019-02-20-gartner-survey-reveals-digital-twins, 2020. Analysis,” in Proc. IEEE Infocom 2019 , pp. 1387-1395, Paris, France,
[5] O. Marai, T. Taleb, and J. Song, “Roads Infrastructure Digital Twin: A May 2019.
Step Toward Smarter Cities Realization,” IEEE Netw., vol. 35, no. 2, [27] Y. Lu, X. Huang, K. Zhang, S. Maharjan, and Y. Zhang,
pp. 136-143, Mar. 2021. “Communication-Efficient Federated Learning and Permissioned
[6] J. Taylor and H. Sharif “Leveraging Digital Twins to Enhance Perfor- Blockchain for Digital Twin Edge Networks,” IEEE Internet Things J.,
mance of IoT in Disadvantaged Networks,” in Proc. IEEE IWCMC, pp. vol. 8, no. 4, pp. 2276-2288, Feb. 2021.
1303-1308, Limassol, Cyprus, Jul. 2020. [28] Y. Lu, X. Huang, K. Zhang, S. Maharjan, and Y. Zhang, “ Low-Latency
[7] W. Sun, P. Wang, N. Xu, G. Wang, and Y. Zhang, “Dynamic Digital Twin Federated Learning and Blockchain for Edge Association in Digital Twin
and Distributed Incentives for Resource Allocation in Aerial-Assisted Empowered 6G Networks,” IEEE Trans. Ind. Inform., vol. 17, no. 7,
Internet of Vehicles,” IEEE Internet Things J., Feb. 2021. pp. 5098-5107, Jul. 2021.
[29] IOTA: A Cryptocurrency for Internet-of-Things. [Online]. Available:
[8] J. Bao, B. Zhao, W. Yu, et al., “OpenSAN: A Software Defined Satellite
http://www.iotatoken.com/, 2019.
Network Architecture,” in Proc. ACM Conference on SIGCOMM, pp.
[30] S. Zhu, Z. Cai, H. Hu, Y. Li, and W. Li, “zkCrowd: A Hybrid
347-348, Chicago, Illinois, USA, 2014.
Blockchain-Based Crowdsourcing Platform,” IEEE Trans. Ind. Inform.,
[9] Z. Tang, B. Zhao, W. R. Yu, et al., “Software Defined Satellite Networks:
vol. 16, no. 6, pp. 4196-4205, Jun. 2020.
Benefits and Challenges,” in Proc. Computing, Communications and IT
[31] N. Aitzhan and D. Svetinovic, “Security and Privacy in Decentralized
Applications Conference, pp. 127- 132, Beijing, China, 2014.
Energy Trading Through Multi-Signatures, Blockchain and Anonymous
[10] P. Bellavista, C. Giannelli, M. Mamei, M, Mendula, and M. Picone, Messaging Streams,” IEEE Trans. Depend. Sec. Comput., vol. 15, no.
“Application-Driven Network-Aware Digital Twin Management in In- 5, pp. 840-852, Oct. 2018.
dustrial Edge Environments,” IEEE Trans. Industr. Inform., Mar. 2021. [32] S. Popov, “The Tangle Version 1.4.3”, IOTA Foundation, Tech, Rep.,
[11] Y. Lu, X. Huang, K. Zhang, S. Maharjan, and Y. Zhang, Apr. 2018.
“Communication-Efficient Federated Learning for Digital Twin Edge [33] EOS Block Producer Voting Guide. [Online] Available: https://medium.
Networks in Industrial IoT,” IEEE Trans. Industr. Inform., vol. 17, no. com/coinmonks/eos-block-producer-voting-guide-fba3a5a6efe0.
8, pp. 5709-5718, Aug. 2021. [34] D. Friedman, D. P. Friedman, and J. Rust, “ The Double Auction Market:
[12] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. Arcas, Institutions Theories and Evidence,” Boulder, CO, USA: Westview
“Communication-Efficient Learning of Deep Networks from Decentral- Press, 1993.
ized Data,” in Proc. Artif. Intell. Stat., pp. 1273-1282, Feb. 2017, [35] F. Kelly, A. Maulloo, and D. Tan, ”Rate Control for Communication
[13] D. Kwon, J. Jeon, S. Park, J. Kim, and S. Cho, “Multiagent DDPG- Networks: Shadow prices, Proportional Fairness and Stability,” J. Oper.
Based Deep Learning for Smart Ocean Federated Learning IoT Net- Res. Soc., vol. 29, no. 3, pp. 237-252, 1998.
works,” IEEE Internet Things J., vol. 7, no. 10, pp. 9895-9903, Apr. [36] A. Krizhevsky and G. Hinton, ”Learning Multiple Layers of Features
2020. from Tiny Images,” Handbook of Systemic Autoimmune Diseases, Apr.
[14] J. Kang, Z. Xiong, D. Niyato, Y. Zou, Y. Zhang, and M. Guizani, 2009.
“Reliable Federated Learning for Mobile Networks,” IEEE Wirel.
Commun., vol.27, no.2, pp. 72-80, Apr. 2020,
[15] X. Huang, S. Leng, S. Maharjan and Y. Zhang, “Multi-Agent Deep B IOGRAPHIES
Reinforcement Learning for Computation Offloading and Interference
Coordination in Small Cell Networks,” IEEE Trans. Veh. Tech., vol. 70,
no. 9, pp. 9282-9293, Sep. 2021.
[16] M. Chen, Z. Yang, W. Saad, C. Yin, H. Poor, and S. Cui, “A Joint
Learning and Communications Framework for Federated Learning Over
Wireless Networks,” IEEE Trans. Wirel. Commun., vol. 20, no. 1, pp.
269-283, Jan. 2021.
[17] T. Le, N. Tran, Y. Tun, M. Nguyen, S. Pandey, Z. Han, and C. Hong,
“An Incentive Mechanism for Federated Learning in Wireless Cellular
Network: An Auction Approach,” IEEE Trans. Wirel. Commun., Mar.
2021.
2327-4662 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: AIT Austrian Institute of Technology. Downloaded on February 16,2022 at 11:55:58 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2021.3126207, IEEE Internet of
Things Journal
14
Li Jiang (jiangli@gdut.edu.cn) received her Ph.D. Yan Zhang [IEEE Fellow’20] (yanzhang@ieee.org)
degree from the School of Information and Com- is currently a Full Professor with the Department
munication Engineering, Beijing University of Posts of Informatics, University of Oslo, Norway. He re-
and Telecommunications (BUPT), China, in 2017. ceived the Ph.D. degree from the School of Electrical
She was also with University of Oslo and Simula and Electronics Engineering, Nanyang Technolog-
Metropolitan Center for Digital Engineering, Nor- ical University, Singapore. He received M.S. and
way, as a visiting Ph.D. student from Aug. 2015 B.S from Beihang University and Nanjing University
to Jul. 2016. She is currently an associate professor of Post and Telecommunications, respectively. His
in the School of Automation, Guangdong University research interests include next-generation wireless
of Technology (GDUT), Guangzhou. Her current re- networks leading to 5G beyond/6G, green and secure
search interests include mobile blockchains, mobile cyber-physical systems (e.g., smart grid and trans-
edge computing, D2D communications, digital twin networks, and resource port). Dr. Zhang is an Editor (or Area Editor, Senior Editor, Associate Editor)
management for B5G and 6G networks. for several IEEE transactions/magazine, including IEEE Communications
Magazine, IEEE Network Magazine, IEEE Transactions on Network Science
and Engineering, IEEE Transactions on Vehicular Technology, IEEE Transac-
tions on Industrial Informatics, IEEE Transactions on Green Communications
and Networking, IEEE Communications Survey and Tutorials, IEEE Internet
of Things Journal, IEEE Systems Journal, IEEE Vehicular Technology Mag-
azine, and IEEE Blockchain Technical Briefs. He is a symposium/track chair
in a number of conferences, including IEEE ICC 2021, IEEE Globecom 2017,
IEEE PIMRC 2016, IEEE SmartGridComm 2015. He is the Chair of IEEE
Communications Society Technical Committee on Green Communications
and Computing (TCGCC). He is an IEEE Vehicular Technology Society
Hao Zheng (2111904331@mail2.gdut.edu.cn) is Distinguished Lecturer during 2016-2020. He is CCF Senior Member, an
currently pursuing his master’s degree with the elected member of CCF Technical Committee of Blockchain, and 2019
School of Automation, Guangdong University of CCF Distinguished Speaker. Since 2018, Prof. Zhang was a recipient of
Technology (GDUT), Guangzhou, China. His cur- the global Highly Cited Researcher Award (Web of Science top 1% most
rent research interests include mobile blockchains, cited worldwide). He is Fellow of IEEE, Fellow of IET, elected member of
UAV networks, machine learning, and resource man- Academia Europaea (MAE), and elected member of Norwegian Academy of
agement for B5G and 6G networks. Technological Sciences (NTVA).