Journal of Network and Computer Applications 167 (2020) 102710

Contents lists available at ScienceDirect

Journal of Network and Computer Applications

journal homepage: www.elsevier.com/locate/jnca

A secure and efficient data sharing scheme based on blockchain in

industrial Internet of Things☆
Jiancheng Chi a , Yu Li a , Jing Huang a , Jing Liu b, ∗ , Yingwei Jin c, ∗∗ , Chen Chen d , Tie Qiu a
a
Tianjin Key Laboratory of Advanced Networking, College of Intelligence and Computing, Tianjin University, Tianjin, 300350, China
b
School of Artificial Intelligence, Hebei University of Technology, Tianjin, 300401, China
c School of Economics and Management, Dalian University of Technology, Dalian, 116024, China
d
State Key Laboratory of Integrated Service Networks, Department of Telecommunication, Xidian University, Xi’an, 710126, China

A R T I C L E I N F O A B S T R A C T

Keywords: Blockchain possesses potential for enabling secure data sharing of industrial big data generated by the Industrial
Data sharing Internet of Things (IIoT). However, the existing data-sharing schemes based on blockchain secure only the data
Blockchain
exchange process between the data-sharing parties; they do not consider the efficiency of data sharing at all. In
Community detection
this study, we propose a data-sharing scheme based on blockchain that comprehensively considers the security
Industrial internet of things (IIoT)
and efficiency of data sharing. In this scheme, a secure data-sharing framework is designed based on identity
authentication and Hyperledger Fabric to ensure the security of data sharing. We also propose a community
detection algorithm to divide the clients into different data-sharing communities according to the similarity of
label data. The scope of data sharing is selected according to the results of the community detection evaluated
by the sharing degree, which can effectively narrow the scope of query-shared data and improve the efficiency
of data sharing. The experimental results show that the proposed scheme is effective for secure and efficient data
sharing among different clients.

1. Introduction first and the biggest problem is that data stored in a cloud service
platform are vulnerable to leakage (Singh et al., 2016). Owing to the
The Industrial Internet of Things (IIoT) connects a large number high sensitivity of industrial big data, the consequences of data leak-
of industrial machines and devices with data perception and collec- age are extremely serious. The second problem is that the volume of
tion capabilities through wired and wireless networks, collecting data industrial big data is large, and therefore, sharing data through a cloud
in real time, making optimal decisions, and optimizing parameter con- service platform directly will consume a significant amount of band-
figuration through data analysis (Li et al., 2017). With the develop- width resources, resulting in increased data-sharing costs (Khan, 2016).
ment of cloud computing, cloud service providers have provided the In addition, shared data should be easily verifiable to prevent tamper-
IIoT with a very convenient platform for the management and control ing of important data, which cannot be guaranteed by cloud service
of a large number of machines and devices (Qiu et al., 2019a). Even platforms.
with low rental fees, the IIoT can obtain sufficient computing and stor- Recently, as one of the underlying technologies of bitcoin,
age resources and services to deploy many functions, such as cloud- blockchain technology has been gradually gaining attention. Blockchain
based automated data collection, storage, duplication, access control, uses cryptography-related methods to generate multiple linked data
and data sharing (Liu et al., 2012; Shen et al., 2017). blocks, with each block containing important information to verify its
However, there exist certain security and privacy issues and con- validity and to generate the next block. Being a distributed database,
cerns when storing and sharing data on cloud service platforms pro- blockchain exhibits the characteristics of non-tamperability, privacy
vided by third parties (Mollah et al., 2017; Chen et al., 2019a). The protection, and decentralization, which are helpful in establishing a

☆
This work is supported by National Natural Science Foundation of China (No. 61672131), the National Key R&D Program of China (No. 2019YFB1703601) and
Tianjin Science and Technology Project (No.18YFCZZC00060 and No.18ZXZNGX00100).
∗
Corresponding author.
∗∗
Corresponding author.
E-mail addresses: [email protected] (J. Liu), [email protected] (Y. Jin).

https://doi.org/10.1016/j.jnca.2020.102710
Received 3 March 2020; Received in revised form 17 April 2020; Accepted 13 May 2020
Available online 9 June 2020
1084-8045/© 2020 Elsevier Ltd. All rights reserved.
J. Chi et al. Journal of Network and Computer Applications 167 (2020) 102710

secure data-sharing mechanism (Ritzdorf et al., 2018; Dinh et al., 2018). platform and completed the data federation of multiple medical service
With the development of blockchain technology, it has become an obvi- providers by designing smart contracts. However, their scheme always
ous trend to apply blockchain for data sharing to various fields, such leads to low transaction throughput and high transaction overhead. The
as medical care (Xia et al., 2017), Internet of vehicles (Chen et al., performance of the alliance chain-based scheme is much better than
2017a, 2019b, 2020; Li et al., 2019), and power industry (Gai et al., that of the public chain-based one, as the former contributes to the user
2019). Existing studies on secure data sharing based on blockchain experience of data sharing. Liang et al. (2017) proposed a user-centered
have achieved some positive results (Kang et al., 2019). However, data-sharing scheme that features privacy protection and identity man-
although previous works have solved the issues of security and privacy agement and implemented it with the alliance chain. However, these
of data sharing by designing a data-sharing framework and a privacy- schemes do not consider the attributes and scope of the data to be
protection model, the attributes and scope of the data to be shared shared. In our proposed framework, Hyperledger Fabric (Thakkar et
are usually not considered. Generally, blockchain can only conduct a al., 2018), the mainstream architecture of alliance chain, is adopted to
transaction query according to the transaction number, address num- ensure data security and improve the performance of data sharing.
ber, or block number. Therefore, the improvement of the efficiency of The community structure in complex networks has been extensively
data sharing while ensuring its security using blockchain technology investigated. Generally, a community structure refers to a community
remains to be a huge challenge (Qiu et al., 2020). with attributes that are the same as or similar to those of a complex net-
In this paper, we propose a secure and efficient data-sharing scheme work, and different community structures have different properties and
based on blockchain and community detection that considers the rel- characteristics in a specific network. At present, community division
evance and sharing scope of the data to be shared. In this scheme, methods are divided into two categories: similarity-based (Chen et al.,
we designed a data-sharing framework based on Hyperledger Fabric 2017b) and modularity-based (Liu and Liu, 2018) methods. Sammarco
that uploads a large amount of data obtained from sensors and pre- etal. (2015) performed community division through trace similarity to
pares them for sharing through the blockchain network. A large num- fully utilize the WLAN monitoring systems and capture as much wire-
ber of clients are divided into multiple communities according to the less traffic as possible. Xia et al. (2016) integrated three different social
correlation and similarity of the collected data, and the data are only dimensions into a unified distance function to represent the closeness
shared among smart factories within the same community, which is of users to other users in a social perception network, also known as
conducive to the dissemination and sharing of useful information and similarity. Williams et al. (2016) proposed a method for dynamic com-
which greatly improves the value density of the data. We also designed munity division detection.
a community detection algorithm to control the specific steps of the
community division and used the sharing degree as the criterion to
3. The data sharing framework based on blockchain and
measure the effectiveness of the community division.
community detection
The main contributions of this paper are as follows:

1. A secure and efficient data-sharing framework based on blockchain 3.1. Framework overview
and community detection was proposed. The framework is com-
posed of the data layer, detection layer, and blockchain layer, which To provide fine-grained data-sharing services, we classified the
perform the processes of identity authentication, data interaction, transactions stored in the blockchain database according to the level of
community detection, blockchain storage, etc. privacy. The privacy levels include public data, community public data,
2. As a part of the data-sharing framework, a community detection and encrypted data in descending order. Here, public data refer to the
algorithm was developed to divide clients into different communi- data that can be seen by all nodes, community public data refer to the
ties based on the label data. Data are shared within a specific scope, data that can be seen by all nodes belonging to the same community,
thus improving the efficiency of data sharing. and encrypted data mainly refer to private data and data that users want
3. The sharing degree and the sum of squared error were jointly used as to buy and sell. We hope that when users share professional data, they
the indices for judging and optimizing the quality of the community will set the data privacy level to public data in the community, so that
detection. the data can be seen by more users who really need them. Therefore,
our main work is to divide the community reasonably by developing a
The remainder of this paper is structured as follows. Section 2 dis- community detection algorithm so that the public data of the commu-
cusses related work done. Section 3 introduces the data-sharing frame- nity can be shared with more users who really need them. Fig. 1 illus-
work based on blockchain. Section 4 illustrates the community detec- trates the data-sharing framework based on blockchain. There are three
tion algorithm in detail. The simulation results are discussed in Section layers in our proposed data sharing framework, including Data Layer,
5, and the concluding remarks of this study are provided in Section 6. Detection Layer and Blockchain layer. Data is collected by Data Layer
and transmitted to Detection Layer. Detection Layer executes commu-
2. Related works nity detection algorithm and divide clients into different communities,
narrowing the scope of data sharing. Blockchain Layer is responsible for
With the maturity of cloud computing technology, many data- keeping the community detection result and transaction records safe.
sharing schemes based on this technology have emerged. For instance,
Wang et al. (2010) proposed the first data-sharing scheme that could
3.1.1. Data layer
provide privacy protection and open auditing methods by utilizing
The data layer includes the perception data generated by large-
cloud storage services. However, the data-sharing schemes based on
scale sensors (Qiu et al., 2019b), including product type and quantity,
cloud computing cannot solve the risk of data privacy leakage caused
machine running state, and other information of many parameters. Per-
by data centralization.
ception data are obtained by sensors and uploaded to the client server
Blockchain has shown its potential to transform traditional indus-
for comprehensive analysis and sharing.
tries owing to its essential features of decentralization, durability,
anonymity, and auditability (Zheng et al., 2017). There are two types
of secure data-sharing schemes based on blockchain, namely, public 3.1.2. Detection layer
chain-based scheme and alliance chain-based scheme. The public chain- The detection layer consists of clients, a client server, and a commu-
based scheme allows any node to join and exit freely, and the data are nity detection server. The client server is mainly responsible for collect-
more secure owing to the characteristic of strong distribution. Azaria et ing the perception data and uploading them to the blockchain network
al. (2016) realized the sharing of patient data based on the Ethereum and to the community detection server. Label data are generated when a

2
J. Chi et al. Journal of Network and Computer Applications 167 (2020) 102710

3. Certificate Authority Server (CAS): The CAS issues a digital certificate

to each client using a public key to prove that the client listed in the
certificate legally owns the public key in the certificate.
4. Community Detection Server (CDS): The CDS is responsible for collect-
ing all tag data, executing community detection algorithm, generat-
ing community detection results and uploading to the blockchain.
5. Endorser Node (EN): The ENk , k ∈ 1, 2, …, is responsible for endors-
ing the transaction proposal submitted by the client server.
6. Committer Node (CN): The CNl , l ∈ 1, 2, …, is responsible for vali-
dating and adding blocks to the blockchain.
7. Order Node (ON): The ON is responsible for sorting and packaging
transactions into blocks.

3.3. Interaction process

The interaction process of the data-sharing framework based on

blockchain and community detection is described as follows and is illus-
trated in Fig. 2. The interaction process consists of four phases, namely
initialization phase, identity authorization phase, signature and verifi-
cation phase, and data sharing phase. In the initialization phase, the
key generation and ID generation of CS and C are mainly completed.
The purpose of the authentication phase is to authenticate both par-
ties before establishing a connection and exchanging information. Sig-
nature and verification phase is responsible for ensuring that data is
not tampered with during transmission. In the data sharing phase, the
client is divided into several communities by the community detection
algorithm, and the sharing degree is used as the index of iterative opti-
mization. The results of community partitioning will be uploaded to the
block chain layer based on Hyperledger Fabric to ensure the tamper-
proof and verifiability of community detection results and transactions.
Fig. 1. The data sharing framework based on blockchain.

3.3.1. Initialization phase

The initialization phase mainly completes the key generation and ID
user joins the affiliate chain. The community detection server is respon-
generation of CS and C related tasks.
sible for collecting all label data, executing the community detection
algorithm, generating the community detection results, and uploading 1. CAS selects a prime number p and two multiplicative groups, i.e.,
to the blockchain network. After the community detection results are G1 and G2 , which are of the order of p. G is a generator of G1 . g is a
successfully written to the blockchain database, the client will execute generator of G2 . f is a bilinear map satisfying f(G1 , G1 ) = G2 .
the smart contract to query the shared data from the community. 2. CAS randomly generates two m × n-dimensional matrices, i.e., Acs
and Ac , and two m-dimensional column vectors, i.e., bcs and bc ,
where R(Acs ) <min{m, n}, which means that the linear equations
3.1.3. Blockchain layer
The blockchain layer is constructed according to the Hyperledger
Acs 𝛼cs = bcs (1)
Fabric framework; it is composed of endorser nodes, order nodes, and
committer nodes. The endorser node is responsible for endorsing the and
transaction proposal submitted by the client server. The order node
Ac 𝛼 c = bc (2)
is responsible for sorting and packaging the transactions into blocks.
The committer node is responsible for validating and adding blocks to have infinite solutions. We denote by 𝛼 cs the solution of linear equa-
the blockchain database. The typical transaction process in fabric is as tion (1) and by 𝛼 c the solution of linear equation (2).
follows: 1) the client side constructs the transaction proposal; 2) the 3. For each client Ci and each client server CSj in the system, CAS
endorsement node simulates the execution of the transaction; 3) the calculates the legal n-dimensional columns 𝛼ci and 𝛼csj , and send 𝛼ci
client sends the transaction to the consensus service; 4) orderer orders and 𝛼csj to the corresponding Ci and Ccs as a proof of identity.
transactions by consensus, generates new blocks, and submits transac- 4. CAS randomly generates two n-dimensional column vectors, i.e., 𝜌c
tions. The data sharing framework in this paper uses the above transac- and 𝜌cs , respectively calculating IDCi and IDCSj by
tion process in the data sharing phase.
IDCi = 𝜌Tc 𝛼ci (3)

3.2. The entities and

IDCSi = 𝜌Tcs 𝛼csi (4)

1. Client (C): The client Ci , i ∈ 1, 2, …, is responsible for collecting all Then, CAS stores Acs , Ac , bcs , bc , 𝜌cs , 𝜌c , IDCi , and IDCSj ; sends Ac , bc ,
kinds of perception data. 𝜌c , and IDCSj to each CSj ; sends IDCi to each Ci ; and then sends Acs ,
2. Client Server (CS): The client server CSj , j ∈ 1, 2, …, is mainly bcs , and 𝜌cs to CDS.
responsible for collecting the perception data from Clients and 5. For each CSj , CAS randomly selects two private keys, i.e., Xj1 and
uploading the perception data to the blockchain and community Xj2 ∈ Zp∗ , and generates two public keys, i.e., Yj1 = Xj1 G and
detection server. Yj2 = Xj2 G. For CDS, CAS randomly selects two private keys,

3
J. Chi et al. Journal of Network and Computer Applications 167 (2020) 102710

Fig. 2. Interaction process of proposed framework.

i.e., Xcds1 and Xcds2 ∈ Zp∗ , and generates two public keys, i.e., If (10) is correct, client Ci and client server CSj have consistent and
Ycds1 = Xcds1 G and Ycds2 = Xcds2 G. The one-way hash function is legal identity parameters. Then, client Ci and the client server CSj
denoted as H ∶ {0, 1}∗ → G1 . CSj and Ci can both obtain the public have authenticated each other successfully.
parameters {p, G1 , G2 , f, G, Yj1 , Yj2 , Ycds1 , Ycds2 , H}

3.3.3. Signature and verification phase

3.3.2. Identity authentication phase Data signature and validation are used to ensure that data are not
The purpose of the authentication phase is to authenticate both par- tampered with during transmission. The signature and verification steps
ties before establishing a connection and exchanging information. between the client and the client server are as follows.

1. If a client Ci wants to communicate with a client server CSj , CSj ran- 1. CSj randomly selects a 𝛽csj ∈ ZP∗ and calculates 𝛾csj by 𝛾csj = 𝛽csj G.
domly selects r ∈ Zp∗ and respectively calculates the authentication Then, 𝛾csj is sent to client Ci . Ci randomly selects a 𝛽ci ∈ ZP∗ and
parameters u1 and u2 by calculates 𝛾ci = 𝛽ci 𝛾csj . Ci randomly generates two parameters, i.e.,
y1 and y2 , respectively calculating 𝜃ci and 𝜇ci by
u1 = rAc (5)
𝜃ci = y1 H (IDci ||𝛾ci ||Mi ) + y2 G (11)
and

u2 = rbc (6) 𝜇ci = y1 𝛽ci , (12)

Then, the authentication message mess1 = {t1 , u1 , H (t1 ||u2 ||IDCSj )} where Mi is the message that needs to be uploaded and the signature
will be delivered to Ci from CSj , where t1 is the timestamp of this key to Ci is SigKeyci = {𝜃ci , 𝜇ci }.
message. 2. Ci sends SigKeyci to CSj . CSj calculates Si by
2. Client Ci receives the authentication message mess1 from CSj and
calculates uci by Si = Xj1 𝜃ci + Xj2 𝛽csj 𝜇ci G (13)

and sends Si back to Ci . Ci calculates Ri by

uci = u1 𝛼ci (7)
Ri = y1−1 (Si − y2 Yj1 ) (14)
Client Ci doesn’t have the parameters Ac and bc . However, the
parameter uci calculated by Ci is equal to u2 received from CSj Finally, Ci obtains the signature 𝜎ci = {IDCi , Mi , Ri , 𝛾ci } of message
according to the equation Mi .
3. CDS receives the message signature 𝜎ci of Mi sent from Ci , and the
uci = u1 𝛼ci = rAc 𝛼ci = rbc = u2 (8)
authenticity of 𝜎ci then needs to be verified. CDS will check the
If (8) is correct, the equation correctness of the equation

H ({(t1 ||u2 ||IDCSj )}) = H ({(t1 ||uci ||IDCSj )}) (9) f (Ri , G) = f (𝛾ci , Yj2 )f (H (IDci ||𝛾ci ||Mi ), Yj1 ) (15)

is obviously correct. That means the client Ci has verified the legality proof :
of client server. Then, client Ci sends the authentication message Ri = y1−1 (Si − y2 Yj1 )
mess2 = {t2 , IDci , H (t2 ||uci ||IDCSj )} to CSj , where t2 is the timestamp
of this message. = y1−1 (Xj1 𝜃ci + Xj2 𝛽csj 𝜇ci G − y2 Xji G)
3. Client CSj receives the authentication message mess2 from Ci and
verifies the correctness of equation = y1−1 (Xj1 (y1 H (IDci ||𝛾ci ||Mi ) + y2 G)

H ({(t2 ||uci ||IDCSj )}) = H ({(t2 ||u2 ||IDCSj )}) (10) + Xj2 𝛽csj y1 𝛽ci G − y2 Xji G)

4
J. Chi et al. Journal of Network and Computer Applications 167 (2020) 102710

= y1−1 Xj1 y1 H (IDci ||𝛾ci ||Mi ) + y1−1 y2 Xj1 G better shared within the community and the value density of the shared
data can be increased.
+ y1−1 Xj2 𝛽csj y1 𝛽ci G − y1−1 y2 Xji G

= Xj1 H (IDci ||𝛾ci ||Mi ) + Xj2 𝛽csj 𝛽ci G 4.1. Similarity calculation

According to the mathematical properties of linear map (Boneh and Each piece of data collected from the client includes label data
Franklin, 2003), and real data. The proposed community detection algorithm needs
to calculate the similarity of label data. The attributes of the label
f (Ri , G) = f (Xj1 H (IDci ||𝛾ci ||Mi ) + Xj2 𝛽csj 𝛽ci G, G)
data include clientId, city, district, communicationProtocol, gatewayType,
= f (Xj2 𝛽ci 𝛽csj G, G)f (Xj1 H (IDci ||𝛾ci ||Mi ), G) machineName, and sensorName, and each attribute is stored as a key-
value pair. Because the value of each key is textual data, the commonly
= f (𝛽ci 𝛽csj G, Xj2 G)f (H (IDci ||𝛾ci ||Mi ), Xj1 G) used numerical similarity calculation methods such as Euclidean dis-
tance are not applicable. Therefore, we use cosine similarity (Fauzi et
= f (𝛾ci , Yj2 )f (H (IDci ||𝛾ci ||Mi ), Yj1 ) al., 2017) to measure the similarity of label data. We denote the vector
space mapping of label data from two clients as X (x1 , x2 , … , xn ) and
If (15) is correct, the message Mi from Ci is validated successfully
Y (y1 , y2 , … , yn ).
and stored by CDS.
When calculating the cosine similarity of a pair of label data DC1
and DC2 , read the key to DC1 and DC2 successively and compare whether
the corresponding values are the same. If they are the same, the vector
3.3.4. Data sharing phase
bits of X and Y corresponding to the label are set to 1. Otherwise, it
In the data-sharing phase, CDS collects and validates data from
means that the same key for DC1 and DC2 corresponds to two different
many clients. With a large amount of data collected, CDS uses the com-
values. Therefore, two vector bits of X and Y are needed to represent
munity detection algorithm to divide clients into communities and uses
this situation, and the two vector bits of X and Y are set to (0, 1) and
the share degree as the index for iterative optimization. The results
(1, 0), respectively. After all the key-value pairs of the pair of label data
of the community division will be uploaded to the blockchain module
are compared, vectors X and Y are constructed and the cosine similarity
based on Hyperledger Fabric to ensure the tamperproofness and verifi-
cosc1 ,c2 (𝜃) of DC1 and DC2 is calculated by
ability of the community detection results.
∑n
1. The community detection server CDS collects label data from all (xi × yi )
cosc1 ,c2 (𝜃) = √ i=1
√ ∈ [0, 1] (16)
clients, verifies the signature, and executes the community detec- ∑n 2 ∑n
i=1 (xi ) × i=1 (yi )2
tion algorithm. The community detection algorithm is discussed in
Section 4.
2. CDS is a node of the blockchain network. When CDS obtains the
4.2. Clustering process
community test results, it initiates a transaction request to store the
results in the blockchain network.
Clustering is used in the community detection algorithm. In this
3. The endorser nodes ENk , k ∈ 1, 2, …, verify the authenticity of the
study, we used an improved K-medoids clustering algorithm to achieve
results. If the verification passes, ENk will endorse the transaction
clustering of clients. The K-medoids (Yu et al., 2018) clustering algo-
from CDS.
rithm is suitable for situations where the distance formula is not
4. If the transaction from CDS obtains enough endorsement from ENk ,
numeric.
the community detection result has taken effect and been sent to the
The detailed clustering process is described in Algorithm 1. Ran-
order node ON.
domly select k points from n sample points of the population as
5. ON sorts the community detection result into blocks and broadcasts
medoids. Assign the remaining n − k points to the class currently best
the blocks to the committer nodes CNl , l ∈ 1, 2, ….
represented by medoids according to the principle of nearest distance
6. All of the CNl complete the final verification and put the blocks from
to the medoids. For all the points in the ith class except for the corre-
ON on the blockchain network.
sponding medoids points, calculate the value of the criterion function in
7. As the nodes of the blockchain network, CS can obtain the commu-
order when it is the new medoids, iterate through all possibilities, and
nity detection result from CNl by invoking the smart contract and Ci
select the point corresponding to the minimum criterion function as the
can obtain its own community information from CS.
new medoids. Repeat the above process until all the medoids points no
8. If a user wants to search a specific record by Ci , the user enters the
longer change or have reached the set maximum number of iterations
tags of interest to form a data querying request. After the querying
and then output k classes.
request is submitted to CS, the similarity between the requested data
and the community center node data of the user is first calculated
Algorithm 1 Clustering
in CS. If the similarity is high, the querying request will be retrieved
from the community by CDS. If it is low, the similarity between Input: K, G
the query request and other community center nodes will be cal- Output: Cfinal
culated, and the community with high similarity will be selected 1: function CLS(K, G)
to retrieve the data in CS. Eventually, the data from the selected 2: /∗Initialize the cluster center∗/
community with greater similarity to the result of the query request 3: RandomCenters(K) = {E}
are retrieved by CDS and shared with Ci , which initiated the data 4: /∗Start allocation client∗/
querying request. 5: for Gi ∈ G, Ej ∈ E do
6: if cos𝜃 (Gi , Ej ) > cos𝜃 Max then
7: Add the Gi to the community of Ej
4. Community detection algorithm 8: end if
(continued on next page)
The purpose of community detection is to cluster users based on
the uploaded data. According to the uploaded data, users with greater
similarity are divided into the same community, so that the data can be

5
J. Chi et al. Journal of Network and Computer Applications 167 (2020) 102710

4.4. Community detection algorithm

Algorithm 1 (continued)
9: end for Based on the above discussion, we propose a community detection
10: /∗Update the cluster center∗/
algorithm in this section. This algorithm executes as follows. CDS reads
11: for commu ∈ {Commu} do
12: Find Ci ∈ commu satisfying Min{S2 (Ci )} the label data for all the clients first. Then, CDS execute the cluster-
13: end for ing algorithm successively to calculate the SSE, SD, and community
14: /∗iteration until the community no changing∗/ detection result of different K values. The local optimal solution of each
15: if Cfinal ! = {Commu} then
iteration is obtained according to the strategy of taking the result cor-
16: Start the next round allocation
17: end if
responding to the maximum SD under the condition of guaranteeing
18: return Cfinal a small SSE first. Finally, the global optimal solution is obtained by
19: end function enough iterations under the constraints of SSE and SD. Algorithm 2
describes the detailed process of the community detection algorithm,
and the variables used in the algorithm are described as follows.

When updating the community, instead of taking the center of grav- • Iter: The number of iterations, i.e., the number of times the clustering
ity of all nodes in the current community as the clustering center, use algorithm is executed.
one node existing in the current community as the clustering center. • Sbest : The local optimal solution of the sharing degree.
The criterion function is to minimize the sum of all other nodes in the • Sfinal : The global optimal solution of the sharing degree.
current community from the clustering center. This will weaken the • Δy: The difference between the function value of the oblique line
effect of outliers to some extent. formed by the beginning and end of the SSE curve and the function
The community detection algorithm we propose is implemented value of the SSE curve.
using the K-medoids clustering algorithm. There are two main evalu- • BestK(): The K value with the highest degree of sharing that is found
ation indices of the clustering algorithm, namely, the sum of squared near the optimal K value.
error (SSE) and the contour coefficient method. However, because the
evaluation method of the contour coefficient method is unstable, we
choose the SSE to evaluate the clustering effect. Algorithm 2 Community Detection Algorithm
∑ ∑
k Input: K, Kmax , Iter, G, D
SSE = |1 − cos 𝜃 |2 (17) Output: Cfinal
i=0 cos 𝜃∈ci
1: Sfinal = 0.0
2: while i = 0; i < Iter; i + + do
4.3. Sharing degree 3: Sbest = 0.0
4: while j = K;j < Kmax ; j + + do
We divide each node into communities according to the similarity of 5: Ccurrent = CLS(K, G)/∗Clustering∗/
the labels, to make the shared data useful to other nodes in the commu- 6: Scurrent = CSD(Cbest , D, G)/∗Calculate SD∗/
nity where the node is located. If the similarity of interest of the label 7: /∗Calculate SSE for each K value∗/
data of community nodes varies greatly, the case may be that some 8: /∗Calculate Δy in the mathematical form of
nodes are not interested in the data shared by some nodes, and even the elbow method∗/
a large degree of sharing is meaningless. Therefore, we denoted the 9: if Δy > ΔyM ax then
sharing degree SD as an evaluation index of the quality of community 10: ΔyM ax = Δy
detection: 11: Cbest = Ccurrent
∑i,j=n,k 12: end if
Q × Ncommuj
i,j=1,1 Ci
SD = (18) 13: end while
Qtotal 14: if BestK() then
Where Q is the number of community public data uploaded by a 15: Cfinal = Cbest
client; Qtotal is the number of data uploaded by all clients; N is the total 16: end if
number of nodes in a community; Ci is the ith client; commuj is the jth 17: end while
community. 18: return Cfinal
When the community public data and the total data amount are
given, the number of clients in each community and the number of pub-
lic data in the community are different in different community detection 5. Experimental simulation
results. The community detection results with the largest proportion of
public data of all communities will be the best, which can be measured 5.1. Data set
by SD. The sharing degree and the error function jointly determine the
quality of the community detection. The smaller the error function is, Based on the real data of a smart factory, we generate 3 data sets
the more similar the nodes in the community are and the higher the according to the number of clients and the proportion of public data in
degree of sharing is, indicating that the public data in the community the community, as shown in Table 1. The data format is JSON string, the
are shared by more nodes. number of clients includes 128 and 256, and the proportion of commu-

Table 1
Dataset description.
DataSet Name Clients Community public Data All The Data Ratio
RandData1 128 344 430 0.8
RandData2 128 162 405 0.4
RandData3 128 41 410 0.1

6
J. Chi et al. Journal of Network and Computer Applications 167 (2020) 102710

nity public data r includes 0.1, 0.4, and 0.8. Each piece of data includes
label data and real data. The attributes of the label data mainly include
clientId, city, district, communicationProtocol, gatewayType, machineName,
and sensorName, and the attributes of the real data include the operat-
ing state parameters of the machines and sensors.

5.2. Performance of community detection algorithm

The proposed community detection algorithm is based on the K-

medoids clustering algorithm. The selection of k directly affects the
performance of the clustering algorithm. If k is selected well, the algo-
rithm will converge. Otherwise, the algorithm will always fall into the
local optimal solution. Through multiple experiments, it was found that,
when the initial clustering center is selected as far as possible each time,
it is difficult to jump out of the local optimum regardless of how many
iterations there are. When random selection is adopted, performance
will greatly increase as the number of iterations increases. Therefore,
we conclude that it is easier to jump out of the local optimal solution by Fig. 4. The dynamics of SD with the change of k.
randomly selecting the initial clustering center. According to previous
research√ experience, we can narrow the selection range of k according
to k < n.
The goal of the community detection algorithm is to ensure the the elbow method.
efficient sharing of data within the community. We must ensure that The elbow method can be used as follows: connect the head point
the public data shared by each client are of interest to other nodes in and tail point of the SSE curve (blue) by a straight line (red), as shown
the same community, to ensure the efficient sharing of data. Gener- in Fig. 3, and calculate the difference ΔY between the red straight line
ally, the clustering effect can be evaluated with SSE and the optimal and the blue SSE curve for each k. The k value corresponding to the
clustering result can be obtained. However, in our scheme, the cluster- maximum ΔY is the potential optimal kop . Then, SD should also be used
ing result evaluated by SSE does not necessarily make the data sharing as an index for choosing k. Select kop , kop − 1, kop + 1 as the candidate
efficient, i.e., the optimal sharing effect will not be obtained if SSE is optimal k and find the one with the greatest SD as the final k. For
the only index considered. Therefore, we need to comprehensively con- instance, according to the elbow method, the potential optimal kop for
sider the two evaluation indices of SSE and SD. The measure we take is clustering is 5 in Fig. 3. Comparing the corresponding SD of 4, 5, and
to increase SD as much as possible to ensure the accuracy of the com- 6, we can obtain the final k as 4, which has the greatest SD.
munity division result. A single consideration of sharing will only make After the optimal k is selected, we need to know how many itera-
the results unsatisfactory and the sharing efficiency will be low. tions can achieve the optimal SD. We selected Dataset1, Dataset2, and
We first choose RandData1 as the experimental object and execute Dataset3 as our experimental objects. For the three data sets, the num-
the community detection algorithm by choosing different values of k. ber of clients is 128; the number of data pieces is 430, 405, and 410; and
The algorithm will continuously add clients with similar attributes to the proportion of community public data is 0.1, 0.4, and 0.8, respec-
the same community and update the cluster center in time during the tively. By constantly adjusting the number of iterations, we improved
iteration process. Figs. 3 and 4 show the change in SSE and sharing the sharing degree as much as possible by reducing SSE. As shown in
degree during clustering. According to the theory of the elbow method, Fig. 5, when r = 0.1, SD increases from 2 to 6; when r = 0.4, SD
SSE will gradually decrease with the increase in k and the decreasing increases from 8 to 24; and when r = 0.8, SD increases from 35 to 64.
degree of k will qualitatively change when the optimal k appears. Before The experimental results showed that the maximum SD can be obtained
the emergence of the optimal k, the value of SSE will decrease rapidly. through about 500 iterations, which indicates that the optimal data-
After the optimal k appears, the decreasing trend of SSE with k will be sharing effect has been achieved by about 500 iterations.
flat. Therefore, the optimal k can be found on the basis of the theory of

Fig. 3. The dynamics of SSE with the change of k. Fig. 5. The dynamics of SD with the iterations of different r.

7
J. Chi et al. Journal of Network and Computer Applications 167 (2020) 102710

of different m is obvious. With the same number of querying requests,

it takes less time when m = 6. This means that, the more the number
of querying labels is entered, the faster the queried data are found in
the community and the smaller the time overhead is. Therefore, as long
as more query labels are given, the shared data needed can be obtained
faster, and this trend will become apparent as m increases.
The performance evaluation of the proposed secure data-sharing
framework generally focuses on the throughput and time cost. Through-
put mainly refers to the number of successful transactions per second,
and the main factors affecting the throughput are server performance;
number of nodes; consensus algorithms; encryption algorithms; trans-
action arrival speed, i.e., the number of transactions that reach the
blockchain system per second; and total business volume. The time
cost mainly includes 1) the network latency for initiating a request and
receiving a response; 2) the execution time of calling the chain code;
3) the time for consensus between nodes, mainly including the time for
endorsement, sorting, and verification; and 4) the time for encryption
and decryption.
Fig. 6. The dynamics of time cost with the querying requests of different m.
In this study, we used a shell script to simulate the interaction pro-
cess with blockchain. The client number was set to 2 and 5, respec-
tively; the request arrival speed was set to 15 and 20, respectively. For
5.3. Performance of proposed data sharing framework
the different numbers of community and query requests, stress tests
were executed. As shown in Fig. 7(a), the time cost decreases with the
The performance experiments of the proposed data-sharing frame-
increase in the number of concurrent processes before the critical value.
work were executed on an ECS server with a dual-core Intel(R) Xeon(R)
As shown in Fig. 7(b), when Concurrent = 2 and ArrivalRate = 15, the
CPU clocked at 2.50 GHz, 8 GB of RAM, and running CentOS 7.6. We
querying throughput is close to 15; when Concurrent = 5 and Arrival-
used Docker to build the data-sharing framework based on blockchain
Rate = 20, the querying throughput is close to 20. This means that the
and community detection, which includes 1 community detection node,
querying requests to the blockchain framework can be handled on time.
1 order node, 1 CA node, 2 client server nodes, 4 peer nodes, and 128
The concurrent number of processes has a great impact on framework
client nodes. The data collected from the clients are stored directly in
performance.
the community node for calculation. The community detection algo-
The impact of the number of communities on performance is illus-
rithm is written in Java, and the smart contract is written in Go.
trated in Fig. 8. When Concurrent = 5 and the number of communities
If a user wants to retrieve a piece of shared data through the pro-
reaches 800, the CPU utilization almost reaches the peak, which indi-
posed framework, the labels of the queried data need to be entered. By
cates that the blockchain network has high performance requirements
changing the number of querying labels m, we can try to find the rela-
for the CPU. In this study, multiple nodes were deployed on a single
tionship between the number of querying requests and the time cost.
server. Therefore, the performance bottleneck of the proposed frame-
When the labels of the queried data are entered, the number of labels
work resulting from the server is very significant. However, with the
m is calculated. At the same time, the similarity between the label data
further development of the server, the performance of the proposed
of the queried data and those of each community center node is cal-
framework will improve significantly. Many studies have shown that
culated and the community with greater similarity is selected for fur-
the performance of the blockchain system based on Hyperledger Fab-
ther retrieval. The greater the similarity between the label data of the
ric has gradually met the business requirements. Therefore, server per-
queried data and those of each community center node, the greater the
formance is no longer the main bottleneck restricting the implementa-
possibility that the queried data will be obtained in this community. As
tion of blockchain, while the secure and trusted services provided by
shown in Fig. 6, the dynamics of time cost with the querying requests
blockchain will become a huge advantage.

Fig. 7. The performance evaluation of proposed secure data sharing framework. (a) Time Cost. (b) Throughput (QPS).

8
J. Chi et al. Journal of Network and Computer Applications 167 (2020) 102710

Chen, D., Yan, Y., Wang, D., Huang, X., 2017. Community detection algorithm based on
structural similarity for bipartite networks. In: 2016 7th IEEE International
Conference on Software Engineering and Service Science (ICSESS), pp. 98–102.
Chen, N., Qiu, T., Zhou, X., Li, K., Atiquzzaman, M., 2019. An intelligent robust
networking mechanism for the internet of things. IEEE Commun. Mag. 57 (11),
91–95.
Chen, C., Hu, J., Qiu, T., Atiquzzaman, M., Ren, Z., 2019. Cvcg: cooperative v2v-aided
transmission scheme based on coalitional game for popular content distribution in
vehicular ad-hoc networks. IEEE Trans. Mobile Comput. 18 (12), 2811–2828.
Chen, C., Wang, C., Qiu, T., Lv, N., Pei, Q., 2020. A secure content sharing scheme based
on blockchain in vehicular named data networks. IEEE Trans. Indust. Inform. 16 (5),
3278–3289.
Dinh, T.T.A., Liu, R., Zhang, M., Chen, G., Ooi, B.C., Wang, J., 2018. Untangling
blockchain: a data processing view of blockchain systems. IEEE Trans. Knowl. Data
Eng. 30 (7), 1366–1385.
Fauzi, M.A., Utomo, D.C., Setiawan, B.D., Pramukantoro, E.S., 2017. Automatic essay
scoring system using n-gram and cosine similarity for gamification based e-learning.
In: Proceedings of the International Conference on Advances in Image Processing.
ACM, pp. 151–155.
Gai, K., Wu, Y., Zhu, L., Xu, L., Zhang, Y., 2019. Permissioned blockchain and edge
computing empowered privacy-preserving smart grid networks. IEEE Internet of
Things J. 6 (5), 7992–8004.
Kang, J., Yu, R., Huang, X., Wu, M., Maharjan, S., Xie, S., Zhang, Y., 2019. Blockchain
Fig. 8. The dynamics of time cost and CPU utilization with the number of for secure and efficient data sharing in vehicular edge computing and networks.
communities. IEEE Internet of Things J. 6 (3), 4660–4670.
Khan, M.A., 2016. A survey of security issues for cloud computing. J. Netw. Comput.
Appl. 71, 11–29.
Li, J., Yu, F.R., Deng, G., Luo, C., Ming, Z., Yan, Q., 2017. Industrial internet: a survey
6. Conclusion on the enabling technologies, applications, and challenges. IEEE Commun. Surv.
Tutor. 19 (3), 1504–1526.
In this paper, we propose a secure and efficient data-sharing scheme Li, M., Zhu, L., Lin, X., 2019. Efficient and privacy-preserving carpooling using
blockchain-assisted vehicular fog computing. IEEE Internet of Things J. 6 (3),
based on blockchain and a community detection algorithm. Under the 4573–4584.
premise of ensuring data security, the proposed scheme provides more Liang, X., Zhao, J., Shetty, S., Liu, J., Li, D., 2017. Integrating blockchain for data
fine-grained data-sharing services by classifying clients using label data. sharing and collaboration in mobile healthcare applications. In: 2017 IEEE 28th
Annual International Symposium on Personal, Indoor, and Mobile Radio
There are four phases to achieve the secure and efficient data shar-
Communications (PIMRC). IEEE, pp. 1–5.
ing,including initialization phase, identity authentication phase, signa- Liu, C., Liu, Q., 2018. Community detection based on differential evolution using
ture and verification phase, data sharing phase. The community detec- modularity density. Information 9 (9), 218.
tion server is the key to the data-sharing scheme. In this scheme, the Liu, X., Zhang, Y., Wang, B., Yan, J., 2012. Mona: secure multi-owner data sharing for
dynamic groups in the cloud. IEEE Trans. Parallel Distr. Syst. 24 (6), 1182–1191.
community detection server receives and analyzes the label data of Mollah, M.B., Azad, M.A.K., Vasilakos, A., 2017. Security and privacy challenges in
all clients, detects the community through cosine similarity and the mobile cloud computing: survey and way ahead. J. Netw. Comput. Appl. 84, 38–54.
K-medoids clustering algorithm, and finally, uploads the community Qiu, T., Li, B., Qu, W., Ahmed, E., Wang, X., 2019. Tosg: a topology optimization scheme
with global small world for industrial heterogeneous internet of things. IEEE Trans.
detection results to the blockchain network. Users can obtain com- Indust. Inform. 15 (6), 3174–3184.
munity results and share data through the blockchain client. We also Qiu, T., Liu, J., Si, W., Wu, D.O., 2019. Robustness optimization scheme with
discuss how to improve the quality of community partitioning based multi-population co-evolution for scale-free wireless sensor networks. IEEE/ACM
Trans. Netw. 27 (3), 1028–1042.
on SSE and SD, as well as the performance advantages of the pro- Qiu, T., Li, B., Zhou, X., Song, H., Lee, I., Lloret, J., 2020. A novel shortcut addition
posed framework in terms of time cost and throughput as evaluated by algorithm with particle swarm for multisink internet of things. IEEE Trans. Indust.
experiments. Experimental simulation results showed that the proposed Inform. 16 (5), 3566–3577.
Ritzdorf, H., Soriente, C., Karame, G.O., Marinovic, S., Gruber, D., Capkun, S., 2018.
scheme can effectively improve the efficiency of data sharing. Toward shared ownership in the cloud. IEEE Trans. Inf. Forensics Secur. 13 (12),
3019–3034.
CRediT authorship contribution statement Sammarco, M., Campista, M.E.M., de Amorim, M.D., 2015. Scalable wireless traffic
capture through community detection and trace similarity. IEEE Trans. Mobile
Comput. 15 (7), 1757–1769.
Jiancheng Chi: Conceptualization, Methodology, Software. Yu Li: Shen, J., Zhou, T., Chen, X., Li, J., Susilo, W., 2017. Anonymous and traceable group
Data curation, Writing - original draft. Jing Huang: Visualization, data sharing in cloud computing. IEEE Trans. Inf. Forensics Secur. 13 (4), 912–925.
Investigation. Jing Liu: Supervision. Yingwei Jin: Software, Valida- Singh, S., Jeong, Y.-S., Park, J.H., 2016. A survey on cloud computing security: issues,
threats, and solutions. J. Netw. Comput. Appl. 75, 200–222.
tion. Chen Chen: Writing - review & editing. Tie Qiu: Writing - review Thakkar, P., Nathan, S., Viswanathan, B., 2018. Performance benchmarking and
& editing. optimizing hyperledger fabric blockchain platform. In: 2018 IEEE 26th International
Symposium on Modeling, Analysis, and Simulation of Computer and
Telecommunication Systems (MASCOTS), pp. 264–276.
Declaration of competing interest Wang, C., Wang, Q., Ren, K., Lou, W., 2010. Privacy-preserving public auditing for data
storage security in cloud computing. In: 2010 Proceedings IEEE Infocom. IEEE, pp.
The authors declare that they have no known competing financial 1–9.
interests or personal relationships that could have appeared to influence Williams, M.J., Whitaker, R.M., Allen, S.M., 2016. There and back again: detecting
regularity in human encounter communities. IEEE Trans. Mobile Comput. 16 (6),
the work reported in this paper. 1744–1757.
Xia, F., Liu, L., Jedari, B., Das, S.K., 2016. Pis: a multi-dimensional routing protocol for
References socially-aware networking. IEEE Trans. Mobile Comput. 15 (11), 2825–2836.
Xia, Q., Sifah, E.B., Asamoah, K.O., Gao, J., Du, X., Guizani, M., 2017. Medshare:
trust-less medical data sharing among cloud service providers via blockchain. IEEE
Azaria, A., Ekblaw, A., Vieira, T., Lippman, A., 2016. Medrec: using blockchain for Access 5, 14757–14767.
medical data access and permission management. In: 2016 2nd International Yu, D., Liu, G., Guo, M., Liu, X., 2018. An improved k-medoids algorithm based on step
Conference on Open and Big Data (OBD). IEEE, pp. 25–30. increasing and optimizing medoids. Expert Syst. Appl. 92, 464–473.
Boneh, D., Franklin, M., 2003. Identity-based encryption from the weil pairing. SIAM J. Zheng, Z., Xie, S., Dai, H., Chen, X., Wang, H., 2017. An overview of blockchain
Comput. 32 (3), 586–615. technology: architecture, consensus, and future trends. In: 2017 IEEE International
Chen, C., Qiu, T., Hu, J., Ren, Z., Zhou, Y., Sangaiah, A., 2017. A congestion avoidance Congress on Big Data (BigData Congress). IEEE, pp. 557–564.
game for information exchange on intersections in heterogeneous vehicular
networks. J. Netw. Comput. Appl. 85, 116–126.

9
J. Chi et al. Journal of Network and Computer Applications 167 (2020) 102710

Jiancheng Chi received his bachelor’s degree in 2017 from Yingwei Jin received the Doctoral degree from the Dalian
School of Mechanical Engineering, Dalian University of Tech- University of Science and Technology in 2004. He is a Pro-
nology (DLUT), Dalian, China. He is currently pursuing the fessor with the School of Management, Dalian University of
Doctor’s degree of Computer Science and Technology in Col- Technology, China. He has published over 30 technical papers
lege of Intelligence and Computing, Tianjin University (TJU), in international journals and conferences. He participated in
Tianjin, China. He is a member of Tianjin Advanced Network- one key R&D project of Ministry of Science and Technology of
ing Key Laboratory (TANK Lab) and his researches focus on the People’s Republic of China, presided over two projects of
the Industrial Internet of Things, the Industrial Big Data, and National Natural Science Foundation of China and published
the Industrial Edge Computing based on Artificial Intelligence. one academic monograph and nearly 80 papers at home and
abroad. His research interests include computer network and
security, Internet technology, and artificial intelligence.

Yu Li received his B.S. degree from School of Mechanical Engi- Chen Chen received the B.Eng., M.Sc. and Ph.D. degrees in
neering, Yanshan University (YSUT), Qinhuangdao, China in telecommunication from Xidian University, Xi’an, China, in
2018. He is currently pursuing the M.S. degree in College of 2000, 2006, and 2008, respectively. He is currently an Asso-
Intelligence and Computing, Tianjin University (TJU), Tianjin, ciate Professor with the Department of telecommunication in
China. He is a member of Tianjin Advanced Networking Key Xidian University, and a member of the State Key Laboratory
Laboratory (TANK Lab) and his researches focus on the Inter- of Integrated Service Networks in Xidian University. He is also
net of Things, the Blockchain Technology, and the Machine the Director of the Xi’an Key Laboratory of Mobile Edge Com-
Learning. puting and Security, and the Director of the intelligent trans-
portation research laboratory in Xidian University. He was a
visiting professor at the department of EECS in the Univer-
sity of Tennessee and the department of CS in the University
of California. He serves as General Chair, PC Chair, Work-
shop Chair or TPC Member of a number of conferences. He
has authored/co-authored 2 books, over 90 scientific papers
Jing Huang received Bachelor degree in Computer Science in international journals and conference proceedings. He has
and Technology in 2003 and the Second Bachelor degree in contributed to the development of 5 copyrighted software sys-
Higher Education Management in 2005 from the school of tems and invented over 80 patents. He is also a senior mem-
education, Tianjin University. Then received Master degree in ber of China Computer Federation (CCF) and member of ACM,
Vocational and Technical Education from the school of educa- Chinese Institute of Electronics.
tion, Tianjin University in 2010. Currently she serves as direc-
tor of the office of scientific research and civil military inte- Tie Qiu received Ph.D degree in computer science from Dalian
gration in the College of intelligence and computing, Tianjin University of Technology in 2012. He is currently a Full Pro-
University, China. She is in charge of Ideological and polit- fessor at School of Computer Science and Technology, Tian-
ical education, scientific research project management, sys- jin University, China. Prior to this position, he held assistant
tem maintenance and development and other related research professor in 2008 and associate professor in 2013 at School
works. She has published more than 10 scientific papers and of Software, Dalian University of Technology. He was a visit-
invented 1 patent. ing professor at electrical and computer engineering at Iowa
State University in U.S. (2014-2015). He serves as an asso-
Dr. Jing Liu received Ph.D. degree in computer science from ciate editor of IEEE Transactions on SMC: Systems, area edi-
Communication University of China in 2012. She is currently tor of Ad Hoc Networks (Elsevier), associate editor of IEEE
a Full Professor at School of artificial intelligence, Hebei Uni- Access Journal, Computers and Electrical Engineering (Else-
versity of Technology, China. She was a visiting professor at vier), Human-centric Computing and Information Sciences
industrial engineering at Iowa University in U.S. (2016–2017). (Springer), a guest editor of Future Generation Computer Sys-
She serves as an series editor of Atlantis Highlights in Intelli- tems. He serves as General Chair, Program Chair, Workshop
gent Systems, a guest editor of International Journal of Indus- Chair, Publicity Chair, Publication Chair or TPC Member of
trial Engineering. She has authored/co-authored 3 books, over a number of international conferences. He has authored/co-
20 scientific papers in international journals and conference authored 9 books, over 100 scientific papers in international
proceedings, such as Computers & Industrial Engineering, journals and conference proceedings, such as IEEE/ACM ToN,
Neuro computing etc. She has contributed to the development IEEE TMC, TKDE TII, TIP, TCY, TITS, TVT, IEEE Communica-
of 8 copyrighted software systems and invented 12 patents. tions Surveys & Tutorials, IEEE Communications, INFOCOM,
She is a senior member of China Computer Federation (CCF). GLOBECOM etc. There are 10 papers listed as ESI highly cited
At the same time, she actively participated in the transforma- papers. He has contributed to the development of 3 copy-
tion of industry and research. As the chairman of an industrial righted software systems and invented 14 patents. He is a
Internet company, she led the team to explore industrial big senior member of China Computer Federation (CCF) and a
data, industrial artificial intelligence, etc. Senior Member of IEEE and ACM.

10