Machine learning in/for blockchain: Future and

challenges
arXiv:1909.06189v3 [cs.CR] 8 Dec 2020

Fang Chen∗, Hong Wan†, Hua Cai‡, and Guang Cheng§

Abstract

Machine learning and blockchain are two of the most noticeable technologies
in recent years. The first one is the foundation of artificial intelligence and
big data, and the second one has significantly disrupted the financial indus-
try. Both technologies are data-driven, and thus there are rapidly growing
interests in integrating them for more secure and efficient data sharing and
analysis. In this paper, we review the research on combining blockchain and
machine learning technologies and demonstrate that they can collaborate
efficiently and effectively. In the end, we point out some future directions
and expect more researches on deeper integration of the two promising tech-
nologies.
Keywords: Blockchain, Bitcoin, deep learning, machine learning, rein-
forcement learning.

1 INTRODUCTION
A blockchain is a shared, distributed public ledger that stores transaction
data in a chain of sequential blocks (Dinh & Thai, 2018). The data (block)
are time-stamped and validated before adding to the chain. Each block
contains information from the previous one. The mathematical structure
for storing data makes it nearly impossible to fake (MIT Technology Re-
view Editors, 2018). Thanks to the legacy of cryptocurrency, the term
∗
Ph.D. student, Department of Industrial Engineering, Purdue University.
†
Associate Professor, Department of Industrial and System Engineering, North Car-
olina State University.
‡
Assistant Professor, Department of Industrial Engineering, Purdue University.
§
Corresponding Author. Professor, Department of Statistics, Purdue Univer-
sity. Guang Cheng gratefully acknowledges NSF DMS-1712907, DMS-1811812, DMS-
1821183,and Office of Naval Research, (ONR N00014-18-2759).

1
”blockchain” has transformed from a cryptography terminology to a buzz
word. Many people believe that cryptocurrency IS blockchain. This is incor-
rect. While blockchain is the foundation of cryptocurrency, the applications
of the blockchain technology are much wider. Scenarios involving data vali-
dating, auditing, and sharing can all consider applying blockchains.
In this paper, we review the research on combining blockchain and ma-
chine learning technologies and demonstrate that they can collaborate ef-
ficiently and effectively. Machine learning is a general terminology that
includes variety of methods, machine learning, deep learning and reinforce-
ment learning. These methods are the core technology for big data analysis
(Buhlmann et al., 2019). As a distributed and append-only ledger system,
the blockchain is a natural tool for sharing and handling big data from
various sources through the incorporation of smart contracts (i.e., a piece
of code that will execute automatically in certain conditions). More specifi-
cally, blockchain can preserve data security and encourage data sharing when
training and testing machine learning models. Also, it allows us to utilize
distributed computing powers (for example, IOT), for developing on-time
prediction models with various sources of data. This is especially impor-
tant for deep learning procedures which require a tremendous amount of
computational power. On the other hand, blockchain systems will generate
a huge amount of data from different sources, and the distributed systems
are harder to monitor and control than the centralized ones. Efficient data
analysis and forecasting of the system behaviors are critical for optimal
blockchain mechanism designs. In addition, machine learning can facilitate
the data verification process and identifying malicious attacks and dishonest
transactions in the blockchain. The interdisciplinary research on combining
the two technologies is of great potential.
In this paper, we review articles that are either using machine learn-
ing techniques to study the blockchain system/structure itself or imple-
menting blockchain techniques to improve machine learning, e.g., collab-
orative/distributed learning. The reviewed papers are summarized in Table
1 below. For papers that apply machine learning and blockchain techniques
separately to various areas, we do not include them in our review but list
some of them in Table 2 below. In the rest of this paper, we first review basic
structure and terminology of blockchain in Section 2. The review is by no
means exhaustive, but sufficient for Sections 3, 4, and 5 that introduce how
different machine learning methods can be incorporated into the blockchain
system. Our work is concluded by Section 6 to discuss potential research
directions and challenges that are arose from ongoing and future fusion of
machine learning and blockchain.

2
 Method Application Paper
Supervised/Unsupervised Learning Yin & Vatrapu (2017), Jourdan et al. (2018),
Transaction Entity Classification
without Deep Methods Akcora et al. (2020)
Jourdan et al. (2018), Shah & Zhang (2014),
Bitcoin Price Prediction
Akcora et al. (2019), Abay et al. (2019), Dey et al. (2020)
Supervised Learning Harris & Waggoner (2019), Chen et al. (2018),
Privacy and Security Preserving
with Deep Methods Zhu, Li, & Yu (2019)
Computation Power Allocation Luong et al. (2018)
Mcnally, Roche, & Caton (2018), Lahmiri & Bekiros (2019),
Cryptocurrency Price Prediction
Alessandretti et al. (2018)
Reinforcement Learning IoT Network Liu, Lin, & Wen (2018)
Eyal & Sirer (2014), Sapirshtein, Sompolinsky & Zohar (2017),
Bitcoin Mining
Wang, Liew, & Zhang (2019)

Table 1: Summary of Papers in the Review

Area Exemplary Sources

Mamoshina et al. (2017), Juneja & Marefat (2018), Okalp et al. (2018),
Zheng et al. (2018), Firdaus et al. (2018), Wang et al. (2018),
Healthcare
Vyas, Gupta, & Yadav (2019), Bhattacharya et al. (2019),
Agbo, Mahmoud, & Eklund (2019), Khezr et al. (2019).
Liu, Lin, & Wen (2018), Xiong & Xiong (2019), Lee & Ryu (2018),
IoT Related Qin et al. (2019), Ozyilmaz, Dogan, & Yurdakul (2018), Singla, Bose, & Katariya (2018),
Shen et al. (2019), Li et al. (2019), Rathore, Pan, & Park (2019), Ferrag & Maglaras (2020).

Table 2: Summary of Some Less Relevant Papers

2 REVIEW ON BLOCKCHAIN
A blockchain, literally speaking, is just a chain of digital blocks. Each block
contains a certain amount of data; and the chain connects these data to
form a distributed database. A newly created block includes multiple trans-
actions collected from nodes and broadcasts to every node on the network.
It can be accepted and added to the blockchain by nodes that have the same
consensus protocol. Each added block includes information of the previous
block in the chain. Hence, if the block is changed, all blocks before this
block will be invalid as well. The strategies to reach agreement of the new
block (consensus) vary in different types of blockchain. The mathematical
structure of the blockchain implies two essential properties: (i) the data (in
block) is immutable (MIT Technology Review Editors, 2018); (ii) the dis-
tributed network with consensus allows users to communicate directly with
each other and download a copy of the current ledger, which means that
there is continuous monitoring and redundancy of the data in the network.
Therefore, the blockchain is more robust to individual outrages and attacks.
Depending on who can access to the blockchain and who can validate the
data, the blockchain can be classified into public chains, private chains, and
consortium chains (Zheng et al., 2018). The comparison of three different
types of blockchains is shown in the Table 3.

3
Attribute Public Private Consortium
Who run/manage the chain All miners One organization/user Selected users
Permission to Access No Yes Yes
Security Nearly impossible to fake Could be tampered Could be tampered
Efficiency Low High High
Centralized No Yes Partial
Example Bitcoin, Ethereum IBM HyperLedger Quorum

Table 3: Comparison of three types of blockchains

In what follows, we use the bitcoin system, which is the most known
blockchain application, as an example to demonstrate how blockchain works
in detail. Typically, an end-to-end blockchain-based transaction needs to
be validated at two different levels, the node level and the block level. The
transaction is first verified between two nodes (Zheng et al., 2018). Then
a unique digital signature, which is a hash wrapping all information of the
transaction, is created. The digital signature that represents the transaction
is submitted to the transaction pool and is waited to be added to a new block.
Before the new block is accepted by the blockchain network, it is required
to be validated by other miners on the network through the Proof-of-Work
(PoW) consensus protocol. The PoW process includes aggregating a set of
transactions to the new block and finding a hash value that is lower than
the target value (Ghimire & Selvaraj, 2018). The new block is only accepted
by the network if transactions are valid and unspent. Other nodes continue
working on creating the new block using the hash from the previous block
(Nakamoto, 2008).
Since the probability that finding a new valid block is extremely low and
the PoW process requires a huge amount of computing power and a high con-
sumption of the electricity, miners tend to collaborate with each other and
form a mining pool. After participating in a mining pool, individual miner
could receive a steady reward and significantly lower the risk. On the other
hand, the mining pools usually charge membership fees to each participant
and allocate rewards to each miner according to their own rewards shar-
ing mechanisms (Bhaskar & Lee, 2015). Some common reward allocation
mechanisms in practice are Pay-Per-Last-N-Shares (PPLNS) (Qin, Yuan, &
Wang, 2019) and Full-Pay-Per-Share (FPPS) (Zhu et al., 2018). Another
popular public chain is Ethereum (Wood, 2014), which allows users to send
not only digital coins but also smart contracts (Wohrer & Zdun, 2018). In
order to reduce the energy consumption on the validation, Ethereum plans
to switch its consensus protocol from the PoW to the Proof-of-Stake (PoS)
gradually (Saleh, 2020).

4
3 SUPERVISED/UNSUPERVISED LEARNING
WITHOUT DEEP METHODS
In this section, we review several applications of machine learning for the
blockchain. Specifically, Section 3.1 reviews three studies regarding trans-
action entities classification (Yin & Vatrapu, 2017; Jourdan et al., 2018;
Akcora et al., 2020) with different purposes. One focuses on the recognition
of cybercriminal entities using supervised learning (Yin & Vatrapu, 2017) as
well as topological data analytics (TDA) method (Akcora et al., 2020), while
another on the recognition of common categories of entities for most trans-
actions (Jourdan et al., 2018). Section 3.2 reviews Bitcoin price prediction
from different perspectives such as probabilistic graphic models (Jourdan et
al., 2018), Bayesian regression (Shah & Zhang, 2014) and feature selection
on the blockchain topological structure using Granger causality and TDA
(Akcora et al., 2019; Abay et al., 2019; Dey et al., 2020).

3.1 Transaction entity classification

In the Bitcoin network, it is crucial to recognize entities behind those po-
tentially illegal ones. The study of identifying entities behind addresses is
called address clustering (Harrigan & Fretter, 2016). Yin & Vatrapu (2017)
apply supervised learning to classify entities of transactions that may in-
volve in cybercriminal activities. The classification model is trained based
on 854 observations with categorical identifiers and then applied to study
10000 uncategorized observations that take 31.62% of unique addresses and
28.99% of total coins in the overall Bitcoin blockchain. The categorical
identifiers represent 12 classes of entities, five of which are related to cyber-
criminal activities. Thirteen classifiers from the Python machine learning
package “scikit-learn” are applied. By comparing accuracy scores of all clas-
sifiers, it is found that Random Forests (77.38%), Extremely Randomised
Forests(76.47%), Bagging (78.46%) and Gradient Boosting (80.76%) stand
out as the best four classifiers. After further comparing precision, recall,
and f1 score of these classifiers, bagging and gradient boosting stand out,
which are then applied to analyze the 10000 observations. The classifica-
tion outcome shows that 5.79% (3.16%) addresses and 10.02% (1.45%) coins
are from cybercriminal entities according to the bagging method (gradient
boosting method).
Bitcoins are found to be a common way to make the ransomware pay-
ment. In order to detect addresses related to ransomware payment, Akcora
et al. (2020) apply a topological data analysis (TDA) approach to generate

5
the bitcoin address graph by first grouping similar addresses into nodes and
then putting common addresses between two nodes into the set of edges.
The TDA is an approach commonly used for dimension reduction. It rep-
resents the data set in a graph by first dividing data to sub-samples based
on different filtration criteria and then clustering similar points in each sub-
sample. The Bitcoin transaction graph model is a directed graph, denoted
as G = (V, E, B), where V is the set of vertices, E is a set of edges and
B = {Address, Transaction} is a set of node types. By using six graph fea-
tures extracted for each address, a TDA Mapper method is applied to create
six filtered cluster tree graphs. After calculating the number of ransomware
addresses in each cluster, denoted as V , a suspicion score is assigned to a
new address. The suspicion scores of addresses in the cluster are set to be 0
initially. It increments by one if inclusion and size thresholds are satisfied as
follow: (1) the inclusion threshold, denoted as 1 , times the total amount of
labeled ransomware addresses is less than V ; (2) the size threshold, denoted
as 2 , times the number of labeled ransomware addresses in the cluster is
greater than the number of all addresses in the cluster. Suspicious addresses
are then filtered by a quantile threshold, denoted as q, when their suspicious
scores are higher than the quantile threshold. The result indicates that the
best TDA model with 1 = 0.05, 2 = 0.35, q = 0.7 outperforms random
forest (RF), and XGBoost in new ransomware addresses prediction.
Jourdan et al. (2018) are interested in classifying entities of transactions
into four most common categories: Exchange, Service, Gambling, Mining
Pool, based on data collected from 97 sources (Ermilov, Panov, & Yanovih,
2017). The goal of classification is to assist in selecting an appropriate
prediction model that is built according to categories of transactions (Jour-
dan et al., 2018). The applied classification method is a gradient boosted
decision tree algorithm along with a Gaussian process based optimization
procedure that determines optimal hyperparameters. Table 4 concludes that
accuracy’s in Exchange, Gambling, and Service categories are high. How-
ever, the accuracy in the Mining Pool category is poor. This may indicate
that mining activities may not be appropriate as an independent label.

3.2 Bitcoin price prediction

UTXOs record the number of Bitcoins in transactions, which enables us to
track buying and selling information to predict the Bitcoin price. Another
contribution of Jourdan et al. (2018) is to forecast the value of UTXOs
by creating probabilistic graphical models. The first model is called the
Block-transaction address model (BT-A) that is a stationary graphic model

6
 Category Accuracy F1 Precision
Exchange 0.94 0.92 0.91
Gambling 0.95 0.97 1.00
Mining 0.50 0.67 1.00
Service 0.95 0.88 0.83
Overall 0.92 0.91 0.92

Table 4: Classification Performance (Jourdan et al., 2018)

Metric BE-TA BE-TA BE-TA BE-TA BT-A

E S G M All
MSE 1.22 -0.3 -0.02 0.06 1.12
RMSE 125 53.3 1.15 5.19 90.5
MAE 15.6 0.94 0.20 2.42 7.47
RMAE 1.82 1.74 1.86 1.93 1.69
NRMSE 1.34 1.28 1.42 1.22 1.29

Table 5: BT-A and BT-EA Performance (Jourdan et al., 2018)

of a Bitcoin block with conditional dependency structures. As an extension

of BT-A, a Block-transaction entity-address model (BT-EA) is further de-
veloped by adding a categorical entity to each address. In terms of MSE,
RMSE, MAE, RMAE, simulation results in Table 5 show that this exten-
sion significantly outperforms the BT-A model in all categories except for
Exchange.
The dependence structure of the BT-A model to obtain the output
UTXOs values, denoted as Vo,u , is illustrated in Figure 1. Here is some
explanation. The BT-A model starts with computing the number of avail-
able UTXOs for the ith input address Ai , denoted as kA U T XO . For each
i
input address, the number of UTXOs used in a transaction is uniformly
drawn from 1 to kA U T XO with the corresponding UTXO value, denoted as
i
Vi,u . The total input value of a transaction is calculated by
Psumming the
input UTXOs value of each input address, denoted as Vt = Vi,u , and the
value of an output UTXO is uniformly drawn from 1 to total transaction
value minus validation fee.
To predict the Bitcoin price, Shah and Zhang (2014) apply the Bayesian
regression for the “latent source model” that is a nonparametric model for
time series binary classification. The latent source model framework is de-
scribed in the Chen, Nikolov, & Shah (2013) and Bresler, Chen, & Shah
(2014), which latent sources are time series with binary labels. Specifically,

7
 Figure 1: Block-transaction Address Model (Jourdan et al., 2018)

the model is described as K distinct unknown latent sources, s1 , ..., sK gener-

ated from a latent distribution over {1, ..., K} with probabilities {µ1 , ..., µk };
K latent distributions, denoted as P1 , ..., PK . Labeled data are generated
from sample index T ∈ {1, ..., K} with P (T = k) = µk . The model that
predicts y given x refers to the equation below.
T
X
P (y|x) = P (y|x, T = k)P (T = k|x)
k=1
(3.1)
T  
X 1 2
= Pk (y)exp − ||x − sk ||2 µk
2
k=1

where T ∈ {1, ..., K} is the sample index with P (T = k) = µk

Due to lacks of information of latent parameters, empirical data is used as
proxy for estimating P (y|x). The expectation of P (y|x) can be estimated
as follows:
Pn 1 2
i=1 yi exp(− 4 ||x − xi ||2 )
E[y|x] = Pn 1 2
(3.2)
i=1 exp(− 4 ||x − xi ||2 )
The future average price change is determined by price changes over three
periods of historical data: previous 30 minutes sample, 60 minutes sample
and 120 minutes sample, denoted as ∆pj , j = 1, 2, 3. Each ∆pj is calculated

8
by (3.2). Then ∆p over a 10-second period is formulated as
3
X
∆p = w0 + wj ∆pj + w4 r (3.3)
j=1

• w0 , w1 , w2 , w3 , w4 are weights to be estimated.

• r = (vb − va )/(vb + va ), where vb , va are the top 60 orders of total

buying and selling volume.
We would like to point out that in order to apply formula (3.3), it is crucial
to verify the stationarity of the price data, which was unfortunately not
done in the referenced paper. The trading strategy for each user is designed
as “buy one bitcoin when ∆p > t; sell one bitcoin when ∆p < −t; otherwise
holding the current number of bitcoin when −t ≤ ∆p ≤ t.” Here, t is a
pre-specified threshold. The designed prediction model is trained by data
gathered from Okcoin before May 2014 and is tested by data after that. It is
found that increasing t leads to an increase of the average profit per trade.
Besides using the Bayesian regression to predict the bitcoin price, the
selection of input features is also important to the performance of the predic-
tion. To better characterize input features, Akcora et al. (2019) introduce
a concept of graphic chainlet, which describes the local topological features
of Bitcoin blockchain, to explore impacts of the Bitcoin blockchain struc-
ture on Bitcoin price formation and dynamics. A transaction-address graph
representation of the Bitcoin blockchain is shown in Figure 2. Circle ver-
tices represent input and output addresses. A square vertex indicates the
transactions and edges stand for UTXOs (a transfer of Bitcoins). A chain-
let model represents x input UTXOs and y output UTXOs involving in a
transaction, denoted as Cx→y . All chainlet and chainlet clusters clustered
by various criteria are evaluated by the Granger causality test (Granger,
1969). The result concludes that the split chainlet cluster defined as when
y < x < 20, individual chainlet (e.g., C1→7 , C6→1 , C3→3 ), extreme chain-
lets (e.g., C20→2,3,12,17 ), clusters according to Cosine Similarity (e.g., C9→11 ,
C3→17 , C8→14 , C1→1 ) are significant to the Bitcoin price formation and
dynamics. A price prediction model is further developed using significant
chainlets.
Chainlet model studies topological features from a single transaction
aspect and only takes the number of input and output UTXOs into ac-
count. Abay et al. (2019) extend the chainlet model to a new graphic
model “ChainNet” that further considers topological features from the as-
pects of the number of distinct chainlets and the amount of coins transferred

9
 Figure 2: A Transaction-Address Graph

by chainlets. More specifically, from the perspective of all transactions, an

occurrence matrix is created to count the number of distinct chainlets among
all transactions. An amount matrix records the sum of Bitcoins transferred
for distinct chainlets. By considering both occurrence and the amount bit-
coins transferred in a transaction, an occurrence matrix with a threshold,
denoted as O ,  ∈ {0, 10, 20, 30, 40, 50}, is created to count the number of
distinct Ci→j that is larger than . Different thresholds result in different
values of O , which are considered as Filtration Features (FL) input in the
prediction model. Betti sequences and Betti derivatives for the blockchain
network are also considered as features in the model. A sliding prediction
approach associated with parameters of prediction horizon, window length
and training length is applied to train the time series prediction model. Ac-
cording to simulation results, ChainNet adopts Betti model features and FL
features for short and long term prediction, respectively, to obtain a better
performance.
Besides considering the effects of features of the Bitcoin topological struc-
ture on Bitcoin price formation and dynamics, topological features of other
types of cryptocurrencies may also affect the Bitcoin price. Dey et al. (2020)
evaluate the Bitcoin price formation and dynamics using the Chinalet model
according to the topological features on the joint of Bitcoin and Litecoin.
Specifically, the occurrences of distinct chainlets in Bitcoin and Litecoin, de-
noted as Ox→yB L
and Ox→y respectively, are considered. The amounts of coins
transferred in Bitcoin and Litecoin, denoted as AB L
x→y and Ax→y respectively,
are also included. Granger causality tests (Granger, 1969) with 1 to 5 lag ef-
fects are applied to assess the significance of chainlets. The result concludes
that the occurrence of chainlets in the Litecoin (O3→3 L , OL , OL , OL )
4→4 4→5 3→6
is significant to the price for all five lag effects in the Granger causality.
B
Also, occurrence and amount of chainlets in the Bitcoin (O20→2,3,12 B ,
, O1→7
AB B
20→12,20 , A3→4 ) are also important to the price for all five lag effects.
Although there are other studies related to Bitcoin price prediction using
machine learning methods, e.g., Greaves & Au (2015), Jiang & Liang (2017),

10
Jang & Lee (2018), and Sun, Liu, & Sima (2018), it is hard to include all
papers in the review. As a result, we will move on to review more articles
in prediction of cryptocurrency price using deep learning in Section 4.

4 SUPERVISED LEARNING WITH DEEP METH-

ODS
In this section, we turn to the application of deep learning. In Section 4.1,
three privacy-preserving collaborative learning frameworks (Harris & Wag-
goner 2019; Chen et al., 2018; Zhu, Li, & Yu, 2019) are reviewed. In Sec-
tion 4.2, we review a deep learning work (Luong et al., 2018) that allocates
computation resource to assist mobile blockchain mining. In Section 4.3, we
focus on cryptocurrency price prediction (Mcnally, Roche, & Caton, 2018;
Lahmiri & Bekiros, 2019) and digital portfolio management using Recur-
rent Neural Network (RNN) and Long-Short Term Memory (LSTM) models
(Alessandretti et al., 2018).

4.1 Decentralized privacy-preserving collaborative learning

Harris & Waggoner (2019) build a decentralized collaborative learning frame-
work with blockchain. The new designed framework extended by the previ-
ous two frameworks (Abernethy & Frongillo, 2011; Waggoner, Frongillo, &
Abernethy, 2015) is designed to collaboratively build a dataset and train a
predictive model. The framework starts with letting the provider define a
loss function and upload 10 out of 100 partial dataset with corresponding
hashes. By using the smart contract that initially contains a model, other
participants add their own data or uploading an update along with a de-
posit of 1 unit of currency until the end condition set by the provider is met.
The provider uploads the rest of 90 partial datasets to evaluate participants’
models. The better model tends to receive more rewards in the end.
Chen et al. (2018) propose a framework called “Learning Chain” to
preserve user’s privacy by applying a decentralized version of the Stochastic
Gradient Descent (SGD) algorithm and a differential privacy mechanism.
The proposed framework contains three phases: blockchain initialization;
local gradient computation; global gradient aggregation. In the first phase,
a peer-to-peer network is set up with computing nodes and data holders.
The second phase involves each data holder Pk retrieving the current model
from the block t, denoted as wt , and computing its own local gradient.
A differential privacy mechanism is then applied to generate a hidden local

11
gradient, denoted as ∇gk (wt )∗ , by adding a noise factor to the local gradient.
The message broadcasts a pseudo-identity of Pk , normalized hidden local
gradient, denoted as ∇gbk (wt )∗ , together with the norm of its un-normalized
version to computing nodes on the network. In the final phase, after solving
Proof-of-Work (PoW), the winner node selects top l-nearest local normalized
gradients according to the cosine distance between each normalized local
gradient and the sum vector of ∇gk (wt )∗ to update the global gradient. The
predictive model is updated by wt+1 = wt + η∇J(wt ), where ∇J(wt ) is the
updated global gradient.
“Learning Chain” is trained and tested in three different data sets: syn-
thetic data set; Wisconsin breast cancer data set; MNIST data set; using
the Ethereum blockchain framework. There exists a trade-off between pri-
vacy and accuracy in the sense that decreasing the privacy budget leads to
an increase of test errors on all data sets. This proposed model is further
compared with the “Learning ChainEX”, which is implemented with higher
differential privacy and has similar test error.
Zhu, Li, & Yu (2019) develop a blockchain-based privacy-preserving
framework to secure the share of updates in federated learning. The Fed-
erated Learning algorithm is developed by McMahan et al., (2017), which
allows each mobile device to compute and upload updates to the global pre-
dictive model based on their local data sets. A security issue arises when
there exist Byzantine devices in the network. In this case, the blockchain
transaction mechanism is adopted to ensure the security of sharing and
updating changes. Specifically, model updates are written in a blockchain
transaction by nodes. Along with the digital signature of a node, a trans-
action broadcasts to other nodes information, including changes of hyper-
parameters and weights, public keys (participants’ addresses). Other nodes
validate the transaction and test updates according to their local data sets.
If most nodes confirm that the performance score of the updated model is
higher than the existing model under their local data sets, the updates are
implemented into the current model.

4.2 Computing power allocation

Luong et al. (2018) develop a deep learning-based auction algorithm for edge
computing resources allocation to support mobile mining activities. The de-
signed framework enables mobile device miners to submit their bid valuation
profiles to one Edge Computing Service Provider (ECSP) for buying addi-
tional computing power. The valuation profile for miner i, denoted as vi ,
is drawn from a distribution that assigns a higher value vi when its block

12
size divided by initial computing capacity is larger. The ECSP evaluates all
valuation profiles and maximizes its revenue in the following steps.
An allocation rule is applied to map transformed valuation profiles, de-
noted as v i := φi (vi ), to assignment probabilities using a Softmax function.
The winner miner i will pay the price pi := φ−1 i (ReLU(maxi6=j v j )). In the
end, the loss function of ECSP is defined as
N
(w,β) (w,β)
X
R(w,
b β) = − gi (vs )pi (vs ) (4.1)
i=1

where stochastic gradient descent (SGD) is applied. Here, gi is the assign-

ment probability and N is the number of miners. The above designed deep
learning (DL) based auction mechanism is empirically compared to a regu-
lar auction mechanism. It is found that DL-based auction achieves higher
revenue and converges to the optimal value faster than other mechanisms.

4.3 Cryptocurrency price prediction

For forecasting Bitcoin price, Mcnally, Roche, & Caton (2018) compare per-
formances of two deep learning algorithms, i.e., Recurrent Neural Network
(RNN) and Long-Short Term Memory (LSTM). It is interesting to note
that two hidden layers with 20 nodes per layer are sufficient in both mod-
els. Specifically, the RNN model adopts the tanh fcuntion as its activation
function while LSTM applies tanh and sigmoid functions for different gates,
which result in longer training time. The data set used to train and test
LSTM and RNN models is the bitcoin price from Aug 19th, 2013 to July
19th, 2016. Features including the opening price, daily high, daily low, the
closing price, hash rate, and mining difficulty are used in the model. The
importance of features is evaluated by the Boruta algorithm, which is a wrap-
per built around the random forest classification algorithm. The traditional
time series model, AutoRegression Integrated Moving Average (ARIMA),
is empirically compared with these deep learning models. The simulation
results show that LSTM, RNN, and ARIMA have similar accuracy, which
are 52.78%, 50.25%, and 50.05%. However, deep learning models have much
lower RMSE values. In addition, the LSTM model is capable of recognizing
long-term dependencies in contrast to the RNN model.
In contrast with other studies mainly for predictive models, Lahmiri &
Bekiros (2019) instead conduct a chaotic time series analysis before build-
ing deep learning models. Hence, their first step is to calculate the largest
Lyapunov exponent (LLE) and then apply detrended fluctuation analysis

13
(DFA) to detect chaos characteristics of cryptocurrency price data without
having the assumption of stationarity. Then a deep neural network (DLNN)
model with LSTM implementation (Hochreiter & Schmidhuber, 1997) and
a generalized regression neural network (GRNN) model (Specht, 1991) are
built to predict three types of cryptocurrency: Bitcoin, Digital Cash, and
Ripple price. The number of data samples obtained for the model is 3006
Bitcoin, 1704 Digital Cash ,and 1357 Ripples. The authors create a many-
to-many sequence prediction, which utilizes the first 90% observations for
training and the last 10% observations for testing and out-of-sample forecast-
ing. According to Figure 3 whose x-axis represents the time horizon and the
y-axis represents the price, positive Hurst exponent (HE) value indicates
long-memory features of data, and negative LLE value indicates training
data is chaos. As a result, a short-term prediction model would be suitable
for data. The simulation results claim that the LSTM model outperforms
the GRNN model in all three cryptocurrencies’ price predictions. Although
the RMSE of the LSTM model is still high, the model demonstrates a similar
trend to real price changes for all three cryptocurrencies.
Besides cryptocurrency price prediction, Alessandertti et al. (2018) ex-
plore a portfolio analysis by forecasting daily prices of 1681 types of cryp-
tocurrencies. Three models are developed to predict the prices of every kind
of cryptocurrency. For each type c, the target is the return of investment
(ROI) at each time ti ∈ {0, ..., 895}, which is expressed as:

price(c, ti ) − price(c, ti − 1)
ROI(c, ti ) = (4.2)
price(c, ti − 1)

Features considered are price, market capitalization, market share, rank, and
volume. The first model is an ensemble of regression trees using XGboost,
which features of each type of cryptocurrency are paring with prices of each
type of cryptocurrency. The second model is a regression model by consid-
ering features of all kinds of cryptocurrency as a whole paired with prices
of each type of cryptocurrency. The third model adopts RNN with LSTM
implementation with the second model’s features and target paring strategy.
All models are one-step ahead forecasting. A portfolio is constructed based
on the predicted prices. Model hyperparameters are optimized by maximiz-
ing either sharp ratio or geometric mean of the total return. The result
concludes that all three models generate profits, and the optimization of
parameters using the sharp ratio metric achieves a higher return. Another
conclusion is that the first two models implementing gradient boosting with
decision trees have higher accuracy for the short-term (5-10 days), while the

14
Figure 3: Chaotic Analysis and Prediction Result (Lahmiri & Bekiros, 2019)

15
third model adopting LSTM has a better prediction performance in the long
term (around 50 days).

5 REINFORCEMENT LEARNING
In this section, we first review a framework that incorporates reinforcement
learning into blockchain to ensure the security of data collection, storage
and processing in the IoT network (Liu, Lin, & Wen, 2018). Secondly, we
review two types of frameworks that study the Bitcoin blockchain mining
activities. The first explores the potential of Bitcoin mining through the
mobile network (Nguyen et al., 2020). While the second formulate a Markov
decision process (MDP) for the blockchain mining activity (Eyal & Sirer,
2014; Sapirshtein, Sompolinsky, & Zohar 2017). The last work in the review
applies a new reinforcement learning algorithm to find the optimal mining
strategy (Wang, Liew, & Zhang, 2019).

5.1 IoT
Liu, Lin, & Wen (2018) propose a framework to secure data collection and
sharing among mobile terminals (MTs) on the IoT network. The framework
consists of two phases: data collection and data sharing. In the first phase,
each MT, denoted as m, adopts multi-agent deep reinforcement learning
(DRL) to maximize efficacy of data collection. The state space is defined
as S = {S1 , S2 , S3 }. Here, S1 = {(xk , y k ), (xc , y c )} is a set of state which
represents coordinates of k Point-of-Interest (PoIs) and c obstacles in the
environment, denoted as Ex × Ey , where x ∈ [0, Ex ], y ∈ [0, Ey ]; S2 stands
for MTs’ coordinates and S3 represents sensing time ht (k) ∈ [0, t] for the
i-th POI. Action space consists of moving direction, denoted as θtm , and
moving distance, denoted as ltm . Thus, it is written as A = {(θtm , ltm ) | θtm ∈
[0, 2π), ltm ∈ [0, lmax )}. The reward rtm is given as

wt bm
t
rtm = (5.1)
αbm
t + κltm

where bmt is the amount of collected data, α, κ are the energy consumption
per collected data and per travelled distance; wt is the achieved geographical
( K ht (k))2
P
fairness, calculated by wt = Pk=1
K Each MT is implemented by four
k k=1 ht (k)2
deep neural networks and actor-critic algorithm is applied to maximize the
reward.

16
 After MTs finish the data collection, they share data through an Ethereum
blockchain network. However, the first step would be to send data to the
certificate authority (CA) for verification. Once CA verifies the ownership
of MTs’ data and checks the consistence of received data and original data
stored in the terminal, a digital signature is generated and sent back to the
MT. As a result, the MT is able to broadcast its transaction request consist-
ing of digital signature of CA, original data and its public key to other nodes
on blockchain network to be further validated. By comparing to randomly
moving MTs, MTs implemented DRL collect much more data but consume
more energy. The blockchain-based data sharing framework can still store
all data sent by MTs even under Dos attack.

5.2 Bitcoin mining

As we discussed in earlier sections, the blockchain mining costs a huge
amount of computing power, so it is nearly impossible to apply the blockchain
to the mobile system. Nguyen et al. (2019) propose a mobile edge comput-
ing (MEC) based blockchain network to assist mobile users (MUs) offloading
mining tasks to the MEC server. Specifically, the state space is defined as
st = {D1t , D0t , g t }, where D1t , D0t are new and buffered transaction data at
the time t separately; g t is the power gain by miner n offloads the task
m to the MEC server. The action space is expressed as at = xtnm , where
xtnm ∈ {0, 1} stands for the nth MU processes m mining tasks locally or of-
floading m tasks to the MEC server, respectively. The goal of a miner is to
maximize the privacy level P t defined in He et al. (2017) and minimize the
sum of cost of energy and time consumption. The system reward is formu-
lated as rt (s, a) = P t (s, a) − C t (s, a), where P (s, a) is the privacy level and
C(s, a) is the sum of the cost of the electricity and the computing power. A
value-based method, Q-learning, and deep q learning are applied to update
the Q value. The result concludes that although the convergence speed for
Q-learning and deep q learning are almost the same, agents trained by the
deep q learning model receive higher total rewards.
Although mining bitcoins could generate a big revenue by selling bit-
coins, the cost of mining is also high due to the high consumption of elec-
tricity. People now are interested in finding the optimal mining strategy
to maximize their profits. Since Bitcoin mining is able to be modeled as a
Markov decision process (MDP) that contains a enormous number of states,
reinforcement learning is applied to study the MDP of bitcoin mining. The
MDP for Bitcoin mining is first proposed by Eyal & Sirer (2014) and then
is extended by Sapirshtein, Sompolinsky, & Zohar (2017). The environment

17
assumes that the block generation time follows Poisson distribution and it is
independently with each other. The new block is created by an honest agent
with probability (1 − α), while the new block is obtained by the adversary
agent, also known as an attacker, with probability α. The adversary may
hide some blocks on its own private chain, but the blockchain is always the
longest public chain. The state space of the MDP is defined as (a, h, f ork),
where a represents the number of blocks on the adversary’s private chain;
h represents the number of blocks on the public chain; fork is an environ-
ment variable that has three values, which are (irrelevant, relevant, active).
(a, h, irrelevant) denotes the case when previous state is (a−1, h) and match
action is feasible, i.e., the last mined block accepted by the chain is mined
by the adversary miner; (a, h, relevant) denotes the case when previous
state is (a, h − 1) and match action is infeasible, i.e., the last mined block
is mined by the honest miner; active refers to the case that the network
is broken into two branches containing the same number of blocks. When
f ork = active, the probability that follows the honest block is γ and the
probability that follows the adversary block is 1 − γ. The action space, de-
fined as A = (Adopt, Override, M atch, W ait), contains four actions. The
Adopt refers to an agent always mines mines the last block on the public
chain and do not have any blocks on its private chain. The Override be-
comes feasible when the number of blocks on the private chain is more than
the number of blocks on the public chain. In other words, all blocks on the
private chain are published to replace the existing public chain. The M atch
action refers to the adversary agent releases the same number of blocks as
the current public chain, which creates a fork on the public chain. The
W ait action is always feasible, which the adversary agent keeps mining on
its private chain and not releasing any new blocks to the public chain. The
transition probability matrix is shown in Table 6. Since the honest agent
is consider as a part of environment, we only focus on finding the optimal
strategy for adversary agents. The number of blocks on the public chain is
considered as rewards. The reward is formulated as two dimensions, which
are the number of blocks mined by honest agents and adversary agents sep-
arately. The reward function then considers relative reward instead of a
absolute reward. The objective function is defined as 5.2.

q a (s, a)
f (s, a) = (5.2)
q a (s, a) + q h (s, a)
Wang, Liew, & Zhang (2019) plan to apply the off-policy based Q-
learning to solve the problem, unfortunately the reason that Q-learning is
used there does not mention in the original paper. Since Q-learning can

18
 State, Action State + 1 Transition Prob. Reward
(a, h, ), adopt (1, 0, irrelevant) α (0, h)
(0, 1, irrelevant) 1−α (0, h)
(a, h, ), override (a − h, 0, irrelevant) α (h + 1, 0)
(a − h − 1, 1, relevant) 1−α (h + 1, 0)
(a, h, irrelevant), wait (a + 1, h, irrelevant) α (0, 0)
(a, h, relevant), wait (a, h + 1, relevant) 1−α (0, 0)
(a, h, active), wait (a + 1, h, active) α (0, 0)
(a, h, relevant), match (a − h, 1, relevant) ˙ − α)
γ (1 (h, 0)
(a, h + 1, relevant) ˙ − α)
(1 − γ)(1 (0, 0)

Table 6: Transition Probability (Sapirshtein, Sompolinsky, & Zohar, 2017)

only solve a linear reward function, the authors propose a new RL multi-
dimensional algorithm based on the off-policy Q-learning. The new algo-
rithm considers two Q-functions, i.e., a pair of (Q(a) (s, a), Q(h) (s, a)). At
a , r h ) from the en-
each time step, the adversary agent observes (st+1 , rt+1 t+1
vironment. Then two Q-functions are updated as follows:
(a)
q (a) (st , at ) ← (1 − β)q (a) (st , at ) + β[(rt+1 + λq (a) (st+1 , a0 )] (5.3)
(h) (h) (h) (h) 0
q (st , at ) ← (1 − β)q (st , at ) + β[(rt+1 + λq (st+1 , a )] (5.4)

where β ∈ (0, 1) is the learning rate, λ is a number close to 1, a0 =

argmaxa f (st+1 , a). The current best action is chosen by the  greedy strat-
egy to maximize the objective function 5.2 with the probability 1 − . A
random action is chosen with the probability . The random selection is in-
volved to avoid trapping at local maximums. The parameter  is determined
by (st ) = exp(− V T(st ) ), where V (st ) is the number of times that the state
was visited and T controls the speed of reducing .
Sensitivity analysis is applied to evaluate the designed optimal strat-
egy and the simulation result is shown in Figure 4. After setting the dis-
counted factor as 1, the paper concludes that the optimal mining strategy
outperforms current mining strategies presented in Eyal & Sirer (2014) and
Sapirshtein, Sompolinsky, & Zohar (2017).

6 CONCLUSION AND FUTURE CHALLENGES

The research we review either applies blockchain in a database to improve
users’ privacy in learning process; or uses machine learning to optimize com-

19
Figure 4: Simulation Result for Different Mining Strategies (Wang, Liew, &
Zhang, 2019)

puter resource allocation or cryptocurrency investment decisions. The ma-

jority can be categorized as applying one technique to another; few is the
actual integration of the two technologies. Hence, it is fair to say the current
research is still very preliminary from an interdisciplinary perspective.
However, we expect new research lines to emerge in the following areas:

• Design “smart agents” with learning abilities to regulate the blockchain

and detect abnormal behaviors. The former is especially important for
consortium chain and private chain that require coordination among
users, while the latter is critical for public chain;

• The learning-based analysis of blockchain-based system is rare. From

financial systems to supply chains, there is an enormous amount of
data available to evaluate the performance of the decentralized struc-
ture of blockchain compared with the traditional centralized one. Learning-
based analysis can shed insights on the mechanism design of the blockchain
structures and provide on-time forecasting models;

• Blockchain to allow anonymously data sharing. With the develop-

ment of IOT and wearable device, the privacy issue catches more and
more attention of users. Combining with data fusion, we can design
multiple-layer blockchain structures that allow sophisticated autho-
rization of data for different users.

• The blockchain mining activity could be considered as an MDP pro-

20
 cess. Although there exist a few works related to finding the optimal
mining strategy using single-agent reinforcement learning, individual
mining is not as popular as pool mining in reality. Specifically, miners
collaborate and compete with each other to mine blocks. A multi-agent
reinforcement learning (MARL) with a mixed setting of collaborative
and competitive agents is more suitable to model the complex pool
mining activity and helps miners find the optimal mining strategies in
the future.

• Cryptocurrency plays an important role, especially in the public chain.

Different chains have their unique cryptocurrency. Now cryptocur-
rency or cryptocurrency portfolio is an investment option similar to
other financial products. Some works have studied cryptocurrency
price prediction using supervised learning techniques, but only a few
of them explore potentials of RL or deep RL. In many cases, RL and
deep RL have a better in financial forecasts, e.g., stock price predic-
tion, since historical data cannot reflect the current market, which
further results in poor prediction performance of future price changes.
We expect that more works adopting RL, deep RL, or inverse RL to
study the investment return of cryptocurrencies emerge soon.

References
[1] Abay, N. C., Akcora, C. G., Gel, Y. R., Kantarcioglu, M., Islambekov,
U. D., Tian, Y., & Thuraisingham, B. (2019). Chainnet: Learning on
blockchain graphs with topological features. In 2019 IEEE International
Conference on Data Mining (ICDM), 946–951.

[2] Abernethy, J. D. & Frongillo, R. M. (2011). A collaborative mechanism

for crowdsourcing prediction problems. In Advances in Neural Informa-
tion Processing Systems 24, J. Shawe-Taylor, R. S. Zemel, P. L. Bartlett,
F. Pereira, and K. Q. Weinberger. (Eds.) Curran Associates, Inc., 2600–
2608.

[3] Agbo, C.C., Mahmoud, Q.H., & Eklund, J.M. (2019). Blockchain Tech-
nology in Healthcare: A Systematic Review. In Healthcare, 7, 56.

[4] Akcora, C. G., Dey, A. K., Gel, Y. R., & Kantarcioglu, M. (2019) Fore-
casting bitcoin price with graph chainlets. In Advances in Knowledge
Discovery and Data Mining. (Cham, 2018), D. Phung, V. S. Tseng, G. I.

21
 Webb, B. Ho, M. Ganji, and L. Rashidi. (Eds). Springer International
Publishing, 765–776.

[5] Akcora, C. G., Li, Y., Gel, Y. R., & Kantarcioglu, M. (2020). Bitcoin-
Heist: Topological Data Analysis for Ransomware Prediction on the Bit-
coin Blockchain. In Twenty-Ninth International Joint Conference on Ar-
tificial Intelligence Special Track on AI in FinTech, 4439 - 4445

[6] Alessandretti, L., ElBahrawy, A., Aiello, L. M., & Baronchelli, A. (2018).
Anticipating cryptocurrency prices using machine learning. In Complexity
2018 .

[7] Bhaskar, N. D. & Lee, K. C. (2015). Chapter 3 - bitcoin mining technol-

ogy. In Handbook of Digital Currency, D. Lee Kuo Chuen. (Ed.) Academic
Press, San Diego, 45 – 65.

[8] Bhattacharya, P., Tanwar, S., Bodke, U., Tyagi, S., & Kumar, N. (2019).
BinDaaS: Blockchain-Based Deep-Learning as-a-Service in Healthcare 4.0
Applications. In IEEE Transactions on Network Science and Engineering

[9] Bresler, G., Chen, G. H., & Shah, D. (2014). A latent source model for
online collaborative filtering. In Advances in Neural Information Process-
ing Systems 27, Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence,
and K. Q. Weinberger, (Eds.) Curran Associates, Inc., 3347–3355.

[10] Buhlmann, P., Drineas, P., Kane, M., & van der Laan, M. (2019).
Handbook of big data. in CRC Press.

[11] Chen, G. H., Nikolov, S., & Shah, D. (2013). A latent source model
for nonparametric time series classification. In Advances in Neural Infor-
mation Processing Systems 26, C. J. C. Burges, L. Bottou, M. Welling,
Z. Ghahramani, and K. Q. Weinberger. (Eds.) Curran Associates, Inc.,
1088–1096.

[12] Chen, X., Ji, J., Luo, C., Liao, W., & Li, P. (2018). When machine
learning meets blockchain: A decentralized, privacy-preserving and secure
design. In 2018 IEEE International Conference on Big Data (Big Data),
IEEE, 1178–1187.

[13] Chung, K.-Y., Yoo, H., Choe, D.-E., & Jung, H. (2019). Blockchain
network based topic mining process for cognitive manufacturing. In Wire-
less Personal Communications, 105, 583–597.

22
[14] Dey, A. K., Akcora, C. G., Gel, Y. R., & Kantarcioglu, M. (2020).
On the role of local blockchain network features in cryptocurrency price
formation. In Canadian Journal of Statistics, 48(3), 561–581.

[15] Dinh, T. N. & Thai, M. T. (2018). Ai and blockchain: A disruptive

integration. In Computer, 51(9), 48–53.

[16] Ermilov, D., Panov, M., & Yanovich, Y. (2017). Automatic bitcoin ad-
dress clustering. In 2017 16th IEEE International Conference on Machine
Learning and Applications (ICMLA), 461–466.

[17] Eyal, I. & Sirer, E. G. (2014). Majority is not enough: Bitcoin mining
is vulnerable. In Commun. ACM, 61(7), 95–102.

[18] Ferrag, M. A. & Maglaras, L. (2020). Deepcoin: A novel deep learning

and blockchain-based energy exchange framework for smart grids. In IEEE
Transactions on Engineering Management, 67(4), 1285–1297.

[19] Firdaus, A., Anuar, N., Faizal, M., Hashem, I., Bachok, S., & Ku-
mar, A. (2018). Root exploit detection and features optimization: Mobile
device and blockchain based medical data management. In Journal of
Medical Systems/, 42.

[20] Ghimire, S. & Selvaraj, H. (2018). A survey on bitcoin cryptocurrency

and its mining. In 2018 26th International Conference on Systems Engi-
neering (ICSEng), 1–6.

[21] Gökalp, E., Gökalp, M. O., Çoban, S., & Eren, P. E. (2018). Analysing
opportunities and challenges of integrated blockchain technologies in
healthcare. In Information Systems: Research, Development, Applica-
tions, Education (Cham, 2018), S. Wrycza and J. Maślankowski, (Eds.)
Springer International Publishing, 174–183.

[22] Granger, C. W. J. (1969). Investigating causal relations by econometric

models and cross-spectral methods. In Econometrica, 37(3), 424–438.

[23] Greaves, A. S. & Au, B. (2015). Using the bitcoin transaction graph to
predict the price of bitcoin. Stanford Publishing.

[24] Harrigan, M. & Fretter, C. (2016). The unreasonable effec-

tiveness of address clustering. In 2016 Intl IEEE Conferences
on Ubiquitous Intelligence Computing, Advanced and Trusted Com-
puting, Scalable Computing and Communications, Cloud and Big

23
 Data Computing, Internet of People, and Smart World Congress
(UIC/ATC/ScalCom/CBDCom/IoP/SmartWorld), 368–373.

[25] Harris, J. D. & Waggoner, B. (2019). Decentralized and collaborative

ai on blockchain. In 2019 IEEE International Conference on Blockchain
(Blockchain), 368–375.

[26] He, X., Liu, J., Jin, R., & Dai, H. (2017). Privacy-aware offloading
in mobile-edge computing. In GLOBECOM 2017 - 2017 IEEE Global
Communications Conference, 1–6.

[27] Hochreiter, S. & Schmidhuber, J. (1997). Long short-term memory. in

Neural Computation, 9(8), 1735–1780.

[28] Jang, H. & Lee, J. (2018). An empirical study on modeling and predic-
tion of bitcoin prices with bayesian neural networks based on blockchain
information. In IEEE Access, 6, 5427–5437.

[29] Jiang, Z. & Liang, J. (2017). Cryptocurrency portfolio management

with deep reinforcement learning. In 2017 Intelligent Systems Conference
(IntelliSys), 905–913.

[30] Jourdan, M., Blandin, S., Wynter, L., & Deshpande, P. (2018). A prob-
abilistic model of the bitcoin blockchain. In CoRR.

[31] Juneja, A. & Marefat, M. (2018). Leveraging blockchain for retraining

deep learning architecture in patient-specific arrhythmia classification. In
2018 IEEE EMBS International Conference on Biomedical Health Infor-
matics (BHI), 393–397.

[32] Khezr, S., Moniruzzaman, M., Yassine, A., & Benlamri, R. (2019).
Blockchain Technology in Healthcare: A Comprehensive Review and Di-
rections for Future Research. In Appl. Sci, 9, 1736.

[33] Khan, A.S., Zhang, X., Lambotharan, S., Zheng, G., AsSadhan, B., &
Hanzo, L. (2020). Machine Learning Aided Blockchain Assisted Frame-
work for Wireless Networks. In IEEE Network, 34(5) 262-268.

[34] Lahmiri, S. & Bekiros, S. (2019). Cryptocurrency forecasting with deep

learning chaotic neural networks. In Chaos, Solitons and Fractals, 118, 35
– 40.

[35] Li, Z., Guo, H., Wang, W. M., Guan, Y., Barenji, A. V., Huang, G. Q.,
McFall, K. S., & Chen, X. (2019). A blockchain and automl approach for

24
 open and automated customer service. In IEEE Transactions on Indus-
trial Informatics, 15(6), 3642–3651.

[36] Liu, C. H., Lin, Q., & Wen, S. (2018). Blockchain-enabled data collec-
tion and sharing for industrial iot with deep reinforcement learning. In
IEEE Transactions on Industrial Informatics, 15(6), 3516-3526.

[37] Luong, N. C., Xiong, Z., Wang, P., & Niyato, D. (2018). Optimal auc-
tion for edge computing resource management in mobile blockchain net-
works: A deep learning approach. In 2018 IEEE International Conference
on Communications (ICC), 1–6.

[38] Mamoshina, P., Ojomoko, L., Yanovich, Y., Ostrovski, A., Botezatu,
A., Prikhodko, P., Izumchenko, E., Aliper, A., Romantsov, K., Zhe-
brak, A., Ogu, I. O., & Zhavoronkov, A. (2017). Converging blockchain
and next-generation artificial intelligence technologies to decentralize
and accelerate biomedical research and healthcare. In Oncotarget, 9(5),
5665–5690.

[39] McMahan, B., Moore, E., Ramage, D., Hampson, S., & y Arcas, B. A.
(2016). Communication-Efficient Learning of Deep Networks from De-
centralized Data. In Proceedings of the 20th International Conference on
Artificial Intelligence and Statistics, A. Singh and J. Zhu, (Eds.), 54 of
Proceedings of Machine Learning Research, PMLR, 1273–1282.

[40] Mcnally, S., Roche, J., & Caton, S. (2018). Predicting the price of
bitcoin using machine learning. In 2018 26th Euromicro International
Conference on Parallel, Distributed and Network-based Processing (PDP),
IEEE, 339–343.

[41] MIT Technology Review Editors (2018). Explainer: What is a

blockchain?

[42] Nakamoto, S. (2008). Bitcoin: A peer-to-peer electronic cash system.

[43] Nguyen, D. C., Pathirana, P., Ding, M., & Seneviratne, A. (2020).
Privacy-preserved task offloading in mobile blockchain with deep rein-
forcement learning. In IEEE Transactions on Network and Service Man-
agement, 1–1.

[44] Özyilmaz, K. R., Doğan, M., & Yurdakul, A. (2018). Idmob: Iot
data marketplace on blockchain. In 2018 Crypto Valley Conference on
Blockchain Technology (CVCBT), 11–19.

25
[45] Qian, Y., Jiang, Y., Chen, J., Zhang, Y., Song, J., Zhou, M., &
Pustišek, M. (2018). Towards decentralized iot security enhancement: A
blockchain approach. In Computers and Electrical Engineering, 72, 266 –
273.

[46] Qin, R., Yuan, Y., & Wang, F.-Y. (2019). A novel hybrid share report-
ing strategy for blockchain miners in pplns pools. In Decision Support
Systems, 118, 91 – 101.

[47] Rathore, S., Pan, Y., & Park, J.H. (2019). BlockDeepNet: A
Blockchain-Based Secure Deep Learning for IoT Network. In Sustainabil-
ity, 11, 3974.

[48] Saleh, F. (2020). Blockchain without waste: Proof-of-stake. In Infor-

mation Systems and Economics eJournal .

[49] Sapirshtein, A., Sompolinsky, Y., & Zohar, A. (2017). Optimal selfish
mining strategies in bitcoin. In Financial Cryptography and Data Security
(Berlin, Heidelberg, 2017), J. Grossklags and B. Preneel, (Eds.), Springer
Berlin Heidelberg, 515–532.

[50] Shah, D. & Zhang, K. (2014). Bayesian regression and bitcoin. In 2014
52nd Annual Allerton Conference on Communication, Control, and Com-
puting (Allerton), 409–414.

[51] Shen, M., Tang, X., Zhu, L., Du, X., & Guizani, M. (2019). Privacy-
preserving support vector machine training over blockchain-based en-
crypted iot data in smart cities. In IEEE Internet of Things Journal,
6(5), 7702–7712.

[52] Singla, K., Bose, J., & Katariya, S. (2018). Machine learning for secure
device personalization using blockchain. In 2018 International Conference
on Advances in Computing, Communications and Informatics (ICACCI),
67–73.

[53] Specht, D. F. (1991). A general regression neural network. In IEEE

Transactions on Neural Network, 2(6), 568–576.

[54] Sun, X., Liu, M., & Sima, Z. (2020). A novel cryptocurrency price
trend forecasting model based on lightgbm. In Finance Research Letters,
32, 101084.

26
[55] Vyas, S., Gupta, M., & Yadav, R. (2019). Converging blockchain and
machine learning for healthcare. In 2019 Amity International Conference
on Artificial Intelligence (AICAI), 709–711.

[56] Waggoner, B., Frongillo, R., & Abernethy, J. D. (2015). A market

framework for eliciting private data. In Advances in Neural Informa-
tion Processing Systems, 28, C. Cortes, N. D. Lawrence, D. D. Lee,
M. Sugiyama, and R. Garnett, (Eds.) Curran Associates, Inc., 3510–3518.

[57] Wang, T., Liew, S. C., & Zhang, S. (2019). When blockchain meets ai:
Optimal mining strategy achieved by machine learning. arXiv eprints.

[58] S. Wang and J. Wang and X. Wang and T. Qiu and Y. Yuan and L.
Ouyang and Y. Guo and F. Wang (2018). Blockchain-Powered Parallel
Healthcare Systems Based on the ACP Approach. In IEEE Transactions
on Computational Social Systems, 5(4), 942 – 950.

[59] Wohrer, M. & Zdun, U. (2018). Smart contracts: security patterns in

the ethereum ecosystem and solidity. In 2018 International Workshop on
Blockchain Oriented Software Engineering (IWBOSE), 2–8.

[60] Wood, G. (2014). Ethereum: A secure decentralised generalised trans-

action ledger.

[61] Xiong, W. & Xiong, L. (2019). Smart contract based data trading
mode using blockchain and machine learning. In IEEE Access, 7, 102331–
102344.

[62] Yin, H. S. & Vatrapu, R. (2017). A first estimation of the proportion of

cybercriminal entities in the bitcoin ecosystem using supervised machine
learning. In 2017 IEEE International Conference on Big Data (Big Data),
3690–3699.

[63] Zheng, X., Mukkamala, R. R., Vatrapu, R., & Ordieres-Mere, J. (2018).
Blockchain-based personal health data sharing system using cloud stor-
age. In 2018 IEEE 20th International Conference on e-Health Networking,
Applications and Services (Healthcom), 1–6.

[64] Zheng, Z., Xie, S., Dai, H.-N., Chen, X., & Wang, H. (2018). Blockchain
challenges and opportunities: a survey. In Int. J. of Web and Grid Ser-
vices, 14, 4.

27
[65] Zhu, S., Li, W., Li, H., Hu, C., & Cai, Z. (2019). A survey: Reward
distribution mechanisms and withholding attacks in bitcoin pool mining.
In Mathematical Foundations of Computing, 1, 393.

[66] Zhu, X., Li, H., & Yu, Y. (2019). Blockchain-based privacy preserv-
ing deep learning. In Information Security and Cryptology (Cham, 2019),
F. Guo, X. Huang, and M. Yung, (Eds.), Springer International Publish-
ing, 370–383.

28