Computer Communications 178 (2021) 37–63

Contents lists available at ScienceDirect

Computer Communications
journal homepage: www.elsevier.com/locate/comcom

Blockchain management and machine learning adaptation for IoT

environment in 5G and beyond networks: A systematic review
Arzoo Miglani a , Neeraj Kumar b,c,d,e ,∗
a Department of Computer Science and Engineering, Thapar Institute of Engineering and Technology, Patiala, Punjab, India
b Department of Computer Science and Engineering, Thapar Institute of Engineering and Technology, Patiala, India
c
Department of Computer Science and Information Engineering, Asia University, Taiwan
d
School of Computer Science, University of Petroleum and Energy Studies, Dehradun, Uttarakhand, India
e
King Abdul Aziz University, Jeddah, Saudi Arabia

ARTICLE INFO ABSTRACT

Keywords: Keeping in view of the constraints and challenges with respect to big data analytics along with security and
Blockchain privacy preservation for 5G and B5G applications, the integration of machine learning and blockchain, two of
Machine learning the most promising technologies of the modern era is inevitable. In comparison to the traditional centralized
Federated learning
techniques for security and privacy preservation, blockchain uses decentralized consensus algorithms for veri-
Internet of Things
fication and validation of different transactions which are supposed to become an integral part of blockchain
Deep learning
5G
network. Starting with the existing literature survey, we introduce the basic concepts of blockchain and
6G machine learning in this article. Then, we presented a comprehensive taxonomy for integration of blockchain
and machine learning in an IoT environment. We also explored federated learning, reinforcement learning,
deep learning algorithms usage in blockchain based applications. Finally, we provide recommendations for
future use cases of these emerging technologies in 5G and B5G technologies.

1. Introduction response time for such large database repositories. Clearly, it is difficult
to process and analyze this huge data to make meaningful correlations
Over the last decade, Internet of Things (IoT) has revolutionized and decisions making. Moreover, all IoT applications are having sen-
the whole world leading to various technological trends starting from sitive information, for which security and privacy preservation are of
Industry 1.0 to Industry 5.0, AR/VR/MR, smart factories, tactile Inter- utmost important. Also, devices are reluctant to transfer their data for
net, smart transportation, smart plants, etc. It is an interconnection of training purposes in an open environment such as the Internet because
various devices monitored and controlled using the Internet in order to of privacy concerns [3]. Data shared by geographically distributed
provide ubiquitous computing services to the end-users. Because of the users may get tampered by different attackers which in return may
constraints such as — heterogeneity of devices, resource constraints, generate inaccurate results. Also, IoT system needs to be autonomous
power storage, security, and data management constant revolutions so that it can learn from the gathered data and make context-based
are foreseen in IoT over the years. Among these, the security and decisions [4]. In such an environment, machine learning (ML) can be
privacy are most crucial keeping in view of the data access restric- an effective tool in understanding the patterns, analyzing, processing,
tions at various levels in different applications [1]. Moreover, with
and making intelligent decisions. Fig. 1(c) shows the global ML market
an increase in the number of IoT devices, the data generated by
share from the year 2017 to 2024 [5]. Technology advancements in
these devices is increasing exponentially in recent years. As per the
ML and deep learning (DL) have changed the way a computer can
report [2], the number of IoT devices connected to the Internet at
process information automatically. The ever-growing market for IoT
the end of Nov. 2019 was 26.6 billion and is expected to reach 75
demands the usage of ML-based models for accuracy and precision
billion by the year 2025. This continuous growth of IoT applications
in the decision-making process. Implementing ML in IoT applications
and the number of devices leads to wireless connectivity to 5G and
Beyond-5G (B5G) of mobile technology. These wireless networks have can significantly improve data analytics and real-time decision-making.
extended network coverage, enhanced quality of service faster speed, Applications of ML in various IoT use-cases (e.g., smart transportation,
and massive bandwidth. However, the conventional databases may face smart grid, etc.) include network optimization, resource allocation,
problems with respect to indexing and searching queries for quick congestion avoidance [6].

∗ Corresponding author at: Department of Computer Science and Engineering, Thapar Institute of Engineering and Technology, Patiala, India.
E-mail addresses: [email protected] (A. Miglani), [email protected] (N. Kumar).

https://doi.org/10.1016/j.comcom.2021.07.009
Received 5 February 2021; Received in revised form 8 July 2021; Accepted 9 July 2021
Available online 14 July 2021
0140-3664/© 2021 Elsevier B.V. All rights reserved.
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Fig. 1. (a) The overall spendings on blockchain based IoT solutions (b) Percentage of startups in different industries focusing on blockchain [9] (c) Global machine learning market
by component.

It is important to note that, ML can be used with 5G and B5G IoT configuration framework [11,12]. Smart contracts in blockchain can
networks to empower self-aggregating communication and intelligent automate negotiations between network service providers and con-
resource orchestration. The use of 5G’s Radio access network (RAN) sumers. Moreover, blockchain can improve the performance of ML
makes the issue of energy efficiency in IoT even more worse as RAN algorithms as it provides digitally signed data from reliable, trusted,
consumes over 80% of the power in the wireless network. ML can help and secure sources. The distributed computing powers can be utilized
to fix 5G’s RAN complexity. For example, autonomous controllers based for developing a better and secure prediction model. Also, in the last
upon Artificial Intelligence (AI) can be used to optimize energy us- few years, automatic decision-making systems have been developed
age [7]. Similarly, predictive models for energy consumption including at rapid growth. In this context, blockchain technology provides a
Markov’s decision process and NN’s can be incorporated with IoT- useful framework to store the record of various activities and de-
enabled devices [8]. Moreover, intelligent predictive algorithms such as cisions by utilizing features of the consensus algorithm and smart
— reinforcement learning optimizes networks independently and pro- contracts. To generate efficient models in ML, a large volume of data
vides decision-making on data being gathered. However, ML enabled is required as it increases the throughput and accuracy. Some of the
5G network faces some security concerns and demands improvement other ways to provide distributed access control are based on fog
to meet the standardization of next network generation. computing, edge computing, and ad-hoc mobile cloud-based on the
Nevertheless, most of the available ML algorithms rely on the trusted third party [13]. An overview of the blockchain for machine
centralized framework which may lead to the security breaches and learning-enabled 5G and B5G networks is shown in Fig. 3. The device
may have a single point of failure. With the centralized authority, layer generates an enormous amount of data which is collected and
threats of privacy preservation, false authentication, data tampering stored securely on blockchain network. Also, this data is sent to the
prevails. Also, the reliability of data is very important for ML algorithms cloud layer for processing and analysis. Communication layer contains
in order to obtain accurate results. Even a small security loophole information regarding network status, Cloud-RAN’s (C-RAN) condition,
in the ML algorithm can generate high false rate for certain events. Base station’s data, available resources, and device performance. In this
Moreover, the computations of ML models are dependent on the trusted layer, network slicing is realized with well-defined service requirements
third party (TTP) (e.g., a cloud service provider) for many security including massive Machine Type Communication (nMTC), enhanced
applications which may raise serious privacy concerns. Hence, there Mobile BroadBand (eMBB), ultra-Reliable Low Latency Communication
is a demand for decentralized framework based ML. To address the (uRLLC). The data with respect to the communication layer is stored
aforementioned challenges, blockchain can be a viable solution as it in the blockchain network. With blockchain, 5G and B5G services
has been widely used in various IoT applications such as — health- can be more scalable as they support efficient solutions for spectrum
care, agriculture, smart grid, etc. Fig. 1(b) represents the percentage sharing and resource management [14]. The last layer is the processing
of startups in different industries focusing on blockchain in the year layer where data is analyzed using learning algorithms including linear
2021 [9]. As per a report in [10], IoT blockchain 50 based spending regression, neural networks (NN’s), SVM, decision trees, etc. After
is expected to reach $573M by 2023 as compared to $174M in the analysis, blockchain is again used as a distributed and immutable ledger
year 2018 (Refer Fig. 1(a)). It eliminates the need for a centralized for the obtained results.
party to manage the overall system. Also, blockchain technology can On the other hand, the adoption of ML in blockchain helps to
provide many benefits to 5G IoT networks including secure authen- analyze the existing issues in blockchain technology, enabling to en-
tication, secure communication, secure network coding, and resource hance the security and privacy of the whole network. For example,

38
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Fig. 2. Adoption of blockchain and ML for 5G and B5G IoT applications.

Fig. 3. An overview of blockchain for ML-enabled 5G and B5G networks.

support vector machine (SVM) provides effective data classification integrated for a secure and efficient decision-making process [15]. This
for blockchain peers and other transactional entities. Moreover, the article discusses the importance of both ML and blockchain in IoT in 5G
supervised ML algorithms such as — random forest, gradient boost, and B5G environments along with our vision towards their integration.
etc. are used to reduce anonymity in the blockchain network. Recently, In the end, we discuss various challenges and future directions to
NN’s are also exploited to predict the price of cryptocurrency. With improve the integration of these technologies for IoT.
various computing models, ML can ease data verification, validation
process and helps in identification of anomalies and malicious attacks
1.1. Review method
in the blockchain network. Resource management, classification of
transactional entities, and managing offloading tasks are some other
The systematic methodology carried out for the review process is
applications of ML for blockchain.
ML and blockchain are two emerging technologies and their inte- represented in Fig. 4(b) and is described as follows:
gration is inevitable. Both ML and blockchain are predicted to revo-
lutionize the world’s economy. The Adoption of blockchain and ML 1.1.1. Review plan
for 5G and B5G IoT applications is represented in Fig. 2. Both of We started with the development of research questions, database
these technologies follow the data-driven approach and hence can be exploration, downloading relevant papers, and evaluation of the same.

39
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Fig. 4. (a) Workflow for identifying primary studies, (b) Survey methodology.

The identified research proposals are first checked for adequate quality can be fitted into IoT were also considered. The search operation was
then it is considered for writing review. In particular, we followed carried out using keywords such as — ML-based blockchain, ML for
the protocol that states: ‘‘review individually and finalize after joint blockchain, ML and blockchain, Blockchain for ML, etc. from 2008
discussion’’. to 2020. Maximum papers are selected from IEEE explore. To refine
the search further, different constraints such as — DL and blockchain,
1.1.2. Research questions DL and Bitcoin, distributed ML were also used (Refer Fig. 4(a)). The
This review focuses on how advantageous it is to combine retrieved papers were categorized as follows:
blockchain and ML technologies. The detailed planning for conducting
the survey process involves framing questions as presented in Table 1. • Blockchain for ML
• ML for blockchain
1.1.3. Sources of data
As a result, we retrieved around 150 articles that were useful for our
The next step was to find relevant literature depending on the
survey. During the initial screening, it has been observed that only
nature and scope of the review. In this step, following databases were
around 90 articles were suitable for writing a good review as other
explored:
articles lack experimental results. Also, for a complete understanding of
• IEEE explore IoT, blockchain and ML individually some survey articles were studied.
• Springer
• Science direct 1.1.5. Conducting review
• ACM digital library The next step was to conduct the review. For this, a detailed study
• Wiley of selected articles was done with a motive to interpret the findings
of each article. After finding the characteristic features of each article,
Additionally, technical reports, patents, weblinks relevant to block-
chain and ML integration are also included. different applications of blockchain for ML and ML for blockchain were
grouped. Also, the key challenges for the integration of blockchain and
1.1.4. Search criteria ML were identified.
Next, search criteria were developed in order to retrieve all related
literature. For this review, we had considered all qualitative literature 1.1.6. Writing review
studies which combined blockchain and ML for the IoT. In fact, the Finally, the last step was to write the review with the structure as
schemes which are not proposed specifically for the IoT domain but mentioned in the previous step.

40
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Table 1
Research questions and their objective.
Q. no. Research question Objective
1 What is blockchain and ML It aims to explore the basic concepts,
technologies? advantages, and challenges of ML
and blockchain so that their
integration can be better understood.
2 Is it feasible to integrate It aims to confirm whether these two
blockchain and ML? technologies can be integrated or not.
3 What are the different areas It checks the areas where
where blockchain and ML are combination of both of these
used? technologies can be proved efficient.
4 What is the purpose of this It aims to check how readers can
survey? gain knowledge regarding state-of-art
literature in the area of ML and
blockchain.
5 How many related survey It aims to check the already existing
articles are already available? related review articles on the
combination of blockchain and ML.
6 What is the novelty of this This question checks how our survey
survey article? will be different from already
existing related work.
7 Who is the audience for this This question aims to check the
review article? journals and their categories where
this review article could be
submitted so that it could be
maximally available to the audience.
8 What should be the taxonomy This question aims to arrange the
of the paper? article in different categories so that
it looks well arranged and increases
readability.
9 How many recent literature This question aims to check whether
studies are available on the the research on the chosen topic is
integration of blockchain and trending or not.
ML?
10 How the related literature This question aims to explore the
articles will be retrieved? available databases for retrieving the
literature work.
11 What are the different datasets The purpose of this question is to get
and testing environments used familiar with different datasets and
in proposals combining programming environment used in
blockchain and ML proposals.

Fig. 5. Organization of the article.

41
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Table 2
List of acronyms in an alphabetical order.

AI Artificial Intelligence
ANN Artificial Neural Networks
ARIMA AutoRegressive Integrated Moving Average
B5G Beyond 5G
BCoT Blockchain Cloud of Things
BNN Bayesian Neural Network
CNN Convolutional Neural Networks
D2D Device-to-Device
DC’s Data centers
DAO Decentralized autonomous organization
DL Deep Learning
DDoS Distributed Denial-of-service
DRL Deep reinforcement learning
ECC Elliptic-curve cryptography
EVM Ethereum virtual machine
GLM Generalized Linear Model
GRNN General Regression Neural Networks
IPFS InterPlanetary File System
IoT Internet of Things
IoV Internet of Vehicle
LLL Lisp like language
LSTM Long Short-Term Memory
MAE Mean Absolute Error
P2P Peer-to-Peer
RMSE Root Mean Square Error
MAPE Mean Absolute Percentage Error
MITM Man-in-the-Middle
MLP Multilayer Perceptron
MNIST Modified National Institute of Standards and Technology database
MSE Mean Square Error
ML Machine Learning
NN Neural Networks
NLP Natural Language Processing
PBFT Practical Byzantine Fault Tolerance
PCA Principle Component Analysis
PoA Proof-of-Activity
PoL Proof-of-Learning
PoS Proof-of-Stake
PoW Proof-of-Work
RL Reinforcement Learning
RNN Recurrent Neural Networks
SDA Stacked Denoising Autoencoder
SVM Support Vector Machine
SVR Support Vector Regression
TTP Trusted Third Party
XAI Explainable AI

1.2. Organization Section 6 discusses concluding remarks for the integration of ML and
blockchain. The structure of the complete article is represented in
The road-map of this survey is described as follows. We begin by Fig. 5.
discussing a high-level view on the need for integration of ML and
blockchain in IoT environment. In Section 1.1, we provided the survey 2. Review of related survey articles
methodology. The list of acronyms used is presented in Table 2. Next,
Section 2, discusses the related survey articles in ML and blockchain Existing literature work reveals that blockchain and ML are sur-
integration. In addition, we discuss how the survey article is different veyed mostly in isolation or with their applications in several vertical
from other related survey articles. Section 3, provides terminologies, domains. Recently, the connections between the two technologies have
basic concepts, and recent trends in the field of ML and blockchain. been observed in many applications. There are several survey articles
In Section 4, we classify the state-of-art literature based on ML and that discuss the concept and compare ML models. Specifically, the
blockchain in two classes, i.e., ML for blockchain and blockchain for survey of ML models for big data analysis can be found in [16–18].
ML. In this section, we also present the comparison of available litera- Meanwhile, multiple notable works such as — [19–21] provide the
ture on blockchain and ML collaboration in tabular form. In Section 5, concepts, advantages, challenges, and future research directions of
we pinpoint various research challenges that should be focused for blockchain technology. The more recent survey articles in the context
future research work in the area of ML and blockchain. In the end, of blockchain applications for IoT have been presented in [22–25]

42
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

whereas authors of [26–28] discuss the applications of ML models in authority. The digital certificate provides the feature of authentication.
various fields of IoT. Nevertheless, a new block before adding to the network needs to be
Several studies were put forward addressing the integration of Arti- validated by the participants of the network. Each validated block
ficial Intelligence (AI) and blockchain. For example, the authors of [29] contains the hash of the immediate previous block so that any record of
presented a review article on the integration of AI and blockchain by the block cannot be changed. Also, the Merkle hash value and append
discussing applications of blockchain for AI as well as AI for blockchain. only structure prevents blockchain from tampering and preserves in-
Likewise, Salah et al. [30] present the review on the literature and sum- tegrity. Additionally, the multi backup facility prevents any tampering
marize the existing blockchain applications and protocols facilitating if a group of peers misbehaves. Blockchain also provides the facility
AI domain. Along with this, open research challenges of implementing of smart contracts to enable automation. Hence, blockchain can boost
blockchain for AI are also discussed by the authors. However, only co-ordination platforms for several components of ML data, algorithm
a few research efforts have been made on the integration of ML and and computing power.
blockchain, in order to provide decision-making service in an intelligent Depending on accessibility, the blockchain network can be catego-
way while assuring security and privacy. For example, Vyas et al. [31] rized into public, private, and consortium network. Popular cryptocur-
discussed the role of blockchain in improving the accuracy of ML results rencies such as — Bitcoin and Ethereum follow the public blockchain
for healthcare applications. However, authors presented a short survey network. A public blockchain allows every node to be a participant of
article and in-depth knowledge cannot be gained with this article. In the network. However, in order to deal with scalability and privacy
the same way, Acheampong [32] presented an overview of the basic issues, consortium and private blockchain were designed. In contrast
concepts of blockchain and ML by discussing the impact of blockchain to the public blockchain, in private blockchain, the owner of the
in ML community. More recently, authors in [33] conducted an inten- blockchain network is a single entity. Here, only selected participants
sive survey that focuses on a specific application of ML for blockchain, (based on some predefined proposal) are allowed to be a part of the
i.e., anomaly detection. Also, this article reviews the application of network. Speed and scalability are the most prominent advantages
blockchain for privacy preservation in learning process. of private blockchain [40]. With the advantage of providing low en-
In contrast, authors of [15] presented a review to discuss the ergy consumption and delay, private blockchain is better to deal with
applications of ML in blockchain technology. Specifically, authors have delay-sensitive and energy-delayed networks. Hyperledger and Ripple
reviewed ML for blockchain applications such as — transaction entity are popular examples of the private blockchain network. Consortium
classification, Bitcoin price prediction, computing power allocations, blockchain is almost similar to private blockchain type except it is
cryptocurrency price prediction, and portfolio management. In another meant to be controlled by multiple entities rather than one [41].
work, Nguyen et al. [34] presented a small section that discusses Quorum cryptocurrency falls under this category.
the efficiency of ML in improving blockchain cloud of things (BCOT) A smart contract is a piece of code written in high level language
framework. Very recently, Rane et al. [35] presented in-depth survey such as Solidity and is executed as per the requirement. It enables au-
on available ML algorithms for predicting Bitcoin prices and concluded tomatic code execution in a trustless environment without verification
that existing schemes only achieve accuracy of 60%–70%. Recently, Liu from a third party. As smart contract resides on a blockchain network
et al. [36] present a survey article that discusses overview, benefits, such that its source code is accessible to everyone on the network. The
applications, open issues, and challenges while combining blockchain most popular platform for smart contract is Ethereum. The applications
and ML. of smart contract are not only limited to cryptocurrency but can be
The earlier review efforts on the integration of ML and blockchain extended to many applications including voting systems, inventory
are summarized in Table 3. It is clear from Table 3 that the proposed management, automation of payments, automation of claims and blind
survey article extensively discusses the companionship between ML and auctions, etc. Next, we will discuss some of the programming languages
blockchain in both aspects, i.e., how blockchain accelerates ML services used to write smart contracts.
and vice-versa along with discussing challenges and future research • Solidity: Solidity [42] is the most popular high-level programming
directions for the integration of both technologies. language used for implementing smart contracts on the Ethereum
platform. The primary purpose of Solidity is to program browser-
3. Preliminaries side UI logic. This language is influenced by C++, python, and
javascript. The main construct of the solidity language is a con-
3.1. Blockchain basics tract which is similar to a class in object-oriented programming
language. Similar to any other programming language solidity
The concept of blockchain was first introduced by Satoshi Nakamoto contracts have functions visibility modifiers and return value to
in 2008 [37]. The cryptocurrencies such as — Bitcoin, Ethereum, Ripple the calling user in the network. One of the different features of
depict the popular applications of blockchain technology. Among all solidity is the modifier which itself is a function wrapping another
cryptocurrencies, Bitcoin is generally referred to as blockchain 1.0 function, however, it is invoked by adding modifier name to the
whereas Ethereum as blockchain 2.0. With the growth in blockchain signature function. It supports inheritance, libraries, and complex
popularity, researchers try to execute more difficult tasks on user defined types.
blockchain. In this context, Ethereum developed a Turing complete • Serpent: Serpent [43] is inspired from the Python language which
language so that various applications can be implemented on the focuses on delivering high productivity and automating tasks.
blockchain. Since then, the expected spending on blockchain solu- Real-time garbage collection and supporting multiple virtual ma-
tions is expected to grow at an estimate of 11.7 billion by the year chines concurrently are two important features of Serpent. The
2022 [38]. The fundamentals of blockchain rely on peer-to-peer (P2P) syntax of Serpent is similar to that of Python, i.e., it also uses in-
network architecture, where data of network is not controlled by a dentation to indicate structure. Serpent is useful for writing code
single centralized authority. Instead, blockchain technology provides a with low-level opcode manipulation. However, unlike solidity,
distributed and decentralized ledger that is maintained by each network Serpent does not have a static type system to be robust against
participant. Decentralization, immutability, anonymity, transparency, bugs.
and security are some important features of blockchain. In order • Vyper: After Solidity, Vyper [44] is the next most popular lan-
to assure trust and security, blockchain leverages the principle of guage for Ethereum virtual machine (EVM) having syntax in-
cryptography [39]. The cryptographically secured structure for data spired from Python. Vyper uses annotations to modify function
storing makes it difficult to breach the security of data. Every node on visibility. Bounds and overflow checking, support for signed in-
the blockchain network gets a digital certificate from the authorized tegers, decimal fixed point numbers, decidability, strong typing,

43
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Table 3
Comparison of related survey articles.
Reference Year Application Blockchain for ML ML for blockchain Challenges
Liu et al. [36] 2020 Communication and ✓ ✓ ✓
network systems
Vyas et al. [31] 2019 Healthcare ✓ × ×
Nguyen et al. [34] 2019 Cloud-of-things × ✓ ✓
Chen et al. [15] 2019 – ∼ ✓ ✓
Salah et al. [30] 2019 – ✓(blockchain for AI) ✓ ×
Rane et al. [35] 2019 Bitcoin × ∼ ✓
Dinh et al. [29] 2018 – ✓(blockchain for AI) ✓(AI for blockchain) ×
Rahouti et al. [33] 2018 Bitcoin × ∼ ✓
Acheampong et al, 2018 Big data ∼ × ×
[32]
The proposed survey – IoT ✓ ✓ ✓

✓ indicates topic is covered, × indicates that topic is not covered and ∼ indicates that topic is partially covered.

and understandable compiler code are the important features The limited resource IoT devices are also unsuitable for highly com-
provided by Vyper. Notably, for achieving auditability and pre- putational PoW consensus algorithm. Hence, the scalability issue of
vention from gas limit attack, Vyper does not provide the feature integrating blockchain and IoT needs an immediate effective solution.
of class inheritance, function overloading, operator overloading, Also, different characteristics of IoT network such as — heterogeneity,
and recursive calling. wireless communication and mobility complicates the security chal-
• LLL: LLL (Lisp like language) is the first low-level language devel- lenge. Moreover, the transparency supported by IoT can affect the
oped after the assembler for EVM and it is a tiny wrapper over privacy of data. Last but not the least, lack of regulations and standards
coding around the assembler itself. LLL provides direct access can influence the future of blockchain and IoT.
to memory in an execution environment and can be easily opti-
mized for speed. Particularly, LLL does not hide highly resource- 3.2. Machine learning basics
constrained nature of EVM and provides efficient use of those
limited resources. ML is a branch of AI that makes programming machines to perform
particular tasks by learning. With time, ML models have been able to
Moreover, with an over-increasing deployment of IoT objects, se- exceed humans in various problems. Particularly, previous experience
curity is of prime concern. Cloud computing has been widely used is used to execute assigned tasks. ML algorithms have proved their sig-
to support IoT for management, processing, and storage. However, its nificance in various areas such as — transportation, image processing,
centralized nature raises security questions. Centralized servers manag- marketing, etc. ML includes various models to solve different types of
ing sensitive IoT data can be shared with anybody without the user’s problems. The most commonly used ML models involve SVM, Artificial
consent, thus leading to privacy breaches [45]. Also, the intermediaries Neural Networks (ANN), decision trees, etc. to name a few. Building
decrease the efficiency of interactions among system components. Also, a new ML model involves two steps, i.e., training and testing in order
with an increase in the number of IoT devices, current centralized to perform tasks of prediction, classification, clustering, etc. on new
devices providing security services including authentication and autho- dataset. Indeed, data is an important source in ML. The data is required
rization will turn into a bottleneck. Moreover, the security vulnerability in preprocessing and training any ML model. First, the ML model is
because of centralization is an easy target for Distributed denial-of- trained with a training dataset. With the increase in size of training
service (DDoS) attacks. Additionally, to ensure data integrity presence data, the efficiency of ML classifier also increases [48]. Next, after
of publically verifiable audits without involving a TTP is desirable. the training phase, the accuracy of the prediction is evaluated with a
In this context, blockchain can mitigate security and privacy risks new dataset. In case of acceptable accuracy, the ML model is deployed
with its capabilities such as — transparency, immutability, anonymity, otherwise it is trained again. In recent, a popular subcategory of ML
decentralization, and operational resilience [4]. The intrinsic features named deep learning (DL) has emerged to imitate the human thinking
of blockchain make it suitable to be applied in different IoT scenarios process. The fundamentals of DL have been originated from cognitive
including smart homes, smart grids, and smart industries, etc. Also, theories that are used to create NN structure. Popular applications of DL
blockchain enables the autonomous function in IoT devices which include object detection, face recognization, and traffic flow prediction
removes the presence of centralized third-party. to name a few [49].
Nevertheless, the existing IoT solutions are having energy-cons- Supervised learning, unsupervised learning, and reinforcement
trained and limited computing capabilities of IoT devices. However, to learning (RL) are three categorizations of learning styles in ML al-
support resource-constrained nature of IoT devices blockchain provides gorithms. In supervised learning, the machine is trained with well
the concept of Simplified payment verification, in which nodes need labeled data, i.e., the data is already mapped with the correct an-
not to store complete blockchain data rather only block headers. In this swer. Next, the machine is fed with a completely new set of data to
context, Le and Mutka [46] proposed a lightweight method to validate generate correct results from analyzing the labeled data from training
blockchain data using bloom filter (probabilistic data structure). Sim- phase. Furthermore, supervised learning is divided into two categories
ilarly, authors in [47] presented a proposal that integrates blockchain that include classification and regression. SVM, decision trees, nearest
with constrained IoT devices. The evaluation of the proposal is carried neighbor, etc. are popular algorithms under this category. In contrast,
out in terms of memory, processing time, and power consumption. unsupervised learning is training the machine with input data that is
However, high computation, storage costs, high energy demands, not labelled or classified. Specifically, the aim is to group unsorted
communication hurdles, mobility of devices, and latency are some data as per similarity and difference such as — pattern detection and
of the challenges faced while integrating blockchain with IoT. In an descriptive modeling. Clustering and association are two categories of
IoT network, devices generate gigabytes of data in real-time. Due to unsupervised learning [50]. K-means clustering and Principle Compo-
lack of storage blockchain might appear unsuitable for IoT networks. nent Analysis (PCA) are popular algorithms under this category. In

44
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Fig. 6. Taxonomy for integration of ML and blockchain.

RL, an agent is employed to interact with the environment in order

to find best outcome by continuously learning from the environment.
RL uses trial-and-error method to train itself when exposed to a certain
environment. Markov’s decision process is a popular example of RL.
Notably, there are vulnerabilities in ML models system with respect to
privacy and security. Security attack in ML mainly includes evasion and
poisoning attacks. Evasion attacks disrupt the entire classification pro-
cess using adversarial examples whereas the poisoning attack destroys
the data while training phase, which can decrease model accuracy [51].
On other hand, the privacy attack on ML model comes from service
providers and third-party entities. Clearly, the development of ML mod-
els empowers to launch new AI services including facial recognition and
words suggestion. Nevertheless, the dataset provided to support these
applications often includes sensitive and private information.
In this discussion, it is important to mention the concept of explain-
able AI (XAI), a ML technique suite that generates more explainable
models, while still maintaining a high level of prediction accuracy.
ML models are considered as block box as the prediction results come
with no justification. The algorithm used for ML results is not able to
explain why a decision has been made. However, the data scientists and
engineers are aware of the process of pre-processing, transformations,
and operations that dataset goes through but the final user is unaware
of it. For applications including, healthcare, military, and finance where Fig. 7. Blockchain for machine learning applications.
trust is a requirement in making decisions, this a major drawback [52].
XAI aims to answer how black box decisions in ML models are made. It
inspects the steps and models used in decision-making. XAI is not only
4.1. Blockchain for machine learning
used for trust verification, but for scientific discovery in future artificial
general intelligence systems. For more details on XAI readers may refer
The applications of blockchain distributed ledger can be utilized be-
to [53–55].
yond cryptocurrency. Blockchain for ML can solve the problem of data
acquisition. With blockchain, the ML algorithm can be fed with highly
4. The integration of machine learning and blockchain in IoT reliable data and thus accurate and trusted results can be achieved.
Also, training ML models with real-data will enhance the accuracy and
efficiency of ML algorithms. The built-in consensus mechanism and
Earlier ML and blockchain had been two different technologies fundamentals of blockchain ensure secure and tamper-proof sharing
without any clear link. Most of the existing literature has studied ML of IoT data. Moreover, the existing client-master type ML models
and blockchain separately. This section provides a review of literature rely on trusted central servers and consider only privacy issues in
available on the integration of both technologies in IoT. Although this linear sharing and ignore privacy in non-linear learning models. In
section divides ML and blockchain integration into two parts. The first the client-master model, an enormous amount of data generated by
subsection reviews the literature and tabulates the existing blockchain IoT devices is collected and stored at one central location whereas, in
applications targeting ML areas whereas the second subsection reviews the distributed multi-party model, data is generated by various parties
ML applications for blockchain. The taxonomy for the integration of ML and stored in a distributed manner. However, the decentralized model
and blockchain is presented in Fig. 6. incurs high communication costs and raises security and privacy issues.

45
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Table 4
Relative comparison of blockchain for ML based proposals.
Reference Blockchain app. for Blockchain type Consensus algorithm Machine learning Evaluation metrics Advantages
ML algo.
[56] Trustless ML – – Neural networks Accuracy, recall, Automated self
contracts precision, F1 score improvement system
for ML agent
[57] Distributed trust in Permissioned PBFT ANN (Stochastic Precision Reduced
ML computations (Hyperledger) Gradient Descent) communication and
storage cost
[58] Verifiable open Public Proof-of-Learning – Accuracy, F score, Energy involved in
repository for ML kappa, RMSE, MAE validation process
models solves useful tasks
[59] Privacy preserving Private PoW – – Deals with
predictive modeling Byzantine general
and sybil attack
problems
[60] Secure access Permissioned – Stacked denoising Accuracy, 95.10% of accuracy
control (Hyperledger) autoencoder sensitivity, is achieved
specificity and
positive predictive
rate
[61] Privacy of data Ethereum, IOTA – Rule mining – Security and privacy
(learning) of data
[62] Trustability and Public, permissioned PoW Associative rule False positive and Easily identify and
automation mining false negative highlight recent
changes and trends

Table 5
Relative comparison of blockchain for ML based proposals.
Reference Blockchain app. for Blockchain type Consensus algorithm Machine learning Evaluation metrics Advantages
ML algo.
[63] Determination of Public PoW Federated Learning – Mitigates DDoS and
relevant data MITM attack
[64] Privacy preservation Public PoW SVM Accuracy, efficiency, Secure data sharing
scalability
[65] Privacy preservation Ethereum PoW Stochastic gradient Privacy budget, Mitigates Byzantine
descent Byzantine attacks attacks
[66] Privacy preserved Permissioned Proof of training Random tree, MAE Secure data
data sharing Quality random forest and management,
gradient boosting Achieving
decision tree differential privacy
[67] Privacy preservation – – Deep learning Accuracy Mitigates effect of
(federated learning) Byzantine devices
[68] Tamper-proof ML Public – CNN Object recognition Prevents against
model accuracy adversarial
tampering and layer
perturbation attack
[69] Distributed ML Permissioned – Differentially Computation time, Low computational
computation (Hyperledger) private stochastic transaction latency complexity and low
gradient descent transactional latency
[70] Distributed ML in Public PBFT Stochastic gradient Convergence attack, High scalability
5G networks descent take over attack,
computation
overhead
[71] Privacy-preservation Public PoW Federated learning Average learning Improved latency
completion latency and reduced block
propagation delays
[72] Privacy-preservation Public Modified version of Federated learning Robustness Robust than typical
PBFT and PoW federated learning
system

The transparency feature supported by blockchain, assures ML users in Tables 4, and 5. Next, we will discuss various literature works
confidentiality and privacy of data. As discussed, more amount of data supporting applications of blockchain for ML.
available for training improves the overall throughput and produces
4.1.1. Trustless ML contracts
a more effective and reliable system. Clearly, blockchain in ML can
A potential use of blockchain for ML is in developing a better ML
result in much safer data and better ML models. The applications of model. Notably, the smart contract facility offered by blockchain can
blockchain targeting ML areas are presented in Fig. 7. The tabulated offer rewards in a trustless manner for training a ML model. By utilizing
comparison of blockchain based ML literature studies is presented this, the organizations which are efficient in solving ML problems

46
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Fig. 8. An illustration of trustless machine learning contracts.

can directly construct their skillset. This can also provide a trustless large scale computations for a multi agent network prove expensive.
reward for the creation of better ML models. The proposal introduced For this, the authors have used a lossy compression technique that
by [56] implements the concept of trustless ML contract and it is reduces the communication and storage cost of the blockchain network.
defined in 3 phases. In the first phase, a dataset, an evaluation function, Similarly, authors of [62] established a link between ML and blockchain
amount of reward, and request for best ML model is submitted by technology in order to solve trustability and automation issues of ML
the reward giver/buyer. In the second phase, the provided dataset is by using association rule mining.
downloaded by ML model providers/practitioners and each provider
works independently in order to train the ML model. After training, the 4.1.3. Verifiable open repository of ML models
providers submit their model. In the last phase, the winner is selected. Among all research work on consensus algorithms, Proof-of-Work
Moreover, such a proposal can be utilized for raising funds transpar-
(PoW) is the widely accepted technical consensus algorithm use to
ently for IoT applications such as — medical research. In addition, it
settle among all participating nodes. However, the PoW consensus
can achieve automated self-improvement for AI agents. Unfortunately,
algorithm proves costly and environmental unfriendly due to the high
this proposal [56] does not require identity and reputation validation
computations involved in it. After PoW many other consensus algo-
for creating a new transaction and hence raises security concerns. Also,
rithms such as — Proof-of-Stake (PoS), Proof-of-Activity (PoA) were
this proposal works only for Ethereum blockchain. Fig. 8 represents an
introduced in order to reduce computations while mining blocks. In
illustration of trustless ML contracts.
this context, the authors of [58] introduced a cryptocurrency named
‘‘WekaCoin’’ that is based on Proof-of-Learning (PoL) consensus algo-
4.1.2. Distributed trust in ML computations
rithm. PoL is inspired by open-source ML competitions (e.g. Kaggle and
Notably, it is difficult to trust results from trained ML models having
CodaLab). Among all network nodes, some nodes called trainers upload
open source code and open data in an IoT environment. In fact, multi
ML models on blockchain network for tasks that were submitted by
agent socio-technical systems (which work collaboratively on some
tasks, share models and data for local computations) due to the involve- other nodes called suppliers. (The model initiator may upload their
ment of independent agents face trust issues in computations from other model on a Interplanetary file system (IPFS) system and in return
agents. In such cases, validation at each individual step of computations receives checksum hash.) The uploaded models are then tested for
and verification of the integrity of computations by other agents is data that was not considered by trainers while training. The validator
required. Another matter to be considered in the context of ML is that nodes which are selected randomly are then supposed to rank these
these algorithms lack trustability and automation. As ML algorithm models and add the information to the block. The trainer nodes having
relies on data that is mutable, so it is difficult to trust the results from the best model are rewarded with WekaCoins by supplier nodes. This
these algorithms. The system administrator can manipulate the data way blockchain can be used for generating verifiable ML models.
source that in return changes the result. Also, existing ML models are The flowchart for the understanding of PoL algorithm is presented in
mostly controlled by human beings so it is difficult to automate the Fig. 9. The main advantage of this protocol is that the computations
ML algorithms. Hence, there is a need for developing an environment involved in the validation process solve useful tasks as well as creates
having trust and transparency in computations for collaborative op- a validated open repository for ML models and datasets. However,
erations. To solve this problem, zero-knowledge proof, Elliptic-curve the authors have not discussed the prevention of collusion among
cryptography (ECC), etc. are some cryptographic techniques that are suppliers, trainers, and validators. In contrast to the permissionless
effective in the verification and validation of computations [73,74]. blockchain, authors of [69] developed privacy preserving distributed
In this context, Raman et al. [57] proposed a model for verification ML model based on permissioned blockchain network. This is, however,
and validation of computations in a permissioned blockchain network a first attempt to propose a distributed ML model for a permissioned
for multi-agent socio-technical system. Authors have demonstrated the blockchain network. Decentralized ML allows machines to perform
usage of blockchain in developing trust for recording and validating intelligent decision-making on data securely stored on the blockchain
audit at each step of computations. However, due to lack of scalability network without involving any TTP. The decentralized ML technique

47
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Fig. 9. Flowchart for PoL algorithm.

allows algorithms or ML models to run directly on connected mobile de- and Sybil attack problem and preserves privacy while predicting the
vices. This distributed technology is smart contract based marketplace modeling process. In this context, Kou et al. [59] have presented
that connects developers, clients, and data owners by facilitating all Modelchain, a private blockchain that enabled privacy preserving pre-
stakeholders in a way to create a middle-man free ML infrastructure. dictive modeling for the healthcare industry. Instead of relying on
The authors demonstrated that the impact of proposed error based only PoW protocol, the authors have designed a new algorithm on
aggregation rule supports high resilience and mitigates collusion attack. the top of PoW named proof of information to increase the efficiency
However, latency and bandwidth are the major drawback of distributed and accuracy of ML model. Unfortunately, the proof of information
ML [75]. To improve network condition, 5G technology can be adopted algorithm proves inefficient to deal with the scalability of the network.
as it enables high availability. In this direction, to ensure byzantine The result section demonstrated that Modelchain provides a secure and
resilience for distributive learning in five networks, authors in [70] privacy preserving interoperability framework. Unfortunately, privacy
have proposed a blockchain based secure computing framework. By preservation is provided but the authors of [59] did not consider the
using a sharding based blockchain, authors have prevented arbitrary basic requirements for differential privacy as differential privacy based
attacks on learning convergence. ML has to consider the fact that how many times a ML model can
However, authors of [76] pointed out that ML programs cannot be trained without any privacy breach. Subsequently, Chen et al. [65]
be stored with blockchain because of the certain limitations of smart proposed another decentralized ML system called ‘‘Learningchain’’ that
contracts. The authors pointed out that smart contracts cannot process takes both linear and non-linear learning models in account without
high computational tasks. With the blockchain mining process, when relying on the central server. Here, differential privacy based methods
output corresponding to any input is expected to be recorded via are also designed to preserve the privacy of data. Differential privacy or
smart contracts, honest miners then execute the program to verify
cryptographic solutions have proved to be efficient for preserving user’s
the correctness of results. In case of a computationally high process,
data privacy [78–80]. This model is implemented on the Ethereum
adversarial nodes can skip and carry forward to verify the new block.
platform and a stochastic gradient descent algorithm is used to design
This way adversarial nodes can get a chance of adding new blocks
a predictive model over blockchain. The proposal works in 3 phases. In
as honest participants are busy with the execution of smart contracts.
the first phase, a P2P network is initialized. In the second phase, data
Moreover, the smart contract cannot carry randomized computations
holders calculate their local gradients as per predefined common loss
as with randomization honest nodes can have inconsistent output.
function and predictive model using differential privacy methods. Next,
Besides, as ML computations are costly and randomized, so ML tasks
computed gradients are broadcasted in the network using differential
are difficult to execute with blockchain. To address this challenge, the
privacy scheme for learning models. After reaching a consensus, local
authors of [76] have used a game theory approach that empowers
gradients are aggregated by the authority holder using Learningchain.
randomized computations on the top of blockchain. Here, a simple
incentive mechanism is designed in order to execute the program with Three different datasets were used for training and testing purposes, i.e.,
crowdsourcing in a blockchain environment. synthetic dataset, Wisconsin breast cancer dataset, and Modified Na-
tional Institute of Standards and Technology database (MNIST) dataset.
It is concluded in results that there is a trade-off between privacy and
4.1.4. Privacy preservation
Another matter to be considered in the context of ML is the privacy accuracy as lowering the privacy budget increases test errors.
preservation of data. For example, ML healthcare predictive modeling With the growing trend of DL models, many DL models are designed
has proved beneficial in national healthcare research and biomedical to be run on client devices such as — IoT devices or smart devices.
discoveries. However, data disclosure of patients to these third-party Although this technique demands enough memory and disk space to
cloud services leads to privacy attacks. The available distributed pri- run the models in real-time. Also, because of privacy concerns, it is not
vacy preserving predictive models are dependent on the central server recommended to upload client data on a centralized machine for pro-
to execute the modeling process [77]. Institutional policies, single cessing and executing ML algorithms. Along the same line of thought,
point of failure, mutable disseminate data, and trust issues are some to preserve privacy while uploading client ML data, authors of [61]
associated risks with the existing client–server architecture. Moreover, proposed another work. Singla et al. [61] proposed a blockchain-based
any participating node cannot leave or join the network for a short system that stores client device profiles in a shared household to predict
period of time in order to avoid any recovery issues. The state-of-art user activity. Here, the main aim is to enable automatic customization
research has adopted blockchain technology in order to deal with the of each client using blockchain decentralized security and privacy.
above-mentioned risks. The characteristics of blockchain technology The personalization feature of each device is computed using rule
make it suitable to deal with centralized privacy preservation mod- mining. However, this proposal is based on the assumption that client
els. Blockchain avoids a single point of failure, Byzantine General, preferences are not changing.

48
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Fig. 10. (a) Centralized ML architecture (b) An illustration of federated learning.

Similarly, to solve the challenge in collaborative data sharing among generates an update to the existing global model located at the server.
multiple parties in IoT applications, Lu et al. [66] proposed a pri- Next, the server combines these models by aggregating them and the
vacy preserving data sharing model using differential privacy methods. whole process is repeated until global model training is completed. The
However, rather than sharing raw data directly, the federated learning primary benefit of federated learning is the decoupling of the training
algorithm is utilized into permissioned blockchain network through phase from the requirement of direct access to raw training data. The
which only data model is shared over decentralized multiple parties. In process of federated learning based model is represented in Fig. 10(b).
a centralized ML model, participants upload their data on central cloud Therefore, it minimizes training and privacy risk. However, the usage
server. The server performs all computational tasks for training on the of a single central server is vulnerable to a single point of failure.
data as shown in Fig. 10(a). This model involves high risks of privacy Moreover, there is no reward service for distributed devices. Notably,
attacks. Also, communication overhead is created between participants the devices with more data samples should be given reward as it con-
and the cloud server. In contrast, federated learning enables ML models tributes more to global training. With blockchain, verified local updates
to be computed on distributed mobile devices. This technique helps and exchanges can be enabled along with providing corresponding
ML models to be trained on the devices where data is produced. This rewards proportional to the size of training sample size. The illustration
way the privacy of data is ensured as data of a particular device does of blockchain based federated learning has been represented in Fig. 11.
not leave its data production place. This technique is disrupting the Unfortunately, the federated learning technique fails to provide security
centralized way of data training. In federated learning, each device has in case of the presence of Byzantine nodes. If an attacker, pretends
its local training dataset that is never seen by the server and each device to be a real data holder and breaks down the security of system,

49
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Fig. 11. An illustration of blockchain based federated learning.

such an attacker is called Byzantine attacker. In another work, Zhu preservation, and further signatures are taken to measure similarity
et al. [67] also presented a blockchain based privacy preserving method among data entries. In an alternative work, the authors of [64] also
for securing updates and achieving consensus in federated learning. proposed blockchain based model named ‘‘secureSVM’’ for privacy
Here, blockchain technology is adopted to deal with Byzantine devices preserved sharing of data while training ML algorithms. Here, IoT data
in the network. In particular, only updates are added in blockchain generator encrypts data on the local device by their private key, and
transaction records. Along with broadcasting digital signatures of a this encrypted data is stored on blockchain. The experimental result
node, other information such as — hyper-parameters, difference in proves that incorporating blockchain with SVM classifier improves the
weights, and public ID’s are also broadcasted. The other participants accuracy of the system model.
of the network validate the broadcasted transactions as per their local
datasets. If majority of the participants approve that the performance 4.1.5. Cryptographic security on ML data
score of the updated model is greater than the existing models then Classification of IoT data with black-box concept, questions the
updates are added to the model. Similarly, Doku et al. [63] also
type of data being collected. Hence, the system needs to attain con-
integrated blockchain technology and federated learning to improve
fidentiality, integrity, anonymity, and secure access on data. Authors
the quality of data. Here, the hash of mobile device data is stored on
of [60] have used blockchain in retraining stacked denoising autoen-
blockchain whereas data still remains on the user’s device, and only
coder (SDA) algorithm for arrhythmia classification. Retraining is used
the locally analyzed results will be shared with ML practitioners via a
to solve non-stationary nature of ECG data because it enables deep net-
secure network. In addition to this, incentives will be provided to data
work in learning any new distribution at specific time intervals whereas
owners.
SDA has the feature of taking different relevant features from data
Additionally, in order to enhance the security of federated learn-
samples. Here, patient data stored on external storage that is collected
ing, the authors of [71] proposed a framework based on blockchain
by retraining SDA algorithm are securely accessed using blockchain
in order to verify and exchange local learning models. This scheme
based access control manager in real-time. A scenario of blockchain
aims to activate on-device ML involving any centralized server. A
reward mechanism is also proposed for user and miner node par- based secure access control on ML data has been represented in Fig. 12.
ticipation. Additionally, authors have evaluated end-to-end average More recently, Goel et al. [68] experimentally investigated the role of
learning completion latency. In a closely related work, authors of [72] blockchain in providing authenticity to each block of Convolutional
proposed federated learning with multi-access edge computing and Neural Network (CNN) model. In CNN, each convolution layer is re-
blockchain technology. Here, edge devices are employed to provide ferred to as a block and the authors pinpointed the accountability of
resources to mobile devices and also to act as blockchain nodes. Here, each block for correct output. To this end, blocks of CNN are kept in
a separate channel is dedicated for learning of every global model random order and neighbor blocks have the information regarding the
in the blockchain network. Unfortunately, in this proposal user de- next legitimate block. Indeed, hiding the architecture of the network
vices are dependent on the integrity of corresponding edge nodes for from attacker, mitigates the threat of white box adversarial attack.
sending transactions to blockchain networks. Additionally, no reward Also, this scheme enhances transparency between blocks and the entire
mechanism for user and miner nodes is designed by authors. network. Unfortunately, the complexity of the system is quite high.
Also, authors of [81] leveraged suite of ML to support data exchange Another potential application of blockchain for ML is in providing
on the blockchain via smart contract for a distributed data vending anonymity. As discussed earlier, if the data is stored anonymously, it is
architecture. Particularly, data embedding and distance metric learning hard to link the true identity of the person. Authors of [82] pointed that
approaches of ML research are used to enable retrieval of smart con- the facility of pseudo anonymity provided by blockchain can encourage
tracts without affecting the integrity of private data. Here, the signature the use of ML on anonymous dataset. Researchers can now use massive
of data entry is generated using data embedding procedure with privacy datasets for their research in order to improve the prediction results

50
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Fig. 12. Blockchain based secure access control on ML data.

of healthcare system. However, along with anonymity, encrypting data

could enhance security of the system. To address this challenge, homo-
morphic encryption was introduced that has the ability to execute ML
operations on encrypted data [83].

Summary and insights

Section 4.1 focuses on various applications of blockchain tech-
nology targeting ML areas for IoT environment. Incorporating
blockchain technology in ML provides reliable sharing of data
for different tasks of ML including prediction, forecasting,
voice, and speech recognization to name a few. However, we
have made several observations after reviewing and tabulating
the literature. Clearly, with trustless ML contracts, trustless
rewards can be provided to the best ML model. However,
there are some risks in the proposal that need to be deal
with. For example, the organizer may deny to reveal the
testing dataset which may stop submitters for their work as no
evaluation function would be available then. Moreover, based
on selection criteria, the reward money can be claimed by the
first submitter fulfilling evaluation criteria. Hence, the reward
mechanism can be evenly distributed in order to incentivize
more participation. Also, it has been observed that nodes are Fig. 13. Machine learning applications for blockchain.
still reluctant to host data on IPFS blockchain data storage. So,
future work should consider these problems before designing
revised trustless ML contracts. 4.2. ML for blockchain
Additionally, it has been observed that most of the proposals
have leveraged public blockchain which makes data gen-
For blockchain, ML can solve issues of uncertain and complex
eration speed slow. Hence, a fast data stream situation in
blockchain is another important topic of research. Moreover, features. In a blockchain environment, the data gathered from IoT
it has been observed that researchers have not considered the sensors can be analyzed and monitored at multiple points by ML
confidentiality of ML data in the blockchain network. models for efficient decision-making [98]. Following an emerging trend
rendered by the adoption of ML in blockchain, Swan [99] introduced
a new term called ‘‘blockchain thinking’’ that enables accommodating

51
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Table 6
Relative comparison of blockchain and DRL based proposals for resource management and computational offloading.
Reference Application Blockchain type Consensus Machine learning Evaluation metrics Advantages
scenario algorithm algo.
[84] Smart Grid Public – DRL Energy cost Significant energy
cost savings
[85] Any wireless Consortium PBFT DRL Average system Intelligent resource
network utility management,
reliable
orchestration,
flexible network
[86] Internet-of-Vehicles – PBFT DRL Scalability, latency, Improved
security transaction
throughput
[87] Any mobile edge – PoW DRL Computational Reduced offloading
computing network latency, energy latency, minimized
consumed, privacy energy consumption
level
[88] IoT network Private PoW DRL Access control Trustworthy access
(Ethereum) performance, control, reduced
computational offloading latency,
latency, energy minimized energy
consumed consumption
[89] Any mobile edge – PoW/PoS DRL and genetic Average cost, task Maximized
computing network algorithm drop rate, average long-term rewards,
transmit time speedup
convergence,
accommodate highly
dynamic
environment,
reduced execution
cost, avoids data
congestion
[90] Any mobile edge Public and PoW Bayesian RL, BNN payoffs of the Scalable, minimal
computing network consortium miners, convergence convergence time
time and complexity

PBFT: Practical Byzantine Fault Tolerance, PoW: Proof-of-Work, PoS: Proof-of-Stake, DRL: Deep Reinforcement Learning.

thinking on blockchain network. The main aim is to utilize the frame- transparent data structures. However, to increase the efficiency of the
work of blockchain for initiating thinking machines. In such a type network, ML models can be experimented with blockchain. In particu-
of framework, input involves sensor data. Further, the input data is lar, deep reinforcement learning (DRL) has been extensively used with
processed at a specific location to generate output that includes storing blockchain to achieve resource management tasks. DRL technique has
information to memory or taking a specific action. This process involves the capability to handle dynamic and large dimensional features of IoT.
‘‘personal thinking chains’’ that signify backup of full human mind files. The main concept behind DRL is that similar to a biological agent,
To implement the blockchain thinking process IPFS could be relevant an artificial agent may learn from interaction with its surroundings to
as it eases P2P file serving system [100]. Notably, the research work take further decisions. By interacting with the environment, the agent
of ML is entirely data-driven. This data can be shared via a central gathers experience to optimize objectives served in the form of cumu-
resource or a distributed file system. Using central repository will be lative rewards. For example, authors in [86] have used DRL method
inefficient with the increase in the number of users. On the other hand, for maximizing transactional throughput of the blockchain network. In
IPFS is a distributed file system to store data files in a decentralized particular, DRL selects block producers, block size, and block interval
manner. Also, each file in an IPFS is assigned a unique fingerprint to adjust the dynamic features of the Internet of Vehicles (IoV) scenario.
called cryptographic hash. IPFS will disseminate data files with a list of Also, in order to achieve resource management for tasks such as —
trusted nodes and the data will be available to other users using content content caching, computation offloading, spectrum sharing, etc., the
identifiers. authors in [85] have utilized DRL. Specifically, this scheme has utilized
Literature provides many studies to use applications of ML for DRL for the Device-to-Device (D2D) caching scheme that matches the
blockchain into next generation IoT which are discussed as follows. The caching supply and demand pairs to maximizes the network utilities
relative comparison of ML for blockchain-based studies is presented in of consortium blockchain enabled framework. Notably, DRL based
Tables 6, 7, 8, 9, and 10. Some of the ML applications with blockchain caching scheme optimizes bandwidth between caching requester and
are represented in Fig. 13. provider. It has been demonstrated in the results that cumulative
average system utility has been improved. However, this proposal has
4.2.1. Resource management and computational offloading not discussed the mining procedure.
Resource management is the process of scheduling and allocating Meanwhile, when embedded with smart contracts, ML helps to
resources in order to maximize efficiency of the IoT system. Energy con- minimize the energy expense in cloud data centers (DC’s) as discussed
sumption, transparency, operational expenditure, request scheduling, by the authors of [84]. Here, the smart contract facility of blockchain
latency, content caching, and security are some of the issues involved in migrates the requests and virtual machines to the cloud DC’s with
the realization of resource management process [101]. To address this minimum load, and RL method based request migration is used for
challenge, few secure and flexible resource management framework energy cost minimization as this method does not require any prior
has been developed in literature by integrating blockchain and ML. A knowledge. Fig. 14 represents the blockchain and ML empowered
blockchain based platform possesses the capability to store all records resource management scenario for smart grid networks. Here, all com-
of transactions related to resource management in a distributed and putation intensive tasks including caching, billing, demand-response

52
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Table 7
Relative comparison ML models for price prediction.
ML model Superset model Strengths Weaknesses
BNN Stochastic neural • Predicts better in case of high • Difficult to scale for large problems
network volatility • Computationally expensive
• Can handle complex relevant
features in the evaluation
• Can update analysis on introduction
of new information
• Predicts consistent trends
LSTM Recurrent neural • Selectively remembers patterns for • Significant computations required
network a long time duration. • Requires longer time for training
• Can handle time series data for
both long and short term duration
SVM Supervised ML • Do not assume any underlying • Not good for linear patterns
relationship about data form • Time consuming for large datasets
• Can handle even unstructured data
like trees, images
• Scales well to high dimensional
data
ANN DL • Support multivariate scenario • Complex training
• Robust • Problem of vanishing gradient
• Can handle stochastic and
non-linear relationship
ARIMA Time series model • Produce good accuracy for linear • Cannot handle highly volatile
time series data natured data
• Uni-variate model • Captures only linear relationship
Regression Supervised ML • Feature selection procedure • Can only handle linear relationship
model removes incorrect values
• Provides simple estimation
procedure
GLM Statistical linear • Can solve linearization problem in • Can suffer substantial precision
regression Bitcoin dataset losses
GRNN Radial basis neural • Can work well with large sample • Responses must be independent
network size
Random forest Supervised ML • Flexible • Decision tree construction is time
• Efficient for large datasets consuming

Table 8
Relative comparison of blockchain and ML based proposals for price prediction.
Reference Blockchain type Consensus Machine Evaluation Advantages Comparative
algorithm learning algo. metrics Model
[91] Public PoW RNN and LSTM RMSE LSTM achieves ARIMA
good accuracy
[92] Public PoW BNN RMSE, MAPE Addresses recent Linear
highly volatile regression, SVR
Bitcoin price
[93] Public PoW SVM, random Sensitivity, Achieved –
forest, binomial specificity, accuracy of
GLM precision and around 98.7%
accuracy
[94] Public PoW Linear MSE Achieved –
regression, SVM, accuracy of
ANN 55.1% with ANN
[95] Public PoW Linear RMSE and MAE Achieved [91]
regression, accuracy of 99%
random forest, with linear
gradient descent regression
[96] Public PoW LSTM RMSE LSTM improved GRNN
accuracy over
GRNN
[97] Public PoW Logistic Accuracy RNN performs –
regression, SVM, best among
ARIMA, RNN logistic
regression, SVM,
ARIMA and RNN

management, etc. are implemented at edge layer of the network due edge devices. Moreover, mobile base stations used to transfer data to
to resource constraints. Notably, learning capable ML agents employed
edge devices also have ML models running on them for scheduling
on edge devices are responsible for implementing effective caching,
computational offloading, scheduling, and real-time decisions on the computational or storage requests.

53
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Fig. 14. Blockchain and ML empowered resource management framework.

Table 9
Relative comparison of ML for blockchain based proposals.
Reference ML app. for Blockchain type Consensus algorithm Machine learning Evaluation metrics Advantages
blockchain algo.
[102] Resource Public PoW Neural network Revenue of edge Revenue of service
management computing service provider is
provider, number of improved
iterations
[103] Computational Public PoW Cross-Silo Federated Accuracy Rewards are
offloading learning propagated
accurately in secure
manner
[104] Classification of Public PoW Gradient boosted MSE, RMSE, MAE, Improved accuracy
transactions decision tree NRMSE
[105] Classification of Public PoW Random forest, Precision, recall, Cybercriminal
transactions extremely F1-score activities can be
randomized forest, detected
bagging and
gradient Boosting
[106] Data availability Ethereum PoW Similarity learning Gas cost for Flexibility in
deploying smart joining/leaving the
contract network, dispute
handling
[107] De-anonymizing Public PoW Logistic regression, Accuracy, F1-score, 95% of classification
Bitcoin entity type gradient boosting Precision accuracy is obtained
[108] Fraud detection in Public PoW Trimmed k-means Accuracy Improved fraud
Bitcoin detection rate
[109] Reducing anonymity Public PoW K-nearest neighbors, Accuracy, precision, Gradient boosting
of the network random forest, extra recall, F1 score performs best with
trees, AdaBoost, 77% accuracy
decision trees,
bagging classifier,
gradient boosting

54
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Table 10
Relative comparison of ML for blockchain based proposals.
Reference ML app. for Blockchain type Consensus algorithm Machine learning Evaluation metrics Advantages
blockchain algo.
[110] Anomaly detection Consortium PoW Supervised ML Research still in Prevents majority
progress attack
[111] Anomaly detection Public PoW k-means clustering, Dual evaluation Supports fraud
Mahalonobis detection in
distance based transaction system
method and SVM
[112] Anomaly detection Public PoW Laws of power Dual evaluation Supports fraud
degree and local detection in
outlier factors transaction system
[113] Anomaly detection Public (Ethereum) PoW Unsupervised ML False positive rate Data integrity
[114] Anomaly detection Permissioned – Gaussian graphical Negative sample Data security
probabilistic model accuracy, positive
sample accuracy
[115] Smart contract Public PoW CNN Accuracy Limited labor cost
security
[116] Smart contract Public PoW LSTM Accuracy, recall Maintains analysis
security score, precision time as smart
score contract grows in
complexity
[117] Data quality Public – – – Secure data sharing
classifier
[118] Classify blockchain Private PoW Bidirectional LSTM, Accuracy Neural network
peers LSTM, SVM, performed better
k-nearest neighbor, than conventional
decision tree method

Another perspective application of ML for blockchain is in of- offloading decisions, the major challenge is to achieve convergence and
floading approaches for the mobile blockchain networks. With the accuracy of deep NN. Also, there is a need to develop effective resource
introduction of mobile technology, the blockchain network can now allocation on mobile blockchain. To address this challenge, authors
be easily used with mobile devices so that more flexible blockchain of [102] designed a multilayer NN supported auction mechanism for re-
applications for IoT can be developed. However, with mobile systems, source allocation in mobile edge computing environment. The auction
resource-constrained IoT devices face difficulty while mining blocks. mechanism assures that edge resources are allocated to those miners
In this context, mobile edge computing facilitates high computational who value resources the most. Simulation results demonstrated that the
tasks for mobile devices. However, there is a challenge of effectively proposal converges quickly to a solution where the profit of the service
allocating available edge computing resources to miners. Mobile de- provider is higher than the proposal provided by the authors of [126].
vices can offload their high computational tasks to the assigned mobile Recently, Asheralieva and Niyato [90] proposed a Bayesian RL and
edge/cloud server. With a motive to enhance the performance of the DL based approach to make interactions among miners in blockchain
system, literature contains multiple offloading approaches. For exam- network with mobile edge computing. In particular, a game theory
ple, convex optimization model, and game theory approaches has been based approach is used by miners to offload its block operation to any
used by authors of [119–123] that minimizes task execution latency. of the base stations with mobile edge computing server. In contrast, the
Nevertheless, these methods fail for highly complex online models and authors in [103] have used federated learning to deal user equipment
also they demands prior knowledge about the system. To solve this privacy issues as edge node transactions are mostly based on centralized
issue, RL can be used where a learning agent is employed to derive approach. Federated learning builds ML models without centralizing
an optimal solution for computational offloading via trail-and-error the training data on a central server. Here federated learning facil-
method. Moreover, this solution does not require prior system statistics itates user equipment’s to train their data locally without exposing
knowledge. However, for high dimensional computational offloading the data for optimizing system model. In contrast, blockchain and
challenges, RL solution also gets fail due to high dimensions of state smart contract facility are used to secure transactions cross silo FL in
and action space as pointed by work in [124,125]. To deal with high B5G network. The tabulated comparison of above-discussed studies on
dimensional data, the use of DRL is beneficial and some literature work resource management and computational offloading has been presented
has demonstrated the scalability and offloading efficiency of DRL in in Table 6.
blockchain based edge computing applications. DRL can achieve an
optimal offloading strategy based on past experiences of offloading. 4.2.2. Predicting cryptocurrency price
Both of the proposals in [87,88] were designed to preserve users’ Bitcoin [127] introduced by Satoshi Nakamoto is the first world’s
privacy and to achieve security as an optimization problem. By using most popular cryptocurrency and is accepted by 111 countries world-
DRL method, performance metrics including computational latency, wide. As a valuable cryptocurrency, Bitcoin provides an opportunity
energy consumed, and privacy level were analyzed proving feasibility for price prediction because of its volatility and open nature [128].
of the proposed scheme with reduced offloading latency and minimum The price of Bitcoin was around $7202 in late 2019, compared to
energy consumption. about $3468 in January 2019 [129]. Researchers and stakeholders of
The above-discussed offloading approaches are designed only for the financial sector are trying to figure out the reason for changing
mining tasks whereas data processing tasks are ignored. In contrast, trends in the cryptography market. Similar to stock market prediction,
the work in [89] has discussed computational offloading for both Bitcoin price prediction can be represented as a model for time series
mining and data processing tasks combining DRL and genetic algo- prediction. However, conventional time series approaches are based on
rithms. Additionally, Markov decision process has been used to handle linear assumptions and are effective in the case of seasonal and noisy
the dynamic environment. However, to implement DRL method for data [130]. The absence of seasonality and the high volatile nature of

55
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Bitcoin blockchain network makes these traditional time series models Notably, previous literature work has focused on analyzing Bit-
unsuitable for Bitcoin price prediction. Nevertheless, for time-series coin prices without taking into account its non-linear relation with
prediction of uncertain data, some non-linear methods such as — ANN, blockchain variables. Further, the authors of [92] have concluded
Bayesian Neural Network (BNN), and SVM have gathered interest from that an ML model only trained with the Bitcoin price index results
researchers. Generally, ML based price prediction models have been in poor predictive performance. Differently, Barro’s Bitcoin pricing
evaluated on the following evaluation metrics: model [133] has been considered by authors for empirical study. In
Mean Absolute Error (MAE) this proposal, the blockchain variables such as — average block size,
transactions per block, median confirmation time, hash rate, difficulty,
1 ∑| ̂
𝑛
|
𝑀𝐴𝐸 = |𝜙 − 𝜙i | (1) miners revenue, and the number of confirmed transactions are used for
𝑛 𝑖=1 | i |
training of model that analyzes Bitcoin price by using BNN’s and results
Root Mean Square Error (RMSE) are compared with those obtained using Support Vector Regression
[ ]1 (SVR) and linear regression model. It is observed that both training and
1 ∑( ̂
𝑛
)2 2
testing phases show poor performance with SVR model. Notably, rather
𝑅𝑀𝑆𝐸 = 𝜙 − 𝜙i (2)
𝑛 𝑖=1 i than training model with only Bitcoin price index, BNN considers non-
linear effect of blockchain information and other macro economical
Mean Absolute Percentage Error (MAPE)
factors affecting the price of Bitcoin whereas regression model can only
𝑛 |̂ |
1 ∑ ||𝜙i − 𝜙i ||
handle linear relationships. Although as an advantage, the feature ex-
𝑀𝐴𝑃 𝐸 = (3) traction procedure of regression model removes incorrect values which
𝑛 𝑖=1 𝜙i
results in better prediction model. Similarly, Madan et al. [93] chose 26
Mean Square Error (MSE) features related to the Bitcoin network along with daily Bitcoin prices.
Some of these features include average confirmation time, block size,
1 ∑( ̂
𝑛
)2
𝑀𝑆𝐸 = 𝜙 − 𝜙i (4) difficulty, estimated transaction volume, and number of transactions,
𝑛 𝑖=1 i
etc. To predict the Bitcoin price, authors have leveraged SVM, random
where 𝜙̂ i = estimated value, 𝜙i = true objective value and 𝑛 = total forest, and binomial generalized linear model (GLM) algorithm and
number of data points. achieved prediction accuracy of around 97% without cross-validating
Relatively few studies have been conducted on estimating time- that however limits the generalizability of results. Results demonstrate
series of Bitcoin price using ML model. In this context, to deal with that the random forest algorithm performs best as it is based on the
uncertain and non-linear data, DL has been proved to be an effective non-parametric decision tree. However, the precision value for random
solution. For example, for the first time, the authors of [96] used DL forest is lower than that of binomial GLM as it also possesses the ca-
for price prediction of cryptocurrency. Other than Bitcoin, DL tech- pability to solve linearization problems for Bitcoin dataset. In addition,
niques are applied to predict the price of Ethereum, Ripple, and digital Greaves and AU [94] developed another Bitcoin price prediction model
cash cryptocurrency. For result analysis, the Long short-term memory by leveraging SVM and ANN and conclude that accuracy with ANN
(LSTM) model is compared with the generalized regression neural is best, i.e., 55%. Authors have used historical time delta of 1 h, 1
network model (GRNN). LSTM is a subtype of recurrent neural network day, 1 weak, and 1 month to develop features for supervised learning.
(RNN) and it is designed to deal with long-term dependency problems. Total Bitcoin passing through, net Bitcoin flow, number of transactions,
LSTM follows recurrent topology whereas GRNN has a parallel and and closeness centrality are the collected features for predicting price.
memory based system and attains fast learning with a large sample Simultaneously, they concluded that net Bitcoin flow, and number of
size. However, the prediction results of LSTM are better over GRNN transactions are the most informative Bitcoin features. Another effort to
for RMSE. Rather than just presenting a predictive model, the authors analyze features that highly relate with Bitcoin price change is carried
have also conducted a chaotic time series analysis. Similarly, Mcnally out in [95] by using linear regression, random forest and gradient de-
et al. [91] predict Bitcoin blockchain price using both LSTM, and RNN scent models. Here, authors have taken features from the dataset such
methods reporting price prediction accuracy of LSTM to be better than as — number of wallets, unspent transaction output, block size, and
RNN. Here, both NN models, i.e., RNN and LSTM are experimented with some others. The performance result of the proposal has been evaluated
two hidden layers having 20 nodes per layer. The dataset used for train- using RMSE and MAE. Likewise, Velankar et al. [134] predicted Bitcoin
ing purposes is considered from Aug. 19, 2013 to July 19, 2016. The price using Bayesian regression, and random forest method. Block size,
result section proves that RNN, LSTM, and Autoregressive Integrated total Bitcoins, day high, day low, number of transactions, and trade
Moving Average (ARIMA) all have almost similar accuracy, i.e., 50.25, volume are the set of selected parameters to be fed to the predictive
52.78, and 50.05 respectively. ARIMA model, however, implements network. On the same line of thought, Mangal et al. [97] experimented
time series data having linear nature. As Bitcoin data is volatile in with logistic regression, SVM, ARIMA, and RNN and concluded that
nature, so ARIMA cannot generate accurate results as compared to RNN has the most accuracy among all.
RNN and LSTM. Here, the DL models are trained with only considering The strengths and weaknesses of each ML methodology for price
Bitcoin price index. Likewise, the authors of [131] demonstrated the prediction is provided in Table 7 whereas the relative comparison of
impact of LSTM for Bitcoin price prediction by opting for 10 neurons the above discussed studies for price prediction has been presented in
in the hidden layer. In contrast, Jang and Lee [92] conducted a study on Table 8.
predicting Bitcoin price by using a BNN. BNN is based on the Bayesian Notably the existing literature on cryptocurrency price prediction
theory for neural networks. BNN’s have applications in various fields is not designed for the IoT environment. However, applications of IoT
such as — pattern recognization, Natural Language Processing (NLP), network includes payment transactions to be made between nodes. In
image recognization, traffic flow prediction, etc. [132]. Similar to a a blockchain based IoT network, payments are realized with digital
model of Multilayer Perceptron (MLP), a BNN consists of an input layer, cryptocurrency and hence the discussed studies on cryptocurrency price
an output layer, and one or multiple hidden layers. While training prediction could be applied to IoT networks as well.
model, backpropagation method updates the weight of neurons at each
layer with current error propagated backward by output layer to the 4.2.3. Anomaly detection/ attack prevention
previous layer. In addition to backpropagation method, delta rule is With the popularity of blockchain, the risk of security issues such
used to minimize the sum of errors. By utilizing the backpropagation as 51% attack (majority attack), double spending attack, etc. also
method, BNN can handle exclusive OR (XOR). Also, the regularization increases as discussed in [135,136]. Due to propagation delay in
term of BNN prevents overfitting problems in training data. blockchain network, the double spending attack might happen when a

56
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

participant tries to indulge in more than one transaction with the same Notably, an online ML security system that detects abnormal clients
number of cryptocoins. On the other hand, the majority attack happens in the network appears to be a topic that is understudied. To this end,
when more than 50% of the network’s participants control the network Bogner [113] proposed an online unsupervised ML method for fraud
and conspire to take control over the ledger. Moreover, the open nature detection that is optimized for interoperability. Different from other
and public design of Bitcoin system allow any user to be a participant. approaches, research of Bonger involves visualization techniques along
The goal of ML models is to learn insights, outliers, classify, and detect with an interactive querying system meant for manual expert analysis.
patterns in large data repository, so it can be used for blockchain attack The proposal is evaluated using public Ethereum blockchain network.
detection. On the other hand, authors in [115] focused on the security of
Moreover, with blockchain technology, ML algorithms can train, Ethereum smart contracts. As smart contracts are open in nature, any
learn, and can take decisions on local system in a decentralized net- vulnerability present in the contract is visible to anybody on the net-
work. Hence, processing data locally can prevent security and pri- work. For example, the decentralized autonomous organization (DAO)
vacy issues to some extent. Various authors have used ML models is a smart contract and due to some vulnerabilities in its code, it was
for anomaly detection in blockchain networks. Both supervised and hacked losing $150 million [139]. Here, in [115], authors have utilized
unsupervised ML algorithms have been employed to design intrusion CNN model for automatic feature extraction along with learning and
detection and prevention system. To detect isolate malicious in the detecting compiler bugs in smart contracts. They translated the byte-
network, various ML models are utilized by literature studies such as code of Solidity in RGB color code which is further transformed into a
— SVM, k-means clustering, etc. For example, Dey [110] has discussed fixed size encoded image. Next, the encoded image is fed to CNN for
the issue of majority attack in the blockchain network. Specifically, detecting bugs. In the similar direction, Tann et al. [116] utilized the
the majority attack is a concern in consortium blockchain (e.g., Hy- LSTM model to detect new attack trends for the smart contracts. LSTM
perledger) as it involves business parties collaboration. To solve the performs a two-class classification and reduces detection loss function
problem faced by majority attack, authors of [110] have proposed an to maximize classification accuracy, and to detect security threats in
approach based on supervised ML model and algorithmic game theory. smart contracts. Authors have leveraged the fact that smart contracts
Supervised ML algorithms are leveraged to classify whether the attack are sequential in nature, so, they can be easily used to update the LSTM
will take place or not. However, this work is still in progress, and model for future contracts.
simulation results or any proof have not been demonstrated by the
authors. In contrast to the supervised ML approach, another effort for 4.2.4. Reduction in anonymity of the network
detecting anomalies in Bitcoin network is made by Pham and Lee [112] Another potential application of ML for blockchain is to reduce the
using 3 unsupervised ML methods that include k-means clustering, anonymity of the network. Notably, blockchain network is assumed to
Mahalanobis distance based method, and SVM (on two Bitcoin trans- attain a high degree of anonymity as in blockchain each participant
action graphs). The dataset used for training includes 6,336769 users is referred by its public key address. However, the authors of [140]
with 37,450461 transactions and 12 features (including in-degree, out- claim that it is possible to cluster Bitcoin addresses and map them to
degree, average in-transactions, balance, etc.) are extracted. On the real-word identity. In the same context, Harlev et al. [109] conducted
other hand, the same authors in their research in [112] use laws of a study to probe the true depth of participant’s anonymity using a
power degree and local outlier factors on the two graphs produced by supervised ML approach. Firstly, the addresses are clustered where they
Bitcoin network to detect anomalies. In a closely related work, authors predict the category of yet unidentified Bitcoin addresses based on how
of [114] proposed an unsupervised statistical ML approach to detect addresses are controlled by a single entity using behavioral intelligence-
anomalies on blockchain based sensor data belonging to condition based clustering and co-spend clustering. Next, the identified clustered
management of the industrial asset. Following a trend rendered by the are categorized into one of the predefined categories, i.e., exchange,
adoption of unsupervised ML for anomaly detection, authors of [108] gambling, hosted wallet, merchant services, mining pool, mixing, ran-
used trimmed k-means clustering for cybercrime detection in Bitcoin somware, and scam, etc. The primary dataset used for simulation
network. Compared to other approaches on fraud detection, k-means includes transactional data which has details about each transaction.
clustering provides better results in terms of detection rate. Similarly, Here, seven different ML algorithms are used to analyze the transac-
Scicchitano et al. [137] proposed an anomaly detection system using tional data involving k-nearest neighbor, random forests, extra trees,
an unsupervised encoder decoder DL model which is trained with Adaboost, decision trees, gradient boosting, and bagging classifier. The
aggregated information extracted by analyzing blockchain network result section concludes that the gradient boosting method performs
activities. best among all. In contrast, Jourdan et al. [107] experimentally obtain
Besides, in order to prevent human trafficking and drug sale involv- lower value for parameter F1-score by using gradient boosting method.
ing Bitcoin, Portnoff et al. [138] proposed another ML based classifier Also, their methodology involves a complex step for hyper-parameter
that categories ads by the person paid for the ads. The ML classifier uti- optimization. In a closely related work, authors of [141] et al. presented
lizes stylometry that takes two ads as input and differentiates whether a method to break the Bitcoin anonymity concept via entity charac-
the ads are published by the same or different users. The flowchart terization. Here, the cascade of classifiers is used which first involves
for ML based anomaly detection in blockchain network is presented entity classification using address and motif’s and next step uses this
in Fig. 15. Firstly, the IoT data provider collects the data from IoT output for input of next classification step. Experiments are conducted
sensors and sends it for the data preparation phase which involves data and compared using Adaboost, Random forest, and gradient boost
preprocessing (transforming the dataset into machine readable format) models. However, there is a disadvantage that this approach is not
and feature extraction. Further, the data analysis phase is carried which able to characterize entities with normal user behavior. The proposal
involves training data with selected ML algorithms. Here, the weights is although able to detect six entity classes, i.e., Exchange, Gambling,
and biases are adjusted in order to get more correct predictions. Finally, Market, Mining Pool, Mixer, and Service. The general procedure of
the trained model is tested against never seen dataset for anomaly entity characterization process of Bitcoin is represented in Fig. 16.
detection. Differently, in the research [63], authors have leveraged
the concept of sharding to solve scalability issues. While implementing 4.2.5. Classification of blockchain related data
the concept of sharding, the blockchain network is divided into inter- Classification of data is very important for decision-making tasks
est groups and each group has its own ledger to verify transactions. [142]. Popular classification algorithms include k-nearest neighbor
Dividing the network improves network efficiency by empowering based methods, decision tree methods, NN based networks, multivari-
parallelism. Proof of Common Interest consensus algorithm is used ate discriminant analysis method, and SVM method [143]. ML classi-
to validate data that is directed to the relevant interest group. The fication methodology has been used with blockchain for data quality
proposal mitigates DDoS, MITM, and data leakage attacks. and transaction entity classification which is discussed as follows:

57
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Fig. 15. Flowchart for machine learning based anomaly detection in blockchain network.

Fig. 16. Entity characterization process of Bitcoin.

• Blockchain data quality classification: With the increase in the sometimes lead to slow degradation of the system. This personal
revolution of IoT technology, the usage of health specific ap- health data is secured with blockchain network by many re-
plications such as — smart bands, smart watches, etc. has also searchers [144]. Moreover, to check the validity of continuous
been increased. However, the presence of malicious nodes can and dynamic generated data by sensors, authors of [117] have

58
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

used ML techniques. Apart from previous roles in blockchain,

Summary and insights
another role named data validator has been introduced who
Section 4.2 reviews various ML applications for blockchain
is responsible for validating and certifying the quality of data
networks such as — to optimize resource allocation, to im-
generated from sensors. Here, the quality classifier for health data prove cryptocurrency price prediction, to detect anomalies
classified the input data with predefined features and removes of the network, and to classify blockchain related data. The
meaningless data and noise. As an example, take the case of increasing storage size on blockchain demands more resources.
smart watch readings for 24 h. The data validation algorithm With data sharding and pruning solutions, ML can help
can differentiate sleep related data from other workout data. blockchain networks in taking better decisions for data stor-
However, the predefined rules depending on the choice of owner age. Also, ML in blockchain networks can identify malicious
decide whether to classify sleep related data as high quality or activities by developing and training ML models. However,
noise. Fig. 17 presents an illustration of ML based blockchain data it has been observed that learning based examination of the
quality validation process in a healthcare network. The serving blockchain systems has not been exploited much in literature
data is validated before it is piped to the blockchain network. work. Moreover, for protecting wallet privacy, the applications
The data analyzer is responsible for computing predefined set of of clustering techniques that are proposed to address a broad
statistics sufficient to define the data. ML agent works at data val- range of blockchain applications have not utilized any string
idator module that trains the system using schema and constrains search mechanism such as — bloom filter. By using the string
API’s. The system can also classify the data into categories using search mechanisms, storage complexity and searching time
ML classifiers. In another work, authors of [145] utilized ML to complexity for various validation and verification operations
can be reduced significantly. Also, the proposed clustering pro-
analyze data of a blockchain based credit card scoring system. The
posals can be extended to increase the relatively low sample
blockchain based transaction data is sent to ML agent that extracts
size of clusters along with adding more cluster categories to
features from data and then applies binary classification model
differentiate effectively between clusters.
to categorize customers that would not be able to the requisite
amount in a destined amount of time. These classification results
are next considered by the bank to decide whether credit request
5. Technical challenges in integration of ML and blockchain for
is to be initiated for a particular customer or not.
IoT
Another matter to be considered in the context of centralization
is trading IoT generated data with a TTP. To this end, the authors
While the previous sections have presented a study on the integra-
of [106] presented a data trading model utilizing blockchain,
tion of blockchain and ML, this section discusses challenges that need
smart contract, and similarity learning. Here, arbitration insti- to be considered for future research.
tution having responsibility of maintaining smart contract uses
ML services to solve any dispute over availability of data for • Confidentiality is still not fully preserved with blockchain as any
data purchasers utilizing classification and clustering data solu- node can trace transactions and it is observed that only a few re-
tions. In particular, similarity learning (distance metric learning) search studies have focused on the lack of confidentiality feature
is adopted to validate the distance between features of actual of blockchain for ML data. Moreover, blockchain standards and
regulations are yet to be finalized.
data and declared data. Distance metric learning has been used
• Here, it is worth mentioning the problem of data storage as nodes
extensively for classification and clustering problems.
of the blockchain network keep the copy of every transaction of
• Classification of blockchain peers/ transaction entities: Public
the network. This increasing database size could be difficult to
blockchains are open and can be joined by anybody in the net-
handle in future. Hence, the issue of scalability for blockchain
work. In such a case, there is a possibility that some participants platforms should be focused to popularize the applications of
may misbehave for personal interest while the majority of the par- blockchain for ML. As a solution, the usage of emerging mecha-
ticipants behave legally. Clearly, it is hard to study the behavior nisms such as — sidechains or childchains should be encouraged
of participants manually. To address this problem, the authors in research. Moreover, PoW computations can prove costly in
of [118] present an approach to classify behavior patterns of terms of resource utilization and transaction time. So, models
participants into predefined categories by using LSTM based DL should be developed that do not consume unnecessary compu-
approach. The transaction amount is extracted as a feature to clas- tational power. Also, the existing ML models demand creation
sify participants. Based on the transaction amount, participants of custom datasets having specific variables. Moreover, these are
are classified into three categories, i.e., stable transaction history, not able to satisfy the various service requirements of complex
medium jitter history, and high jitter transaction history. With networks. Hence, it is challenging to scale for development of ML
a similar motive, the authors of [104] classified the entities of models with the ever increasing IoT data.
transaction in 4 categories, i.e., exchange, service, gambling, and • It is also observed from the tabulated comparison of available
mining pool. Here, gradient boosted decision tree algorithm with literature that most of the research work is based on permissioned
blockchain. However, the vulnerability of 51% attack is easy to
a Gaussian process based optimization is used as a classification
launch in permissioned blockchain to which none of the studies
method. The results concluded that the accuracy in the classifica-
have considered. Also, the use of permissioned blockchain limits
tion of categories exchange, gambling, and service categorization
the access of an enormous amount of data that can be required
is high as compared to the mining pool category. Additionally,
for ML system to process accurate decision-making. To address
authors of [105] presented a supervised ML approach in order to this problem, blockchain platforms and IoT resources should be
classify entities of transactions engaged in cybercriminal activity. equipped with a Trusted Execution Environment (TEE) [146].
To train the classification model 854 categorical observations • Federated learning adopted by many researchers faces the issue
with 12 classes and 10 000 non-categorical identifiers are con- of communication bandwidth. Undoubtedly, the mobile device
sidered. It is concluded in results that random forest, extremely has enough computing resources in order to implement federated
randomized forest, bagging, and gradient boosting are the best learning. Unfortunately, the bandwidth of wireless communi-
four classifiers. cation is not adequate. So, research has been shifting gradu-
ally towards computational resources to wireless communication.

59
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

Fig. 17. An illustration of ML based blockchain data quality validation.

Methods such as — deep gradient compression should be used to • A large number of research options are available for future re-
decrease the communication bandwidth [147]. search related to designing of application-level consensus proto-
• Also, in public blockchain data is publicly available and accessible cols that consider proofs with quality of learning models, space
for all readers which is indeed a privacy concern. However, and time efficient search strategies, and quality of optimiza-
using private blockchain can limit the exposure of large amount tion [30].
of data which is obviously necessary for ML model to perform
accurate decision-making. Along with privacy, security is another 6. Conclusion
concerning factor as this technology suffers attacks in the applica-
tion layer. Also, the consensus mechanisms can be compromised In this paper, we reviewed the current state-of-art related to the
collaboration of ML and blockchain. We presented an overview of
depending on the hashing power of the miner. Nevertheless, ML
blockchain technology and how this decentralized technology can solve
algorithms provide detection of various attacks in blockchain
the privacy issues related to ML. Moreover, we provide an overview
networks but still, there are challenges for using ML algorithms
of ML technology and discussed key applications, applicability of
in detecting malicious threats. For example, for a large dataset
blockchain features for ML. The literature review shows that blockchain
having malicious data the security solution for detecting ma-
and ML collaborated applications are still in infancy and there are
licious behavior has to deal with high dimensionality of data
many research challenges that need to be addressed. However, the
for pre-processing. In such case, ML model has to first perform
current research is a foundation for an interdisciplinary perspective.
dimensionality reduction step. Moreover, it is impossible to train In the future, we will implement one of these techniques in future IoT
a ML model with a large dataset in real-time so it is challenging applications to check its performance with respect to other applications
to detect online attacks in dynamic networks. using various performance evaluation metrics.
• 5G and B5G are an example of a heterogeneous network designed
for a wide range of IoT devices. The enormous amount of data Declaration of competing interest
generated from these devices can put heavy weight on ML model
for decision making leading to limited performance. In this con- The authors declare that they have no known competing finan-
text, blockchain can solve the security issues to some extent but cial interests or personal relationships that could have appeared to
network performance at issue will still be a problem. influence the work reported in this paper.

60
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

References [25] S. Aggarwal, R. Chaudhary, G.S. Aujla, N. Kumar, K.-K.R. Choo, A.Y. Zomaya,
Blockchain for smart communities: Applications, challenges and opportunities,
[1] F. Khan, A. ur Rehman, Z. Yanliang, S. Mastorakis, H. Song, M.A. Jan, K. Dev, A J. Netw. Comput. Appl. (2019).
secured and reliable continuous transmission scheme in cognitive HARQ-aided [26] S. Durga, R. Nag, E. Daniel, Survey on machine learning and deep learning
internet of things, IEEE Internet Things J. (2021). algorithms used in internet of things (IoT) healthcare, in: 2019 3rd International
[2] Open energy information, 2016, Available: http://en.openei.org/datasets/ Conference on Computing Methodologies and Communication, ICCMC, 2019,
dataset/commercial-and-residential-hourly-load-profiles-for-all-tmy3-locations- pp. 1018–1022, http://dx.doi.org/10.1109/ICCMC.2019.8819806.
in-the-united-states, (Accessed: Jul. 2016). [27] F. Hussain, S.A. Hassan, R. Hussain, E. Hossain, Machine learning for resource
[3] Q. Feng, D. He, S. Zeadally, M.K. Khan, N. Kumar, A survey on privacy management in cellular and IoT networks: Potentials, current solutions, and
protection in blockchain system, J. Netw. Comput. Appl. 126 (2019) 45–58. open challenges, IEEE Commun. Surv. Tutor. (2020) 1, http://dx.doi.org/10.
[4] P.K. Sharma, N. Kumar, J.H. Park, Blockchain-based distributed framework 1109/COMST.2020.2964534.
for automotive industry in a smart city, IEEE Trans. Ind. Inf. 15 (7) (2018) [28] N. Majumdar, S. Shukla, A. Bhatnagar, Survey on applications of internet
4197–4205. of things using machine learning, in: 2019 9th International Conference on
[5] Imachine learning market research report- global forcast to 2024, 2020, Cloud Computing, Data Science Engineering, Confluence, 2019, pp. 562–566,
Available: https://www.marketresearchfuture.com/reports/machine-learning- http://dx.doi.org/10.1109/CONFLUENCE.2019.8776951.
market-2494/, (Accessed: Jan. 2020). [29] T.N. Dinh, M.T. Thai, Ai and blockchain: A disruptive integration, Computer
[6] F. Zantalis, G. Koulouras, S. Karabetsos, D. Kandris, A review of machine 51 (9) (2018) 48–53.
learning and IoT in smart transportation, Future Internet 11 (4) (2019) 94. [30] K. Salah, M.H.U. Rehman, N. Nizamuddin, A. Al-Fuqaha, Blockchain for AI:
[7] K. Awahara, S. Izumi, T. Abe, T. Suganuma, Autonomous control method using Review and open research challenges, IEEE Access 7 (2019) 10127–10149,
AI planning for energy-efficient network systems, in: 2013 Eighth Interna- http://dx.doi.org/10.1109/ACCESS.2018.2890507.
tional Conference on Broadband and Wireless Computing, Communication and [31] S. Vyas, M. Gupta, R. Yadav, Converging blockchain and machine learning for
Applications, IEEE, 2013, pp. 628–633. healthcare, in: 2019 Amity International Conference on Artificial Intelligence,
[8] W. Ejaz, M. Naeem, A. Shahid, A. Anpalagan, M. Jo, Efficient energy man- AICAI, 2019, pp. 709–711, http://dx.doi.org/10.1109/AICAI.2019.8701230.
agement for the internet of things in smart cities, IEEE Commun. Mag. 55 (1) [32] F.A. Acheampong, Big data, machine learning and the blockchain technology:
(2017) 84–91. An overview, Int. J. Comput. Appl. 975 (2018) 8887.
[9] Beyond the hype: the real impact of blockchain on economy, 2021, [33] M. Rahouti, K. Xiong, N. Ghani, Bitcoin concepts, threats, and machine-learning
Available: https://appinventiv.com/blog/real-impact-of-blockchain-technology- security solutions, IEEE Access 6 (2018) 67189–67205.
on-economy/, (Accessed: May 2021). [34] D.C. Nguyen, P.N. Pathirana, M. Ding, A. Seneviratne, Integration of blockchain
[10] Industrial & iot blockchain market report 2019-2023, 2020, Available: and cloud of things: Architecture, applications and challenges, 2019, arXiv
https://iot-analytics.com/sample-industrial-iot-blockchain-market-report-2019- preprint arXiv:1908.09058.
2023/, (Accessed: Jan. 2020). [35] P.V. Rane, S.N. Dhage, Systematic erudition of bitcoin price prediction us-
[11] M.A. Togou, T. Bi, K. Dev, K. McDonnell, A. Milenovic, H. Tewari, G.-M. ing machine learning techniques, in: 2019 5th International Conference on
Muntean, DBNS: A distributed blockchain-enabled network slicing framework Advanced Computing & Communication Systems, ICACCS, IEEE, 2019, pp.
for 5G networks, IEEE Commun. Mag. 58 (11) (2020) 90–96. 594–598.
[12] M.A. Togou, T. Bi, K. Dev, K. McDonnell, A. Milenovic, H. Tewari, G.-M. [36] Y. Liu, F.R. Yu, X. Li, H. Ji, V.C. Leung, Blockchain and machine learning for
Muntean, A distributed blockchain-based broker for efficient resource provi- communications and networking systems, IEEE Commun. Surv. Tutor. 22 (2)
sioning in 5g networks, in: 2020 International Wireless Communications and (2020) 1392–1431.
Mobile Computing, IWCMC, IEEE, 2020, pp. 1485–1490. [37] S. Nakamoto, A. Bitcoin, A peer-to-peer electronic cash system, 2008,
[13] A. Kumari, S. Tanwar, S. Tyagi, N. Kumar, R.M. Parizi, K.-K.R. Choo, Fog data Bitcoin.–URL: https://Bitcoin.Org/Bitcoin.Pdf.
analytics: A taxonomy and process model, J. Netw. Comput. Appl. 128 (2019) [38] Blockchain - statistics & facts, 2019, Available: https://www.statista.com/
90–104. topics/5122/blockchain/, (Accessed: Jan, 2019).
[14] R.L. Kumar, Q.-V. Pham, F. Khan, M.J. Piran, K. Dev, Blockchain for securing [39] R. Chaudhary, A. Jindal, G.S. Aujla, S. Aggarwal, N. Kumar, K.-K.R. Choo,
aerial communications: Potentials, solutions, and research directions, Phys. BEST: Blockchain-based secure energy trading in SDN-enabled intelligent
Commun. (2021) 101390. transportation system, Comput. Secur. 85 (2019) 288–299.
[15] F. Chen, H. Wan, H. Cai, G. Cheng, Machine learning in/for blockchain: Future [40] A. Jindal, G.S. Aujla, N. Kumar, SURVIVOR: A blockchain based edge-as-a-
and challenges, 2019, arXiv preprint arXiv:1909.06189. service framework for secure energy trading in SDN-enabled vehicle-to-grid
[16] M. Mamdouh, M.A. I. Elrukhsi, A. Khattab, Securing the internet of things and environment, Comput. Netw. 153 (2019) 36–48.
wireless sensor networks via machine learning: A survey, in: 2018 International [41] G. Bansal, A. Dua, G.S. Aujla, M. Singh, N. Kumar, Smartchain: A smart
Conference on Computer and Applications, ICCA, 2018, pp. 215–218, http: and scalable blockchain consortium for smart grid systems, in: 2019 IEEE
//dx.doi.org/10.1109/COMAPP.2018.8460440. International Conference on Communications Workshops, ICC Workshops, 2019,
[17] S. Athmaja, M. Hanumanthappa, V. Kavitha, A survey of machine learning algo- pp. 1–6, http://dx.doi.org/10.1109/ICCW.2019.8757069.
rithms for big data analytics, in: 2017 International Conference on Innovations [42] Solidity, 2021, Available: https://docs.soliditylang.org/en/develop/, (Accessed:
in Information, Embedded and Communication Systems, ICIIECS, 2017, pp. 1–4, May. 2021).
http://dx.doi.org/10.1109/ICIIECS.2017.8276028. [43] An introduction to serpent, 2021, Available: https://www.cs.cmu.edu/~music/
[18] J. Qiu, Q. Wu, G. Ding, Y. Xu, S. Feng, A survey of machine learning for big serpent/doc/serpent.htm, (Accessed: May. 2021).
data processing, EURASIP J. Adv. Signal Process. 2016 (1) (2016) 67. [44] Vyper, 2021, Available: https://vyper.readthedocs.io/en/latest/, (Accessed:
[19] Q.E. Abbas, J. Sung-Bong, A survey of blockchain and its applications, in: May. 2021).
2019 International Conference on Artificial Intelligence in Information and [45] D. He, M. Ma, S. Zeadally, N. Kumar, K. Liang, Certificateless public key
Communication, ICAIIC, 2019, pp. 001–003, http://dx.doi.org/10.1109/ICAIIC. authenticated encryption with keyword search for industrial internet of things,
2019.8669067. IEEE Trans. Ind. Inf. 14 (8) (2017) 3618–3627.
[20] T.T.A. Dinh, R. Liu, M. Zhang, G. Chen, B.C. Ooi, J. Wang, Untangling [46] T. Le, M.W. Mutka, A lightweight block validation method for resource-
blockchain: A data processing view of blockchain systems, IEEE Trans. Knowl. constrained iot devices in blockchain-based applications, in: 2019 IEEE 20th
Data Eng. 30 (7) (2018) 1366–1385, http://dx.doi.org/10.1109/TKDE.2017. International Symposium on‘‘ a World of Wireless, Mobile and Multimedia
2781227. Networks’’, WoWMoM, IEEE, 2019, pp. 1–9.
[21] T.T. Huynh, T.D. Nguyen, H. Tan, A survey on security and privacy issues of [47] M. Pincheira, M. Vecchio, Towards trusted data on decentralized IoT applica-
blockchain technology, in: 2019 International Conference on System Science tions: Integrating blockchain in constrained devices, in: 2020 IEEE International
and Engineering, ICSSE, 2019, pp. 362–367, http://dx.doi.org/10.1109/ICSSE. Conference on Communications Workshops, ICC Workshops, IEEE, 2020, pp.
2019.8823094. 1–6.
[22] S. Cho, S. Lee, Survey on the application of blockchain to IoT, in: 2019 [48] C. Sun, A. Shrivastava, S. Singh, A. Gupta, Revisiting unreasonable effective-
International Conference on Electronics, Information, and Communication, ness of data in deep learning era, in: Proceedings of the IEEE International
ICEIC, 2019, pp. 1–2, http://dx.doi.org/10.23919/ELINFOCOM.2019.8706369. Conference on Computer Vision, 2017, pp. 843–852.
[23] J. Xie, H. Tang, T. Huang, F.R. Yu, R. Xie, J. Liu, Y. Liu, A survey of [49] A. Miglani, N. Kumar, Deep learning models for traffic flow prediction in
blockchain technology applied to smart cities: Research issues and challenges, autonomous vehicles: A review, solutions, and challenges, Veh. Commun. 20
IEEE Commun. Surv. Tutor. 21 (3) (2019) 2794–2830, http://dx.doi.org/10. (2019) 100184.
1109/COMST.2019.2899617. [50] N. Kumar, N. Chilamkurti, J.H. Park, ALCA: agent learning–based clustering
[24] F. Alkurdi, I. Elgendi, K.S. Munasinghe, D. Sharma, A. Jamalipour, Blockchain algorithm in vehicular ad hoc networks, Pers. Ubiquitous Comput. 17 (8) (2013)
in IoT security: A survey, in: 2018 28th International Telecommunication 1683–1692.
Networks and Applications Conference, ITNAC, 2018, pp. 1–4, http://dx.doi. [51] H. Bae, J. Jang, D. Jung, H. Jang, H. Ha, H. Lee, S. Yoon, Security and privacy
org/10.1109/ATNAC.2018.8615409. issues in deep learning, 2018, arXiv preprint arXiv:1807.11655.

61
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

[52] R. Gupta, S. Tanwar, S. Tyagi, N. Kumar, M.S. Obaidat, B. Sadoun, Habits: [75] D. Bertsekas, Tsitsiklis, Parallel and Distributed Computation, Vol. 3, Prentice
Blockchain-based telesurgery framework for healthcare 4.0, in: 2019 Interna- Hall, 1989, pp. 191–233.
tional Conference on Computer, Information and Telecommunication Systems, [76] Y. Lu, Q. Tang, G. Wang, On enabling machine learning tasks atop public
CITS, IEEE, 2019, pp. 1–5. blockchains: A crowdsourcing approach, in: 2018 IEEE International Conference
[53] F.K. Došilović, M. Brčić, N. Hlupić, Explainable artificial intelligence: A on Data Mining Workshops, ICDMW, 2018, pp. 81–88, http://dx.doi.org/10.
survey, in: 2018 41st International Convention on Information and Commu- 1109/ICDMW.2018.00019.
nication Technology, Electronics and Microelectronics, MIPRO, IEEE, 2018, pp. [77] D. He, N. Kumar, J.-H. Lee, Privacy-preserving data aggregation scheme against
0210–0215. internal attackers in smart grids, Wirel. Netw. 22 (2) (2016) 491–502.
[54] W. Samek, K.-R. Müller, Towards explainable artificial intelligence, in: Explain- [78] Q. Wang, S. Hu, M. Du, J. Wang, K. Ren, Learning privately: Privacy-preserving
able AI: Interpreting, Explaining and Visualizing Deep Learning, Springer, 2019, canonical correlation analysis for cross-media retrieval, in: IEEE INFOCOM
pp. 5–22. 2017-IEEE Conference on Computer Communications, IEEE, 2017, pp. 1–9.
[55] A.B. Arrieta, N. Díaz-Rodríguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, [79] Y. Rahulamathavan, R.C.-W. Phan, S. Veluru, K. Cumanan, M. Rajarajan,
S. García, S. Gil-López, D. Molina, R. Benjamins, et al., Explainable artificial Privacy-preserving multi-class support vector machine for outsourcing the data
intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward classification in cloud, IEEE Trans. Dependable Secure Comput. 11 (5) (2013)
responsible AI, Inf. Fusion 58 (2020) 82–115. 467–479.
[56] A.B. Kurtulmus, K. Daniel, Trustless machine learning contracts; evaluating and
[80] H. Kim, S.-H. Kim, J.Y. Hwang, C. Seo, Efficient privacy-preserving machine
exchanging machine learning models on the ethereum blockchain, 2018, arXiv
learning for blockchain network, IEEE Access 7 (2019) 136481–136495.
preprint arXiv:1802.10185.
[81] J. Zhou, F. Tang, H. Zhu, N. Nan, Z. Zhou, Distributed data vend-
[57] R.K. Raman, R. Vaculin, M. Hind, S.L. Remy, E.K. Pissadaki, N.K. Bore, R.
ing on blockchain, in: 2018 IEEE International Conference on Internet of
Daneshvar, B. Srivastava, K.R. Varshney, A scalable blockchain approach for
Things (IThings) and IEEE Green Computing and Communications (Green-
trusted computation and verifiable simulation in multi-party collaborations, in:
Com) and IEEE Cyber, Physical and Social Computing (CPSCom) and IEEE
2019 IEEE International Conference on Blockchain and Cryptocurrency, ICBC,
Smart Data, SmartData, 2018, pp. 1100–1107, http://dx.doi.org/10.1109/
IEEE, 2019, pp. 277–284.
Cybermatics_2018.2018.00201.
[58] F. Bravo-Marquez, S. Reeves, M. Ugarte, Proof-of-learning: A blockchain
[82] M. Hanley, H. Tewari, Managing lifetime healthcare data on the
consensus mechanism based on machine learning competitions, in: 2019
blockchain, in: 2018 IEEE SmartWorld, Ubiquitous Intelligence Computing,
IEEE International Conference on Decentralized Applications and Infrastruc-
Advanced Trusted Computing, Scalable Computing Communications, Cloud
tures, DAPPCON, 2019, pp. 119–124, http://dx.doi.org/10.1109/DAPPCON.
Big Data Computing, Internet of People and Smart City Innovation
2019.00023.
(SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), 2018, pp. 246–251,
[59] T.-T. Kuo, L. Ohno-Machado, Modelchain: Decentralized privacy-preserving
http://dx.doi.org/10.1109/SmartWorld.2018.00077.
healthcare predictive modeling framework on private blockchain networks,
2018, arXiv preprint arXiv:1802.01746. [83] M. Van Dijk, C. Gentry, S. Halevi, V. Vaikuntanathan, Fully homomorphic
[60] A. Juneja, M. Marefat, Leveraging blockchain for retraining deep learning encryption over the integers, in: Annual International Conference on the Theory
architecture in patient-specific arrhythmia classification, in: 2018 IEEE EMBS and Applications of Cryptographic Techniques, Springer, 2010, pp. 24–43.
International Conference on Biomedical Health Informatics, BHI, 2018, pp. [84] C. Xu, K. Wang, M. Guo, Intelligent resource management in blockchain-based
393–397, http://dx.doi.org/10.1109/BHI.2018.8333451. cloud datacenters, IEEE Cloud Comput. 4 (6) (2017) 50–59, http://dx.doi.org/
[61] K. Singla, J. Bose, S. Katariya, Machine learning for secure device person- 10.1109/MCC.2018.1081060.
alization using blockchain, in: 2018 International Conference on Advances [85] Y. Dai, D. Xu, S. Maharjan, Z. Chen, Q. He, Y. Zhang, Blockchain and deep
in Computing, Communications and Informatics, ICACCI, 2018, pp. 67–73, reinforcement learning empowered intelligent 5G beyond, IEEE Netw. 33 (3)
http://dx.doi.org/10.1109/ICACCI.2018.8554476. (2019) 10–17, http://dx.doi.org/10.1109/MNET.2019.1800376.
[62] T. Wang, A unified analytical framework for trustable machine learning and [86] M. Liu, Y. Teng, F.R. Yu, V.C.M. Leung, M. Song, Deep reinforcement learning
automation running with blockchain, in: 2018 IEEE International Conference on based performance optimization in blockchain-enabled internet of vehicle, in:
Big Data, Big Data, 2018, pp. 4974–4983, http://dx.doi.org/10.1109/BigData. ICC 2019 - 2019 IEEE International Conference on Communications, ICC, 2019,
2018.8622262. pp. 1–6, http://dx.doi.org/10.1109/ICC.2019.8761206.
[63] R. Doku, D.B. Rawat, C. Liu, Towards federated learning approach to determine [87] D.C. Nguyen, P.N. Pathirana, M. Ding, A. Seneviratne, Privacy-preserved task
data relevance in big data, in: 2019 IEEE 20th International Conference on offloading in mobile blockchain with deep reinforcement learning, 2019, arXiv
Information Reuse and Integration for Data Science, IRI, 2019, pp. 184–192, preprint arXiv:1908.07467.
http://dx.doi.org/10.1109/IRI.2019.00039. [88] D.C. Nguyen, P.N. Pathirana, M. Ding, A. Seneviratne, Secure computation
[64] M. Shen, X. Tang, L. Zhu, X. Du, M. Guizani, Privacy-preserving support vector offloading in blockchain based iot networks with deep reinforcement learning,
machine training over blockchain-based encrypted IoT data in smart cities, IEEE 2019, arXiv preprint arXiv:1908.07466.
Internet Things J. 6 (5) (2019) 7702–7712, http://dx.doi.org/10.1109/JIOT. [89] X. Qiu, L. Liu, W. Chen, Z. Hong, Z. Zheng, Online deep reinforcement learning
2019.2901840. for computation offloading in blockchain-empowered mobile edge computing,
[65] X. Chen, J. Ji, C. Luo, W. Liao, P. Li, When machine learning meets IEEE Trans. Veh. Technol. 68 (8) (2019) 8050–8062, http://dx.doi.org/10.
blockchain: A decentralized, privacy-preserving and secure design, in: 2018 1109/TVT.2019.2924015.
IEEE International Conference on Big Data, Big Data, IEEE, 2018, pp. [90] A. Asheralieva, D. Niyato, Bayesian Reinforcement learning and Bayesian deep
1178–1187. learning for blockchains with mobile edge computing, IEEE Trans. Cogn.
[66] Y. Lu, X. Huang, Y. Dai, S. Maharjan, Y. Zhang, Blockchain and federated Commun. Netw. (2020).
learning for privacy-preserved data sharing in industrial IoT, IEEE Trans. Ind. [91] S. McNally, J. Roche, S. Caton, Predicting the price of bitcoin using ma-
Inf. (2019) 1, http://dx.doi.org/10.1109/TII.2019.2942190. chine learning, in: 2018 26th Euromicro International Conference on Parallel,
[67] X. Zhu, H. Li, Y. Yu, Blockchain-based privacy preserving deep learning, in: Distributed and Network-Based Processing, PDP, IEEE, 2018, pp. 339–343.
International Conference on Information Security and Cryptology, Springer,
[92] H. Jang, J. Lee, An empirical study on modeling and prediction of bitcoin prices
2018, pp. 370–383.
with Bayesian neural networks based on blockchain information, IEEE Access
[68] A. Goel, A. Agarwal, M. Vatsa, R. Singh, N. Ratha, DeepRing: Protecting deep
6 (2018) 5427–5437, http://dx.doi.org/10.1109/ACCESS.2017.2779181.
neural network with blockchain, in: Proceedings of the IEEE Conference on
[93] I. Madan, S. Saluja, A. Zhao, Automated bitcoin trading via machine learn-
Computer Vision and Pattern Recognition Workshops, 2019.
ing algorithms, 20, 2015, URL: http://Cs229.Stanford.Edu/Proj2014/Isaac%
[69] H. Kim, S. Kim, J.Y. Hwang, C. Seo, Efficient privacy-preserving machine
20Madan.
learning for blockchain network, IEEE Access 7 (2019) 136481–136495, http:
[94] A. Greaves, B. Au, Using the bitcoin transaction graph to predict the price of
//dx.doi.org/10.1109/ACCESS.2019.2940052.
bitcoin, No Data (2015).
[70] S. Zhou, H. Huang, W. Chen, P. Zhou, Z. Zheng, S. Guo, Pirate: A blockchain-
based secure framework of distributed machine learning in 5g networks, IEEE [95] M. Saad, A. Mohaisen, Towards characterizing blockchain-based cryptocurren-
Netw. 34 (6) (2020) 84–91. cies for highly-accurate predictions, in: IEEE INFOCOM 2018 - IEEE Conference
[71] H. Kim, J. Park, M. Bennis, S.-L. Kim, Blockchained on-device federated on Computer Communications Workshops, INFOCOM WKSHPS, 2018, pp.
learning, IEEE Commun. Lett. (2019). 704–709, http://dx.doi.org/10.1109/INFCOMW.2018.8406859.
[72] U. Majeed, C.S. Hong, FLchain: Federated learning via MEC-enabled blockchain [96] S. Lahmiri, S. Bekiros, Cryptocurrency forecasting with deep learning chaotic
network, in: 2019 20th Asia-Pacific Network Operations and Management neural networks, Chaos Solitons Fractals 118 (2019) 35–40.
Symposium, APNOMS, IEEE, 2019, pp. 1–4. [97] N. Mangla, A. Bhat, G. Avabratha, N. Bhat, Bitcoin price prediction using
[73] Q. Jiang, N. Kumar, J. Ma, J. Shen, D. He, N. Chilamkurti, A privacy-aware two- machine learning.
factor authentication protocol based on elliptic curve cryptography for wireless [98] A. Pieroni, N. Scarpato, L. Di Nunzio, F. Fallucchi, M. Raso, Smarter city: smart
sensor networks, Int. J. Netw. Manage. 27 (3) (2017) e1937. energy grid based on blockchain technology, Int. J. Adv. Sci. Eng. Inf. Technol.
[74] D. He, N. Kumar, N. Chilamkurti, J.-H. Lee, Lightweight ECC based RFID 8 (1) (2018) 298–306.
authentication integrated with an iD verifier transfer protocol, J. Med. Syst. [99] M. Swan, Blockchain thinking: The brain as a dac (decentralized autonomous
38 (10) (2014) 1–6. organization), in: Texas Bitcoin Conference, Chicago, 2015, pp. 27–29.

62
A. Miglani and N. Kumar Computer Communications 178 (2021) 37–63

[100] A. Fadaeddini, B. Majidi, M. Eshghi, Secure decentralized peer-to-peer train- [122] W. Chen, Z. Zhang, Z. Hong, C. Chen, J. Wu, S. Maharjan, Z. Zheng, Y. Zhang,
ing of deep neural networks based on distributed ledger technology, J. Cooperative and distributed computation offloading for blockchain-empowered
Supercomput. (2020) 1–15. industrial internet of things, IEEE Internet Things J. 6 (5) (2019) 8433–8446,
[101] J.A. Dias, J.J. Rodrigues, N. Kumar, V. Korotaev, G. Han, REMA: A resource http://dx.doi.org/10.1109/JIOT.2019.2918296.
management tool to improve the performance of vehicular delay-tolerant [123] N. Kumar, S. Misra, J.J. Rodrigues, M.S. Obaidat, Coalition games for spatio-
networks, Veh. Commun. 9 (2017) 135–143. temporal big data in Internet of Vehicles environment: A comparative analysis,
[102] N.C. Luong, Z. Xiong, P. Wang, D. Niyato, Optimal auction for edge com- IEEE Internet Things J. 2 (4) (2015) 310–320.
puting resource management in mobile blockchain networks: A deep learning [124] R.S. Sutton, A.G. Barto, Reinforcement Learning: An Introduction, MIT Press,
approach, in: 2018 IEEE International Conference on Communications, ICC, Cambridge, MA, 2011.
2018, pp. 1–6, http://dx.doi.org/10.1109/ICC.2018.8422743. [125] K. Arulkumaran, M.P. Deisenroth, M. Brundage, A.A. Bharath, Deep rein-
[103] S. Rahmadika, M. Firdaus, S. Jang, K.-H. Rhee, Blockchain-enabled 5G edge forcement learning: A brief survey, IEEE Signal Process. Mag. 34 (6) (2017)
26–38.
networks and beyond: An intelligent cross-silo federated learning approach,
[126] W. Vickrey, Counterspeculation, auctions, and competitive sealed tenders, J.
Secur. Commun. Netw. 2021 (2021).
Finance 16 (1) (1961) 8–37.
[104] M. Jourdan, S. Blandin, L. Wynter, P. Deshpande, A probabilistic model of the
[127] S. Nakamoto, et al., Bitcoin: A peer-to-peer electronic cash system, in: Working
bitcoin blockchain, in: Proceedings of the IEEE Conference on Computer Vision
Paper, 2008.
and Pattern Recognition Workshops, 2019.
[128] M. Briere, K. Oosterlinck, A. Szafarz, Virtual currency, tangible return: Portfolio
[105] H.S. Yin, R. Vatrapu, A first estimation of the proportion of cybercriminal enti-
diversification with bitcoin, J. Asset Manage. 16 (6) (2015) 365–373.
ties in the bitcoin ecosystem using supervised machine learning, in: 2017 IEEE
[129] Bitcoin (btc) and united states dollar (usd) year 2019 exchange rate history,
International Conference on Big Data, Big Data, IEEE, 2017, pp. 3690–3699. 2019, Available: https://freecurrencyrates.com/en/exchange-rate-history/BTC-
[106] W. Xiong, L. Xiong, Smart contract based data trading mode using blockchain USD/2019, (Accessed: Dec. 2019).
and machine learning, IEEE Access 7 (2019) 102331–102344, http://dx.doi. [130] C. Chatfield, M. Yar, Holt-Winters forecasting: some practical issues, J. R. Stat.
org/10.1109/ACCESS.2019.2928325. Soc. D 37 (2) (1988) 129–140.
[107] M. Jourdan, S. Blandin, L. Wynter, P. Deshpande, Characterizing entities in [131] K. Struga, O. Qirici, Bitcoin price prediction with neural networks, in: RTA-CSIT,
the bitcoin blockchain, in: 2018 IEEE International Conference on Data Mining 2018, pp. 41–49.
Workshops, ICDMW, 2018, pp. 55–62, http://dx.doi.org/10.1109/ICDMW.2018. [132] N. Kumar, R. Iqbal, S. Misra, J.J. Rodrigues, Bayesian Coalition game for
00016. contention-aware reliable data forwarding in vehicular mobile cloud, Future
[108] P. Monamo, V. Marivate, B. Twala, Unsupervised learning for robust Bitcoin Gener. Comput. Syst. 48 (2015) 60–72.
fraud detection, in: 2016 Information Security for South Africa, ISSA, IEEE, [133] R.J. Barro, Money and the price level under the gold standard, Econ. J. 89
2016, pp. 129–134. (353) (1979) 13–33.
[109] M.A. Harlev, H. Sun Yin, K.C. Langenheldt, R. Mukkamala, R. Vatrapu, Breaking [134] S. Velankar, S. Valecha, S. Maji, Bitcoin price prediction using machine
bad: De-anonymising entity types on the bitcoin blockchain using supervised learning, in: 2018 20th International Conference on Advanced Communication
machine learning, in: Proceedings of the 51st Hawaii International Conference Technology, ICACT, 2018, pp. 144–147, http://dx.doi.org/10.23919/ICACT.
on System Sciences, 2018. 2018.8323676.
[110] S. Dey, Securing majority-attack in blockchain using machine learning and [135] I.-C. Lin, T.-C. Liao, A survey of blockchain security issues and challenges, IJ
algorithmic game theory: A proof of work, in: 2018 10th Computer Science Netw. Secur. 19 (5) (2017) 653–659.
and Electronic Engineering, CEEC, 2018, pp. 7–10, http://dx.doi.org/10.1109/ [136] M. Rosenfeld, Analysis of hashrate-based double spending, 2014, arXiv preprint
CEEC.2018.8674185. arXiv:1402.2009.
[111] T. Pham, S. Lee, Anomaly detection in bitcoin network using unsupervised [137] F. Scicchitano, A. Liguori, M. Guarascio, E. Ritacco, G. Manco, A deep learning
learning methods, 2016, arXiv preprint arXiv:1611.03941. approach for detecting security attacks on blockchain, in: ITASEC, 2020, pp.
[112] T. Pham, S. Lee, Anomaly detection in the bitcoin system-a network perspective, 212–222.
[138] R.S. Portnoff, D.Y. Huang, P. Doerfler, S. Afroz, D. McCoy, Backpage and
2016, arXiv preprint arXiv:1611.03942.
bitcoin: Uncovering human traffickers, in: Proceedings of the 23rd ACM SIGKDD
[113] A. Bogner, Seeing is understanding: anomaly detection in blockchains with
International Conference on Knowledge Discovery and Data Mining, ACM, 2017,
visualized features, in: Proceedings of the 2017 ACM International Joint
pp. 1595–1604.
Conference on Pervasive and Ubiquitous Computing and Proceedings of the
[139] What was the DAO? 2021, Available: https://www.gemini.com/cryptopedia/
2017 ACM International Symposium on Wearable Computers, ACM, 2017, pp.
the-dao-hack-makerdao, (Accessed: May. 2021).
5–8.
[140] F. Reid, M. Harrigan, An analysis of anonymity in the bitcoin system, in:
[114] T. Idé, Collaborative anomaly detection on blockchain from noisy sensor data,
Security and Privacy in Social Networks, Springer, 2013, pp. 197–223.
in: 2018 IEEE International Conference on Data Mining Workshops, ICDMW, [141] F. Zola, M. Eguimendia, J.L. Bruse, R. Orduna Urrutia, Cascading machine
IEEE, 2018, pp. 120–127. learning to attack bitcoin anonymity, in: 2019 IEEE International Confer-
[115] T.H.-D. Huang, Hunting the ethereum smart contract: Color-inspired inspection ence on Blockchain, Blockchain, 2019, pp. 10–17, http://dx.doi.org/10.1109/
of potential attacks, 2018, arXiv preprint arXiv:1807.01868. Blockchain.2019.00011.
[116] W.J.-W. Tann, X.J. Han, S.S. Gupta, Y.-S. Ong, Towards safer smart contracts: A [142] M.W.L. Moreira, J.J.P.C. Rodrigues, V. Korotaev, J. Al-Muhtadi, N. Kumar,
sequence learning approach to detecting security threats, 2018, arXiv preprint A comprehensive review on smart decision support systems for health care,
arXiv:1811.06632. IEEE Syst. J. 13 (3) (2019) 3536–3545, http://dx.doi.org/10.1109/JSYST.2018.
[117] X. Zheng, R.R. Mukkamala, R. Vatrapu, J. Ordieres-Mere, Blockchain-based 2890121.
personal health data sharing system using cloud storage, in: 2018 IEEE 20th [143] M.Y. Kiang, A comparative assessment of classification methods, Decis. Support
International Conference on E-Health Networking, Applications and Services, Syst. 35 (4) (2003) 441–454.
Healthcom, IEEE, 2018, pp. 1–6. [144] J. Vora, P. Italiya, S. Tanwar, S. Tyagi, N. Kumar, M.S. Obaidat, K.-F. Hsiao,
[118] H. Tang, Y. Jiao, B. Huang, C. Lin, S. Goyal, B. Wang, Learning to classify Ensuring privacy and security in e-health records, in: 2018 International
blockchain peers according to their behavior sequences, IEEE Access 6 (2018) Conference on Computer, Information and Telecommunication Systems, CITS,
71208–71215. IEEE, 2018, pp. 1–5.
[119] S. Li, N. Zhang, S. Lin, L. Kong, A. Katangur, M.K. Khan, M. Ni, G. Zhu, Joint [145] S. Chakraborty, S. Aich, S.J. Seong, H. Kim, A blockchain based credit analysis
admission control and resource allocation in edge computing for internet of framework for efficient financial systems, in: 2019 21st International Conference
things, IEEE Netw. 32 (1) (2018) 72–79. on Advanced Communication Technology, ICACT, 2019, pp. 56–60, http://dx.
[120] X. He, H. Xing, Y. Chen, A. Nallanathan, Energy-efficient mobile-edge com- doi.org/10.23919/ICACT.2019.8701926.
putation offloading for applications with shared data, in: 2018 IEEE Global [146] V.A. Siris, D. Dimopoulos, N. Fotiou, S. Voulgaris, G.C. Polyzos, IoT resource
Communications Conference, GLOBECOM, IEEE, 2018, pp. 1–6. access utilizing blockchains and trusted execution environments, in: 2019
[121] J. Kang, Z. Xiong, D. Niyato, P. Wang, D. Ye, D.I. Kim, Incentivizing consensus Global IoT Summit, GIoTS, 2019, pp. 1–6, http://dx.doi.org/10.1109/GIOTS.
propagation in proof-of-stake based consortium blockchain networks, IEEE 2019.8766403.
[147] Y. Lin, S. Han, H. Mao, Y. Wang, W.J. Dally, Deep gradient compression:
Wirel. Commun. Lett. 8 (1) (2019) 157–160, http://dx.doi.org/10.1109/LWC.
Reducing the communication bandwidth for distributed training, 2017, arXiv
2018.2864758.
preprint arXiv:1712.01887.

63