Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

7650 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 17, NO.

11, NOVEMBER 2021

A Sharding Scheme-Based Many-Objective


Optimization Algorithm for Enhancing Security in
Blockchain-Enabled Industrial Internet of Things
Xingjuan Cai , Shaojin Geng , Jingbo Zhang, Di Wu, Zhihua Cui , Wensheng Zhang,
and Jinjun Chen , Senior Member, IEEE

Abstract—While the industrial Internet of Things (IIoT) including logistics, medical, manufacturing, and energy indus-
can support efficient control of the physical world through tries. It collects industrial data through various smart objects
large amounts of industrial data, data security has been a to achieve troubleshooting, malicious behavior detection [2],
challenge due to various interconnections and accesses.
Blockchain technology can support security and privacy and other functions. For this reason, protecting the security and
preservation in IIoT data with its trusted and reliable secu- privacy of large amounts of data is the most important issue
rity mechanism. Sharding technology can help improve the in IIoT. However, there are several challenges that must be
overall throughput and scalability of blockchain networks. overcome before the formal adoption of IIoT across various
However, the effectiveness of sharding is still challenging industrial sectors. Furthermore, as a kind of data security protec-
due to the uneven distribution of malicious nodes. By aim-
ing to improve the performance of blockchain networks tion [3], blockchain technology [4] can effectively reduce data
and reduce the possibility of malicious node aggrega- system risks by enabling anonymous and trustful transactions
tion, in this article, we propose a many-objective optimiza- in a decentralized and trustless environment, which is slowly
tion algorithm based on the dynamic reward and penalty changing the entire industrial field [5].
mechanism (MaOEA-DRP) to optimize the shard valida- The term “blockchain” first appeared in a white paper on
tion validity model. Then, an optimal blockchain sharding
scheme is obtained. Compared with other state-of-the-art bitcoin published by Nakamoto [6]. The blockchain is a col-
many-objective optimization algorithms, MaOEA-DRP per- lection of stored data. Each block contains all the transaction
forms better on the DTLZ test suite. The simulation results information and blocks information of the current network for
demonstrate that our proposed algorithm can significantly a period. Each block points to the previous block and these
improve the throughput and validity of sharding for better
blocks are chained together. The characteristics of distributed,
security in the blockchain-enabled IIoT.
traceable, nontamperable, district centralization, and so on make
Index Terms—Blockchain, industrial Internet of Things the blockchain technology widely used by different enterprises.
(IIoT), many-objective optimization algorithm, privacy As an indispensable and vital part of the blockchain [7], there
preservation, scalability, security.
are many consensus algorithms. The bitcoin blockchain uses
the proof of work (POW) mechanism [8], which is a scheme to
I. INTRODUCTION grab decision-making power through the workload. In addition,
HE industrial Internet of Things (IIoT) [1] applies the the proof of stake mechanism [9] was developed to optimize
T Internet of Things technology to modern industrial sectors, the problem of resource consumption in POW. The practical
byzantine fault-tolerant (PBFT) algorithm [10] is a solution to
the consensus of nodes in a network and is used to address the in-
Manuscript received July 29, 2020; revised November 3, 2020; ac-
cepted December 27, 2020. Date of publication January 14, 2021; date
efficiency of the original byzantine fault-tolerant algorithm. The
of current version July 26, 2021. Paper no. TII-20-3640. (Corresponding authors demonstrated that honest nodes can reach an agreement
Author: Zhihua Cui.) on command in situations where the number of honest nodes is
Xingjuan Cai, Shaojin Geng, Jingbo Zhang, Di Wu, and Zhihua Cui
are with the School of Computer Science and Technology, Taiyuan
larger than 3f and the number of malicious nodes is f or less,
University of Science and Technology, Taiyuan 030024, China (e-mail: that is, 3f + 1 ≥ n.
xingjuancai@163.com; shaojin_geng@163.com; world4609@126.com; In order to satisfy the needs of modern business scenarios, the
wuxiaodou942@163.com; zhihua.cui@hotmail.com).
Wensheng Zhang is with the State Key Laboratory of Intelligent
scalability and transaction throughput of the blockchain [11] are
Control and Management of Complex Systems, Institute of Automa- urgent problems to be addressed. The existing technologies are
tion, Chinese Academy of Sciences, Beijing 100864, China (e-mail: generally divided into three categories: sidechain, sharding, and
wensheng.zhang@ia.ac.cn).
Jinjun Chen is with the Computer Science and Software Engineer-
directed acyclic graphs (DAGs). Sidechain technology [12] is
ing, Swinburne University of Technology, Hawthorn, VIC 3122, Australia a cross-blockchain solution that achieves the transfer of digital
(e-mail: jinjun.chen@gmail.com). assets between two blockchains and improves the scalability of
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TII.2021.3051607.
the blockchain through fusion. The DAG [13] changes the data
Digital Object Identifier 10.1109/TII.2021.3051607 structure of the blockchain. It replaces the original blockchain

1551-3203 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
CAI et al.: SHARDING SCHEME-BASED MANY-OBJECTIVE OPTIMIZATION ALGORITHM FOR ENHANCING SECURITY 7651

structure with a DAG structure and changes synchronous ac- the blockchain network into multiple small network modules
counting to asynchronous accounting. It has the features of that can process transactions in parallel. All small network
fast transaction speed, robust scalability, and high safety factor. modules can simultaneously execute the consensus algorithm
Meanwhile, sharding [14] is an internal distribution scheme of in the blockchain to improve the throughput of the blockchain
the blockchain that divides the blockchain network into several and reduce the delay. This also allows some transactions to be
small transaction networks to achieve parallel processing, which processed and stored on a single shard. In actual application
greatly improves the throughput performance of the blockchain. scenarios, the throughput of the blockchain will increase linearly
However, single-shard invalidation can be caused by malicious with the increasing number of nodes. This is the main research
nodes. content of the present article. Transaction sharding refers to
The most common algorithm in sharding is the PBFT algo- dividing different transactions into different shards, which can
rithm, which is a proven deterministic consensus algorithm for greatly reduce the calculation overhead by allocating some
achieving final data consistency in peer to peer (P2P) networks. highly related transactions to the same shard. State sharding,
It allows the existence of malicious nodes with propagated also called data sharding, classifies the data of the entire net-
messages but it will still not affect the security of blockchain work according to the state, and each shard only processes its
decisions when its proportion is less than one-third of all nodes. own data. However, no effective solution for the problem of
Although the blockchain sharding scheme has contributed to mutual communication between heterogeneous nodes has been
improving the throughput, there remains the issue that a single obtained.
shard occupied by malicious nodes results in shard invalidation The sharding technology was first proposed by Wang et al.
when the total number of malicious nodes is less than one-third [18] and it can usually be applied to databases and cloud archi-
of all nodes in the sharding scheme. Existing methods [15] that tecture. Compared with the traditional nonsharded blockchains,
improve the effectiveness of sharding by increasing the size of the sharding technology allows communication, data storage,
nodes or changing the number of shards do not address this and computational overload to be processed in parallel between
problem fundamentally. Hence, the present article combines shards, which can significantly improve the transaction through-
an intelligent optimization algorithm with a shard validation put of the entire blockchain network. Saino et al. [19] paid
validity model to ensure the effectiveness of sharding. The special attention to the load balancing and cache performance of
contributions of this article are summarized as follows. sharding but ignored the sharding failure problem. Bugday et al.
1) A shard validation validity model with four objectives [20] used adaptive learning methods to allocate nodes to shards,
is designed to address the problem of shard invalidation. which helps reduce the possibility of single-shard failure due
The model addresses shard invalidation probability (SIP), to a large number of malicious nodes. However, the sharding
delay, throughput TH, and the load of malicious nodes scheme should also guarantee the maximum throughput [21].
(LMN). The proposed model not only reduces the prob- Yun et al. [22] proposed a shard allocation scheme of the trust
ability of shard invalidation but also ensures the highest value model (TBSD). This scheme prevents the aggregation
throughput. of malicious nodes by combining the node trust model and
2) We propose a many-objective optimization algorithm genetic algorithm. But this is a single-objective optimization
based on the dynamic reward and penalty mechanism problem, and thus, it cannot guarantee the optimal performance
(MaOEA-DRP). The dynamic reward and penalty mech- of the sharding scheme. Liu et al. [23] designed a special
anism dynamically combine the diversity function and three-tiered sharding blockchain network that can significantly
the convergence function to increase the selection pres- improve transaction efficiency and system scalability. However,
sure and make the population closer to the real Pareto sharding effectiveness is still worth considering. In addition to
frontier (PF). At the same time, the weights of the two ensuring that the sharding is effective, an optimal sharding allo-
functions are dynamically set to classify individuals in cation scheme also needs to improve the sharding performance,
the population, thereby making individuals with different such as throughput. In our opinion, the sharding scheme is a
performances evolve iteratively. many-objective optimization problem [24]. Preventing sharding
The rest of this article is organized as follows. Section II failure is certainly an important topic, but how to improve the
introduces related sharding studies and our problem statement. throughput of the system and reduce the running time while
The proposed shard validation validity model is described in ensuring the effectiveness of sharding is the focus of this article.
Section III. Section IV proposes the many-objective optimiza-
tion algorithm. Section V conducts two experiments to verify the
efficiency of our proposed method. Finally, Section VI concludes B. Problem Statement
the article. The combination of IIoT and blockchain technology [25] has
become an important solution to protect the data security. The
II. RELATED WORK AND PROBLEM STATEMENT most common method is to use the IIoT terminal data center
as a blockchain miner node. A consensus mechanism is added
A. Sharding Technology
to compete for the qualification of writing blocks. Eventually,
As an important solution to blockchain scalability problems, transactions and data transmission between IIoT mobile terminal
the sharding technology [16] has broad application prospects. It devices are recorded in the blockchain, and the decentralization
is composed of three categories: network sharding, transaction of the blockchain is used to effectively guarantee IoT data
sharding, and state sharding [17]. Network sharding divides security.
7652 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 17, NO. 11, NOVEMBER 2021

shard points, and the consensus can be reached. In scheme 2,


the two malicious nodes (nodes 6 and 8) are assigned to the
same shard, resulting in the proportion of malicious nodes in the
shard exceeding one-third, and the failure to reach a consensus
invalidates the shard. Therefore, with the number of network
nodes increasing, how to obtain an optimal sharding scheme so
that each sharding does not fail is an urgent problem to be solved.

III. SHARD VALIDATION VALIDITY MODEL


In view of the invalidation of shards in the blockchain, this
Fig. 1. Blockchain application model for IIoT.
section designs a shard validation validity model to make mali-
cious nodes more evenly distributed among multiple shards. The
sharding optimization scheme has the following four objectives:
1) minimizing the SIP;
2) minimizing the delay;
3) maximizing the throughput;
4) balancing the LMN.
These are elaborated in the following part.

A. Shard Invalidation Probability


The shard validity probability is the most important problem
Fig. 2. Sharding scheme leads to the aggregation of malicious nodes. to be addressed. In order to prevent the shard failure, we reduce
SIP. When calculating the probability, we found that a node can
only appear in two identities, namely, the malicious node and
It can be seen from Fig. 1 that user data and computing tasks the honest node, and these two nodes are independent of each
are generally transferred to edge notes (EN) using distributed other. This is a probability model that conforms to the binomial
units (DU) in IIoT [26]. However, because user equipment (UE) distribution. Therefore, we calculate the invalidation probability
is in a harsh environment for a long time, it is impossible to of shards by the following equation:
ensure that the task to be migrated for each device has not been
tampered with. In order to avoid computing tasks across the 
s 
ki
SIP = Ckbi fi b (1 − fi )k − b (1)
entire network from being tampered with, we adopt a method i = 1 b=ki /3
that combines IIoT and blockchain networks. The blockchain
nodes are composed of all UE and DU. The UE layer con- fi = Mi /ki (2)
sists of ordinary nodes, and the DU layer contains consensus
where s is the number of shards, and ki is the total number of
nodes. The ordinary nodes only transmit and receive ledger
nodes in the ith shard. fi refers to the proportion of malicious
data, while consensus nodes perform consensus operations and
nodes in relation to the total number of sharding nodes in the
generate blocks. Malicious consensus nodes tamper with the
ith shard. Mi is the number of malicious nodes in the ith
data, causing excessive computing overhead and even paralysis
shard. This is a probabilistic calculation consistent with the
of the IIoT. This article assumes that each attacked UE will send
binomial distribution because in a sharding, only if the number
the wrong calculation task to the DU, and therefore, we model
of malicious nodes exceeds one-third of the total number of
the DU layer with multiobjective optimization combined with
sharding nodes, will the sharding become invalid. Therefore,
the blockchain sharding technology and propose an effective
we calculate the failure probability of the shards in each case
blockchain sharding scheme.
with more than one-third of the nodes and sum them up.
In general, network sharding is generally considered to be
an on-chain capacity expansion technology. On the condition
B. Delay (TD)
that there are some fixed number of nodes across the network,
the more network shards, the fewer the number of nodes on the Since sharding divides the blockchain into multiple chains for
chain within each shard. This can result in a higher efficiency of parallel execution, there is a delay problem between multiple
the single-shard consensus and increased transaction efficiency chains because of the length of execution. Therefore, we intend
of the entire network; however, the security of the blockchain to minimize the delay and obtain the execution time of each
network will be reduced. In this article, we are mainly consider- shard.
ing the network sharding scheme based on the PBFT consensus First, the leader node sends a message to other nodes. After
algorithm. accepting this message, these nodes need to send the message
Fig. 2 shows two sharding schemes. In scheme 1, there is to other nodes. Through this Byzantine consensus mechanism,
one malicious node in each of the two shards. When we have it is ensured that the shards are not invaded by malicious nodes.
the Byzantine fault tolerance within a shard, the proportion of Therefore, for node i and node j, the transmission time is 2 ×
malicious nodes is less than one-third of the total number of D(i → j)/v. Minimizing the maximum transmission time can
CAI et al.: SHARDING SCHEME-BASED MANY-OBJECTIVE OPTIMIZATION ALGORITHM FOR ENHANCING SECURITY 7653

reduce the delay. The equation is as follows: A. Dynamic Reward and Penalty Mechanism
max(Dis(l → m) + 2 × D(m → n)) The dynamic reward and penalty mechanism dynamically
T = (3) combines the diversity function and the convergence function to
v
s increase the selection pressure and make the population closer
T D = max Ti (4) to the real PF. At the same time, the weights of the two functions
i=1
are dynamically set to classify the individuals in the population;
where Dis(l → m) represents the distance from the leader
this makes the individuals with different performances evolve
node to the mth node. D(m → n) is the distance between the
iteratively according to the reward and penalty mechanism. The
mth node and nth node. v indicates that the propagation rate
calculation is
of electromagnetic waves in free space is the speed of light
3 × 105 km/s. Considering the actual situation where malicious DRP = δ × C(pi ) + ζ × D(pi ) (7)
nodes will attack the blockchain network, we add Dis(l → m)
when modeling the delay. D(m → n) is used to simulate an where C(pi ) and D(pi ) represent the convergence function and
attack from a malicious node, and the result of the attack will diversity function of the population, respectively, and δ and ζ are
eventually increase the delay. two weight vectors. As the number of iterations increases, the
diversity and convergence functions are dynamically rewarded
and punished by these two weights. Specifically, the conver-
C. Throughput
gence function C(pi ) represents the convergence ability of the
While reducing the SIP, we should also ensure that the population and calculates the Euclidean distance from the in-
throughput TH of the system is maximized. The parallel exe- dividual f  (pi ) to the ideal point z ∗ = {z1min , z2min , . . . , zM
min
}.
cution of multiple shards can improve system throughput. The The greater the value of C(pi ), the closer the individual is to the
throughput value is the number of transactions executed per ideal point, which enhances the selection pressure. This can be
second. We simulate it as the average number of nodes in the expressed by
shard. The more nodes there are, the more transactions can be
executed. The equation can be expressed as C(pi ) = 1 − f  (pi ) − z ∗  . (8)

s
ki The normalized diversity function D(pi ) calculates the Eu-
TH = (5) clidean distance from each individual to its nearest neighbor and
T
i=1 i can be expressed by (9). The greater the distance, the better the
where Ti is the execution time of the ith shard. The greater the diversity of the population.
throughput, the better the system performance. E(pi ) − Emin
D(pi ) = . (9)
Emax − Emin
D. Load of Malicious Nodes
Here, E(pi ) is the minimum Euclidean distance between indi-
By reducing SIP, we can control the number of malicious vidual i and other individuals. Emin and Emax are the minimum
nodes in each shard. It is necessary to ensure that malicious nodes and maximum distances, respectively.
do not affect the security of the shard, but we also cannot make The design of the weighting factor will greatly affect the
the number of malicious nodes in the shard too small because performance of the algorithm, and its value directly affects
this would cause an increase in the number of malicious nodes the effectiveness of the diversity function and the convergence
in other shards and make the shard invalid. To scatter malicious function. Therefore, we have designed dynamic weights. meanc
nodes in different shards as much as possible and prevent shard is the mean convergence distance and is used to divide the
failure, we follow individual into two parts: the area near to and the area far from the

 s ideal point. By comparing the individual values D(pi ) with mean
1 N 2
LMN =  (ki − ) (6) diversity distance meand , we can describe crowded parts and
s i=1 s sparse areas. P A(pi ) is the mapping distance from the individual
to the ideal point, and P B(pi ) is the vertical distance from the
where N is the total number of nodes. Through this equation, individual to the reference line
we can evenly distribute the malicious nodes among different  
 
(f (pi ) − z ∗ ) · z 
T
shards as much as possible to prevent the case of having too
many or too few malicious nodes in a single shard. P A(pi ) = (10)
z
 
 z 
IV. MAOEA-DRP ALGORITHM  ∗
P B(pi ) = f (pi ) − (z + P A(pi ) ) . (11)
z 
In this section, the dynamic reward and punishment mecha-
nism is proposed on the basis of the balanceable fitness estima- Here, P A(pi ) and P B(pi ) are used to evaluate the conver-
tion (BFE) method [27]. This is an improved strategy that can gence toward the effective frontier and measure the diversity of
dynamically adjust the weight value of the diversity function the population, respectively. meanP A and meanP B are the mean
and the convergence function, and its effect is better than that of values of P A(pi ) and P B(pi ), respectively. The calculations
the original method. are expressed in (12) and (13), respectively, where N is the
7654 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 17, NO. 11, NOVEMBER 2021

TABLE I
PARAMETER SETTINGS FOR DTLZ TEST SUITE

Algorithm 1: The Main Framework of MaOEA-DRP.


Input: Population p, the reference point z
Fig. 3. Objective space is divided into four types by meanC , meanD ,
meanP A , and meanP B .
Output: Population P
1: Cmin = min each objective (P )
2: While(t < tmax )
3: Q = Dynamic reward and penalty mechanism//Refer
population size
to (7)

N 4: for i = 1:N
meanP A = P A(pi )/N (12) 5: q1 = Q(randi )
i=1 6: q2 = Q(randi )

N 7: Compare q1 and q2
meanP B = P B(pi )/N . (13) 8: Choose a larger individual put in Q
i=1 9: End for
10: G = Cross operator and mutation operator (Q)
The individuals located at the edge of the coordinate axis that
11: P = Reference point strategy (Q, G)//Refer to
is close to the real front surface but far away from the ideal
non-dominated sorting genetic Algorithm III
point can be found. Fig. 3 shows an example in the 2-D space to
(NSGA-III)
describe the classification. In case 1, these individuals are close
12: Cmin = min each objective (Cmin , P )
to the ideal point and in the sparse region; therefore, the ζ value
13: End while
should reward the diversity and distance the same as δ. In case
2, the individuals are close to ideal points but are in relatively
crowded areas, and ζ is set between 0.6 and 1.0 to penalize their
software is MATLAB, the system RAM is 16 GB, and the
diversity distance. Similar to in the first case, the individuals in
processor is Intel(R) Core(TM) i7-9750H CPU@2.60 GHz.
the third case are close to the edge of the PF and are relatively
sparse, and thus they all need to be rewarded. Finally, case 4 has
the worst individuals, which should be punished for convergence A. Comparison Experiments
and diversity distance. First, the parameter settings for the DTLZ test function are
introduced in Table I, including the population size, number of
B. Framework of MAOEA-DRP decisions, and number of objectives. The value of parameter k
is different for each test suite. k = 5 for DTLZ1, k = 10 for
The detailed flow of our proposed algorithm is expressed
DTLZ2-6, and k = 20 for DTLZ7. The number of decisions is
in Algorithm 1 through pseudocode. First, we initialize the
determined by M − 1 + k, where M represents the number of
population and calculate the objective function value. Parents
objectives.
with better performance can produce high-quality offsprings,
In order to verify the influence of the improved BFE strategy
which can accelerate population evolution in a better direction.
on the algorithm performance, the average BFE function values
Therefore, the dynamic reward and penalty mechanism is used
are analyzed. It can be seen from Fig. 4 that the BFE strategy
to select better performing parents. At the same time, a cross op-
produced the worst performance and its average value tended
erator and a mutation operator are executed to produce superior
to 0.25. For MaOEA-DPR, the average value of BFE gradually
offspring. Finally, the parents and offspring are combined, and
increased with the number of iterations until it reached about
the best solution is selected by using a reference point strategy.
0.45; the overall performance was better.
The best solution is passed to the next generation. The population
The many-objective optimization algorithms used for com-
is output until the final iteration.
parisons are NSGA-III [28], Two_Arch2 [29], VaEA [30],
EFRRR [31], and KnEA [32]. The characteristics of the DTLZ
V. SIMULATION EXPERIMENTS
test function are presented in [31]. The maximum number of
First, we conduct experiments to compare our proposed algo- iterations is 10 000, and each algorithm runs 20× independently.
rithm with other many-objective optimization algorithms. Then, The probability of cross and mutation is set to 1/20. The inverted
we conduct comparisons regarding the shard validation. All generational distance (IGD) [33] indicator is an important per-
experiments are run under Windows 10 OS. The simulation formance indicator for evaluating many-objective optimization
CAI et al.: SHARDING SCHEME-BASED MANY-OBJECTIVE OPTIMIZATION ALGORITHM FOR ENHANCING SECURITY 7655

performance.
|Z|
1  |P |
IGD(P, Z) = min d(zi , pj ). (14)
Z i=1j =1
Table II lists the average and variance of IGD obtained by
the six algorithms after 20 independent run times. Symbols
“+/−/=” reflect the comparison result, indicating that the com-
pared algorithm has superior, inferior, or equal performance
compared with our proposed MAOEA-DPR algorithm. The best
results are shown in black font.
We can observe that MAOEA-DPR has a total of 19 optimal
performance solutions. The performance difference is not ob-
Fig. 4. Average values of the BFE function.
vious on DTLZ4. The reason for this be that these algorithms
have little difference when dealing with a series of nonuniform
many-objective optimization problems. At the same time, in
the DTLZ2 test function with concave features, our algorithm
performed poorly on six and eight objectives. It also performed
poorly on DTLZ7. The reason for this be that DTLZ7 is a mixed,
nonconnected, multimodel model, and our proposed algorithm
is not suitable for solving this type of model. Our algorithm
showed better performance in other test functions, and overall,
it has the best performance.

B. Comparison Experiments Combining


Proposed Model
To verify the effectiveness of our proposed model, we com-
bined six algorithms with a model addressing shard validation.
First, we initialized 600 nodes. The 3σ principles of the state
distribution satisfy the number of malicious nodes not higher
than 1/3 1−P (μ−σ<X<μ + σ) = 31.7%.
Fig. 5 shows the nondominated frontier for six different
algorithms, which can see the conflict between the various
objectives clearly. Our objectives are to minimize SIP, minimize
the delay, maximize the throughput, and balance the LMN.
As can be seen from Fig. 5, the convergence performance of
NSGA-III, EFRRR, and MaOEA-DRP is better than those of
other algorithms for the first objective. Meanwhile, the diversity
performance of the second objective is better; furthermore, the
nodes in MaOEA-DRP are more evenly distributed. KnEA and
MaOEA-DRP algorithms can better show the conflict, and the
performance of convergence and diversity is better, too. In
summary, it can be concluded that the MaOEA-DRP algorithm
has the best nondominated frontier.
Fig. 5. Nondominated frontier on the sharding validation The solution obtained by the many-objective optimization
validity model for six different algorithms. (a) NSGA-III. (b) algorithm is a population, and it is impossible to intuitively
Two_Arch2. (c) Vector angle based evolutionary algorithm (VaEA).
(d) Ensemble fitness ranking with a ranking restriction (EFRRR). (e) compare the performance of different algorithms. This makes
Knee point driven evolutionary algorithm (KnEA). (f) MaOEA-DPR. the problem of selecting solutions in the field of intelligent opti-
mization a difficult problem. However, in practical applications,
since all the obtained solutions are nondominated solutions with
a similar performance, we only need to select them according to
the specific needs or preferences of users or enterprises. In this
algorithms and it can be expressed by (14). In this equation, P article, we first select the individual with the largest throughput
represents the individual, and Z is the number of reference points in each algorithm of the obtained population and compare the
that are uniformly distributed along the PF in space. d(zi , pj ) performance in regards to other objectives on the basis of en-
represents the Euclidean distance between the individual and suring the maximum throughput. From the comparison results
the reference point. The smaller the IGD value, the better the in Table III, we can see that MaOEA-DPR not only shows
7656 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 17, NO. 11, NOVEMBER 2021

TABLE II
COMPARISON EXPERIMENT FOR DIFFERENT MANY-OBJECTIVE OPTIMIZATION ALGORITHM ON DTLZ TEST FUNCTION
CAI et al.: SHARDING SCHEME-BASED MANY-OBJECTIVE OPTIMIZATION ALGORITHM FOR ENHANCING SECURITY 7657

TABLE III [5] X. Wang et al., “Survey on blockchain for Internet of Things,” Comput.
COMPARISON OF OTHER OBJECTIVE FUNCTION VALUES WHEN Commun., vol. 136, pp. 10–29, 2019.
THROUGHPUT IS MAXIMUM [6] S. Nakamoto, “Bitcoin: A peer-to-peer electronic cash system,” 2008.
[Online]. Available: https://bitcoin:org/bitcoin:pdf
[7] M. U. Hassan, M. H. Rehmani, and J. Chen, “DEAL: Differentially private
auction for blockchain based microgrids energy trading,” IEEE Trans. Serv.
Comput., vol. 13, no. 2, pp. 263–275, Mar./Apr. 2020.
[8] M. Vukolić, “The quest for scalable blockchain fabric: Proof-of-work vs.
BFT replication,” in Proc. Int. Workshop Open Problems Netw. Secur.,
Zurich, Switzerland, 2016, pp. 112–125.
[9] J. Kang, Z. Xiong, D. Niyato, P. Wang, D. Ye, and D. I. Kim, “Incentivizing
consensus propagation in proof-of-stake based consortium blockchain
networks,” IEEE Wireless Commun. Lett., vol. 8, no. 1, pp. 157–160,
TABLE IV
Feb. 2019.
COMPARISON RESULTS OF THE MEAN FOR EACH OBJECTIVE IN POPULATION
[10] Z. Tian, M. Li, M. Qiu, Y. Sun, and S. Su, “Block-DEF: A secure digital
evidence framework using blockchain,” Inf. Sci., vol. 491, pp. 151–165,
2019.
[11] M. El-Hindi, C. Binnig, A. Arasu, D. Kossmann, and R. Ramamurthy,
“BlockchainDB: A shared database on blockchains,” Proc. VLDB Endow-
ment, vol. 12, no. 11, pp. 1597–1609, 2019.
[12] A. Singh, K. Click, R. M. Parizi, Q. Zhang, A. Dehghantanha, and K.-K. R.
Choo, “Sidechain technologies in blockchain networks: An examination
and state-of-the-art review,” J. Netw. Comput. Appl., vol. 149, 2020,
Art. no. 102471.
[13] B. Wang, M. Dabbaghjamanesh, A. Kavousi-Fard, and S. Mehraeen,
advantages in the throughput but also in the other objectives. “Cybersecurity enhancement of power trading within the networked mi-
crogrids based on blockchain and directed acyclic graph approach,” IEEE
The solution obtained by our algorithm is the best. Trans. Ind. Appl., vol. 55, no. 6, pp. 7300–7309, Nov./Dec. 2019.
In order to compare the performance of obtained populations [14] X. Feng et al., “Pruneable sharding-based blockchain protocol,” Peer-to-
more fairly, we took the mean value of each objective for Peer Netw. Appl., vol. 12, no. 4, pp. 934–950, 2019.
[15] L. Luu, V. Narayanan, C. Zheng, K. Baweja, S. L. Gilbert, and P. Saxena,
comparison. It can be seen from Table IV that the performance “A secure sharding protocol for open blockchains,” in Proc. ACM Conf.
is the best with regards to the last three objectives, but the Comput. Commun. Secur., 2016, pp. 17–30.
first objective is slightly worse than the NSGA-III and VaEA [16] S. Li, M. Yu, C.-S. Yang, A. S. Avestimehr, S. Kannan, and P. Viswanath,
“PolyShard: Coded sharding achieves linearly scaling efficiency and
algorithms. But overall, our algorithm still performs well, thus security simultaneously,” IEEE Trans. Inf. Forensics Secur., vol. 16,
verifying the effectiveness of our proposed model. pp. 249–261, Jul. 2020.
[17] H. Chen and Y. Wang, “SSChain: A full sharding protocol for public
blockchain without data migration overhead,” Pervasive Mobile Comput.,
VI. CONCLUSION vol. 59, 2019, Art. no. 101055.
[18] G. Wang, Z. J. Shi, M. Nixon, and S. Han, “SoK: Sharding on blockchain,”
The sharding technology is promising to address the ex- in Proc. 1st ACM Conf. Adv. Financial Technol., New York, NY, USA,
pansion problem in blockchain technology. In this article, we 2019, pp. 41–61.
intended to address the contradiction between the throughput [19] L. Saino, I. Psaras, E. Leonardi, and G. Pavlou, “Load imbalance and
caching performance of sharded systems,” IEEE/ACM Trans. Netw.,
and validity of the sharding technology. A many-objective op- vol. 28, no. 1, pp. 112–125, Feb. 2020.
timization algorithm was proposed to optimize the shard vali- [20] A. Bugday, A. Ozsoy, and H. Sever, “Securing blockchain shards by using
dation validity model. The experimental results showed that the learning based reputation and verifiable random functions,” in Proc. Int.
Symp. Netw., Comput. Commun., Istanbul, Turkey, 2019, pp. 1–4. doi:
algorithm can address the conflict between throughput and shard 10.1109/ISNCC.2019.8909175.
validity for better security in the blockchain-enabled IIoT. [21] S. Woo, J. Song, S. Kim, Y. Kim, and S. Park, “GARET: Improving
This article provided an important reference for future throughput using gas consumption-aware relocation in ethereum sharding
environments,” Cluster Comput., vol. 23, pp. 2235–2247, 2020.
blockchain research. Additionally, our proposed algorithm can [22] J. Yun, Y. Goh, and J.-M. Chung, “Trust-based shard distribution scheme
be used to address other optimization problems in blockchain for fault-tolerant shard blockchain networks,” IEEE Access, vol. 7,
technology to enhance security for IIoT. In this article, the pp. 135164–135175, 2019.
[23] C. Liu, Y. Xiao, V. Javangula, Q. Hu, S. Wang, and X. Cheng, “Norma-
number of nodes was a fixed value; therefore, our future work Chain: A blockchain-based normalized autonomous transaction settlement
will address the problem of dynamic nodes. system for IoT-based e-commerce,” IEEE Internet Things J., vol. 6, no. 3,
pp. 4680–4693, Jun. 2019.
REFERENCES [24] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithm
based on decomposition,” IEEE Trans. Evol. Comput., vol. 11, no. 6,
[1] W. Liang, M. Tang, J. Long, X. Peng, J. Xu, and K.-C. Li, “A secure fab- pp. 712–731, Dec. 2007.
ric blockchain-based data transmission technique for industrial Internet- [25] O. Alkadi, N. Moustafa, B. Turnbull, and K.-K. R. Choo, “A deep
of-Things,” IEEE Trans. Ind. Informat., vol. 15, no. 6, pp. 3582–3592, blockchain framework-enabled collaborative intrusion detection for pro-
Jun. 2019. tecting IoT and cloud networks,” IEEE Internet Things J., to be published,
[2] Z. Cui, L. Du, P. Wang, X. Cai, and W. Zhang, “Malicious code detec- doi: 10.1109/JIOT.2020.2996590.
tion based on CNNs and multi-objective algorithm,” J. Parallel Distrib. [26] M. Banerjee, C. Borges, K.-K. R. Choo, J. Lee, and C. Nicopoulos, “A
Comput., vol. 129, pp. 50–58, 2019. hardware-assisted heartbeat mechanism for fault identification in large-
[3] K. Gai, Y. Wu, L. Zhu, M. Qiu, and M. Shen, “Privacy-preserving energy scale IoT systems,” IEEE Trans. Dependable Secure Comput., to be
trading using consortium blockchain in smart grid,” IEEE Trans. Ind. published, doi: 10.1109/TDSC.2020.3009212.
Informat., vol. 15, no. 6, pp. 3548–3558, Jun. 2019. [27] Q. Lin et al., “Particle swarm optimization with a balanceable fitness
[4] Z. Cui et al., “A hybrid blockchain-based identity authentication scheme estimation for many-objective optimization problems,” IEEE Trans. Evol.
for multi-WSN,” IEEE Trans. Serv. Comput., vol. 13, no. 2, pp. 2410–251, Comput., vol. 22, no. 1, pp. 32–46, Feb. 2018.
Mar./Apr. 2020.
7658 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 17, NO. 11, NOVEMBER 2021

[28] K. Deb and H. Jain, “An evolutionary many-objective optimization algo- Zhihua Cui received the Ph.D. degree in con-
rithm using reference-point-based nondominated sorting approach, Part trol theory and engineering from Xi’an Jiaotong
I: Solving problems with box constraints,” IEEE Trans. Evol. Comput., University, Xi’an, China, in 2008.
vol. 18, no. 4, pp. 577–601, Aug. 2014. He is currently a Professor with the School
[29] H. Wang, L. Jiao, and X. Yao, “Two_Arch2: An improved two-archive of Computer Science and Technology, Taiyuan
algorithm for many-objective optimization,” IEEE Trans. Evol. Comput., University of Science and Technology, Taiyuan,
vol. 19, no. 4, pp. 524–541, Aug. 2015. China. His research interests include compu-
[30] Y. Xiang, Y. Zhou, M. Li, and Z. Chen, “A vector angle-based evolutionary tational intelligence, stochastic algorithm, and
algorithm for unconstrained many-objective optimization,” IEEE Trans. combinatorial optimization.
Evol. Comput., vol. 21, no. 1, pp. 131–152, Feb. 2017. Dr. Cui is the Editor-in-Chief for the Interna-
[31] Y. Yuan, H. Xu, B. Wang, B. Zhang, and X. Yao, “Balancing convergence tional Journal of Bioinspired Computation.
and diversity in decomposition-based many-objective optimizers,” IEEE
Trans. Evol. Comput., vol. 20, no. 2, pp. 180–198, Apr. 2016.
[32] X. Zhang, Y. Tian, and Y. Jin, “A knee point driven evolutionary algorithm
for many-objective optimization,” IEEE Trans. Evol. Comput., vol. 19,
no. 6, pp. 761–776, Dec. 2015.
[33] E. Zitzler, L. Thiele, M. Laumanns, C. M. Fonseca, and V. G. da Fonseca,
“Performance assessment of multiobjective optimizers: An analysis and
review,” IEEE Trans. Evol. Comput., vol. 7, no. 2, pp. 117–132, Apr. 2003. Wensheng Zhang received the Ph.D. degree in
pattern recognition and intelligent systems from
the Institute of Automation, Chinese Academy of
Xingjuan Cai received the Ph.D. degree in con- Sciences (CAS), Beijing, China, in 2000.
trol theory and engineering from Tongji Univer- He is currently a Professor with Machine
sity, Shanghai, China, in 2017. Learning and Data Mining and the Director of
She is currently a Professor with the School the Research and Development, Institute of Au-
of Computer Science and Technology, Taiyuan tomation, CAS, Beijing. His current research in-
University of Science and Technology, Taiyuan, terests include computer vision, pattern recogni-
China. Her interests include cloud computing, tion, artificial intelligence, and computer–human
and bioinspired computation and applications. interaction.

Shaojin Geng Shaojin Geng is currently


working on task scheduling strategy in Jinjun Chen (Senior Member, IEEE) received
multi-cloud environment toward the M.S.
the Ph.D. degree in information technology
degree with the School of Computer Science
from the Swinburne University of Technology,
and Technology, Taiyuan University of Science
Hawthorn, VIC, Australia, in 2007.
and Technology, Taiyuan, China.
He is currently a Professor with the
His research interests include cloud
Swinburne University of Technology. His
computing, computational intelligence, and
research results have been published by
combinatorial optimization. various conferences and journals including
various IEEE/ACM transactions. His research
interests include scalability, data systems, cloud
computing, data privacy, and related various
research topics.

Jingbo Zhang is currently working on com-


pressed sensing of hyperspectral image based
on many–objective optimization toward the M.S.
degree with the School of Computer Science
and Technology, Taiyuan University of Science
and Technology, Taiyuan, China.
His research interests include computational
intelligence and combinatorial optimization.

Di Wu is currently working on eelection strategy


of many-objective optimization algorithm toward
the M.S. degree with the School of Computer
Science and Technology, Taiyuan University of
Science and Technology, Taiyuan, China.
Her research interests include computational
intelligence and algorithm optimization.

You might also like