Cluster Computing

https://doi.org/10.1007/s10586-022-03582-7 (0123456789().,-volV)(0123456789().
,- volV)

Blockchain for deep learning: review and open challenges

Muhammad Shafay1 • Raja Wasim Ahmad1,2 • Khaled Salah1 • Ibrar Yaqoob1 • Raja Jayaraman3 •

Mohammed Omar3

Received: 13 October 2021 / Revised: 11 January 2022 / Accepted: 3 March 2022

 The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2022

Abstract
Deep learning has gained huge traction in recent years because of its potential to make informed decisions. A large portion
of today’s deep learning systems are based on centralized servers and fall short in providing operational transparency,
traceability, reliability, security, and trusted data provenance features. Also, training deep learning models by utilizing
centralized data is vulnerable to the single point of failure problem. In this paper, we explore the importance of integrating
blockchain technology with deep learning. We review the existing literature focused on the integration of blockchain with
deep learning. We classify and categorize the literature by devising a thematic taxonomy based on seven parameters;
namely, blockchain type, deep learning models, deep learning specific consensus protocols, application area, services, data
types, and deployment goals. We provide insightful discussions on the state-of-the-art blockchain-based deep learning
frameworks by highlighting their strengths and weaknesses. Furthermore, we compare the existing blockchain-based deep
learning frameworks based on four parameters such as blockchain type, consensus protocol, deep learning method, and
dataset. Finally, we present important research challenges which need to be addressed to develop highly trustworthy deep
learning frameworks.

Keywords Deep learning
AI
Machine learning
Federated learning
Blockchain
Ethereum
Smart contracts

Security
Transparency

1 Introduction deep learning techniques have assisted health physicians in

diagnosing COVID-19 patients using the dataset of CT and
The potential of deep learning has been witnessed in almost X-ray images [4, 5]. Apart from deep learning applications
all industrial sectors. For example, in the healthcare sector, in the healthcare industry, it has been employed by security
deep learning models are used by physicians to correctly officers at airports to identify and verify banned items in
diagnose the disease of the patient from the symptoms. passengers’ luggage or safeguarding software from vul-
During the recent pandemic caused by the spread of nerabilities [6–8]. Using biometric security and face
coronavirus disease (COVID-19), deep learning models recognition features, deep learning models can assist the
have been employed to predict the disease spread rate in a authorities in recognizing any physical dangers in real-
particular region and assist the authorities in managing the time. The efficacy and efficiency of a deep learning system
pandemic using the forecasted results [1–3]. Also, novel basis on the quality of the data used during the model
training phase [9]. The majority of the deep learning
techniques have considered centralized storage and pro-
& Ibrar Yaqoob cessing for training the model that is prone to a single point
[email protected] of failure and data alteration by the adversaries. Any
1
Department of Electrical Engineering and Computer Science, alteration of the data used for deep learning operations can
Khalifa University, Abu Dhabi 127788, UAE corrupt the training model. Blockchain is a decentralized
2
College of Engineering and Information Technology, Ajman technology that can efficiently handle data integrity,
University, Ajman, UAE security, and confidentiality [2, 10, 11, 11]. The integration
3
Department of Industrial & Systems Engineering, Khalifa of blockchain with deep learning can bring several benefits,
University, Abu Dhabi 127788, UAE e.g., automated and trusted decision making, efficient data

123
 Cluster Computing

market management, data security, better model building capable of generating human-like text in online web ser-
for prediction purposes, model sharing, and enhancement vices based on intelligent interpretation of the input data
of the robustness of the deep learning-based systems. [14]. Apart from these primary benefits of deep learning, it
The data collection stage which involves data acquisi- has the potential in terms of efficient data processing for
tion, labeling, and improvement, is of utmost important as sensory data analytic and optical character recognition
it can significantly affect the quality and performance of (OCR)-based applications. Traditional deep learning sys-
the developed deep learning models. Data collection tems have leveraged cloud-based servers to store and
ensures that large-sized, diverse, and high-quality training manage large-sized data to train the algorithms. Cloud
data is gathered from multiple sources to enable the computing can speed up the training process of deep
training models to perform well on real-world testing data learning algorithms by leveraging clusters of GPUs and
[9]. The model trained on a small amount of data is often CPUs for quickly executing compute-intensive tasks
overfit and offers limited performance on testing data [13, 15, 16]. For instance, in telehealth and telemedicine-
[12, 13]. The quality of an algorithm is highly affected by based services, wearable devices can collect and transfer
the quality of the supplied data; hence models trained using large-sized healthcare data to the cloud servers through a
high-quality data offer higher predictive accuracy. Figure 1 trusted edge server [2]. In the next phase, such healthcare
enlists important applications and benefits of deep learning data is analyzed on resource-rich cloud servers to identify
techniques along with a description of the role of each the patterns in the data using deep learning models.
application. The support of deep learning in image analysis However, many of the traditional systems are incapable to
and recognition, as highlighted in Fig. 1 enables identifying fully exploit their potential because of the hindrance cre-
objects, humans, or any action in the image. Such image ated by centralized architecture that is followed by the
analysis techniques can be useful for detecting the number existing systems. Furthermore, centralized-based data
of faces in an image or locating the nearby area of an storage and processing raise the risk of a single point of
autonomous vehicle using image segmentation techniques. failure problem. Since data has the utmost importance in a
Similarly, voice recognition techniques are mostly used for deep learning system, hence such data requires high pro-
controlling smartphones or smart homes by recognizing tection against any external or internal attack. Furthermore,
voice-based commands. Text prediction models are often the data saved on a centralized system will be less reliable
used in web services to predict the text in received emails, because it is susceptible to changes and fraud.
summarize messages, or paraphrase essays. Not only is a Blockchain technology has the potential to efficiently
prediction made, but the deep learning models are also handle the aforementioned problems caused by centralized-

Fig. 1 Key benefits of deep

learning techniques in various
fields

123
Cluster Computing

based data storage and processing. Blockchain is a revo- More specifically, it assists in identifying the owner of the
lutionary technology that maintains a shared ledger of data machine learning algorithms, datasets, the source of data,
among participants in a decentralized network. It ensures participants, the base model, and processes involved during
that all ledger copies maintained by the participants are model creation using the record of immutable transactions
verified and proven consistent [3]. The differentiating stored on the blockchain. The blockchain features such as
features of blockchain assist in enhancing the robustness of consensus algorithm, data immutability, and cryptographic
the deep learning models by protecting the data against hash functions assure that attacks on AI models including
several types of attacks by adversaries. By design, block- data, model, and algorithm poisoning are not possible
chain represents a tamper-proof and tamper-resilient tech- [24, 25]. Existing studies have thoroughly explored the role
nology that assists in tracking the data to ensure that it has and applications of blockchain technology in AI-related
not been tampered with since its creation [3, 17, 18]. The domains. For instance, the studies discussed in [26] and
main advantages of novel blockchain technologies include [27] have briefly discussed the role of blockchain in AI and
data immutability, transparency, security, provenance, machine learning fields, respectively. To the best of our
traceability, and operational visibility, and such benefits are knowledge, there exists no survey/review article that has
envisioned by the decentralized and Peer-to-Peer (P2P) thoroughly explored the role of blockchain in the deep
architecture of blockchain. The business processes on the learning domain to date. To fill this research gap, we have
blockchain platform can be automated using self-executing thoroughly reviewed the literature to highlight the strengths
smart contracts that eliminate the role of third parties in and weaknesses in the existing frameworks. It has proposed
executing the services [19]. The smart contract enables the a taxonomy to classify the literature based on a set of
development of a cheap, swift, and reliable system for deep parameters. It has investigated the reviewed literature
learning applications. The consensus algorithms imple- based on chosen parameters. Finally, open research chal-
mented by the existing blockchain platforms ensure data lenges along with guidelines are presented. The key con-
integrity [18, 20–22]. Existing research works have tributions of this paper are as follows:
employed private, public, and consortium blockchains to
• We devise a taxonomy to categorize and classify the
propose systems for various types of deep learning-based
existing literature related to blockchain-based deep
applications. Among these three categories, public block-
learning frameworks based on seven important
chain platforms are vulnerable to inference attacks as
parameters.
transactions, pseudonymous addresses, and other user data
• We present insights into the state-of-the-art blockchain-
are publicly available [3]. However, private and consor-
based deep learning frameworks by highlighting their
tium platforms preserve data privacy in a much better way
strengths and weaknesses.
than public platforms.
• We compare the blockchain-based deep learning
Machine learning models are created, used, and trained
frameworks based on important parameters.
by different entities. Blockchain technology enables
• We discuss several research challenges that can affect
establishing the provenance of machine learning models,
the performance, accuracy, and prediction quality of
thus leading to trusted Artificial intelligence (AI) systems.
existing blockchain-based deep learning frameworks.
Blockchain technology presents a robust system and can
incentivize the participants who share their data (data Figure 2 presents an organogram to show the organi-
trading) which is used to train machine learning models. zation of the paper. Section 2 discusses the necessary
Storing machine learning-related data on the blockchain background about blockchain and deep learning models.
network reduces the chances of errors in the model, as the Section 3 presents and discusses a detailed taxonomy that
blockchain network will not have duplicate, missing, or classifies the existing literature into several categories.
noisy data, which is a fundamental necessity for AI-based Section 4 provides a detailed insightful discussion about
models. Combining the values of blockchain and machine existing studies related to blockchain-assisted deep learn-
learning techniques offers many benefits to the applica- ing frameworks. Section 5 highlights and discusses several
tions, such as data authenticity assurance, data usage open challenges in this field of research. Finally, Sect. 6
management, audit trails, and introducing new values to the concludes the paper.
business processes through automation enabled by smart
contracts. For instance, the machine learning models
embedded in smart contracts can automatically recommend
a request for authorities to recall expired medicines.
Blockchain can capture the evolution of machine learning
models as it progresses and records the various stages of
the model during its creation, updating, or usage [23, 24].

123
 Cluster Computing

Fig. 2 Organogram of the paper

2 Background consensus algorithms, and decentralization assist in

improving business effectiveness [31]. Figure 3 presents an
This section briefly discusses the key features of block- overview of the blockchain benefits that make it different
chain and deep learning and the benefits of integrating from the other technologies.
them in terms of data security, automatic decision making, Blockchain technology can assist in protecting the
and enhanced robustness. Electronic health records (EHR) and Personal health
records (PHR) of patients. EHR and PHR systems contain
2.1 Blockchain technology information related to diagnoses, medications, allergies,

Blockchain stores information in a way that makes it

extremely difficult for hackers to change, manipulate, or
delete data. By design, it is a decentralized technology that
is based on P2P architecture for storing and processing
transactions and data. It consists of many nodes which
verify and store the transactions in the form of blocks. Each
block in the chain stores a set of transactions and it is
ensured that the existing blocks are correctly linked to the
newly created block to form the chain of blocks. After
adding the block to its local chain by a miner, the newly
added block is propagated to all participating nodes to
ensure data consistency [28, 29]. The decentralized con-
sensus protocol ensures that the blockchain transactions are
validated and agreed upon by the miner nodes. For
instance, the Proof-of-work and Proof-of-Stake are con-
sensus protocols implemented by many blockchain plat-
forms and they can keep the blockchain secure from any
internal or external data hacking attack. Smart contracts,
which represent an electronic program, are another
important feature of blockchain technology and executes
only when predetermined criteria are met [3, 17, 30]. Smart
contracts are aimed at reducing the risk and cost of busi-
nesses. The key characteristics and features of blockchain
technology such as data immutability, smart contracts,
Fig. 3 Advantages of Blockchain technology

123
Cluster Computing

medical disease history, test results, etc. EHR contains data The deep learning models are trained by using a large
about a patient from all clinicians, hospitals, test centers, amount of labeled data.
laboratories, or other healthcare organizations. On the other The data representation techniques assist the machine
hand, the PHR is a patient-centric system that is managed learning algorithms in learning the globalized data
by patients in a controlled and private environment sequences. Therefore, data quality plays a critical role,
[17, 32]. It can ensure that the data is controlled and specifically in machine learning algorithms, to make the
managed by patients, and it can be shared with other users correct decisions using time series data. More specifically,
in compliance with the patient’s consent management if the data consists of inadequate descriptions, irrespective
policy. The consent management policies are implemented of how elaborate the algorithm is, the model will not per-
through self-executing smart contracts. However, block- form well on such data. Hence, feature engineering is
chain becomes a costly technology if appropriate tech- considered as it may assist the reconstruction of data by
niques are not implemented to manage large-sized exploiting the sets of features from the raw data [43]. In the
healthcare data [17, 32]. To take full advantage of block- case of a deep learning algorithm, the models are intelli-
chain in many healthcare applications, the pointers and gent and automatically extract the high-level latent space
linkers can play a valuable role to minimize the data size. features from the basic form of the data. By design, the
In addition, decentralized storage systems are also capable deep learning models consist of multiple layers. The lower-
to securely store large-sized data and avoid single-point-of- level layers are responsible for extracting lower-level fea-
failure-related problems. The examples of most widely tures, while higher layers extract more abstract features
used decentralized storage systems in healthcare sector from the input data. The number of layers in DL affects the
include InterPlanetary File System (IPFS) [33], Cassandra accuracy and security.
[34], SWARM [35], Storj [36], OrbitDB, and Skeps, to Tuning the deep learning model’s hyper-parameters
name a few [33]. plays a critical role in achieving optimal performance for
the given application [44, 45]. In terms of scalability, the
2.2 Deep learning and federated learning important parameter to consider is the choice of network
topology on which the model is deployed. For example, a
Deep learning encompasses inside AI and machine learn- deep learning model, instantiated within the client-server
ing. In existing deep learning approaches, a model learns architecture, provides a scalable solution because the
the latent space representation of the most basic form of model is trained only at the server end, and each node will
data, i.e, images, text, and speech signals. Figure 4 shows procure the trained model instance to perform its desired
the relationship between deep learning, AI, and machine task [46]. Here, if the need arises to re-train (or fine-tune)
learning. Deep learning allows the hardware to execute the DL model, it can be effectively performed at the server
many applications with human-like accuracy or even in end, and this newly tuned instance can be broadcast to each
some cases the better accuracy than humans. To highlight of the nodes within the network (by analyzing the block-
the widespread applications of deep learning in different chain timestamps) to perform the underlying job. Further-
fields, image classification [37], object detection [38], self- more, deploying the model on the server end also provides
driving cars [39], disease prediction [40], and voice control ease in tuning the model’s intrinsic hyper-parameters (only
[41] are well known use cases of deep learning techniques. once), which can then be generalized across the whole
network. On the contrary, deploying the DL model in a
peer-to-peer (P2P) network is not a very scalable solution
because each node (within the network) will have to
explicitly train/re-train its model instance to perform the
desired task. This explicit re-training (or fine-tuning) pro-
cess can be exhaustive (especially for performing a similar
or inter-related task across each node)—also, it’s highly
dependent on the computational capacity of each node.
Deep learning model analyzes the data patterns in the
Internet traffic and learns from them to train a model to
recognize cybersecurity threats. There are many types of
security threats over the Internet that cybersecurity teams
are facing nowadays. For instance, malware, data breaches,
social engineering, phishing, Denial-of-service (DOS), and
Fig. 4 Relationships between AI, machine learning, and deep learning insider attacks steal the users’ private data for harmful
[42] purposes. Deep learning has the potential to combat the

123
 Cluster Computing

aforementioned cybersecurity threats. It assists in detecting database in which all network nodes can hold and exchange
the intrusion in the system and preventing it from occurring data in a manner that is decentralized and verifiable.
in the future. Deep learning models analyze the data traffic
and verify the transaction signatures to detect the intrusion. Table 1 summarizes some of the key features of
It immediately notifies the user about the intrusion detected blockchain and deep learning that assist in improving deep
(through blockchain smart contracts). Furthermore, deep learning-based applications [26]. Figure 5 highlights the
learning models can recognize suspicious activities within four key aspects, scenarios, and categories that take
the system based on the data traffic analysis to identify bad advantage of such integration. By design, blockchain is a
actors or malware. Natural Language Processing (NLP) fault-tolerant technology that secures the data, while deep
which is one of the subtypes of deep learning detects social learning focuses on utilizing such data to train the models
engineering-related data theft threats [47, 48]. and make accurate predictions. The data immutability
Federated learning methods aim to address the data feature of blockchain ensures that data leakage cannot
confidentiality and security challenges in deep learning happen, thus protecting the deep learning model or data
methods. Data has become the new oil in the recent era against several types of attacks or data noise issues. As a
owing to ongoing advancements in deep learning. It is result, the predictions made by such models are more
forecasted that smartphone users will grow to 4.3 billion in trustworthy and accurate. The collaboration of blockchain
number by the year 2023, and they will generate a huge and deep learning opens doors towards automation of
amount of data [49]. This data can be gathered from several tasks where data requires careful handling and high
powerful sensors such as GPS, cameras, and microphones. security. Deep learning and blockchain together can pro-
On the other hand, deep learning models require a large vide a stable, permanent, and decentralized infrastructure
amount of data to update the models and optimize the user for the critical data that deep learning-driven applications
experience at the expense of data leakage and privacy would acquire, process, and employ. The following is a
issues. Federated learning (FL) [43] represents a system summary of the benefits associated with the integration of
that does not rely on a central authority to collect and blockchain technology with deep learning algorithms:
process the users’ data. Also, traditional systems require
• Data Security: Being a decentralized technology, the
high bandwidth while uploading and downloading data to a
information stored on the blockchain is highly secure.
central server for model generation. As a result, the central
Private blockchain platforms are deployed for storing
server suffers from a large communication overhead, thus
and processing private and confidential information. It
necessitating the use of a decentralized server model to
is required to keep the private keys of the nodes secret
generate models in a resource-efficient way [50]. The
as these keys can be the only way to access the
unique characteristics of federated learning models have
blockchain data. Deep learning algorithms can run on
motivated several industries to employ federated learning-
stable data provided by the blockchain, thereby result-
based systems for predictive text, voice recognition,
ing in more trustworthy, accurate, and reliable decision-
cybersecurity threat identification, enhancement of the
making [51].
performance of devices in IoT applications, and disease
• Automatic Decision Making: Blockchain is a well-
diagnosis.
known technology that processes transactions on a P2P
basis. It makes it easy to verify the decisions made by
2.3 Blockchain-based deep learning
the deep learning models through the traceability
feature. It also ensures that the documents have not
The re-usability and trusted sharing of deep learning
been tampered with during the human-assisted auditing
models is an essential requirement that can be fulfilled by
phase [24, 52].
blockchain technology. Similarly, auditability, data verifi-
cation, attestation of results, provenance, traceability of
ownership, usage, and guarantee of fairness are the main Table 1 A summary of the deep learning and blockchain features that
assists in improving deep learning-based applications [26]
motivations behind the integration of blockchain and deep
learning [24]. Deep learning models are fed with a large Blockchain Deep learning Potential outcomes
data of diverse examples and such data is used by the
Immutable Scalable Flexibility in learning strategies
models to learn the features and produce an output with
Transparent Layered Collaborative model udate
probability vectors in place. Even though deep learning
Integrity Resource Enhanced scalability
models perform exceptionally well on raw data, but the intensive
quality of the data still matters regarding the prediction for Cybersecurity Data intensive Upgraded data security
many real-world scenarios. The blockchain is a global

123
Cluster Computing

Fig. 5 Benefits resulted from the integration of deep learning and blockchain technology

• Cumulative Judgements: In many cases, the autono- and differences among the state-of-the-art blockchain-
mous digital agent makes decisions based on the data based deep learning frameworks. Given below is a brief
gathered related to a particular scenario. Deep rein- introduction to the selected parameters along with their
forcement learning and swarm robotics are examples of technological details.
such agent-based decision-making system [53–55]. The
voting-based approach can assist the robots to make 3.1 Blockchain type
decisions based on the data collected by swarm robotics
on the blockchain. This parameter classifies the existing studies into three
• Enhanced Robustness: In few cases, the accuracy of categories based on the blockchain platforms selected by
decisions made by the deep learning models surpasses the existing blockchain-based deep learning frameworks.
human-level accuracy. Hence, the highly accurate deep Many of the services and applications enabled by the deep
learning model increases the trust of the stakeholders in learning methods have time-related constraints; therefore,
the decisions. Also, being backed by decentralized the time-specific modalities of the blockchain assists in
technology, the robustness of the deep learning-based improving such services. Based on the design, character-
system can be ensured. The integration of deep learning istics, and policies, blockchain platforms considered by
with blockchain can be valuable in business forums state-of-the-art deep learning frameworks can be classified
where the parties can work in a trust-less and automatic into public, private, and Consortium/Federated categories
environment [56]. as discussed below.

3.1.1 Public blockchain

3 Taxonomy of blockchain-based deep
The public blockchain platform leveraged by the existing
learning frameworks
blockchain-assisted deep learning frameworks allows per-
missionless or unrestricted access to the distributed ledger
Automated decision-making, data security, accurate fore-
by the users or machine learning devices. Users access the
casting, efficient data market management, and enhance-
ledger copy that is distributed among all nodes within the
ment of the robustness of the system are the key benefits
public blockchain network and performs transactions.
that can be achieved through the unionization of block-
Public blockchain platforms maintain transaction anon-
chain and deep learning techniques. This section presents a
ymity because of decentralized data storage and process-
thematic taxonomy to classify the existing literature related
ing. Furthermore, public blockchain platforms are secure
to the unionization of blockchain and deep learning tech-
against several types of attacks; therefore, they assist the
niques based on a set of parameters. The identified
deep learning models in coming up with the correct and
parameters as shown in Fig. 6 highlight the commonalities
trustworthy results [28, 30, 57].

123
 Cluster Computing

Fig. 6 A taxonomy of blockchain for deep learning frameworks

3.1.2 Private blockchain the deep learning models used for decision-making in
several applications areas are categorized into five major
The private blockchain platforms leveraged by the block- categories. Below is a brief overview of deep learning
chain-assisted deep learning frameworks are controlled and models that have used blockchain-based data to generate
managed by a single entity. The private platforms are patterns and make decisions.
permissioned where the authority lies within the control-
ling entity [58]. As the identities of the validators and 3.2.1 Convolution neural network
nodes are known to the central authority, hence the private
network requires relatively lesser complex mathematical Convolution Neural Network (CNN) which is also known
calculations to verify the transactions. As a result, the as ConvNet processes an image to identify the objects,
private platform’s transaction execution speed is higher assign weights to the objects, and classify them according
than the public platform. to the context. It also enables detecting object instances in
the processed image [60]. The existing blockchain-based
3.1.3 Consortium/federated blockchain deep learning frameworks have employed CNN to classify
images, detect objects, and segment the instances in vari-
Consortium blockchain platforms leveraged by the existing ous use cases. The advantage of selecting CNN in block-
blockchain-assisted deep learning frameworks hold the chain-based studies is the minimum preprocessing time
characteristics of both private and public blockchain plat- required by the algorithm because of choosing adapt-
forms. A consortium blockchain functions as a permis- able filters to determine the characteristics of the image.
sioned network and, multiple heterogeneous groups can
have the authorization role, unlike private networks where 3.2.2 Recurrent neural network
a single authority is responsible for controlling and
managing the network [59]. In general, anyone on the A CNN model outperforms on images data used as input.
blockchain network can access the content on the block- However, Recurrent Neural Network (RNN) uses sequen-
chain, but only a limited authorized group of users can tial or time-series data to generate patterns [61]. The
append the data to the ledger. Also, the transaction vali- famous applications of RNN for blockchain-based solu-
dation rate of consortium platforms is faster than public tions include voice or speech recognition, speech-to-text
blockchain platforms. conversion, voice search, and natural language processing
(NLP). Also, the input data is independent of each other in
3.2 Deep learning models CNN; whereas, the previous inputs are linked and influence
the output in RNN models. Long Short-Term Memory
A deep learning model processes the collected data and (LSTM) [62] and Gated Recurrent Units (GRU) [63] are
makes patterns useful in decision-making in various use upgraded versions of RNN to address the shortcoming of
cases. Based on the configuration of neural network layers, RNN and are widely employed for accurate forecasting.

123
Cluster Computing

3.2.3 Generative adversarial networks (GAN) task using the same type of input data and generates a
competitive model. The suppliers and training data are the
The generative model learns the patterns in an unsuper- vital components in the BlockML architecture. Suppliers
vised manner and is capable of generating unique data. are those entities that publish a machine learning task,
More specifically, it is a form of generative modeling that reward, test data, hash, and training data to IPFS. Once the
employs deep learning techniques such as convolutional model training by the miners has been completed, miners
neural networks. By design, the GAN model consists of a immediately publish the results on the blockchain, and the
generator and a discriminator network. The generator is supplier releases the test dataset on IPFS. Based on the
responsible for producing new examples, whereas the dis- model results generated for the test data, the miners fetch
criminator learns to classify the data as real or fake [64]. the competing solution and the highest-ranked solution is
considered the winner. The block of the successful miner
3.2.4 Deep reinforcement learning (DRL) will be appended to the blockchain.

DRL is inspired by theories of human behavior that basis 3.3.2 Committee consensus mechanism
on behavioral ecology and enables expert systems to
understand the data more precisely. Intelligent agents take The consensus algorithm in this category updates the glo-
actions in an environment consisting of DRL models to bal model using the federated learning-based approach.
learn. Furthermore, agents are implicitly validated or The Committee consensus mechanism (CCM) updates the
penalized depending on their behavior. Behaviors that lead model locally and then adds it to the blockchain [69]. The
to the desired result are rewarded, thus termed as rein- blockchain part of federated learning is divided into two
forced learning-based model [65]. parts. The first part of the blockchain stores the randomly
initialized model at the time of system initialization. The
3.2.5 Geometric deep learning second part of the blockchain is responsible to store the
model updates generated by the participating nodes in the
It is a deep learning variant that focuses on developing federated learning-based systems. These nodes essentially
neural networks that basis on non-euclidean data [66]. A obtain the model present in the blockchain and undergo the
graph is a specific example of non-euclidean data. The data training locally. Afterward, the local gradient is verified
modeling can be done with fewer efforts and resources and added to the new block; hence, the blocks continue to
while using graph-based data. The graphs are input to the grow in number.
geometric deep learning models rather than the data in the
traditional form to generic neural networks. In a nutshell, 3.3.3 WekaCoin
geometric deep learning has the potential to extract more
fine granular details from the data. WekaCoin is a cryptocurrency that provides a decentral-
ized database of machine learning models which are pub-
3.3 Deep learning specific consensus protocols licly accessible. The block of WekaCoin holds three items
such as transactions, the hash of the previous block, and the
DL-specific consensus algorithms are aimed at minimizing information about the machine learning contest used to
the convergence time of the consensus process in validat- verify the block in the previous iteration. By registering a
ing the transactions and assisting the agents in multiagent- model transaction on the chain, the newly created model
based systems to reach an agreement while using minimal can be uploaded to the WekaCoin network [70]. The deep
resources. The traditional consensus algorithms such as learning model trainers only store the hash value of their
PoW are slow and less energy efficient. The existing models at this point. Deep learning model trainers publish
blockchain-based studies have proposed various DL- their models to IPFS once the test data is available. The
specific algorithms such as BlockML, Committee algo- advantage of this approach is the minimum energy
rithm, WekaCoin, and Proof-of-learning consensus algo- requirement as only relevant tasks are solved by miners,
rithms. These algorithms are briefly discussed below. the development of a library of machine learning algo-
rithms, and datasets that are publicly available.
3.3.1 BlockML
3.3.4 Proof-of-learning
The BlockML protocol aims at minimizing the compute-
intensive operations in the neural network’s training phase The proof-of-learning algorithm practices a nonconvex
using a supervised approach [67]. Unlike the existing optimization approach of potentially large-sized neural
systems (e.g., [68]), BlockML forces the miners to solve a networks and exhibits NP-hardness characteristics because

123
 Cluster Computing

of computational complexity in solving the asymmetry speed, congested route, heading angle, and road state with
challenge. In this system, a Secure mapping layer (SML) is other vehicles. The vehicles make different decisions, such
introduced for tamper prevention [71]. In the Proof-of- as optimal route finding based on such data. The V2I model
learning algorithm, the consensus nodes provide the pro- enables vehicles to share the captured data with roadside
cessing power and strive to train the model for the issued units, such as RFID readers and cameras, lane markers, and
task. Once a model has been created that fulfills the min- streetlights, to improve public safety, efficiency, and con-
imal training exactness, the successful miner broadcasts the venience [74].
new block and declares its success.
3.4.3 Traffic management
3.4 Application areas
Road congestion is one of the biggest issues in urban areas
This parameter specifies the key application areas that are and can result in large financial and ecological losses due to
focussed on by the state-of-the-art blockchain-assisted increased carbon dioxide emissions [75, 76]. One of the
machine learning frameworks to ensure data integrity. The most widely used approaches for forecasting traffic inten-
main categories include healthcare, IoVs, traffic manage- sity using real-time data is crowdsourcing. However, cur-
ment, and safety and protection. rent crowdsourcing models are impractical due to
centralized-based data storage and human safety concerns.
3.4.1 Healthcare
3.4.4 Safety and protection
Healthcare is another field in which patient data is of
extreme importance. The healthcare data that comprises Blockchain technology is well-known and practicable in
clinicians reviewing images and scans is used for research various fields due to the security feature it provides in the
purposes for training deep learning models for the predic- terms of the immutability of the data. The differentiating
tion of communicable and non-communicable diseases features of blockchain technology and decentralized data
such as COVID-19 and cancer, identifying new ways of storage characteristics can assist in securing various
diagnosing illness, improving the quality of healthcare application areas. However, the 51% attack on the block-
services, and improving clinical trials [2]. The healthcare chain that can affect data integrity is still possible [77].
data consists of records related to the diseases, disease Data can be protected from such attacks using machine
symptoms, and medical profile of the patient that assist the learning approaches. Signatures represent the malicious
deep learning models in predicting the medical condition of code’s fingerprint and aid in detecting and identifying
a future patient. The high security and precision of such malware.
healthcare data is the utmost requirement of the healthcare
industry as it directly affects human life. Storing such data 3.5 Services
on centralized systems makes them vulnerable to attacks.
Hence, blockchain technology, in this case, can play a vital This parameter highlights the primary service of block-
role in preserving all the data as it is inherently decen- chain-based deep learning solutions targeting various
tralized and provides a solution to data security threats. application areas such as healthcare, vehicular communi-
Blockchain assures that the data is safe from any planned cation, and IoT. Existing approaches aim for many objec-
or accidental loss. tives including privacy preservation , traffic violation
prediction, anomaly detection, cellular traffic management,
3.4.2 Internet of vehicles and forking prevention.

The rise in the popularity of the Internet of Things (IoT) as 3.5.1 Privacy preservation
an instrumentation technology for vehicles has rapidly
increased the networking capabilities of existing devices Privacy protection assures that the data about an entity
[72]. The vehicles in modern transportation systems are cannot be disrupted or viewed by unauthorized users.
usually equipped with devices, sensors, and intelligent Blockchain technology has the potential to preserve the
software to get connected and exchange data with each privacy of the data used in deep learning models [78].
other. By introducing blockchain on the Internet of Vehi- Privacy protection can be ensured through proxy re-en-
cles (IoV) network, the data can be securely shared among cryption and highly profound data encryption schemes.
the entities as highlighted in Fig. 7 [73]. The V2V com-
munication mode of intelligent transportation systems
enables a vehicle to share data about a vehicle’s position,

123
Cluster Computing

Fig. 7 Blockchain-based system

for Vehicle-to-Vehicle (V2V)
and Vehicle-to-Infrastructure
(V2I) communication

3.5.2 Violation prediction 3.5.5 Forking prevention

Traffic violations can be of many types including over- Forking refers to the blockchain state where the ledger is
speeding, driving while intoxicated, illegal lane changes, or divided into two potential paths or a situation where more
failure to stop at a red light. The blockchain and deep than two blocks have the same block height. Blockchain
learning amalgam play a key role in predicting the viola- and federated learning models can prevent forking by
tions made by the drivers. For instance, the deep learning- enabling the miners to verify the blocks correctly. For
based classifier can use highly secure blockchain data to instance, by broadcasting an acknowledgment signal in
predict traffic incidents [79]. Such data can be immutably blockchain-assisted deep learning models, the occurrence
recorded on the blockchain and can be used by the road- of forking can be prevented successfully [82]. On receiving
ways department for road network planning, insurance the acknowledgment signal from all the miners, the gen-
companies for damage assessment, and law enforcement erated block is considered valid.
agencies for lawmaking.
3.5.6 EHR forecasting
3.5.3 Anomaly detection
EHR refers to the digital version of the patient-related data
Anomaly detection deals with the identification of data that such as lab results, medical history, allergies, and pre-
deviates from normal behavior. Based on blockchain-based scribed medicines. Usually, patients use an open frame-
data and transactions, the encoder-decoder network in deep work to regulate, manage, and share their medical data with
learning can learn the normal behavior of a real-world family, friends, and physicians. Data integrity, data pro-
phenomenon [80]. It can immediately detect an anomaly tection, and secrecy are all major concerns in such a set-
based on the analysis of data. Furthermore, such anomalies ting. EHR forecasting using deep learning techniques assist
can be distinguished by the model as well. in predicting health situations in a particular region and
facilitating health care service in that region [83].
3.5.4 Data traffic management
3.6 Data types
Predicting data traffic using deep learning methods in
cellular networks can improve the reliability and efficiency The attributes of this parameter define the type of data that
of the network. The current 4G LTE systems can be con- can be accepted and processed by the deep learning models
figured to the blockchain-based cellular network by cus- backed by innovative blockchain technology. The deep
tomizing the access points as blockchain nodes to manage learning models supported by blockchain, in general, take
the cellular traffic. Every newly created block can be data in the form of images and text.
transmitted to all nodes of the blockchain called Block-
chain-radio access network (B-RAN) [81]. 3.6.1 Image data

CNN accepts the image as an input and extracts deep and

shallow features to classify the data. Storing and protecting

123
 Cluster Computing

the images used by the CNN models for training purposes 4 Blockchain-based deep learning
by employing blockchain can increase the trust of users on frameworks
data [84]. To optimally utilize the capacity of the ledger,
blockchain can store an immutable hash of the data stored This section presents a detailed review of state-of-the-art
on a third-party storage system such as IPFS. In this way, blockchain-based deep learning solutions for various
the data can be tracked via the immutable hash of the application areas. It compares existing blockchain-based
images stored on the blockchain. deep learning framework schemes based on parameters
selected from the literature.
3.6.2 Textual data
4.1 Review of blockchain-based deep learning
Textual data requires less storage capacity compared to frameworks
image data. However, such data need to be optimally stored
on the blockchain to ensure platform scalability. Most deep This subsection presents a review of existing blockchain-
learning-based systems have considered textual data for based deep learning frameworks mainly targeting health-
traffic or disease prediction. care, vehicular networks, cellular traffic management, and
blockchain safety and protection from adversarial attacks.
3.7 Deployment goal Data analysis The patient-centric data management and
resource-friendly design of deep learning models has
Operational auditability, data verification, attestation of become an essential element particularly in pharmacoge-
results, provenance, traceability of ownership, usage, and nomics research. The work discussed in [85] has proposed
guarantee of fairness can be achieved by integrating a decentralized system that provides pharmacogenomics
blockchain and deep learning [24]. The deployment goal data to deep learning models to predict ovarian cancer. The
parameter highlights the main services that can be achieved blockchain platform is used for sharing healthcare data,
by converging blockchain and deep learning technology. patient records, and ovarian cancer predictions made by the
The main goals include model provenance and decision- model with the participating organizations. The authors of
making. the proposed research have validated their methods by
comparing them with the CRYPTO?? standards. These
3.7.1 Trusted AI models standards are helpful to evaluate the schemes based on the
amount of time it takes to encrypt and decrypt data.
Blockchain technology assists in establishing the data However, the proposed model has been tested using a small
provenance of the deep learning models used by various number of examples per class and makes predictions for
industries for data classification and prediction. The pre- unseen data belonging to the learned classes.
dicted results generated by the trusted AI models are highly Provenance data for AI models An AI model can lead to
reliable and correct. The model provenance enabled by erroneous or misleading results particularly when the
blockchain technology can assist in verifying the trust- source of the employed deep learning model is unverifiable
worthiness of a deep learning model. or unknown. For instance, a deep learning model can be fed
with poisoned data to compromise the malware detection
3.7.2 AI decisions sharing functionality of the malware detection model. By tracking
and tracing the data related to the history of an AI model
Blockchain technology assists to securely share the data ranging from its creation, the data used in training, the
between the untrusted participants. The AI decision-shar- owner of the model, and the processes involved during
ing parameter describes the predictions made by the AI model creation can assist in protecting the model from
models based on time series data and secure sharing of the several data positioning attacks. Additionally, the avail-
prediction among the potentially untrusted participants. For ability of verifiable training data significantly increases the
instance, in IoV, the AI model can identify a road accident trust of the users in the AI models. The study presented in
and share this information with all the other vehicles reg- [24] has proposed a blockchain-based architecture for
istered to the network to avoid traffic congestion. maintaining provenance data for AI models. The main
classes implemented by the proposed system to generate
highly trusted AI models include participants, datasets,
models, operations, and compute pipelines or projects. The
record of immutable transactions stored on the blockchain

123
Cluster Computing

has assisted the users in successfully maintaining the data train the deep learning models to classify arrhythmia dis-
provenance of the generated AI models. ease. Arrhythmia is a disease related to irregularities in the
Model prediction In the healthcare field, cross-institu- heartbeat of the patient. In a study proposed in [91], two-
tional health data sharing for research purposes is indis- layered stacked denoising auto-encoders (SDA) are used to
pensable. It is invaluable to assure full compliance with the extract the features and to detect anomalous heartbeats
health data sharing rules defined by the regularities for data during the first phase. Furthermore, the SDA is retrained in
privacy and security preservation. Ensuring the security of the testing phase as well to control the false-positive rate in
the data is one of the main challenges faced by the the output classification. The data gathered after retraining
healthcare industry. The study discussed in [68] has high- can be stored in the hardware and can be accessed by using
lighted the importance of sharing deep learning models the pointers that are stored in a time-stamped ledger.
between the peers of the blockchain network instead of Combined cooperative positioning A Global positioning
sharing data to preserve data privacy and confidentiality. system (GPS) is the most widely used system that assists in
The study has proposed a novel consensus algorithm called tracking and tracing vehicles. The issues in GPS-based
proof of information [68], which is well suited for health- vehicle position estimation include errors in range esti-
care model prediction. The proposed study has employed mation, wrong installation, a low battery of GPS enabled
an incremental federated learning model for implementa- devices, and impossible vehicle lane-level navigation. A
tion purposes. The incremental learning approach trains the cooperative positioning approach that aims at optimizing
model using a single dataset in the first round. The second the GPS-based position estimation is useful to improve the
round considers a new dataset that belongs to a similar lane-level vehicle positioning accuracy [92]. The cooper-
class but different unseen examples. The errors are calcu- ative positioning approach suggested in [92] involves a
lated for each model trained by each member and the blockchain-based system for combining cooperative posi-
model with the least error is broadcasted to all participants. tioning (CP) with the IoV to increase GPS data validity and
Finally, through blockchain, all participants receive this consistency. Through decentralized blockchain technology,
model as it exhibits least errors as the latest model. data can be efficiently and securely shared among the
Data filtration Deep learning methods are highly cap- vehicles that are part of the vehicular network. The coop-
able of accurately predicting the causes or sources of dis- erative vehicles in a vehicular network can share the
ease and ensures that high-quality data is used during the trained deep learning models to estimate the error in
training process. In the research work proposed by Zheng vehicle estimated position. The vehicles can evaluate the
et al. [86], a deep learning technique supported by relatively accurate location based on the traffic sign and
immutable blockchain technology is employed to classify share the updated model with other vehicles. The proposed
healthcare data into high-quality and low-quality classes study overlooked the cost and security analysis to highlight
and aligning such data with certified consumers. It was the feasibility of their proposed research.
noticed that the majority of healthcare data comes from Crowdsensing In a traditional vehicle-to-vehicle (V2V)
wearable devices at a continuous dynamic rate; such communication network, the data privacy of the partici-
datasets can be highly valuable to build disease prediction pating vehicles can be affected by the centralized-based
models. The proposed study has mainly focused on data storage and processing. To ensure that the vehicle
dynamic data as this type of data is easily accessible via exchanges data in a trustworthy way, blockchain proves to
wearable devices. Sophisticated machine learning tech- be a viable solution. The work presented in [93] has pro-
niques are used to analyze the trend in the accumulated posed a blockchain-based system to handle the network
data to ensure the validity of data with high-performance latency in centralized systems. The proposed research has
validation patterns. The proposed study has assured that also solved data protection problems in the vehicular
data is collected from a reliable device to build deep crowdsensing frameworks by employing 5G-enabled IoV
learning models. [93]. A deep reinforcement learning-based algorithm has
Disease classification Deep learning models are famous been used to deal with block mining in blockchain net-
for their feature extraction abilities and can play an works. The system consists of a registration authority that
important role in future medical research. Deep learning authenticates and authorizes the roadside units, which acts
can analyze medical images like X-rays, MRI scans, CT as the miner nodes within the blockchain network. The
scans, etc., and determine any health risks and flag vehicles that are involved in data exchange are also reg-
anomalies in the CT scan images. Deep learning approa- istered to the blockchain networks. The participating enti-
ches are being utilized to predict various diseases such as ties collect the data and share it with the group head who is
pneumonia [87], multi-retinal disease, class imbalance in responsible for refining the data and transferring it to the
coronary heart disease, and cardiovascular disease [88–90]. server via base stations.
Juneja et al. [91] have employed blockchain platform to

123
 Cluster Computing

Errors reporting As discussed in [92], GPS-based improved traffic management. A consortium blockchain is
techniques do not provide high lane-level positioning used to keep the distributed ledger running by employing a
accuracy for vehicles. To handle this issue, Li et al. [94] small number of preselected miners [96].
proposed an edge and deep learning-based error sharing Computations offloading Mobile edge-cloud computa-
framework to improve the accuracy of position estimation tion offloading (MECCO) [101] presents a viable solution
of vehicles. The proposed study has also secured the data to overcome the challenges related to resource limitation of
published by participants by employing a blockchain mobile devices by identifying and offloading resource-in-
platform for vehicle position error reporting and sharing. In tensive tasks of an application to the mobile edge-servers.
the proposed framework, the evolution of positioning error Although MECCO has significantly augmented mobile
is evaluated using the deep learning-based prediction battery lifetime, preserving the privacy and security of the
model. The role of blockchain was limited to a trusted traffic data is overlooked [102]. As a result, the MEECO
bridge between the existing vehicles and edge computing can be fully exposed to a variety of risks and attacks.
nodes. Furthermore, the deep neural network-based error Nguyen et al. [99] has proposed a blockchain and deep
correction protocol was deployed on the edge server to reinforcement learning-based solution to handle the secu-
exploit the low latency and higher computing capacity of rity and privacy challenges in the MECCO system. The
edge servers while training the deep learning model. To proposed system has employed blockchain-based smart
reduce the block confirmation time, the proposed study has contracts to increase trust among participants of the edge-
employed a delegated proof-of-stake (DPoS) consensus cloud computing network.
algorithm instead of PoW which is resource unfriendly by Channel protection Singh et al. [98] has proposed a
design. blockchain and deep learning-based approach that has
Crowdsourcing The researchers in [95] have employed a implemented Zero Knowledge Proof (ZKP) for validating
crowdsourcing technique that uses a blockchain network to the registered machines. The proposed method registers the
identify and report an unwanted situation that occurred on machines such as a drone and verifies them using ZKP
the road such as road jamming due to vehicle accidents. before the transactions from such machines can be granted.
The proposed system has employed a smart contract that The proposed study selects a miner node using a novel
enables vehicles to register and leave the network freely selection algorithm that involves a deep Boltzmann
and anytime. The proposed system has chosen a proof-of- machine. The blockchain assures that data integrity will not
authority consensus algorithm for data validation while be compromised during device-to-anything (D2X) com-
considering the resource constraint nature of participating munication. Also, ZKP-based validation assists in limiting
entities/devices. On encountering an unwanted incident on the chances of network hacking by malicious drones. In the
the road, the user can employ the DApp interface to report proposed study, at the time of machine registration, the role
the event to the blockchain. Furthermore, to assure that the of each drone like the ordinary drone or miner drone is
same participants do not transfer the same data to the explicitly mentioned. In a nutshell, the proposed research
blockchain for economic rewards, only the first participant has produced a model for safe device-to-device (D2D) and
who shares the information about the jammed route can D2X communications.
receive tokens as a reward. It encourages users to exchange Utility-based blockchain security The deep learning
information quickly to claim the rewards. The proposed model can assist in safeguarding the blockchain from
study has used LSTM to predict the likelihood of traffic several types of attacks by adversaries. In decentralized
jams at a specific moment, location, and point. technology, a double-spending problem that represents a
Traffic flow prediction: Federated learning is a decen- flaw in existing cryptocurrency systems can occur, and it
tralized machine learning approach that trains an algorithm can enable the user to spend the same token two times.
on various devices that have local data samples. A Feder- According to Rosenfeld’s [103] model, the number of
ated learning network supported by blockchain technology verification required to keep the attacker’s effectiveness in
ensures the preservation of the privacy of the client data. double-spending below 10%, 1%, and 0.1% is 2, 4, and 6,
Blockchain has a vital role to play in such systems to respectively. In [100], a utility function-based approach has
ensure that the malicious participant can not share the false been proposed, in which utility function is calculated based
update; thus, the efficacy of the global model cannot be on the value of the product or service being traded against
compromised. Also, the federated learning framework the cryptocurrency [100]. The evaluation of such a utility
supported by blockchain enhances the accuracy in traffic function depends on the utilization of smart systems to
flow prediction and improves traffic management with infer the value of the product or service. Relying on the
enhanced user security and privacy. The researchers uti- cost of the commodity/service, this utility feature will
lized this technology along with Gated Recurrent Unit influence the attacker’s decision regarding attacking the
(GRU) neural network to process live traffic data for blockchain. This utility function once fed into machine

123
Cluster Computing

learning-based classification algorithms can assist in high-level design of a blockchain, deep learning, and IPFS-
determining whether or not an intrusion is likely to occur. based system that can automate healthcare, telecommuni-
As discussed in this section, it can be comprehended that cation, and transportation management services.
the majority of the deep learning systems utilize block-
chains for secure storage of data and easy accessibility. 4.2 Comparison of existing frameworks
Contrarily, some of the methods discussed the employment
of deep learning models to protect the blockchain from This subsection compares existing blockchain-based deep
foreign attacks. learning frameworks based on several parameters. The
Figure 8 discusses a federated learning-based system main parameters that are considered for the analysis
[24, 25] that securely shares the models constructed by the include solution category, blockchain type, consensus
devices based on the local data. Federated learning is a approach, deep learning methods, data set used by the deep
decentralized type of machine learning. In a traditional learning model, study strengths, and limitations. Table 2
machine learning approach, the data is captured at a cen- compares the research works discussed above in different
tralized device and used to learn and train the models. categories where deep learning and blockchain are playing
However, storing data on centralized devices affects the their part.
users’ data privacy. Federated learning is a decentralized The majority of the literature related to blockchain-as-
approach that allows users to train their models locally sisted machine learning frameworks have considered
using private data, thus preserving data privacy. After healthcare, the IoV, cellular traffic management, and
training the model locally, the training results (not data) are blockchain safety and protection fields. The frameworks
shared by each participating device with the centralized discussed in [68, 85], and [86] focused on creating machine
coordinator to update the global model. The centralized learning-based models for disease prediction and data fil-
coordinator will now update its central machine learning tration in the healthcare industry. However, the frame-
model based on the training results shared by each device. works discussed in [92, 93], and [94] have focused on
This central machine learning model is represented as a employing machine learning-based models and blockchain
global updated model in figure 8. This global model can be technology for position estimation and error reporting for
accessed by the participating devices. The role of block- vehicles that are part of the established IoV. The works
chain in this scenario is to coordinate between participants discussed in [95] and [96] have focused on employing
to share the local and global models. Figure 9 shows a

Fig. 8 Blockchain-based federated learning

123
 Cluster Computing

Fig. 9 Highlighting the system components and participants and describing the outcomes resulting from the integration of deep learning with
blockchain technology in various fields

machine learning-based models and blockchain technology presented in [24] has discussed how blockchain platforms
for cellular traffic management. can be fully adopted to verify the genuineness of an AI
Among [68, 85], and [86] frameworks, the study dis- model based on the model provenance data. The data about
cussed in [86] has considered LSTM deep learning model; the AI models stored on the blockchain in [24] includes
whereas [85] and [68] has considered one-shot learning training datasets, training results, model owner credentials,
with Siamese neural network and incremental learning, the source of data, and participants.
respectively. Moreover, the frameworks discussed in [85]
and [86] have selected the human protein atlas and EHR
based data set for model training purposes, respectively. 5 Research challenges and opportunities
Among [92, 93], and [94] studies, the work discussed in
[92] has considered Deep Boltzmann Machine deep This section briefly discusses several research challenges
learning models. On the other hand, [92] has considered and opportunities regarding the integration with deep
DNN and Reinforcement learning deep learning model. learning and blockchain technology. The major issues that
The MIT-BIH database dataset is considered for model hinder the successful integration of these technologies
training in [92]. The work discussed in [95] and [96] has include platform scalability, assurance of secure data
considered custom datasets for model training purposes. exchange, transaction execution latency, platform interop-
Also, [95] and [96] has followed CNN and DNN deep erability support, the large volume of data collected from
learning models for data classification. sensors and wearable devices, secure economical models,
The existing frameworks have implemented various and computationally expensive consensus protocols. In the
services related to privacy prevention, violation prediction, following section, the aforementioned challenges and
anomaly detection, forking prevention, and EHR forecast- opportunities are discussed.
ing in healthcare, cellular data traffic, and vehicular com-
munication networks [68, 85, 86, 92, 94, 95]. In a vehicular 5.1 Platform scalability
communication network, blockchain can be useful to
securely share the decision made by an AI model such as A major issue faced by architects designing deep learning
road accident detection or traffic jamming with the applications is the availability of multiple variants and
authorized participating entities. Such data is useful to configurations of blockchain technology. A scalable
efficiently manage road traffic by evenly distributing the blockchain platform can efficiently handle the large vol-
traffic on the roads [104]. In comparison to the works ume and velocity of transactions generated by diverse
presented in [68, 85, 86, 92, 94], and [95], the work users. A sizeable blockchain network would necessitate a

123
Cluster Computing

Table 2 A comparison and analysis of state-of-the-art blockchain-based deep learning frameworks

Category Blockchain Consensus Deep learning Dataset Study strengths Study limitations
type protocol method used

Ovarian cancer N/A N/A One-shot Human Requires less training Performance is not as good as
prediction [85] Learning Protein time as only one DNN
Atlas example per class is
required
Data exchange Private Proof-of- Incremental N/A The considered model Possibility of happening of
[68] Information Learning can learn new classes on catastrophic forgetting
a pre-trained network
EHR prediction N/A N/A LSTM EHR-based Superior performance on Longer training time and
[97] dataset sequential data excessive memory
requirement
Arrhythmia N/A N/A SDA? Sigmoid MIT-BIH Deals well with noise Requires large amount of data
classification Dataset and random variations in for better results
[91] the data
Miner Node N/A Zero- Deep Boltzmann N/A Data labelling is not Expensive in terms of memory
Selection [98] Knowledge Machine required and CPU cycles
Proof
Communication Private N/A DNN? N/A Implemented method Not suitable for simple
security [99] Reinforcement can solve complex problems
Learning problems that
conventional methods
cannot
Securing N/A Proof-of- N/A Game Data is fully secured This approach is not practical
blockchain Work Theory-
[100] based
Utility
Function
Traffic jam Public Proof-of- ANN?LSTM Historic ANN are fault tolerant Black box behavior makes it
prediction [95] Authority Traffic as information is impossible to develop a
Data ? distributed over all the relation between dependent
Custom nodes and indepen- dent variables
Dataset
Traffic flow Consortium Delegated GRU N/A Lesser expensive in Lesser learning ability
prediction [96] PBFT terms of memory compared to LSTM
requirement as
compared to LSTM
Incident N/A N/A CNN Custom Excellent feature Large dataset is required for
prediction [81] Dataset extraction cap- ability training and noise in data can
and computationally cause misclassification
efficient solution
GPS correction Public Delegated DNN Custom Reduced need of feature Not prone to data redundancy
[94] PoS Dataset engineering and inconsistency

comparable amount of accounts to implement deep learn- To carefully handle the scalability challenges of existing
ing-based services targeting healthcare [68], road traffic blockchain platforms, the compression algorithms having
jam [95], and traffic management in cellular networks. The lightweight design, high compression ratio, and resource
deployment of blockchain on such a massive scale will lead inexpensive nature should be integrated into the existing
to several problems mainly related to demand of users for blockchain-assisted deep learning solutions [105]. How-
internet connectivity, data velocity, speed, and volume of ever, many of the current compression algorithms are
transactions generated by participants. Considering the unable to fully provide the appropriate ratio necessary to
storage and computational requirements of the ever-grow- bring down the economic cost of the large-scale deploy-
ing blockchain ledger, the number of blocks and transac- ment to provide deep learning-based services using
tions to be added to the blockchain must be reduced blockchain.
considerably to satisfy the anticipated requests by the users.

123
 Cluster Computing

5.2 Data validity and secure sharing possible structural enhancement. Deep learning algorithms
not only solve the real-world object detection and classi-
Data sharing between the participants of the healthcare fication problems in various domains, but these models can
industry such as doctors, patients, and nurses can be kept also provide a way to compress the data with a higher
anonymous owing to state-of-the-art encryption and compression ratio [105, 109, 110]. However, blockchain
decryption techniques implemented by private blockchain represents an ever-growing ledger as new blocks keep on
platforms. With the exponential rise of IoT and wearable adding to the existing network. As the size of the block-
devices in healthcare and vehicular networks, consumers chain increases, the efficiency of the network is profoundly
are greatly concerned about data privacy, security, and affected. Deep learning approaches can be employed for
confidentiality. A multi-layer blockchain architecture that compressing the data as well as assisting in minimizing the
supports data fusion and allows advanced analytical redundant data. Since data stored on the blockchain is
authentication for user groups can assist in securely sharing permanent, hence, the growing size of the ledger is a big
data between the participants. Moreover, during the recent concern and needs to be addressed properly.
pandemic caused by COVID-19, the verification and secure
sharing of immunity passports [2, 106, 107] between the 5.4 Platform throughput and latency
authorized users through a blockchain platform has become
a must to meet requirements in the healthcare industry for Blockchain latency and throughput are important perfor-
developing a viable, secure, and fault-tolerant solution. mance metrics that have a significant impact on the tasks’
More specifically, the potential of blockchain in validating quality of service (QoS) [111]. In the blockchain, latency is
and archiving immutable data in real-time opens up the the time taken by the network to verify and execute a
possibility to ensure the authenticity of the data. The transaction in order to store it on the ledger. In the deep
blockchain platform serves as the basis for many research learning context, latency refers to the amount of time
projects. Particularly, it allows companies and organiza- required by the model to process a single data unit. High
tions, along with technologists and experts in sharing transaction execution latency is one of the major research
knowledge and validating the data in new systems in a challenges that should be addressed properly as it can
trusted and reliable way. However, the IoT devices used to affect the performance of processes that require quick
collect vehicular data can be faulty or inappropriately decision-making. For example, in the vehicular network,
deployed. As a result, the data generated by such devices is the decision-making of vehicles requires fast transaction
often incorrect, misleading, and unreliable. Metadata processing to minimize traffic jams or accidents. Forking
binding [108] is a technique that requires the IoT devices to [112] is another problem that occurs frequently as a result
calculate and send their accuracy and precision level along of long block propagation delays between miners. Block-
with collected data to the blockchain. Such metadata can be chain throughput refers to the number of transactions
analyzed by the deployed smart contracts to verify the executed in a unit of time. Every blockchain platform
validity of the data before storing it on the blockchain. offers different transaction execution throughput and
Also, in a few cases, a reputation-aware system can be latency. For instance, the Ethereum blockchain platform
advantageous in calculating the trust level of participating offers a throughput of 16.5 transactions per second
users based on their behaviors. Smart contracts can be [17, 30]. On the other hand, private blockchain platforms
deployed to ensure that data from only highly reputed such as Hyperledger Fabric can execute several thousand
entities is recorded on the blockchain. transactions in one second. Blockchain platforms with high
throughput are more suitable for deep learning-based
5.3 Structural enhancement and storage applications. A private blockchain network is more cen-
capacity tralized and comprises a limited number of nodes that are
distributed locally, so private platforms are much faster
Deep learning is regarded as a resource expensive and data- than public platforms. The consensus protocols imple-
intensive approach that assists in solving various real-world mented by the public platforms are one of the main reasons
problems. Blockchain technology aims to ensure that the for the low transaction execution speed on the public
data stored on it is highly secure, immutable, verifiable, blockchain platforms.
transparent, and visible to the authorized stakeholders. The
deep learning-based approach which can systematically 5.5 Cryptocurrencies, deep learning-based
evaluate the performance of blockchain platforms is eco- consensus protocols, and regulations
nomically desirable. Such an approach can provide a means
to analyze the current structure of the blockchain for The cryptocurrency represents a virtual currency that can-
not be double-spent or counterfeited due to blockchain

123
Cluster Computing

network decentralization and data security that is ensured methodology can be used for storing the perceptive infor-
using encryption and decryption techniques which are mation, which comes at the cost of enhanced delays. The
based on one-way hashing functions. In comparison to interoperable blockchain platforms enable the participants
conventional physical currencies, cryptocurrency is a of healthcare and vehicular communication networks to
decentralized token that is not regulated by any govern- share data and information uninterruptably, securely,
ment agency. Due to the unlimited advantages of cryp- quickly, and seamlessly [2, 32]. The platform interoper-
tocurrency, it has become an investment option similar to ability is affected by many factors such as the choice of
the stock exchange. Supervised deep learning models can blockchain-supported languages, consensus protocols,
be used to predict the future of cryptocurrencies. A few cryptographic hashing algorithms, and the type of data
studies [113–115] have employed machine learning and being used by the participants.
deep reinforcement learning techniques to predict the price
of crypto. However, due to the technical incompetence of 5.7 Secure economical models
deep learning models in text classification and prediction,
the majority of the deep learning models remain imperfect The emerging deep learning models have shown unlimited
and do not provide fruitful results. Hence, there is a opportunities and potential in various fields due to tech-
potential research gap in the field of crypto price prediction nological advancements in existing computing technolo-
that can be fulfilled using deep learning-based models. gies and storage systems. More specifically, in the
Moreover, many deep learning applications require the healthcare sector, deep learning models are used to classify
models to make quick predictions as the lag in the outcome normal patients from unhealthy ones or predict the spread
may cause unwanted consequences. Currently, the generic of a disease in a particular community or region. On the
blockchain consensus algorithms require time on the scale other hand, blockchain technology is employed for ensur-
of seconds [116]. Deep learning-specific consensus proto- ing data security and integrity. In a network of vehicles, the
cols can be designed based on the proofs related to the smart vehicles which are equipped with deep learning
quality of the data, optimization techniques, nature of the models can share the data with the neighboring vehicles
learning models, and total convergence time of the model. using blockchain to avoid traffic-jamming-related issues
Lack of standards and regulatory frameworks is another [117]. However, such models have shown limitations in
challenge to the existing blockchain technology and it terms of high cost and less resource efficiency. Currently,
affects blockchain adaptability to deep learning frame- existing deep learning models require high-performing
works [2, 3, 3]. Extensive research is needed in this computing devices for training purposes [118]; and, the
direction to propose standards and regulatory frameworks blockchain is an expensive storage medium. More research
for blockchain technology. is needed to propose cost-efficient, resource-friendly, fast,
and high-performance-based blockchain-assisted deep
5.6 High-speed computing/storage devices learning frameworks.
and platform interoperability

Every node in a blockchain network maintains an up-to- 6 Conclusion

date copy of the ledger, thereby enabling fault tolerance in
the system. The escalating storage requirements of block- In this paper, we have reviewed the state-of-the-art
chain nodes can affect the scalability of the blockchain blockchain-based deep learning frameworks. We presented
network. The leveraging of centralized-based high-speed the key features of blockchain and deep learning along with
data storage and processing devices can lead to several a detailed discussion on the benefits resulted from their
security concerns. The potential of such high-speed storage integration. The successful integration of deep learning
devices can be exploited when they are employed as with blockchain can facilitate in terms of data security and
blockchain nodes. However, this method is not economi- privacy to the existing systems and enhance the QoS in
cally feasible due to high configuration cost of nodes. Also, several applications mainly related to healthcare, block-
the failure of centralized storage devices can lead to the chain security, data traffic management, and vehicular
breakdown of the deep learning services. The spike in communication in urban areas. We devised a taxonomy to
management costs and complicity is another downside of categorize the reported literature in several categories
using external high-speed storage devices. Distributed file based on seven parameters such as blockchain type, deep
system such as IPFS [33] or Swarm [35] provides a learning models, deep learning specific consensus proto-
decentralized storage system, but the sensitive data can be cols, services, application areas, deployment goals, and
compromised as data on such storage systems are available data types. The critical aspects of existing blockchain-
to the public. However, the encryption-decryption based deep learning frameworks are analyzed through a

123
 Cluster Computing

comprehensive analysis of the reported frameworks. 2. Ahmad, R.W., Salah, K., Jayaraman, R., Yaqoob, I., Ellahham,
Finally, we identified and discussed several technological S., Omar, M.: The role of blockchain technology in telehealth
and telemedicine. Int. J. Med. Inf. 148, 104399 (2021)
and social challenges and barriers that require further 3. Ahmad, R.W., Hasan, H., Yaqoob, I., Salah, K., Jayaraman, R.,
research to unlock the full potential of blockchain in deep Omar, M.: Blockchain for aerospace and defense: opportunities
learning-based systems. Our concluding remarks along and open research challenges. Comput. Ind. Eng. 151, 106982
with the key recommendations include: (2021)
4. Oh, Y., Park, S., Ye, J.C.: Deep learning COVID-19 features on
• Data traceability, immutability, and integrity features of CXR using limited training data sets. IEEE Trans. Med. imaging
39(8), 2688–2700 (2020)
blockchain technology can assist in identifying the
5. Shuja, J., Alanazi, E., Alasmary, W., Alashaikh, A.: COVID-19
volume and type of data collected to train deep learning open source data sets: a comprehensive survey. Appl. Intell.
models. However, the existing blockchain-based sys- 51(3), 1296–1325 (2021)
tems are incapable of efficiently handling data quality 6. Shafay, M., Hassan, T., Velayudhan, D., Damiani, E., Werghi,
N.: Deep fusion driven semantic segmentation for the automatic
problems, particularly in the healthcare and transporta-
recognition of concealed contraband items. In: SoCPaR, 2020,
tion industries. pp. 550–559
• The key performance metrics such as system through- 7. Hassan, T., Shafay, M., Akçay, S., Khan, S., Bennamoun, M.,
put, execution latency, and block propagation time, data Damiani, E., Werghi, N.: Meta-transfer learning driven tensor-
shot detector for the autonomous localization and recognition of
volume, conflicting interests of participants, and smart
concealed baggage threats, Nov 2020. https://www.mdpi.com/
contract vulnerabilities can significantly impact the 1424-8220/20/22/6450
effectiveness of existing blockchain-based deep learn- 8. Berman, D.S., Buczak, A.L., Chavis, J.S., Corbett, C.L.: A
ing systems. survey of Deep learning methods for cyber security. Information
10(4), 122 (2019)
• Private blockchain platforms ensure data privacy
9. Lawrence, S., Giles, C.L.: Overfitting and Neural networks:
through private channels and access control policies. conjugate gradient and backpropagation. In: Proceedings of the
However, public blockchain platforms are prone to data IEEE-INNS-ENNS International Joint Conference on Neural
privacy leakage problems because they have a zero- Networks. IJCNN 2000. Neural Computing: New Challenges
and Perspectives for the New Millenniu, vol. 1. IEEE, 2000,
access control policy. Public blockchain platforms can
pp. 114–119
successfully capture the evolution of deep learning 10. Al Ridhawi, I., Aloqaily, M., Jararweh, Y.: An incentive-based
models as it progresses and records the model state mechanism for volunteer computing using blockchain. ACM
during its creation, updating, or usage stages. Trans. Internet Technol. 21(4), 1–22 (2021)
11. Tan, L., Xiao, H., Yu, K., Aloqaily, M., Jararweh, Y.: A
• The efficiency of blockchain-based applications is
blockchain-empowered crowdsourcing system for 5g-enabled
highly affected by increasing the size of the blockchain smart cities. Comput. Stand. Interfaces 76, 103517 (2021)
network. Deep learning approaches can be employed 12. Narayan, S., Tagliarini, G.: An analysis of underfitting in MLP
for compressing the data as well as minimizing the networks. In: Proceedings. 2005 IEEE International Joint Con-
ference on Neural Networks, 2005., vol. 2. IEEE, 2005,
redundant data.
pp. 984–988
13. Shuja, J., Bilal, K., Alasmary, W., Sinky, H., Alanazi, E.:
Acknowledgements This work was supported by the Khalifa Applying machine learning techniques for caching in edge
University of Science and Technology under Award RCII-2019-002- networks: a comprehensive survey, arXiv preprint arXiv:2006.
Research Center for Digital Supply Chain and Operations Manage- 16864, 2020
ment and CIRA-2019-001. 14. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.,
Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell A.,
Data availability Not applicable. et al.: Language models are few-shot learners. arXiv preprint
arXiv:2005.14165, 2020
15. Wang, L., Von Laszewski, G., Younge, A., He, X., Kunze, M.,
Declarations Tao, J., Fu, C.: Cloud computing: a perspective study. New Gen.
Comput. 28(2), 137–146 (2010)
Conflict of interest The authors have not disclosed any competing 16. Shiraz, M., Gani, A., Ahmad, R.W., Shah, S.A.A., Karim, A.,
interests. Rahman, Z.A.: A lightweight distributed framework for com-
putational offloading in mobile cloud computing. PLoS ONE
9(8), e102270-9 (2014)
References 17. Ahmad, R.W., Hasan, H., Jayaraman, R., Salah, K., Omar, M.:
Blockchain applications and architectures for port operations
and logistics management. Res. Transp. Business Manag.
1. Ayyoubzadeh, S.M., Ayyoubzadeh, S.M., Zahedi, H., Ahmadi, (2021). https://doi.org/10.1016/j.rtbm.2021.100620
M., Kalhori, S.R.N.: Predicting COVID-19 incidence through 18. Bach, L.M., Mihaljevic, B., Zagar, M.: Comparative analysis of
analysis of google trends data in Iran: data mining and deep blockchain consensus algorithms. In: 41st International Con-
learning pilot study. JMIR Public Health Surv. 6(2), e18828 vention on Information and Communication Technology, Elec-
(2020) tronics and Microelectronics (MIPRO). IEEE 2018,
pp. 1545–1550 (2018)

123
Cluster Computing

19. Zheng, Z., Xie, S., Dai, H.-N., Chen, W., Chen, X., Weng, J., 40. Hassan, T., Hassan, B., El-Baz, A., Werghi, N.: A dilated
Imran, M.: An overview on smart contracts: Challenges, residual hierarchically fashioned segmentation framework for
advances and platforms. Fut. Gen. Comput. Syst. 105, 475–491 extracting Gleason tissues and grading prostate cancer from
(2020) whole slide images, 2021
20. Ren, W., Hu, J., Zhu, T., Ren, Y., Choo, K.-K.R.: A flexible 41. Sheridan, T.B.: Human-robot interaction: status and challenges.
method to defend against computationally resourceful miners in Hum Factors 58(4), 525–532 (2016)
blockchain Proof-of-work. Inf. Sci. 507, 161–171 (2020) 42. Mark, R.: Publish Date: May 27, The difference between Arti-
21. Vashchuk, O., Shuwar, R.: Pros and cons of consensus algorithm ficial intelligence, Machine learning and Deep learning. https://
Proof-of-stake: difference in the network safety in Proof-of- www.intel.la/content/www/xl/es/artificial-intelligence/posts/dif
work and proof-of-stake. Electron. Inf. Technol. 9(9), 106–112 ference-between-ai-machine-learning-deep-learning.html
(2018) 43. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas,
22. Singh, P.K., Singh, R., Nandi, S.K., Nandi, S.: Managing smart B.A.: Communication-efficient learning of deep networks from
home appliances with Proof-of-authority and blockchain. In: decentralized data. In: Artificial intelligence and statistics.
International Conference on Innovations for Community Ser- PMLR, 2017, pp. 1273–1282
vices. Springer, Berlin, 2019, pp. 221–232 44. Hassan, T., Akcay, S., Bennamoun, M., Khan, S., Werghi, N.: A
23. Jerry Cuomo, P.: How blockchain adds trust to AI and IoT, Aug novel incremental learning driven instance segmentation
2020. https://www.ibm.com/blogs/blockchain/2020/08/how- framework to recognize highly cluttered instances of the con-
blockchain-adds-trust-to-ai-and-iot/ traband items. IEEE Transactions on Systems, Man, and
24. Sarpatwar, K., Vaculin, R., Min, H., Su, G., Heath, T., Gana- Cybernetics: Systems, To be published, 2022
pavarapu, G., Dillenberger, D.: Towards enabling trusted arti- 45. Hassan, T., Hassan, B., Akram, M.U., Hashmi, S., Taguri, A.H.,
ficial intelligence via blockchain. In: Policy-based autonomic Werghi, N.: Incremental cross-domain adaptation for robust
data governance. Springer, Berlin, 2019, pp. 137–153 retinopathy screening via Bayesian deep learning. IEEE Trans.
25. Shinde, R., Patil, S., Kotecha, K., Ruikar, K.: Blockchain for Instrum. Measur. 70, 1–14 (2021)
securing ai applications and open innovations. J. Open Innov. 46. Hassan, T., Aslam, S., Jang, J.W.: Fully automated multi-reso-
7(3), 189 (2021) lution channels and multithreaded spectrum allocation protocol
26. Salah, K., Rehman, M.H.U., Nizamuddin, N., Al-Fuqaha, A.: for iot based sensor nets. IEEE Access 6, 545–556 (2018)
Blockchain for AI: review and open research challenges. IEEE 47. Mahdavifar, S., Ghorbani, A.A.: Application of deep learning to
Access 7, 127–149 (2019) cybersecurity: a survey. Neurocomputing 347, 149–176 (2019)
27. Tanwar, S., Bhatia, Q., Patel, P., Kumari, A., Singh, P.K., Hong, 48. Salloum, S.A., Alshurideh, M., Elnagar, A., Shaalan, K.:
W.-C.: Machine learning adoption in blockchain-based smart Machine learning and deep learning techniques for cybersecu-
applications: the challenges, and a way forward. IEEE Access 8, rity: a review. AICV 2020, 50–57 (2020)
474–488 (2020) 49. P. by S. O’Dea and A. 6, Smartphone users 2026, Aug 2021.
28. Nakamoto, S.: Bitcoin: a peer-to-peer electronic cash system, https://www.statista.com/statistics/330695/number-of-smart
Decentralized Business Review, p. 21260, 2008 phone-users-worldwide/
29. Yeow, K., Gani, A., Ahmad, R.W., Rodrigues, J.J., Ko, K.: 50. He, C., Tan, C., Tang, H., Qiu, S., Liu, J.: Central server free
Decentralized consensus for edge-centric internet of things: a federated learning over single-sided trust social networks, arXiv
review, taxonomy, and research issues. IEEE Access 6, preprint arXiv:1910.04956, 2019
1513–1524 (2017) 51. Marr, B.: Artificial intelligence and blockchain: 3 major benefits
30. Wood, D.D.: Ethereum: a secure decentralised generalised of combining these two mega-trends, Mar 2018. https://www.
transaction ledger, 2014 forbes.com/sites/bernardmarr/2018/03/02/artificial-intelligence-
31. Sookhak, M., Jabbarpour, M.R., Safa, N.S., Yu, F.R.: Block- and-blockchain-3-major-benefits-of-combining-these-two-
chain and smart contract for access control in healthcare: a mega-trends/?sh=604fcaa04b44
survey, issues and challenges, and open issues. J. Netw. Comput. 52. Campbell, D.: Combining AI and blockchain to push frontiers in
Appl. 178, 102950 (2021) healthcare, Nov 2018. https://www.macadamian.com/learn/com
32. Ahmad, R.W., Salah, K., Jayaraman, R., Yaqoob, I., Ellahham, bining-ai-and-blockchain-in-healthcare/
S., Omar, M.: Blockchain and COVID-19 pandemic: applica- 53. Castelló Ferrer, E.: The blockchain: a new framework for
tions and challenges, IEEE TechRxiv, 2020 robotic Swarm systems, Advances in Intelligent Systems and
33. Benet, J.: IPFS - Content Addressed, versioned, P2P file system, Computing, p. 1037–1058, Oct 2018. http://dx.doi.org/10.1007/
2014 978-3-030-02683-7_77
34. Lakshman, A., Malik, P.: Cassandra: a decentralized structured 54. Janson, S., Merkle, D., Middendorf, M.: A decentralization
storage system. SIGOPS Oper. Syst. Rev. 44(2), 35–40 (2010). approach for swarm intelligence algorithms in networks applied
https://doi.org/10.1145/1773912.1773922 to multi Swarm PSO. Int. J. Intell. Comput. Cybern. 1(1), 25–45
35. Hartman, J., Murdock, I., Spalink, T.: The Swarm scalable (2008). https://doi.org/10.1108/17563780810857112
storage system. In: Proceedings of 19th IEEE International 55. Hassan, K., Tahir, F., Rehan, M., Ahn, C.K., Chadli, M.: On
Conference on Distributed Computing Systems (Cat. relative-output feedback approach for group consensus of clus-
No.99CB37003), 1999, pp. 74–81 ters of multiagent systems. IEEE Trans. Cybern. 1–12, 2021
36. Wilkinson, S., Boshevski, T., Brandoff, J., Buterin, V.: Storj: a (2021)
peer-to-peer cloud storage network, 2014 56. Magazzeni, D., McBurney, P., Nash, W.: Validation and veri-
37. Chan, T.-H., Jia, K., Gao, S., Lu, J., Zeng, Z., Ma, Y.: PCANet: fication of smart contracts: a research agenda. Computer 50(9),
a simple deep learning baseline for image classification? IEEE 50–57 (2017)
Trans. Image Process. 24(12), 5017–5032 (2015) 57. Open source P2P digital currency. https://litecoin.org/
38. Zhao, Z.-Q., Zheng, P., Xu, S.-T., Wu, X.: Object detection with 58. Dinh, T.T.A., Wang, J., Chen, G., Liu, R., Ooi, B.C., Tan, K.-L.:
deep learning: a review. IEEE Trans. Neural Netw. Learn. Syst. BLOCKBENCH: A Framework for Analyzing Private Block-
30(11), 3212–3232 (2019) chains, 2017
39. Daily, M., Medasani, S., Behringer, R., Trivedi, M.: Self-driving
cars. Computer 50, 18–23 (2017)

123
 Cluster Computing

59. Li, Z., Kang, J., Yu, R., Ye, D., Deng, Q., Zhang, Y.: Consor- SIGSAC Conference on Computer and Communications Secu-
tium blockchain for secure energy trading in industrial internet- rity, 2019, pp. 2561–2563
of-things. IEEE Trans. Ind. Inf. 14(8), 3690–3700 (2018) 79. Philip, A.O., Saravanaguru, R.K.: Secure incident & evidence
60. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classifi- management framework (SIEMF) for internet of vehicles using
cation with deep convolutional neural networks. Commun. deep learning and blockchain. Open Comput. Sci. 10(1),
ACM 60(6), 84–90 (2017). https://doi.org/10.1145/3065386 408–421 (2020)
61. Rumelhart, D.E., McClelland, J.L.: Learning Internal Repre- 80. Scicchitano, F., Liguori, A., Guarascio, M., Ritacco, E., Manco,
sentations by Error Propagation, pp. 318–362 (1987) G.: A deep learning approach for detecting security attacks on
62. Hochreiter, S., Schmidhuber, J.: Long short-term memory. blockchain. In: ITASEC, 2020, pp. 212–222
Neural Comput. 9(8), 1735–1780 (1997). https://doi.org/10. 81. Kurri, V., Raja, V., Prakasam, P.: Cellular traffic prediction on
1162/neco.1997.9.8.1735 blockchain-based mobile networks using LSTM model in 4G
63. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evalu- LTE network. Peer-to-Peer Netw. Appl. 14(3), 1088–1105
ation of gated recurrent neural networks on sequence modeling. (2021)
In: NIPS 2014 Workshop on Deep Learning, December 2014 82. Kim, H., Park, J., Bennis, M., Kim, S.-L.: Blockchained on-
64. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde- device federated learning. IEEE Commun. Lett. 24(6),
Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative 1279–1283 (2019)
Adversarial Nets, Ser NIPS’14, pp. 2672–2680. MIT Press, 83. Vora, J., Nayyar, A., Tanwar, S., Tyagi, S., Kumar, N., Obaidat,
Cambridge (2014) M.S., Rodrigues, J.J.: BHEEM: a blockchain-based framework
65. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., et al.: for securing electronic health records. In: IEEE Globecom
Human-level control through deep reinforcement learning. Workshops (GC Wkshps). IEEE 2018, pp. 1–6 (2018)
Nature 518(7540), 529–533 (2015). https://doi.org/10.1038/ 84. Omaar, J., Schwerin, S., McMullen, G.: Forever isn’t free: The
nature14236 cost of storage on a blockchain database. With assistance of
66. Bronstein, M.M., Bruna, J., LeCun, Y., Szlam, A., Van- Simon Schwerin, McMullen Greg. https://medium.com/ipdb-
dergheynst, P.: Geometric deep learning: going beyond Eucli- blog/forever-isnt-freethe-cost-of-storage-on-a-blockchain-data
dean data. IEEE Signal Process. Mag. 34(4), 18–42 (2017) base-59003f63e01, vol. 2, no. 28, p. 2019, 2017
67. Merlina, A.: BlockML: a useful proof of work system based on 85. Abraham, M., Vyshnavi, A., Srinivasan, C., Namboori, P.:
machine learning tasks. In: Proceedings of the 20th International Healthcare security using blockchain for pharmacogenomics.
Middleware Conference Doctoral Symposium, 2019 J. Int. Pharm. Res. 46, 529–533 (2019)
68. Kuo, T.-T., Ohno-Machado, L.: Modelchain: decentralized pri- 86. Zheng, X., Mukkamala, R.R., Vatrapu, R., Ordieres-Mere, J.:
vacy-preserving healthcare predictive modeling framework on Blockchain-based personal health data sharing system using
private blockchain networks, arXiv preprint arXiv:1802.01746, cloud storage. In: 2018 IEEE 20th International Conference on
2018 e-Health Networking, Applications and Services (Healthcom).
69. Li, Y., Chen, C., Liu, N., Huang, H., Zheng, Z., Yan, Q.: A IEEE, 2018, pp. 1–6
blockchain-based decentralized federated learning framework 87. Liang, G., Zheng, L.: A transfer learning method with deep
with committee consensus. IEEE Netw. 35(1), 234–241 (2020) residual network for pediatric pneumonia diagnosis. Comput.
70. Bravo-Marquez, F., Reeves, S., Ugarte, M.: Proof-of-learning: a Methods Progr. Bomed. 187, 104964 (2020)
blockchain consensus mechanism based on machine learning 88. Dutta, A., Batabyal, T., Basu, M., Acton, S.T.: An efficient
competitions. In: 2019 IEEE International Conference on convolutional neural network for coronary heart disease pre-
Decentralized Applications and Infrastructures (DAPPCON). diction. Expert Syst. Appl. 159, 113408 (2020)
IEEE, 2019, pp. 119–124 89. Hasan, N.I., Bhattacharjee, A.: Deep learning approach to car-
71. Lan, Y., Liu, Y., Li, B.: Proof of Learning (PoLe): Empowering diovascular disease classification employing modified ECG
machine learning with consensus building on blockchains, 2020 signal from empirical mode decomposition. Biomed. Signal
72. Salah, K., Alfalasi, A., Alfalasi, M., Alharmoudi, M., Alzaabi, Process. Control 52, 128–140 (2019)
M., Alzyeodi, A., Ahmad, R.W.:‘Iot-enabled shipping container 90. Arunkumar, R., Karthigaikumar, P.: Multi-retinal disease clas-
with environmental monitoring and location tracking. In: IEEE sification by reduced deep learning features. Neural Compu.
17th Annual Consumer Communications & Networking Con- Appl. 28(2), 329–334 (2017)
ference (CCNC). IEEE 2020, pp. 1–6 (2020) 91. Juneja, A., Marefat, M.: Leveraging blockchain for retraining
73. Wang, C., Cheng, X., Li, J., He, Y., Xiao, K.: A survey: deep learning architecture in patient-specific arrhythmia classi-
applications of blockchain in the internet of vehicles. EURASIP fication. In: 2018 IEEE EMBS International Conference on
J. Wirel. Commun. Netw. 2021(1), 1–16 (2021) Biomedical & Health Informatics (BHI). IEEE, 2018,
74. Dimitrakopoulos, G., Demestichas, P.: Intelligent transportation pp. 393–397
systems. IEEE Vehicular Technol. Mag. 5(1), 77–84 (2010) 92. Rohani, M., Gingras, D., Gruyer, D.: A novel approach for
75. Ni, J., Lin, X., Zhang, K., Shen, X.: Privacy-preserving real-time improved vehicular positioning using cooperative map matching
navigation system using vehicular crowdsourcing. In: IEEE 84th and dynamic base station DGPS concept. IEEE Trans. Intell.
Vehicular Technology Conference (VTC-Fall). IEEE 2016, Transp. Syst. 17(1), 230–239 (2015)
pp. 1–5 (2016) 93. Wang, S., Sun, S., Wang, X., Ning, Z., Rodrigues, J.J.: Secure
76. Xie, B., Chen, Y., Xu, M.: Evaluating urban traffic jam based on crowdsensing in 5G internet of vehicles: When deep reinforce-
a urban cell transmission model (UCTM). In: 2012 12th Inter- ment learning meets blockchain. In: IEEE Consumer Electronics
national Conference on ITS Telecommunications. IEEE, 2012, Magazine, 2020
pp. 211–215 94. Li, C., Fu, Y., Yu, F.R., Luan, T.H., Zhang, Y.: Vehicle position
77. Siddiqui, S.T., Ahmad, R., Shuaib, M., Alam, S.: Blockchain correction: a vehicular blockchain networks-based GPS error
security threats, attacks and countermeasures. Adv. Intell. Syst. sharing framework. IEEE Trans. Intell. Transp. Syst. 22(2),
Comput. 1097, 51–62 (2020) 898–912 (2020)
78. Awan, S., Li, F., Luo, B., Liu, M.: Poster: a reliable and 95. Hassija, V., Gupta, V., Garg, S., Chamola, V.: Traffic jam
accountable privacy-preserving federated learning framework probability estimation based on blockchain and deep neural
using the blockchain. In: Proceedings of the 2019 ACM networks. IEEE Trans. Intell. Transp. Syst. 22, 7 (2020)

123
Cluster Computing

96. Liu, Y., James, J., Kang, J., Niyato, D., Zhang, S.: Privacy- 114. Jiang, Z., Liang, J.: Cryptocurrency portfolio management with
preserving traffic flow prediction: a federated learning approach. deep reinforcement learning. In: Intelligent Systems Conference
IEEE Internet Things J. 7(8), 7751–7763 (2020) (IntelliSys). IEEE 2017, 905–913 (2017)
97. Bhattacharya, P., Tanwar, S., Bodke, U., Tyagi, S., Kumar, N.: 115. Lahmiri, S., Bekiros, S.: Cryptocurrency forecasting with deep
Bindaas: blockchain-based deep-learning as-a-service in learning chaotic neural networks. Chaos Solitons Fractals 118,
healthcare 4.0 applications. IEEE Trans. Netw. Sci. Eng. 8, 2 35–40 (2019)
(2019) 116. Sankar, L.S., Sindhu, M., Sethumadhavan, M.: Survey of con-
98. Singh, M., Aujla, G.S., Bali, R.S.: A deep learning-based sensus protocols on blockchain applications. In :2017 4th
blockchain mechanism for secure Internet-of-Drones environ- International Conference on Advanced Computing and Com-
ment. IEEE Trans. Intell. Transp. Syst. 22(7), 4404–4413 (2021) munication Systems (ICACCS). IEEE, 2017, pp. 1–5
99. Nguyen, D., Pathirana, P., Ding, M., Seneviratne, A.: Secure 117. Yang, F., Wang, S., Li, J., Liu, Z., Sun, Q.: An overview of
computation offloading in blockchain based IoT networks with internet of vehicles. China Commun. 11(10), 1–15 (2014)
deep reinforcement learning. IEEE Trans. Netw. Sci. Eng. 118. Strubell, E., Ganesh, A., McCallum, A.: Energy and policy
(2021). https://doi.org/10.48550/arXiv.1908.07466 considerations for deep learning in nlp. arXiv preprint arXiv:
100. Dey, S.: Securing majority-attack in blockchain using machine 1906.02243, 2019
learning and algorithmic game theory: a proof of work. In: 10th
Computer Science and Electronic Engineering (CEEC). IEEE
2018, pp. 7–10 (2018)
101. Rimal, B.P., Van, D.P., Maier, M.: Mobile-edge computing vs. Muhammad Shafay received his
centralized cloud computing in fiber-wireless access networks. undergraduate degree in Elec-
In: 2016 IEEE Conference on Computer Communications trical Engineering from PIEAS,
Workshops (INFOCOM WKSHPS). IEEE, 2016, pp. 991–996 Pakistan. He is currently pursu-
102. Zhang, J., Chen, B., Zhao, Y., Cheng, X., Hu, F.: Data security ing a Master’s degree in elec-
and privacy-preserving in Edge computing paradigm: survey trical engineering and computer
and open issues. IEEE Access 6, 209–237 (2018) science with Khalifa University,
103. Rosenfeld, M.: Analysis of Hashrate-based double spending. UAE. His research interests
arXiv preprint arXiv:1402.2009, 2014 include deep learning, block-
104. Rehman, A., Rathore, M.M., Paul, A., Saeed, F., Ahmad, R.W.: chain, computer vision, and
Vehicular traffic optimisation and even distribution using ant image processing.
colony in smart city environment. IET Intell. Transp. Syst.
12(7), 594–601 (2018)
105. Ni, F., Zhang, J., Noori, M.N.: Deep learning for data anomaly
detection and data compression of a long-span suspension
bridge. Comput. Aid. Civ. Infrastruct. Eng. 35(7), 685–700
(2020) Raja Wasim Ahmad received his
106. Ouellette, M., Shaw, M.L.: For a decentralized vaccine passport, Ph.D. in Computer Science
Health Policy, 2021 from the Center for Mobile
107. Khalid, T., Abbasi, M.A.K., Zuraiz, M., Khan, A.N., Ali, M., Cloud Computing Research
Ahmad, R.W., Rodrigues, J.J., Aslam, M.: A survey on privacy (C4MCCR), Faculty of Com-
and access control schemes in fog computing. Int. J. Commun. puter Science and Information
Syst. 34(2), e4181 (2021) Technology (FSKTM), Univer-
108. Wrona, K., Jarosz, M.: Use of blockchains for secure binding of sity of Malaya, Malaysia. He
metadata in military applications of iot. In: IEEE 5th World completed his masters in com-
Forum on Internet of Things (WF-IoT). IEEE 2019, pp. 213–218 puter science from COMSATS
(2019) University Islamabad (CUI),
109. Abbasi, G.A., Tiew, L.Y., Tang, J., Goh, Y.-N., Thurasamy, R.: Abbottabad campus, in 2011.
The adoption of cryptocurrency as a disruptive force: Deep Currently, he is working as a
learning-based dual stage structural equation modelling and Postdoctoral research fellow at
artificial neural network analysis. PLoS ONE 16(3), e0247582 Khalifa University, Abu Dhabi,
(2021) UAE. He has published several research papers in leading journals
110. Demertzis, K., Iliadis, L., Tziritas, N., Kikiras, P.: Anomaly and conferences. Most of his research contributions are published in
detection via blockchained deep learning smart contracts in top-cited journals such as IEEE Transactions on Emerging Topics in
industry 4.0. Neural Comput. Appl. 32(23), 361–378 (2020) Computing, IEEE Systems Journal, IEEE Access, Journal of Net-
111. Irshad, T., Shan, R.-U., Ahmad, R.W., Khalid, A., Ab Hamid, works and Computer Applications, Journal of Grid Computing,
S.H.: Multi-rat based adaptive quality of service (QOS) man- Cluster Computing, International Journal of Communication Systems,
agement in WBAN. Malays. J. Comput. Sci. 33(4), 252–269 International Journal of Information Management, Journal of Systems
(2020) and Software, Journal of Supercomputing, IET Intelligent Transport
112. da Silva, F.J.C., Damsgaard, S.B., Sorensen, M.A.M., Marty, F., Systems, and Renewable and Sustainable Energy Reviews. Dr. Raja is
Altariqi, B., Chatzigianni, E., Madsen, T.K., Schwefel, H.P.: also serving as a reviewer for several journals and conferences. His
Analysis of blockchain forking on an ethereum network. In: research work is well acknowledged in national and international
European Wireless,: 25th European Wireless Conference. VDE conferences. One of his research articles has received an award of the
2019, pp. 1–6 (2019) best paper in SPECTS symposium held in 2018 in France.
113. Patel, M.M., Tanwar, S., Gupta, R., Kumar, N.: A deep learning-
based cryptocurrency price prediction scheme for financial
institutions. J. Inf. Secur. Appl. 55, 102583 (2020)

123
 Cluster Computing

Khaled Salah is a full Professor Raja Jayaraman is an Associate

at the Department of Electrical Professor in the Department of
and Computer Engineering, Industrial & Systems Engineer-
Khalifa University, UAE. He ing at Khalifa University, Abu
received the B.S. degree in Dhabi, UAE. He received his
Computer Engineering with a Ph.D. in Industrial Engineering
minor in Computer Science from Texas Tech University, a
from Iowa State University, Master of Science degree in
USA, in 1990, the M.S. degree Industrial Engineering from
in Computer Systems Engi- New Mexico State University, a
neering from Illinois Institute of Master and Bachelors in Math-
Technology, USA, in 1994, and ematics from India. His exper-
the Ph.D. degree in Computer tise is in multi-criteria
Science from the same institu- optimization techniques applied
tion in 2000. Khaled has over to diverse applications including
220 publications and 3 US patents, has been giving a number of supply chain and logistics, healthcare, energy, environment and sus-
international keynote speeches, invited talks, tutorials, and research tainability. Raja’s research interests are primarily focused in using
seminars on the subjects of Blockchain, IoT, Fog and Cloud Com- blockchain technology, systems engineering and process optimization
puting, and Cybersecurity. Khaled served as the chair of the Track techniques to characterize, model and analyze complex systems with
Chair of IEEE Globecom 2018 on Cloud Computing. Khaled is an applications to supply chains, maintenance operations planning and
Associate Editor of IEEE Blockchain Tech Briefs, and a member of healthcare delivery. His post doctorial research was centered on
IEEE Blockchain Education Commitee. Khaled is now leading a technology adoption and implementation of innovative practices in
number of projects on how to leverage blockchain for Healthacare, the healthcare supply chains and service delivery. He has led several
5G Networks, Combating Deepfake Videos, Supply Chain Manage- successful research projects and pilot implementations in the area of
ment, and AI. supply chain data standards adoption in the US healthcare system. His
research has appeared in top rated journals including: Annals of
Ibrar Yaqoob (S’16-M’18- Operations Research, IISE Transactions, Energy Policy, Applied
SM’19) is a research scientist at Energy, Knowledge Based Systems, IEEE Access, Journal of Theo-
the Khalifa University, where he retical Biology, Engineering Management Journal and others
had joined as a postdoctoral
fellow in October 2019. Previ- Mohammed Omar is currently a
ously, he worked as a research full Professor and the Founding
professor at the Department of Chair of the Department of
Computer Science and Engi- Engineering Systems and Man-
neering, Kyung Hee University, agement (currently renamed
South Korea, where he com- Industrial and Systems Engi-
pleted his postdoctoral fellow- neering). Prior to joining the
ship. Prior to that, he received Masdar Institute/KUST, he was
his Ph.D. (Computer Science) an Associate Professor and a
from the University of Malaya, Graduate Coordinator with
Malaysia, in 2017. He worked Clemson University, Clemson,
as a research assistant at the Centre for Mobile Cloud Computing SC, USA. He was a part of the
Research (C4MCCR), University of Malaya. His numerous research Founding Faculty Cohort of
articles are very famous and among the most downloaded in top Clemson University research
journals. He has been selected as a highly cited researcher (HCR) park in Greenville, SC, USA.
worldwide for the year 2021 by Clarivate (Web of Science). He is He has over 100 publications in the areas of product lifecycle man-
currently serving as an editor for various IEEE, Elsevier, and Springer agement, knowledge-based manufacturing, and automated testing
journals. He has been involved in a number of conferences and systems, in addition to authoring several books and book chapters. He
workshops in various capacities. His research interests include holds four U.S. and international patents. He was named a Tennessee
blockchain, mobile edge/cloud computing, the Internet of Things, Valley Authority Fellow of two consecutive years during the Ph.D.
computer networks, and big data. degree, in addition to being a Toyota Manufacturing Fellow. His
professional career includes a Postdoctoral service at the Center for
Robotics and Manufacturing Systems CRMS, and a Visiting Scholar
at the Toyota Instrumentation and Engineering Division, Toyota
Motor Company, Japan. His group graduated seven Ph.D. disserta-
tions and over 35 M.Sc. theses. Four Ph.D. students are currently on
academic ranks in U.S. universities. His work has been recognized by
the U.S. Society of manufacturing engineers SME through the
Richard L. Kegg Award. He has also received the SAE Foundation
Award for Manufacturing Leadership. In addition, he has received the
Murray Stokely Award from the College of Engineering, Clemson
University. He has also led an NSF I/UCRC Center and a part of the
DoE GATE Center of Excellence in Sustainable Mobility Systems.
His current research interests include capabilities in composite fab-
rication and manufacturing analytics at a laboratory Masdar City

123
Cluster Computing

Campus. His current research group supported two Postdoctoral and as an Associate Editor for the Journal of Soft Computing (a
Scholar’s Career Planning to become an Assistant Professor at the Springer), handling the areas of decision science, knowledge-based
Texas A&M (TAMUQ), in 2013, and the University of Sharjah, in systems, in addition to his membership on several editorial boards and
2015. He currently serves as an Editor-in-Chief for the Journal of conference organizations. Furthermore, he serves on the Advisory
Material Science Research (Part of the Canadian Research Center), Board of the Strata PJSC (part of Mubadala Aerospace).

123