See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/378741851

Robust integration of blockchain and explainable federated learning for

automated credit scoring

Article in Computer Networks · March 2024

DOI: 10.1016/j.comnet.2024.110303

CITATIONS READS

0 137

4 authors, including:

Zhe Hou Kamanashis Biswas

Griffith University Australian Catholic University
79 PUBLICATIONS 346 CITATIONS 75 PUBLICATIONS 2,038 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Zhe Hou on 06 March 2024.

The user has requested enhancement of the downloaded file.

Robust integration of blockchain and explainable federated learning for
automated credit scoring
Zorka Jovanovica,∗ , Zhe Houa , Kamanashis Biswasb and Vallipuram Muthukkumarasamya
a Griffith University, Australia
b Australian Catholic University, Australia

ARTICLE INFO ABSTRACT

Keywords: This article examines the integration of blockchain, eXplainable Artificial Intelligence (XAI), es-
Automated Credit Scoring pecially in the context of federated learning, for credit scoring in financial sectors to improve the
Blockchain credit assessment process. Research shows that integration of these cutting-edge technologies is
Explainable Artificial Intelligence in its infancy, specifically in the areas of embracing broader data, model verification, behavioural
Decentralised Federated Learning reliability and model explainability for intelligent credit assessment. The conventional credit risk
assessment process utilises historical application data. However, reliable and dynamic transactional
customer data are necessary for robust credit risk evaluation in practice. Therefore, this research
proposes a framework for integrating blockchain and XAI to enable automated credit decisions.
The main focus is on effectively integrating multi-party, privacy-preserving decentralised learning
models with blockchain technology to provide reliability, transparency, and explainability. The
proposed framework can be a foundation for integrating technological solutions while ensuring model
verification, behavioural reliability, and model explainability for intelligent credit assessment.

1. Introduction financial institutions and reduce information asymmetry be-

tween lenders and borrowers, a proposed solution in Zhang
Credit Assessment (CA) is the process that assures the
et al. (2020a) involves leveraging blockchain technology to
development of credit scorecards to assess the creditworthi- establish a credit data-sharing alliance.
ness of the customers and loan applications following the The risk in the existing process is that the customer
policy of the lending institution. Mature banks are looking may be deemed creditworthy based on the limited dataset.
at making the process efficient, transparent and sustainable In Óskarsdóttir et al. (2019) presented that combining call-
to reduce the model risk and provide adequate governance.
detail records with traditional data in credit scoring models
Increasing competition and growing pressure for revenue
significantly increases their performance. At the same time,
generation are setting the requirements for the banks to they may have a high likelihood of defaulting on their credit
explore further effective integration and technologies that obligations. Additionally, some customers may be unfairly
will result in quicker turnaround time while managing the denied due to limited credit history Hurley and Adebayo
authenticity of the data source, transparency and privacy (2016). Ensuring the absence of model bias and discrim-
protection. It is necessary to employ an efficient, transparent,
ination is crucial throughout the scoring process Dastile
traceable, secure, and interpretable modelling process to
et al. (2022). Therefore, there is a risk to the existing credit
ensure accurate credit risk assessment. This approach aims assessment system’s trustworthiness, efficiency and fairness
to minimize model risk, mitigate bias and imperfections, and Hurlin et al. (2022).
deliver reliable and sufficient results. The credit assessment process encounters several chal-
Problem statement: Traditional financial institutions as- lenges that need to be addressed. Firstly, recognising the in-
sess credit applications based on data available to them at creasing challenges arising from liability concerns, sharing
the time of the credit application, such as customers’ credit or broadcasting data across various organizations. Various
scores, existing debt and income. The risk associated with data-sharing regulations, such as the General Data Protec-
tion Regulation (GDPR), limit data sharing opportunities
the customer’s creditworthiness may not be appropriately
across different organizations. Hence, increasingly complex
identified as the customer data comes from a single source
of information provided at the time of the application, such regulatory compliance and governance requirements must
as historical spending patterns. Thus, it does not consider be met Bücker et al. (2022a). Additionally, verifying and
the broader dynamic transactional customer data associated validating the accuracy of customer-provided information
with the customer’s financial behaviour, such as dynamic is crucial, ensuring that it reflects their actual financial
behaviour, including income, expenses, assets, and liabili-
payment behaviour, spending patterns, and financial health.
ties. Investigating the customer’s history of late payments
To mitigate the data scarcity in small and medium-sized
or inconsistent employment records is also important. Tra-
∗ Corresponding author ditionally, credit risk assessment has relied on historical
[email protected] (Z. Jovanovic); application data, but acquiring reliable and dynamic transac-
[email protected] (Z. Hou); [email protected] (K.
Biswas); [email protected] (V. Muthukkumarasamy)
tional data for model development has proven challenging.
ORCID (s): 0000-0002-6787-6379 (V. Muthukkumarasamy) Another essential aspect is the need for secure, transparent,

Jovanovic et al.: Preprint submitted to Elsevier Page 1 of 20

 Blockchain, XAI and Credit Scoring

traceable, explainable, and robust modelling techniques to these advancements aim to optimize credit risk management,
ensure an ethical credit assessment decision process. Evalu- reduce defaults, and strengthen trust in the financial system.
ating the strength and adaptability of the scorecard model The significance of this research is as follows:
is another area of concern. Assessing its robustness and
enabling dynamic updates is crucial to ensure its effective- • A novel credit assessment process is required that
ness over time. Finally, leveraging emerging technologies leverages comprehensive data sources and applies
becomes necessary to construct an intelligent and reliable them to advanced Artificial Intelligence (AI) algo-
scorecard engine that upholds data privacy, security, and im- rithms to provide a more holistic view of the cus-
mutability while enhancing the overall customer experience. tomer’s creditworthiness. The real-world financial
Blockchain can facilitate providing a decentralized credit sectors require a broader range of data sources and
scoring solution, as it trains a single credit scoring model models to ensure well-informed decision-making in
without sharing customer data, as Hassija et al. (2020) the credit assessment process.
suggested. Additionally, Federated Learning (FL) may serve • Enabling multi-source data support in credit mod-
as part of a privacy-preserving machine learning framework, elling promotes collaborative modelling among multi-
allowing multiple parties to collaboratively train a single ple parties while upholding privacy. Moreover, incor-
credit scoring model without sharing their customer data. porating multiparty privacy-preserving protection in
However, the authors in Hassija et al. (2020) consider single- credit modelling carries significant business benefits
model training. by facilitating accurate credit assessment while ensur-
The researchers in Yang et al. (2023) presented a credit ing the privacy of all involved parties.
scoring system that combines explainable federated learning
and blockchain to tackle challenges related to credit model • Trustworthiness and unbiased evaluation are essential
sharing and safeguarding data privacy. Their method elu- for reliable credit assessment processes. However, the
cidates the FL process, suggesting a decentralized Byzan- complex algorithms utilised in FL and blockchain-
tine fault-tolerant stochastic gradient descent algorithm (D- based consensus mechanisms can obscure the ratio-
SGD). From a mathematical perspective, the study integrates nale behind credit assessments, posing challenges in
the Shapley value with DPOS (Delegated Proof of Stake) meeting evolving regulatory requirements around ex-
as a consensus protocol, enabling the algorithm to compute plainability. While the use of blockchain can enable
the contribution values of the involved parties during the trustworthiness and transparency, XAI contributes to
execution of the federated algorithm. fairness in credit scoring.
Imteaj and Amini (2022) introduced a model based on
• Enabling the reliability and impartiality of the credit
FL to anticipate financial distress among borrowers. This
assessment models by incorporating adherence to reg-
approach involves constructing a global machine-learning
ulatory requirements. Therefore, its significance en-
model that evolves from the local models of distributed
ables financial institutions to employ trustworthy, un-
agents. The model achieved prediction accuracy almost in-
biased credit assessment models.
distinguishable from that of a centralized model. However,
there is no interpretability of the model and local model We propose an automated credit decision framework fo-
generation applying Stochastic Gradient Descent (SGD). cusing on the robust integration of blockchain and XAI to
Cheng et al. (2021) proposed SecureBoost, an FL boost- achieve these goals. The primary contributions of this paper
ing model, providing theoretical evidence that the model can be summarised as follows:
achieves accuracy on par with the non-federated boosting
model. However, the model is not interpretable. • Our research explores the fundamental features of
Our motivation is to thoroughly understand customers’ XAI and blockchain for credit scoring. We have con-
creditworthiness and trustworthiness in model prediction ducted an in-depth credit scoring analysis and pre-
and to address the needs for well-informed decision-making sented a taxonomy of the blockchain and XAI, which
in the financial sector while protecting customer privacy. We has not been done before. Our comprehensive taxon-
emphasize collaborative modelling, privacy-preserving pro- omy of blockchain and XAI features highlights their
tection, and adherence to regulatory requirements to ensure importance and insights for use within credit assess-
the accuracy and reliability of credit assessments while re- ment. This can assist researchers and practitioners in
specting all involved parties’ privacy settings. This approach navigating and applying these evolving technologies
provides various benefits, including enhanced collabora- effectively within the domain of credit scoring.
tion among multiple entities, improved privacy protection
through multiparty privacy-preserving measures, and the • Performed a comparative analysis of proposed archi-
development of more accurate credit assessment models. In- tectures that combine blockchain, FL, and XAI tech-
tegrating advanced technologies such as blockchain, FL, and nologies to construct credit scoring systems. We ex-
XAI also fosters technological innovation and creates trust- amine the difficulties of integrating these technologies
worthiness and unbiased credit assessment models. Overall, into credit assessment, addressing fundamental chal-
lenges and examining the integration mechanisms im-
plemented across diverse industrial applications. Our

Jovanovic et al.: Preprint submitted to Elsevier Page 2 of 20

 Blockchain, XAI and Credit Scoring

findings highlight that current solutions primarily fo- 2.1. Blockchain Technology
cus on data storage security and privacy, with limited Blockchain as a distributed ledger technology was intro-
impact on model verification, behavioural reliability duced with Bitcoin (Nakamoto and Bitcoin (2008)) to solve
and explainability in intelligent credit assessment. the double-spending problem using a peer-to-peer network.
The proposed peer-to-peer distributed timestamp server uses
• We present a conceptual framework that combines the Proof-of-work (PoW) system to record a chronological
blockchain, FL, and XAI technologies to establish order of transactions into timestamp blocks. The author
an automated decision-making credit assessment pro- defined an electronic coin as a chain of digital signatures.
cess. By utilising the optimal features of these tech- The hash of the previous transaction and the owner’s pri-
nologies, this framework aims to fulfil the require- vate key are required to sign their transactions digitally.
ments of the banking industry and regulatory stan- The public key is used to verify the sender’s identity. A
dards. The result is a credit scoring system that is both proposed timestamp server consists of the hash of a block
effective and explainable, thereby enhancing reliabil- of timestamped items and the previous timestamp in its
ity and transparency in the decision-making process. hash, constructing a chain. The Proof-of-work system was
After thoroughly examining the qualitative features re- implemented by incrementing a nonce in the block until a
quired for designing an efficient credit scoring framework, value is found that gives the block’s hash the required zero
we performed an initial complexity analysis of the proposed bits, consistent with the SHA-256 algorithm. The Proof-
framework. of-work requires CPU computation for mining the network
The rest of this paper is organised as follows: Section nodes and finding a Proof-of-work for its block. Once the
2 presents the key blockchain concepts. Section 3 reviews node considers the Proof-of-work, it broadcasts the block to
XAI techniques, including the surveyed work on the XAI all nodes. The nodes accept the block only if all transactions
and CA. Furthermore, we analysed the integration of FL, are validated and not spent. A hash of the accepted block is
blockchain and XAI. In Section 4, we present the credit created to be used as a previous hash for the new block in the
assessment’s principal functionalities, including the existing network. The block header contains the hash of the previous
process’s limitations. Section 5, we propose the conceptual block validated and a hash of all transactions contained in
framework that will address identified limitations and pro- the block (Merkle tree) as presented in Figure 2. Privacy is
vide an outlook on future research. Section 6 presents the preserved by keeping the public key anonymous.
analysis of the key characteristics required to build a robust
Block Header Block Header
credit assessment. Finally, Section 7 concludes the paper.
Figure 1 presents an overview of related work integrating Previous Hash Nonce Previous Hash Nonce
blockchain and XAI for credit assessment. Merkle Root Merkle Root

Hash01 Hash23

Hash0 D Hash1 D Hash2 M Hash3 M

Merkle Branch for Tx3

Tx3

Figure 2: Merkle tree of hash data in blockchain.

The history of the blockchain, starting from Blockchain

1.0 to Blockchain 4.0, has been discussed by Tanwar (2022).
Blockchain 5.0 is the latest generation of blockchain that
has been applied together with AI, hyper-converged infras-
tructure, and industry 4.0 technologies for high security,
efficiency, reliability, and scalability Tanwar (2022). Verma
et al. (2022) evaluated the integration of blockchain with
Figure 1: Overview of the integration of technologies and credit
Industry 5.0 focusing on how the technology can enhance
assessment. the security challenges of cyber-physical systems, such as
security, trust and transparency.
A systematic literature review of the blockchain-based
application across multiple domains such as supply chain,
2. Blockchain Features for Credit Assessment business process enactment (Stiehle and Weber (2022),)
financial, healthcare, IoT, privacy and data management has
This section describes the background of blockchain fea-
been analyzed (Casino et al. (2019); Uddin et al. (2021);
tures and the possible contributions to the credit assessment.

Jovanovic et al.: Preprint submitted to Elsevier Page 3 of 20

 Blockchain, XAI and Credit Scoring

Alam et al. (2021); Chowdhury et al. (2020); Wamba and to be resolved for the anticipated integration of blockchain
Queiroz (2020)). Investigating trends in blockchain tech- technology in the banking industry.
nology by applying text mining and clustering for register Blockchain integrates computer technologies, distributed
patents provides insights for researchers and inventors (Ba- data storage, information transmission, consensus mecha-
makan et al. (2021)). nisms, and encryption algorithms. Figure 3 presents the
The taxonomy of the blockchain for business process blockchain’s taxonomy for the credit assessment application.
enactment presented in Stiehle and Weber (2022) is based The taxonomy is organised into four dimensions: type,
on the two characteristics of capabilities and enforced guar- storage, blockchain features and applications in credit as-
antees. The analysis categorized capabilities based on the sessment. A public blockchain is open and permissionless,
different factors of model support, resource allocation and and decentralised. A private blockchain is permissioned,
process flexibility. Model support is a notation for the busi- and access to a network is restricted to authorised partici-
ness process. The authors used resource allocation to differ- pants. A blockchain consortium is semi-decentralised and
entiate various source allocations and examined how differ- permissioned, meaning nodes from multiple organisations
ent approaches impacted process flexibility. Regarding en- collectively own and manage the network. Storage on the
forcement, the authors determine that control flow, resource blockchain can be "on-chain" or "off-chain". On-chain refers
allocation, and data-integrity aspects are enforced on-chain. to the data stored on the blockchain, which means the
The benefits of using blockchain technology to im- network nodes verify it. Off-chain refers to storing data
prove security, transparency, and trust in different appli- outside of the blockchain in a separate system. The features
cations such as Multi-Agent Systems (MAS) (Calvaresi of blockchains are Immutability, Decentralised, Security and
et al. (2019)), energy market (Karandikar et al. (2021)), Cryptography, Distributed Ledgers, Consensus and Smart
identity management (Stockburger et al. (2021); Moinet Contracts. The immutability features of the blockchain
et al. (2017)), multi-organisation collaboration system (Lo ensure data integrity and transparency. Blockchains are
et al. (2021)) and data store (Radha et al. (2021)). decentralised networks without a central authority and are
A detailed survey on blockchain applications for AI thus more resilient against attack. Blockchains use various
shows that adopting blockchain for AI applications is still cryptographic algorithms, such as asymmetric-key algo-
in its infancy, Salah et al. (2019) . There are many research rithms (digital signatures), hashing, public-key cryptog-
challenges to be addressed in areas related to privacy, smart raphy, elliptic curve cryptography, and the Merkel tree.
contract security, trusted oracles, scalability, consensus pro- Asymmetric-key algorithms (digital signatures) are used to
tocols, standardisation, interoperability, quantum computing authenticate transitions and ensure the parties approve them.
resiliency, and governance (Salah et al. (2019), Sachan et al. Hashing uses mathematical algorithms to generate one-way
(2020), El Azzaoui et al. (2020)). functions while ensuring the immutability and integrity of
The integration of FL with blockchain to address ma- the data stored on the network. Public-key cryptography is
chine learning models’ privacy, security and scalability used to authenticate transactions and verify identity in the
challenges in distributed environments has been analysed. network. Elliptic curve cryptography ensures private keys’
Aledhari et al. (2020) and Qu et al. (2022) provided an security and authenticates transactions. The Merkle Tree
overview of the enabling technologies, protocols and ap- verifies the integrity of the transaction data. Consensus in
plications of FL and blockchain-enabled FL, respectively. blockchain refers to the process whereby nodes in distributed
The comprehensive overview of research in blockchain- networks work together to validate and process transactions,
based FL with different consensus mechanisms and privacy- which is essential for the integrity and immutability of the
preserving techniques is presented in Wang and Hu (2021). blockchain. There are several blockchain consensus algo-
The blockchain approach to enhance security and privacy FL rithms, such as Proof of work (PoW), proof of stake (PoS),
for IoT is proposed in Li et al. (2020) and Issa et al. (2023). delegated Proof of stake (DPoS), Proof-of-Authority (PoA),
Issa et al. (2023) discussed the challenges and risks of using Proof of elapsed time (PoET), and Practical Byzantine Fault
centralised storage and deep learning for IoT applications. Tolerance (PBFT).
While FL is a promising solution for preserving data privacy, The blockchain-based credit assessment modelling (Zhang
it still has a challenge of the model vulnerability. Issa et al. et al. (2019); Hassija et al. (2020); Qiao et al. (2022))
(2023) proposed utilizing the blockchain smart contract considers data privacy protection issues. The blockchain is
to safeguard FL and reviewed the blockchain-based FL introduced for storing credit data, which ensures full data
techniques securing IoT systems. traceability of the credit scoring process Yang et al. (2022b),
while the consensus mechanism is used to assess whether the
2.2. Blockchain Technology and Credit credit data is stored according to a predefined set of rules.
Assessment Walambe et al. (2020) proposed a system that leverages
The prospect of integrating blockchain within the bank- blockchain’s secure and immutable nature to store machine
ing and financial sector has been presented in Guo and Liang learning model explanations for credit scoring. The pro-
(2016) and Polyviou et al. (2019). Nowadays, blockchain’s posed system aims to enable local interpretations of the
breakthrough is in data storage and information transmis- global model to be publicly available to customers to access
sion. Regulation, efficiency, and security are the challenges securely. The authors demonstrated the trustworthiness of an

Jovanovic et al.: Preprint submitted to Elsevier Page 4 of 20

 Blockchain, XAI and Credit Scoring

Blockchain

Type Storage Features Application

Models in CA

Public Immutability Privacy Data Security

On-chains

Private Off-chains Decentralised Data store

Consortium Security/ Secure Lending

Cryptography

Distributed XAI store

Ledgers

Consensus Data Source

Authenticity

Smart Data Sharing

Contract

Figure 3: Taxonomy of blockchain technology for credit assessment.

explained model prediction, with the security, reproducibil- Nassar et al. (2020) proposed a framework based on
ity, traceability and transparency of blockchain, providing the principle that critical decisions in complex AI systems
the end-user with a way to securely request an explanation must be subject to consensus among distributed AI and XAI
for the credit-scoring decision. Blockchain tamper-proof agents hosted in trusted oracles. Blockchain can fulfil trust-
characteristics ensure the authenticity of the data and min- worthy AI requirements for resilience to biases and adversar-
imise the impact of false data for credit evaluation mod- ial attacks. Blockchain provides key features for XAI agents:
elling (Zhang et al. (2019)). The blockchain-based frame- Transparency and Visibility, Immutability, Tractability and
work that assists the gathering of information about the cus- Nonrepudiation and Smart Contracts.
tomers from the various financial institutions and calculates
their score based on the consensus of multiple institutions
improves the credit decision process (Chakraborty et al. 3. Use of Explainable AI in Credit Assessment
(2019)). For example, blockchain for the credit evaluation In this section, we first describe the overall XAI tech-
system of traders in the food supply chain has been analysed niques. Following this, we present an overview of XAI meth-
Mao et al. (2018). ods and describe their characteristics for the explainable
A blockchain-based credit score evaluation is proposed credit assessment. Finally, we describe the integration of the
to ensure transparency in the lending process (Patel et al. blockchain and XAI.
(2020); Hassija et al. (2020)). Blockchain and Decentralized
Credit Scoring Model presents a theory to model the optimal 3.1. Background of XAI
investment strategy for different risk vs. return scenarios Machine learning (ML) models are predominantly black
Hassija et al. (2020). In Patel et al. (2020) KiRTi, a deep boxes. The model-agnostic techniques have been developed
learning-based credit recommender, is proposed to automate to explain the predictions of any classifier in an interpretable
loan disbursements and repayments. This work is a step form. Among this area’s best known contributions is the
forward in eliminating the requirement of third-party credit Locally Interpretable Model-Agnostic Explanations (LIME)
rating agencies for credit score generation. (Ribeiro et al. (2016)). LIME constructs locally linear mod-
Cho et al. (2021) designed a Verifiable Credential (VC) els around the predictions of a model to explain it by ap-
model for VC generation and revocation verification for proximating it locally with an interpretable model. These
credit scoring data. Blockchain-authorized data (Zhang contributions fall under model agnostic (MA) and local (L)
et al. (2020a); Zhu (2020)) and model sharing (Yang et al. explanations. Notably, the authors propose algorithms for
(2022a)) enhances the security of credit reporting. individual predictions to solve the "trusting a prediction"
problem known as the LIME algorithm by approximating it

Jovanovic et al.: Preprint submitted to Elsevier Page 5 of 20

 Blockchain, XAI and Credit Scoring

locally with an interpretable mode. Furthermore, the authors and Prakoonwit (2022). Rajbahadur et al. (2022) explored
proposed a Submodular Pick SP-LIME algorithm to select a the impact of feature importance measures on the inter-
set of predictions (and explanations) to solve the "trusting pretability and stability of the classifiers. The Neural-Backed
the model" problem via submodule optimisation. Decision Trees (NBDT) model Wan et al. (2020) trains a
The LIME explanation is obtained by minimising the decision tree to represent a (deep) neural network and main-
following objective function: tains a high level of model interpretability. The authors Hara
and Hayashi (2018) proposed a Bayesian model selection to
𝜉(𝑥) = arg min 𝐿(𝑓 , 𝑔, 𝜋𝑧 ) + Ω(𝑔) (1) improve the model interpretability of tree ensembles.
𝑔∈𝐺

where 𝐿(𝑓 , 𝑔, 𝜋𝑧 ) measures faithfulness of explanation model 3.2. XAI and Credit Assessment
𝑔, in approximating original model 𝑓 in the locality defined The surveyed work on XAI and CA is presented in
by 𝜋𝑧 . 𝐺 represents the class of the of potentially inter- Figure 4 and Table 1. Figure 4 presents the different aspects
pretable models, while 𝜋𝑧 (𝑥) is proximity measure between of classifying XAI methods based on their characteristics.
an instance 𝑧 to 𝑥, knows as locality around 𝑥. Ω(𝑔) penalizes As a result, according to the proposed taxonomy, three
the complexity of the explanation 𝑔. main categories for the XAI are identified: form, scope
SHAP (Shapley Additive exPlanations) value is pro- and applicability. Another important aspect is the form of
posed by Lundberg and Lee (2017) for interpreting and the XAI method: numeric or rule-based. Numerical expan-
understanding the predictions made by the machine learning sion in the form of the importance of a specific feature
models. The proposed SHAP values measure the contri- to the overall performance of a model is called feature
bution of each feature to a model prediction. The feature importance (Altmann et al. (2010)). The explainability pro-
importance explanation technique is a form of ranking the duced by rule-based explanations by exploiting several rule-
importance of each feature in the prediction output by the extraction techniques, such as automated reasoning-based
model to be explained. The SHAP method calculates an models (Bride et al. (2018, 2021); Zhang et al. (2021b)),
additive feature attribution measures that satisfies the set of is known as approximate model. Based on the scope of
required properties (local accuracy, missingness and con- interpretation, if the method explains a specific instance, it is
sistency). The first property local accuracy requires expla- known as local, and if the method explains the whole model,
nation model 𝑔 to at least match the original model 𝑓 output then it is global. An important aspect of separating XAI
for a simplified input. The second property missingness methods is the type of algorithms that could be applied. If
requires features missing in the original input to have no the technique has restricted application to a specific family
attributed impact. The third property consistency requires of algorithms, it is called model-specific. The method used
that if a model changes such that some simplified input’s for any possible algorithms is model agnostic. The recent
contribution increases or stays the same regardless of the work by Wan et al. (2020) presents a Neural backed Decision
other inputs, the input’s attribution should be consistent. The Tree (NBDT), which explores the combination of neural nets
Shapley value for each feature can be calculated using the and decision trees. Such an intersection would preserve high-
following formula: level interpretability while neural networks provide high
accuracy. Recent work by de Lange et al. (2022) presents
the combination of the LightGBM model with SHAP, which
∑ |𝑧′ |!(𝑀 − |𝑧′ | − 1)!
𝜙𝑖 (𝑓 , 𝑥) =
′ ′
[𝑓𝑥 (𝑧 )−𝑓𝑥 (𝑧 ∖𝑖)] (2) enables the interpretation of explanatory variables affecting
′ ′ 𝑀! credit predictions.
𝑧 ⊆𝑥
Table 1 presents the XAI characteristics and broader
where 𝑀 denotes a number of all features, 𝑓 is a model. applications. XAI in credit risk applications is presented
′ ′ ′ ′
|𝑧 | is the number of non-zero entries in 𝑧 , and 𝑧 ⊆ 𝑥 in Moscato et al. (2021), Walambe et al. (2020), de Lange
′
represents all 𝑧 vectors where the non-zero entries are a et al. (2022), Bücker et al. (2022b), Sachan et al. (2020), and
′ ′ ′
subset of non-zero entries of 𝑥 , while 𝑧 ∖𝑖 denotes 𝑧 = 0. Srinivasan et al. (2019).
A comprehensive taxonomy of the XAI method is pre-
sented in Arrieta et al. (2020),Schwalbe and Finzel (2021), 3.3. Blockchain, Federated Learning and XAI
Islam et al. (2022), Sahakyan et al. (2021), Vilone and Longo Federated Learning trains machine learning models on
(2020), and Lampathaki et al. (2021). The XAI is a subsec- multiple datasets distributed across different clients without
tion of AI that focuses on the transparency of the AI sys- data sharing Yang et al. (2022a). FL enables multiple clients
tems’ decision-making. Integrating XAI into cybersecurity to solve machine learning problems under the coordination
intends to improve the AI security system’s trustworthiness, of the central aggregator, which ensures data privacy (Zhang
interpretability and resilience. The XAI methods to tackle et al. (2021a)).
cybersecurity issues have been presented in different areas, Regarding communication delays, the global model in
such as industrial IoT (Khan et al. (2021)), advanced persis- FL involves multiple iterative rounds of model updates from
tent threats (Li et al. (2021)), intrusion detection (Oseni et al. users, engendering significant communication overhead and
(2022)) and autonomous driving (Rjoub et al. (2022)). incurring additional storage costs during network transmis-
The survey of resampling techniques on feature impor- sion, Konečnỳ et al. (2016). The FL is contingent upon the
tance in imbalanced blockchain data is presented in Alarab seamless communication between clients and servers. This

Jovanovic et al.: Preprint submitted to Elsevier Page 6 of 20

 Blockchain, XAI and Credit Scoring

Table 1
Related work on XAI and credit assessment and their characteristics. Yes (✓), No (×), Partial details on explainability (*).

Approximate Model
Feature Importance

Model Agnostic

Model Specific
Global
Local
Reference Method Credit Application
Moscato et al. (2021), Walambe et al. LIME ✓ ✓ ✓ * ✓ × XAI for Credit score
(2020), Bücker et al. (2022b)
Moscato et al. (2021) Anchors ✓ ✓ ✓ * ✓ × XAI for Credit score
Moscato et al. (2021), de Lange et al. SHAP ✓ × ✓ ✓ ✓ × XAI for Credit score
(2022)
Moscato et al. (2021) BEEF ✓ × × ✓ ✓ × XAI for Credit score
Moscato et al. (2021) LORE ✓ * ✓ * ✓ × XAI for Credit score
Ma et al. (2022) MUC ✓ ✓ ✓ ✓ × ✓ Loan application improv-
ments
Srinivasan et al. (2019) ARAEGAN+GM × × ✓ × × ✓ Credit loan denials
Sachan et al. (2020) MAKER × × × × ✓ × Loan underwriting
Fahner (2018) TGAMT × × × ✓ × ✓ Explainability by design
Credit score
Bride et al. (2021) Silas ✓ ✓ × ✓ × ✓ XAI via logical reasoning
on credit data
Zhang et al. (2021b) OptExplain × ✓ × ✓ × ✓ XAI via logical reasoning
on credit data

Feature communication in bandwidth-constrained learning environ-

Importance
ments. Hieu et al. (2020) introduced the application of deep
Form
reinforcement learning in optimizing system parameters for
Approximate minimizing delay, energy consumption and maximizing total
Model
rewards.
A comprehensive and systematic Privacy-Preserving
Local FL (PPFL) review is presented in Yin et al. (2021). The
XAI Scope overview of the main characteristics of the Blockchain-
Based Federated Learning (BCFL) framework, architectural
Global design, deployed platforms and feasible applications for
BCFL is presented in Li et al. (2022a). Li et al. (2022b)
Model
proposed a systematic study on privacy and security in
Agnostic blockchain-based FL methodologies and discussed the in-
Applicability tegration of blockchain with FL in various human-centric
Model applications in IoT and intelligent environments.
Specific Blockchain design that enables recording and secure
incentives for distributed FL model training via Smart Con-
tracts with Class-Sampled Validation ErrorScheme (CSVES)
Figure 4: Overall taxonomy of XAI.
to validate the quality of gradients to determine reward
is proposed in Martinez et al. (2019). The advantages of
this approach encompass increased trust in the FL process
communication involves the transmission of local learning and enhanced incentives for participants during gradient
models and multiple training iterations for model updates, validation. However, potential limitations may arise from
making communication and training efficiency critical for centralised model aggregation, lack of explainability of the
FL performance Wu et al. (2023). To manage upstream trained models, and the impact of new data on the training
communication delay, a Sparse Ternary Compression (STC) process.
framework is proposed by Sattler et al. (2019) extends gradi- The overview of FL and blockchain integration, called
ent sparsification with downstream compression, surpassing FLchain, can potentially transform intelligent mobile edge
federated averaging in various scenarios and advocating for a computing (MEC) networks into a decentralised, secure,
transformative shift toward high-frequency, low-bandwidth and privacy-enhancing system Nguyen et al. (2021). The

Jovanovic et al.: Preprint submitted to Elsevier Page 7 of 20

 Blockchain, XAI and Credit Scoring

article presents four use cases that demonstrate the potential securely. The authors demonstrated the trustworthiness of an
applications of FLchain in edge networks, including edge explained model prediction, with the security, reproducibil-
data sharing, edge content caching, and edge crowdsensing. ity, traceability and transparency of blockchain, providing
However, research lacks a comprehensive evaluation of ex- the end-user with a way to securely request an explanation
periments to assess their effectiveness and limitations in a for the credit-scoring decision. However, the proposed so-
practical setting fully. lution considers only a single AI and XAI method: Ran-
The proposed serverless function for training FL Fed- dom Forest (RF) and Locally Interpretable Model-Agnostic
Less is detailed in Grafberger et al. (2021), which utilised (LIME). FL may be regarded as preserving privacy in credit
serverless technologies, AWS Lambda, Azure functions and assessments. FL trains machine learning models on multi-
Openwhisk to enable FL while providing authentification, ple datasets distributed across different institutions without
authorisation and differential privacy. FedLess supports Lo- data sharing. However, the system’s efficiency proposed in
cal Differential Privacy, a technique that adds noise to the Walambe et al. (2020) depends on the quality of the machine
data before sharing it. The paper introduces a novel approach learning model used for credit scoring. Hence, if the model is
that leverages serverless computing to address the challenges unreliable, the explanation stored on the blockchain may be
of scalability, infrastructure management, and inactive client inaccurate, leading to incorrect credit scoring decisions. The
computing resources in FL. However, it is important to limitation of the proposed system in Walambe et al. (2020)
consider the limitations of FedLess in the specific context is that it does not ensure that the model is reliable and the
of the target domain and requirements. Further research and performance of the model is validated.
evaluation are needed to fully understand the effectiveness Recent research has presented uses of blockchain as a
and limitations of FedLess in other areas, such as credit score distributed data structure with major features summarised as
modelling. immutability, transparency and encryption. The integration
The behaviour attestation method is used to verify the of blockchain and XAI is presented in Table 2. Integra-
consistency of the behaviour of each participating client tion of the blockchain and XAI can be achieved through
during the training process for detecting poisoning attacks decentralised data storage, smart contracts and decentralised
in FL (Mallah et al. (2021)). The authors presented the model learning (Li et al. (2020); Aledhari et al. (2020);
AttestedFL algorithm for defence against untargeted model Qu et al. (2022)). The primary use of blockchain for AI is
poisoning attacks in FL with contributions to reducing attack for secure data storage (Salah et al. (2019); Walambe et al.
effectiveness, increasing accuracy, pattern-based detection, (2020); Patel et al. (2020); Zhang et al. (2019)) and audit
and flexibility in deployment. However, further research trailing of XAI decisions (Malhotra et al. (2021)).
and optimisations are required to explore its efficacy under
different scenarios. 4. Technologies for Integrated Credit
Al Mallah and López (2022) proposed techniques to
address the latency challenges by decoupling the monitor-
Assessment
ing phase from the detection phase in decentralised FL In this section, we first describe the overall concepts of
approaches defences that protect against poisoning attacks credit assessment. Following this, we present an overview
in FL. The blockchain replaced the centralized aggregation of the related work on using XAI and blockchain for credit
of the traditional FL. It divided the blockchain network assessment.
into two types of miners: minersFL responsible for FL, and
minersMON, responsible for monitoring. Workers perform 4.1. Credit assessment fundamentals
the FL and send their local model updates to minersMON, Credit evaluation assesses a borrower’s capacity to be-
responsible for monitoring. The blockchain minersFL nodes come eligible for a loan and the ability to repay. Credit
randomly select a set of reliable workers to continue the FL evaluation is the process that assures the development of
process and calculate the average model using the updated credit scorecards to assess the creditworthiness of the cus-
model from the workers and minersMON. The proposed tomers and loan applications following the policy of the
design does not store the model updates on the blockchain. lending institution. The banks are looking at making the
Instead, the hash value is written on the blockchain and process efficient, transparent and sustainable to reduce the
points towards the model updates. The blockchain stores model risk and provide adequate governance. The increasing
the commitments of all workers on the model updates they competition and growing pressure for revenue generation are
worked on. A Merkle tree is used to authenticate the model requiring banks to explore further effective integration that
updates submitted by the workers. The proposed approach is will result in quicker turnaround time while managing the
designed for resource-contained nodes like mobiles and the authenticity of the data and privacy protection.
Internet of Things (IoT). The book by Thomas et al. (2017) has been recognised
Walambe et al. (2020) proposed a system that leverages as a bible of credit scoring and reviews statistical and op-
blockchain’s secure and immutable nature to store machine erational research methods used in building the scorecard.
learning model explanations for credit scoring. The pro- One of the first credit scoring approaches was developed to
posed system aims to enable local interpretations of the predict companies’ bankruptcy risk (Altman (1968)).
global model to be publicly available for customers to access Credit scoring is one of the earliest financial risk man-
agement tools (Thomas et al. (2017)) and is a method that is

Jovanovic et al.: Preprint submitted to Elsevier Page 8 of 20

 Blockchain, XAI and Credit Scoring

Table 2
A comparison of properties of blockchain, AI and integration mechanisms, Yes (✓), No (×), Insufficient details (*).

Blockchain AI Integration

Decentralised ML

Implementation
Smart Contract
Transparency
Immutability

XAI Method
Traceability

FL Method

Storage
Oracles

Privacy
XAI
FL
Reference
Salah et al. (2019) ✓ × × × × × * ✓ * ✓ ✓ ✓ ×
Hossain et al. (2020) × ✓ ✓ × × ✓ DL ✓ LIMA ✓ ✓ × ✓
Walambe et al. (2020) × ✓ × ✓ × ✓ RF ✓ LIME ✓ × ✓ ✓
Patel et al. (2020) ✓ × ✓ ✓ × ✓ LSTM × * ✓ × ✓ ✓
Nassar et al. (2020) ✓ ✓ ✓ ✓ × ✓ * ✓ * ✓ ✓ × ×
Zhang et al. (2019) ✓ × × × × ✓ Logit × * ✓ ✓ ✓ ✓
Polyviou et al. (2019) ✓ × × ✓ × × * × * × ✓ ✓ ×
Calvaresi et al. (2019) × × ✓ × × × * ✓ * × ✓ × ×
Hassija et al. (2020) ✓ × × × × × * × * PoV ✓ ✓ ✓
Malhotra et al. (2021) ✓ ✓ ✓ ✓ × ✓ SVM ✓ LIME ✓ ✓ ✓ ✓
Verma et al. (2022) ✓ * ✓ ✓ * * * * * * ✓ ✓ ×
Bellagarda and Abu- * * * ✓ ✓ ✓ * ✓ * * * ✓ ×
Mahfouz (2022)
Yin et al. (2021) × × × × × ✓ * × * × ✓ ✓ ×
Chen et al. (2022) × × × × × ✓ ESB-FL × * PoS ✓ ✓
Zhang et al. (2021a) × × × × × ✓ * × * × ✓ ✓ ×
Cheng et al. (2021) × × × × × ✓ RL- × * × × ✓ ✓
SecureBost
Srinivasan et al. (2019) × × × × × ✓ SVM, Naive ✓ ARAEGAN × × × ✓
Bayes +GM
Bride et al. (2021) × × × × × ✓ Silas ✓ Logical Rea- × × × ✓
soning
Sachan et al. (2020) × × × × × ✓ BRB ✓ MAKER × × × ✓
Davis et al. (2022) × × × × × ✓ Optimal Tree, ✓ LIME, × × × ✓
NN, RF SHAP, DiCE

used to predict the probability that a borrower will default data flow for the credit assessment process. The application
or become delinquent and to measure the profitability of data consists of variables such as the applicant’s age, time
granting loans. Traditional credit evaluation methods consist at current/previous residence, time at current/previous job,
of judgmental models, statistical methods, regression anal- housing status, occupation group, income, number of de-
ysis (Turkson et al. (2016)), discriminant analysis (Altman pendents, banking relationship, debt ratio, and credit refer-
(1968); Turkson et al. (2016)), logistic models and pro- ences. Credit references or bureau information consist of the
bit models. Recently, alternative machine learning methods previous defaults, arrears and the customer’s current status
such as artificial neural networks (ANNs) (West (2000)), on other loans, including the number of enquiries, hardship
neural networks (NN) (Yobas et al. (2000); Turkson et al. information and repayment history information. The major
(2016)), bayesian networks (Turkson et al. (2016)), support credit bureau providers are Equifax, Illion and Experian. The
vector machines (SVMs) (Harris (2015)), decision trees (Nie comprehensive credit score is the number that models the
et al. (2011); Dumitrescu et al. (2022); Yobas et al. (2000); data held in the credit bureau and indicates the likelihood of
Turkson et al. (2016)), XGBoost (Xia et al. (2017)) and repaying the money to the credit applicant’s credit bureau.
other methods have been introduced to build credit scoring The bureau has its method for modelling the comprehen-
models. The fairness of AI techniques in the context of sive credit score. The customer application data is used
the credit scoring model has been analysed by Hurlin et al. to perform the calculations related to the serviceability of
(2022). the customer, and that information is used in the scorecard
Credit scoring systems are based on the past perfor- model.
mance of customers, similar to those who will be assessed The credit scoring model uses any characteristic of the
under the scheme. When the customer applies for a loan, the customers that aids prediction in the scoring system. The
financial institutions collect the customer details, known as variables are mainly associated with default risks, such as
application data. Figure 5 presents an illustrative view of the

Jovanovic et al.: Preprint submitted to Elsevier Page 9 of 20

 Blockchain, XAI and Credit Scoring

Bank Bureau
Customer

Customer’s Capacity to Financial Scorecard

details Repay ratios Score
Credit
Checks

Decision

Comprehensive
Approve Refer Decline Credit Score
Score Score Score

Previous
Refer Decline Defaults
Rules Rules
Underwriting
Policy Rules
Behavioural
Message Message Payments
Rules Rules

Review Decision

Figure 5: Overview of the credit assessment process.

previous defaults or arrears or the customer’s current sta- A loan underwriting process or Underwriting Policy
tus on other loans and comprehensive credit score. Other Rules evaluates the information in a loan application follow-
variables present the stability of the consumer, such as time ing the scorecard cut-off score outcome and the policy of the
at address and time at present employment. A different lending institution as shown in Figure 5.
group of variables gives a view of the consumer’s residen- A loan underwriting system containing coded under-
tial status, spouse’s employment, number of children, and writer guidelines decides acceptance or rejection when spe-
number of dependents. A separate set of variables shows the cific default rules in the rule base are triggered. The loan
consumer’s serviceability, such as the Debt to Income ratio. underwriting could be manual or automated. Manual under-
The good/bad flag is created based on the loan repayment writing refers to processing non-standard (higher risk) loans.
history of the accepted population for the scorecard devel- The underwriting system consists of a codified set of rules
opment. Borrowers who have missed payments or gone past based on the policy of the lending institution to assist in a
a certain number of days, usually 90 days, are categorised final lending decision. The key limitation in the existing liter-
as "bad" borrowers, while those who have not are classified ature is that the credit scoring and underwriting process have
as "good" borrowers. The good and bad flags are then used been considered in isolation, while the automated intelligent
to develop a scorecard model. The Kolmogorov–Smirnov credit evaluation should consider both.
statistic determines the cut-off score and measures the dis- Sachan et al. (2020) proposed an XAI decision-support
tance between the cumulative distribution of goods and bads. system to automate loan underwriting by a belief-rule-base
The cut-off score is the maximum distance between the (BRB) system. The solution proposed by the authors aims
distribution of good/bad and is used to predict the good/bad to enhance the efficiency and accuracy of the underwriting
(Dastile et al. (2020)). If the score of a customer is above or process while preserving transparency and fairness.
equal to the cut-off score, then the customer is predicted as An intelligent credit risk scorecard approach based on
a good borrower otherwise, a bad borrower. Subsequently, statistical principles is needed for specific business objec-
the scorecard is applied to the rejected population to predict tives like predicting losses better (Siddiqi (2012)). A deeper
good/bad, known as reject inference (Siddiqi (2017)). view of creating, evaluating, and monitoring scorecards is
presented in Siddiqi (2017).

Jovanovic et al.: Preprint submitted to Elsevier Page 10 of 20

 Blockchain, XAI and Credit Scoring

Credit scoring is a supervised learning problem. Specif- The traditional credit risk assessment process utilises
ically, it is a binary classification problem aiming to classify the application data, while dynamic transactional data has
good and bad borrowers (Dastile et al. (2020)). A systematic recently been used to evaluate the credit application (Zhang
literature survey approach to statistical and machine learning et al. (2020b)). The authors proposed cost-sensitive multiple-
models in credit scoring, identifying literature limitations, instance learning (CSMIL) to build a credit scoring model
proposing a guiding machine learning framework and point- incorporating customers’ dynamic transactional data and
ing to emerging directions have been proposed by Dastile static/personal information. This study is the first to apply
et al. (2020). However, the LIME method covers the explain- a CSMIL model to credit risk assessment and considers
ability of credit scoring methods to a limited extent. the impact of dynamic transactional data and time-series
The most popular technique in credit scoring modelling information. The work presented in Zhang et al. (2020b) is
is Logistic Regression Equation 3. The Logistic Regression limited as it does not include explainability techniques in the
assumes a linear relationship between the log of probability credit scoring model while utilising dynamic transactional
odds and inputs (Thomas (2000)). The logistic regression is data. Furthermore, the model performance may deviate due
sensitive to the correlation between the predicted variables. to the data update; identifying those deviations may require
Thus, it should be ensured that no correlated variables are in model recalibration.
the regression set. Logistic regression is the log of the prob- P2P lending is a business model involving borrowers,
ability odds by a linear combination of the input variables. lenders, and a P2P platform. A P2P platform generally
has a large number of users and frequent transactions. A
benchmarking study of some of the most used credit risk
∑ 𝑚
𝑙𝑜𝑔(
𝑝
) = 𝑤0 + 𝑤𝑖 𝑋𝑖 (3) scoring models to predict if a loan will be repaid in a P2P
1−𝑝 𝑖=1 platform has been analyzed by Moscato et al. (2021). The
authors compared the obtained outcomes concerning the
Where 𝑝 represents the proportional response, 𝑤0 is the state-of-the-art approaches and also evaluated them in terms
intercept, when 𝑋 = 0 intercept is the log of the odds of of their explainability through different XAI tools. Zhang
having the outcome. 𝑿𝑖 are application characteristics and et al. (2020c) proposed a new online integrated credit scoring
weights 𝒘𝑖 are the score of the characteristics. model (OICSM) for P2P lending that integrates gradient-
Equation 3 is considered a linear regression of the non- boosting decision trees and the neural network to make the
linear function of the probability of being a good customer. credit scoring model handle two types of features (numerical
The score 𝑠(𝑥) of the scorecard presented in the Equation 3 and categorical) more effectively and update the model
is the following Equation: online. This is one of the first experiments considering the
problem of the credit scoring model online update to avoid
∑
𝑚
prediction deviation. The limitation of the OICSM scoring
𝑠(𝑥) = 𝑤0 + 𝑤𝑖 𝑋𝑖 (4)
𝑖=1
model is that it does not include XAI techniques to ensure
transparency in the credit scoring model. Furthermore, the
Another important technique in credit scoring modelling traceability of model updates is not considered.
is non-linear regression, known as a probit analysis. The pro-
bit model 𝑵(𝑥) is given as the cumulative normal (standard 4.2. Blockchain and XAI for Credit Assessment
Gaussian) distribution function defined below: Credit assessment requires an efficient, transparent, trace-
𝑥
able, secure, and sustainable process to reduce the model risk
𝑦2
𝑁(𝑥) = √
1
𝑒− 2 𝑑𝑦 (5) and provide adequate governance. The surveyed work on
2𝜋 ∫−∞ integrating CA, blockchain and XAI is presented in Table
3. Table 3 presents limited work that has been done to
The goal is to estimate 𝑵 −1 (𝑝𝑖 ) as a linear function of the examine the integration of blockchain and XAI for the credit
characteristics of the applicant, as follows: assessment process. As discussed previously, the system
proposed in Walambe et al. (2020) relies on the quality of the
∑
𝑚
machine learning model used for credit scoring. If the model
𝑁 −1
(𝑝𝑖 ) = 𝑤0 + 𝑤𝑖 𝑋𝑖 (6)
is unreliable, the explanation stored on the blockchain may
𝑖=1
be inaccurate, resulting in incorrect credit scoring decisions.
The value of 𝑵 −1 indicates that the customer is good if the The proposed mechanism lacks a technique to ensure that
score is above a certain level. Linear programming is used the model is reliable and its performance is validated, thus
as a classification approach for scorecard modelling. limiting its effectiveness.
The popular machine-learning techniques in credit scor- The efficiency of the Credit evaluation has been ad-
ing are Random Forest, Artificial Neural Networks, and dressed in the BACS scheme by Yang et al. (2022b). The
Convolution Neural Networks. Random Forest (RF) is an BACS scheme consists of credit data storage to the blockchain
ensemble of decision trees (Breiman (2001)), such that K to ensure traceability. The random forest model effectively
decision trees are built on different bootstrap samples of the integrated the critical steps of credit data feature extraction,
data. feature selection, credit model construction, and model

Jovanovic et al.: Preprint submitted to Elsevier Page 11 of 20

 Blockchain, XAI and Credit Scoring

Table 3
A comparative analysis of the use of XAI and Blockchain (BC) for credit assessment: Yes (✓), No (×), Minor advancement (*).
Reference XAI for CA BC for CA XAI and BC Key Technologies Considerations
for CA
Qiao et al. (2022, × ✓ * PHE, SMPC algorithm BC for privacy and data
2023a) security
Hassija et al. (2020) × ✓ * Prospect theory for risk BC for secure lending
vs return
Walambe et al. (2020) ✓ ✓ ✓ RF method, LIME XAI with BC to store
explanation in block
Patel et al. (2020) × ✓ * LSTM BC to update the credit
score
Zhang et al. (2020a) × ✓ * PBFT consensus consortium BC for CA
Yang et al. (2022b) × ✓ * BACS BC and AutoML for CA
Yang et al. (2022a) × ✓ * IPFS to store encrypted BC and FL for data
data sharing in CA

evaluation. Blockchain technology as discussed in this ar- to heightened communication overhead and computational
ticle requires a consensus mechanism to determine whether intensity in FL. The EFCS is evaluated using six credit
credit data is stored and used within predefined rules. The datasets from traditional financial institutions and peer-to-
consensus process is divided into the sorting service and the peer lending platforms. The datasets from Germany, Taiwan,
synchronised ledger. This work by Yang et al. (2022b) has and Australia are available through the UCI machine learn-
a few limitations. Firstly, the paper does not explore alter- ing repository. Additionally, P2P lending datasets and credit
native blockchain platforms beyond Fabric Hyper-ledger for card datasets are employed for further validation. Specifi-
ensuring consensus on updates to the model. Additionally, cally, two P2P datasets are collected from China’s pioneering
it does not consider XAI methods other than consensus for P2P lending platforms. This diverse set of datasets from
improving transparency in credit decision outcomes. Finally, various sources enhances the applicability of the evaluation
the study relies solely on historical credit data and does process for EFCS.
not explore the potential benefits of using transaction data Table 3 presents case studies of use cases of the XAI and
to identify early delinquent behaviour. Thus, identifying blockchain for credit assessment.
changes in data, model input assumptions, or scorecard
model performance may also have limitations that have not
5. Proposed Conceptual Framework
yet been explored.
This section presents a conceptual framework based on
4.2.1. Case studies decentralised blockchain as a solution to induce model ver-
Authors in Yang et al. (2023) introduced an explainable ification, behavioural reliability and explainability for intel-
federated learning and blockchain-based credit scoring sys- ligent credit assessment. This framework uses a blockchain-
tem to address credit model sharing challenges and ensure based FL solution to enable AI machine model learning and
data privacy. Their approach explains the FL mechanism, verification of the methods, and it is a machine-learning
proposing a Decentralized Byzantine fault-tolerant Stochas- model built on distributed datasets. FL benefits blockchain
tic Gradient Descent algorithm (D-SGD). Mathematically, with aspects of privacy-preserving data exchange. Our pro-
the study combines the Shapley value with Delegated Proof posed conceptual framework considers blockchain-based FL
of Stake (DPOS) for a consensus protocol. The algorithm through the consortium or private blockchain platform.
calculates the contribution values of the parties in the ex- Robust integration of technologies can be defined as
ecution of the federated algorithm. Evaluation of the pro- facilitating decentralised model learning, verification, and
posed Explainable Federated learning and blockchain-based model aggregation on distributed multisource datasets. This
Credit scoring System (EFCS) includes simulations and involves fostering collaboration among different sources
experiments using the "Give Me Some Credit" dataset from while preserving data privacy and enhancing the overall
Kaggle. The dataset contains 150,000 credit card payments reliability and transparency of the credit scoring system.
and income-related data, with 10,026 default customers. This definition effectively captures the essence of robust in-
The performance assessment encompasses accuracy, pre- tegration in the proposed conceptual framework, highlight-
cision, recall, F1 score, and AUC. The modelling process ing key aspects such as decentralized learning, verification,
involves the coordinating party calculating contributions and collaboration, data privacy, reliability, and transparency,
recording them in the current block of transactions. Training leading to a trustworthy credit scoring decision-making
results reflect aggregated data source outcomes, with each process.
participant iterating locally 20 times before sending the gra- Blockchain enables FL to enhance the process of the
dient. Increased participants lead to longer training times due global model aggregation such that model aggregation is to

Jovanovic et al.: Preprint submitted to Elsevier Page 12 of 20

 Blockchain, XAI and Credit Scoring

be computed by the nodes and miners, leaving the central The participating nodes from banks, bureaus and enter-
aggregation unneeded. prises are required to obtain a user’s certificate for access
Figure 6 illustrates the architecture of the proposed in- to data. Specifically, scorecard clients are the distinct client
telligent automated credit assessment that enables AI model functions that will collaborate in FL. Specifically, Hyper-
learning and blockchain miner verification of the model ledger Fabric incorporates the ciphertext-policy attribute-
while ensuring privacy is protected. The architecture con- based encryption (CP-ABE) access control scheme avoiding
sists of the following roles: unauthorized access, Qiao et al. (2023b). All nodes in the
Hyperledger Fabric network are generally assumed to be
• Scorecard Clients credible. As a result, the consensus mechanism employed
• Federated Learning Local Model Miners by Fabric mainly focuses on ordering transaction proposals
rather than validating them, Qiao et al. (2023b).
• Blockchain Miners verify the models and generate In Thantharate and Thantharate (2023) presented Ze-
XAI models roTrustBlock, a comprehensive blockchain framework for
secure and private health information exchange using the
• Serverless Aggregation Node(s) for the Global Model Hyperledger Fabric. The architecture and consensus proto-
• Distributed ledger for the Global Model, Local Model cols are designed to comply with security and confidentiality
Updates, Re-train model regulations.
Our proposed system uses FL and blockchain to en-
Our conceptual framework considers the architecture able the aggregation of information about customers with-
of Flexible Couple Blockchain-based Federated Learning out compromising customer privacy. To ensure privacy-
(FIC-BCFL), presented in Wang and Hu (2021). The archi- preserving features in credit score modelling, our proposed
tecture of the FIC-BCFL indicates the clients are responsible framework utilises federated model learning, incorporat-
for collecting and training local models. The miners of ing credit application data, customer transaction data, and
the blockchain perform the verification of the local model credit bureau information. The original data associated with
updates. The FL can ensure the parameters of the models are providers, banks, credit bureaus and enterprises are hashed,
stored on the blockchain, and the blockchain miners perform and the associated hash will be stored on the blockchain.
the aggregation of the global model. Privacy protection will be achieved using the SHA256 hash
Our conceptual proposal considers the Predefine Nodes per- algorithm. Our framework process considers the data is
forming the model training for FL. Those nodes are au- stored off-chain, and the hash of the data is stored on the
thorised to perform the model learning and are equipped blockchain to form a unique index to identify the corre-
with computational powers and storage to receive the data sponding off-chain data. Raw data is not shared in our
and train local models. Furthermore, our framework is an proposed framework. Only the model, model parameters,
extension to FIC-BCFL as it incorporates XAI as well. accuracy, and updates will be shared. Data sharing is a
Specifically, future implementation will consider 1) common use case for IPFS due to its high availability and
IPFS for distributed storage of data, the global AI and good performance, Nyaletey et al. (2019), hence supporting
XAI models’ parameters, 2) Consortium blockchain for the the idea of using IPFS in our proposed framework.
system’s logic and state, and 3) FL for the AI model learning.
5.2. Role of Local Model Miners
5.1. Scorecard Clients Our proposed conceptual framework considers the Pre-
Banks use various channels to gather data for credit define Nodes performing the model training for FL. Those
scoring, such as credit reports from bureaus, loan applica- nodes are authorised to perform the model learning and are
tions, and income verification. They may also analyze bank equipped with computational powers and storage to train
account transactions, review public records for legal infor- to receive the data and to train local models. The nodes
mation, and take behavioral data into consideration when associated with the clients are randomly selected to perform
assessing creditworthiness. Some banks even explore social the model training. The clients define the initial models.
media and online presence. Credit scoring models weigh The nodes train the models on the local data and upload
factors differently to calculate credit scores, incorporating hash local model parameters in the on-chain blockchain. The
information on payment history, outstanding debts, and fi- client node updates the model parameters in off-chain IPFS
nancial behaviours. It is important for banks to comply with for the same on-chain hashed local model parameters. The
privacy regulations throughout the data collection process to hash data and hash model generated locally will be stored
ensure the protection of individuals’ sensitive information. on the blockchain and maintained on-chain.
A good credit assessment consists of multisource data, To ensure the privacy of the training model is achieved,
such as banks’, bureaus’ and enterprises’ data. Multi-party we will use Paillie’s Cryptosystem, which is homomorphic
data enables a broader platform to provide a comprehen- encryption used in distributed machine learning.
sive model learning foundation for a good credit evalu-
ation system. The technology we consider is the consor-
tium blockchain, which ensures a strict access mechanism.

Jovanovic et al.: Preprint submitted to Elsevier Page 13 of 20

 Blockchain, XAI and Credit Scoring

Trained Global model

Assemble the Global model

MinerVER MinerXAI MinerVER MinerXAI

Blockchain

Training Local Models with privacy protection

Federated
Learning

Train Re-train

Bank Bureaus
Head Enterprises

Local Model
Updates
Scorecard Clients

Figure 6: Proposed conceptual framework.

5.3. Generation and verification of XAI Models global solution, it is necessary to use Convex optimisation,
The miners validate the local models by invoking the similar to the study presented in Chen and Wang (2014).
Smart Contract through an oracle to access the model param-
eters value in the table off-chain. The smart contact queries
IPFS for the model parameters with the same on-chain min 𝐶(𝒖) s.t. 𝜑𝑖 (𝒖)≤0,
𝒖 (7)
hashed model parameters. The local model’s authenticity is 𝜓𝑗 (𝒖)=0, 𝑖= 1, ..., 𝑙, 𝑗= 1, ..., 𝑚
confirmed by training the selected model on its local data.
Miners use the smart contract to invoke a pair of the local where 𝐶 is the cost function and 𝒖∈ℝ𝐻 is the optimization
hash data and local hash model to obtain data and model (control) variable. The functions 𝐶, 𝜑1 , ..., 𝜑𝑙 are convex
parameters from off-chain IPFS. Off-chain data is used to while the functions 𝜓1 , ..., 𝜓𝑚 are affine (Boyd et al. (2004)).
train and compare the model with the authentic model. A Selected miners will generate the XAI for the respected
range of different AI and XAI models will be considered, models. Our work will consider the proposed framework
such as Logistic Regression (LR), Random Forest (RF), presented in Al Mallah and López (2022). This decoupled
RidgeClassifier, GaussianNB, and SGDClassifier. the monitoring phase from the detection phase in defence
Most of the learning techniques in machine learning against poisoning attacks and replaced the centralised Fed-
belong to Non-convex Training. Training neural networks erated Learning - chief with the workers that collaborate
can pose challenges, particularly due to factors like sensitiv- to train the global model. Al Mallah and López (2022)
ity to initialization, step sizes, mini-batching, and optimizer proposed techniques to address the latency challenges by
selection. As a result, close monitoring and interpretation of decoupling the monitoring phase from the detection phase
the model’s learning process are crucial due to its intricate in decentralised FL approaches defences that protect against
black-box nature. To have model parameters representing the poisoning attacks in FL. The blockchain replaced the cen-
tralized aggregation of the traditional FL. It divided the
blockchain network into two types of miners: minersFL

Jovanovic et al.: Preprint submitted to Elsevier Page 14 of 20

 Blockchain, XAI and Credit Scoring

responsible for FL, and minersMON, responsible for mon- training process. Specifically, we will consider the FedLess
itoring. Workers perform the FL and send their local model (Grafberger et al. (2021)) framework and extend its average
updates to minersMON, responsible for monitoring. The model aggregation Federated Averaging (FedAvg) with op-
blockchain minersFL nodes randomly select a set of reliable timisation. The serverless computing platform to be used is
workers to continue the FL process and calculate the aver- AWS Lambda (AWS (2022)). The hash of the global model
age model using the updated model from the workers, and is to be stored on the blockchain. The miners who performed
minersMON. The proposed design does not store the model the verification were randomly selected for the global model
updates on the blockchain. Instead, the hash value is written aggregation and assembly. Federated Averaging (FedAvg)
on the blockchain and points towards the model updates. is the most common model aggregation technique in FL
The blockchain stores the commitments of all workers on proposed by McMahan et al. (2017), based on averaging the
the model updates they worked on. A Merkle tree is used to model weights across all clients.
authenticate the model updates submitted by the workers. Credit assessment requires an efficient, transparent, trace-
Hence, in our conceptual framework, we will consider able, secure, and sustainable process to reduce the model risk
decoupling the miners performing the model verification and provide adequate governan
minerVER from those miners responsible for generating Convex optimisation is a branch of mathematical op-
model explanation named minerXAI. This improvement will timisation focused on problems where both the objective
enable reliability, transparency and explainability of the function and the constraints are convex. It deals with finding
credit assessment model. The minerVER validates models a convex function’s minimum (or maximum) over a convex
using a k-fold cross-validation technique. Regarding model set. A set Ω ∈ ℝ𝑛 is convex if, for all 𝕩 and 𝕪 in Ω and for
reliability, each local model obtains its own set of metric all 𝜆 in [0, 1] it holds 𝜆𝕩 + (1 − 𝜆)𝕪.
functions after being trained. These metric functions are A convex function has the property that the line segment
used to evaluate a model’s performance based on specific between any two points on the function lie above the function
objectives, and they play a critical role in assessing predic- itself. Formally, it is defined by the following theorem.
tion errors. The SHAP model for explainability is used.
Lemma 1. (Boyd et al. (2004)) A function 𝕗 ∶ ℝ𝑛 → ℝ is
5.4. Functions of Serverless Aggregation Nodes convex if, for all 𝕩 and 𝕪 in the domain of 𝕗 and for all 𝜆 in
We will use serverless FL and blockchain to enhance the interval [0, 1], the following holds
the process of the global model aggregation without a cen- 𝕗 (𝜆𝕩 + (1 − 𝜆)𝕪) ≤ 𝜆𝕗 (𝕩) + (1 − 𝜆)𝕗 (𝕪) (8)
tralized aggregator. The serverless aggregation utilises the
cloud provider to ensure the scalability of the proposed so- The main property of convex optimisation is its ability to
lution. Our approach will employ serverless computing, FL, guarantee a global minimum, meaning it is possible to find
and blockchain to enable privacy-preserving, decentralised the best solution to a the problem rather than just a local
machine learning. In this approach, each client runs a local minimum.
machine learning model using serverless computing, and
the updates from each client are securely aggregated using
blockchain. This approach will enable a decentralised and max 𝒑𝑇 𝒘 − 12 𝒘𝑇 𝑄𝒘
𝒘 (9)
secure model training process without needing a central s.t. 𝐺𝒘 ≤ 𝒉, 𝐴𝒘 = 𝒃
server or data aggregator.
Grafberger et al. (2021) presented the workflow for train- where, 𝒑 is the mean accuracy of all accuracy types for
ing multiple clients using FedLess in a single FL round. each local model 𝑛-dimensional vector. 𝑄 is 𝑛×𝑛 covariance
The FL admin selects the model, registered client functions, matrix of local model accuracies that consists of the accu-
and hyperparameters. The FedLess controller requests a new racy classification score metrics used to measure the classi-
invocation token from the Auth Server and uses it along fication performance of considered classification models. 𝐴
with the credentials to access the parameter server to invoke is 𝑚 × 𝑛 real matrix, 𝐺 is 𝑚 × 𝑛 real matrix, 𝒃 is a real-valued
the clients randomly selected for this round. The clients 𝑚-dimensional vector. Quadratic programming aims to find
validate the signature and authorization of the token and an n-dimensional vector 𝒘 to meet the imposed constraints.
load the latest global model from the parameter server before The variable 𝒘 in our framework symbolises the weights
performing local training, optionally using Local Differen- allocated to each model, reflecting their significance derived
tial Privacy (LDP). Once training is finished, the clients from the accuracy of local data. The weights assigned to
upload their parameters to the parameter server. The FedLess each model emphasise their performance, contributing to a
controller waits until all clients have completed training and compelling ensemble that enhances the overall predictive
starts the model aggregation by invoking the aggregator power of the system. The Algorithm 1 presents local model
function. The aggregator loads the client results, aggregates training. The assembling of the global model is formalised
the parameters, and stores the new global model. Finally, in Algorithm 2.
the controller starts the evaluation, either using the global Regarding the practical implementation of the Algo-
test set or invoking a new selection of clients to evaluate rithm 2, time complexity is critical. The time complexity
their test set. It aggregates the returned metrics to resume the 𝑇 (𝑛) of our proposed Algorithm 2 is linear to the training
time of ML models. For example, if the chosen model is

Jovanovic et al.: Preprint submitted to Elsevier Page 15 of 20

 Blockchain, XAI and Credit Scoring

Algorithm 1 Local Model Training Algorithm 5.5. Model distribution and retraining
procedure T RAIN LOCAL MODEL Λ𝑖 (𝐷𝑖 ) The assembly of the global model will occur as an
Load data set 𝐷𝑖 that includes local features and labels off-chain process. The hash global model parameters and
Validate data weights will be stored on the blockchain. Similar to Li
Select type of classification model Λ𝑖 et al. (2020), we consider the Committee Consensus Mech-
Tuning the hyper-parameters of a model Λ𝑖 anism blocks to store the global model and local updates.
Train a local model Λ𝑖 Communication-based generated mechanisms reach an agree-
Evaluate a vector of model accuracy metric 𝜇𝑖 ment before appending blocks. Selected nodes will validate
Evaluate a mean value 𝜇̄𝑖 of a vector 𝜇𝑖 the updates.
return Λ𝑖 , 𝜇𝑖 , 𝜇̄𝑖 All clients can download the global model parameters
end procedure and weights from the blockchain and continue to use them
in the next round of learning models.
As the client’s data changes, the models may need to
Algorithm 2 Global Model Aggregation Algorithm
be updated. Therefore, retraining will be performed as a fit
procedure AGGREGATE GLOBAL MODEL Γ
method for new data. At the same time, the original model
Request 𝑁 local models
parameters are to be used as a starting point in the retraining
for 𝑖 ← 1, 𝑁 do
process. Re-training of the model will occur once a change
Λ𝑖 , 𝜇𝑖 , 𝜇̄𝑖 ← T RAIN LOCAL MODEL Λ𝑖
of the statistical properties is detected, such as a change in
end for
the Population Stability Index (PSI).
Concatenate accuracy vectors into matrix 𝑋
for 𝑖 ← 1, 𝑁 do
𝑋 ← 𝑐𝑜𝑛𝑐𝑎𝑡(𝑋, 𝜇𝑖 ) 6. Discussion and Analysis
end for Table 4 presents a comprehensive list of key charac-
Create matrix 𝑄 ← 𝑋 𝑇 𝑋 teristics required to build reliable credit scoring modelling.
Create vector 𝑝 ← [𝜇̄1 , … , 𝜇̄𝑁 ]𝑇 Specifically, credit score modelling requires diverse data
Calculate consensus weights 𝒘 to enable collaborative modelling while ensuring privacy,
transparency, and fairness. Our research proposes a novel
max 𝒑𝑇 𝒘 − 21 𝒘𝑇 𝑄𝒘 conceptual framework that integrates these elements previ-
𝒘 ously studied in isolation. Some previous research has ex-
s.t. 𝐺𝒘 ≤ 𝒉, 𝐴𝒘 = 𝒃 plored the use of the blockchain in the context of credit data
sharing (Zhang et al. (2020a); Zhu (2020)), credit evaluation
Compose a global model Γ ← 𝑤1 Λ1 + … + 𝑤𝑁 Λ𝑁
(Zeng et al. (2019); Zhang et al. (2019)) and storing expla-
return Γ
nation on the blockchain (Walambe et al. (2020)). However,
end procedure
our motivation aligns with a similar study Al Mallah and
López (2022), which discussed the use of the blockchain
for model authentication. While our unique research focus
Random Forest, then the time complexity would be 𝑂(𝑁 ⋅
is on the importance of model explainability for the spe-
𝑚 ⋅ 𝑙𝑜𝑔 𝑚 ⋅ 𝑑 ⋅ 𝑘), where 𝑚 is the number of training
cific credit score modelling application, which incorporates
samples, 𝑑 is the dimension (number of features), 𝑘 is the
privacy-preserving decentralised model learning combined
number of trees, and 𝑁 is the number of local models. We
with reliability, transparency, and explainability features of
assume that the input to the convex quadratic programming
the blockchain miners.
problem is much smaller than the size of the dataset, so its
The authors in Al Mallah and López (2022) used the
complexity is subsumed by the training time. The algorithm
blockchain to develop an immutable framework for decen-
for assembling the global model based on the optimisation
tralised, federated model learning. The Merkle tree was
will follow specifications defined in the smart contract.
utilized to store the local model updates to verify the validity
The uniqueness of our proposed approach lies in its non-
of the model updates. However, the study does not consider
iterative and parallel nature, suggesting potential efficiency
the explainability of the models and specific application of
gains over traditional iterative methods. An extension to our
credit scoring.
framework incorporates an integrated evaluation process,
The system proposed by Walambe et al. (2020) relies on
wherein local model prediction accuracy directly contributes
the quality of the machine learning model used for credit
to the assembly of the global model. To augment overall
scoring. If the model is not verified, the explanation stored
model accuracy, we propose an additional enhancement
on the blockchain may be inaccurate, resulting in incorrect
involving utilising XAI model input impact measures and an
credit scoring decisions. The proposed mechanism lacks a
accuracy matrix during the global model assembly. This ex-
technique to ensure diverse data is used in a decentralised
tension aims to provide a more comprehensive and accurate
FL model, and its performance is validated, thus limiting its
credit assessment mechanism.
effectiveness.

Jovanovic et al.: Preprint submitted to Elsevier Page 16 of 20

 Blockchain, XAI and Credit Scoring

The serverless function for training Federated Learning 7. Conclusion

FedLess is detailed in Grafberger et al. (2021), which utilised
Our research investigates the core features of XAI,
serverless technologies, AWS Lambda, Azure functions and
Openwhisk to enable multisource FL while providing model blockchain, and credit scoring. Specifically, we examine
aggregation. However, it is important to consider the limita- recent efforts to integrate XAI, blockchain, and FL for credit
tions of FedLess in the specific context of model verification scoring and identify limitations in these approaches. While
and explainability in the domain of credit scoring. Further these solutions primarily focus on enhancing data storage
security and privacy, we identify the need for combining
research and evaluation are needed to fully understand the
these technologies to ensure model verification, behavioural
effectiveness and limitations of FedLess in other areas, such
as credit score modelling. reliability, and model explainability for intelligent credit
The work presented in Zhang et al. (2020a) explored assessment. To address those challenges and create a reliable
the use of blockchain in credit data sharing. However, this and explainable credit scoring process, we propose a novel
framework that leverages the benefits of blockchain and
research is limited as it does not include model learning,
FL. Our framework’s distinctiveness lies in its holistic
verification, aggregation and explainability techniques in the
credit scoring model while utilising dynamic multisource design, which incorporates privacy-preserving decentralised
data. model learning coupled with the reliability, transparency,
A method to validate the quality of FL model gradients and explainability features of the blockchain.
and to determine reward is proposed in Martinez et al. We have thoroughly examined the qualitative features
necessary for designing an efficient credit scoring frame-
(2019). The advantages of the proposed approach encompass
work. In our future work, we will employ the framework
increased trust in the FL process and enhanced incentives for
participants during gradient validation. However, potential to quantify and evaluate the effectiveness of the proposed
limitations may arise from centralised model aggregation, architecture, including communication delay in a real envi-
lack of explainability of the trained models, and the impact ronment.
of new data on the training process in the domain of credit Our proposed framework has certain limitations that
could be addressed and improved upon in future research.
scoring.
The study is based on Hyperledger Fabric, which is a suitable
The proposed conceptual framework identifies the need
for combining different technologies to ensure model verifi- platform for credit scoring applications due to its scalabil-
cation, behavioural reliability, and model explainability for ity and compliance features. Its modular architecture and
intelligent credit scoring. Specifically, our framework uses permissioned blockchain model facilitate efficient workload
a Blockchain-based Federated Learning solution to enable distribution and make it well-suited for scaling up to meet
the demands of growing networks. By optimizing smart
decentralised model learning, verification of the models and
contracts, computational overhead can be minimized to en-
model aggregation on distributed multisource datasets.
sure effective transaction processing and contract execution.
Risks: Implementing blockchain technology and FL for Its features, such as private channels and access controls,
credit assessment has vulnerabilities that require attention. align with regulatory requirements for data privacy and
Regulatory compliance is a primary challenge that requires confidentiality in credit scoring. Its interoperability and au-
constant monitoring of evolving regulations and ensuring ditability features also support seamless integration with
data privacy, model explainability and reliability across all external systems and compliance with financial regulations,
parties involved. Technologies also face data privacy is- making it a reliable choice for building secure and scalable
sues that demand careful management of sensitive infor- credit scoring applications. Implement privacy-preserving
mation and privacy-preserving techniques. Interoperabil- techniques like zero-knowledge proofs and homomorphic
ity and scalability concerns may arise when integrating encryption to protect sensitive data. However, improvements
these technologies with existing financial systems. The real- related to performance, advanced cryptography, and real-
world challenge of achieving explainability and model inter- world pilot testing will be addressed in future work.
pretability persists due to the decentralized and collaborative
nature of FL and the limited clarity of blockchain trans- CRediT authorship contribution statement
actions. To mitigate these risks, it is essential to focus on
robust security measures, careful technological design, and Zorka Jovanovic: Formal analysis, Methodology, Writ-
ongoing collaboration with industry stakeholders and regu- ing - original draft. Zhe Hou: Conceptualization, Supervi-
lators. By doing so, we may ensure that these technologies sion, Writing - review and editing. Kamanashis Biswas:
can be safely and securely integrated into existing systems Conceptualization, Supervision, Writing - review and edit-
while maintaining data privacy, model explainability and ing. Vallipuram Muthukkumarasamy: Conceptualization,
regulatory compliance. Supervision, Writing - review and editing.

References
Al Mallah, R., López, D., 2022. Blockchain-based monitoring for poison at-
tack detection in decentralized federated learning, in: 2022 International

Jovanovic et al.: Preprint submitted to Elsevier Page 17 of 20

 Blockchain, XAI and Credit Scoring

Table 4
Comparison of credit assessment requirements among different models: Yes (✓), No (×).
Reference Multisource Model Model XAI Model Credit Application
Verification Aggregation
Al Mallah and López (2022) ✓ ✓ ✓ × ×
Walambe et al. (2020) × × × ✓ ✓
Grafberger et al. (2021) ✓ × ✓ × ×
Zhang et al. (2020a) ✓ × × × ✓
Martinez et al. (2019) ✓ ✓ ✓ × ×
Proposed Approach ✓ ✓ ✓ ✓ ✓

Conference on Electrical, Computer, Communications and Mechatron- Chakraborty, S., Aich, S., Seong, S.J., Kim, H.C., 2019. A blockchain
ics Engineering (ICECCME), IEEE. pp. 1–6. based credit analysis framework for efficient financial systems, in: 2019
Alam, S., Shuaib, M., Khan, W.Z., Garg, S., Kaddoum, G., Hossain, M.S., 21st International Conference on Advanced Communication Technol-
Zikria, Y.B., 2021. Blockchain-based initiatives: Current state and ogy (ICACT), IEEE. pp. 56–60.
challenges. Computer networks (Amsterdam, Netherlands : 1999) 198, Chen, B., Zeng, H., Xiang, T., Guo, S., Zhang, T., Liu, Y., 2022. Esb-
108395. fl: Efficient and secure blockchain-based federated learning with fair
Alarab, I., Prakoonwit, S., 2022. Effect of data resampling on feature impor- payment. IEEE Transactions on Big Data , 1–1doi:10.1109/TBDATA.2022.
tance in imbalanced blockchain data: Comparison studies of resampling 3177170.
techniques. Data Science and Management . Chen, S., Wang, Y., 2014. Convolutional neural network and convex
Aledhari, M., Razzak, R., Parizi, R.M., Saeed, F., 2020. Federated learning: optimization. Dept. of Elect. and Comput. Eng., Univ. of California at
A survey on enabling technologies, protocols, and applications. IEEE San Diego, San Diego, CA, USA, Tech. Rep .
Access 8, 140699–140725. Cheng, K., Fan, T., Jin, Y., Liu, Y., Chen, T., Papadopoulos, D., Yang, Q.,
Altman, E.I., 1968. Financial ratios, discriminant analysis and the predic- 2021. Secureboost: A lossless federated learning framework. IEEE
tion of corporate bankruptcy. The journal of finance 23, 589–609. Intelligent Systems 36, 87–98. doi:10.1109/MIS.2021.3082561.
Altmann, A., Toloşi, L., Sander, O., Lengauer, T., 2010. Permutation Cho, K.W., Jeong, B.G., Shin, S.U., 2021. Verifiable credential proof
importance: a corrected feature importance measure. Bioinformatics 26, generation and verification model for decentralized ssi-based credit
1340–1347. scoring data. IEICE Transactions on Information and Systems 104,
Arrieta, A.B., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., 1857–1868.
Barbado, A., García, S., Gil-López, S., Molina, D., Benjamins, R., et al., Chowdhury, M.J.M., Ferdous, M.S., Biswas, K., Chowdhury, N.,
2020. Explainable artificial intelligence (xai): Concepts, taxonomies, Muthukkumarasamy, V., 2020. A survey on blockchain-based platforms
opportunities and challenges toward responsible ai. Information fusion for iot use-cases. Knowledge engineering review 35.
58, 82–115. Dastile, X., Celik, T., Potsane, M., 2020. Statistical and machine learning
AWS, 2022. Aws lambda. URL: https://aws.amazon.com/lambda/. models in credit scoring: A systematic literature survey. Applied Soft
Bamakan, S.M.H., Bondarti, A.B., Bondarti, P.B., Qu, Q., 2021. Blockchain Computing 91, 106263.
technology forecasting by patent analytics and text mining. Blockchain: Dastile, X., Celik, T., Vandierendonck, H., 2022. Model-agnostic counter-
Research and Applications 2, 100019. factual explanations in credit scoring. IEEE Access 10, 69543–69554.
Bellagarda, J.S., Abu-Mahfouz, A.M., 2022. An updated survey on the Davis, R., Lo, A.W., Mishra, S., Nourian, A., Singh, M., Wu, N., Zhang,
convergence of distributed ledger technology and artificial intelligence: R., 2022. Explainable machine learning models of consumer credit risk.
Current state, major challenges and future direction. IEEE Access 10, Available at SSRN .
50774–50793. doi:doi.org/10.1109/ACCESS.2022.3173297. Dumitrescu, E., Hué, S., Hurlin, C., Tokpavi, S., 2022. Machine learning for
Boyd, S., Boyd, S.P., Vandenberghe, L., 2004. Convex optimization. credit scoring: Improving logistic regression with non-linear decision-
Cambridge university press. tree effects. European Journal of Operational Research 297, 1178–1192.
Breiman, L., 2001. Random forests. Machine learning 45, 5–32. El Azzaoui, A., Singh, S.K., Pan, Y., Park, J.H., 2020. Block5gintell:
Bride, H., Cai, C.H., Dong, J., Dong, J.S., Hóu, Z., Mirjalili, S., Sun, Blockchain for ai-enabled 5g networks. IEEE Access 8, 145918–
J., 2021. Silas: A high-performance machine learning foundation for 145935.
logical reasoning and verification. Expert Systems with Applications Fahner, G., 2018. Developing transparent credit risk scorecards more
176, 114806. effectively: An explainable artificial intelligence approach. Data Anal
Bride, H., Dong, J., Dong, J.S., Hóu, Z., 2018. Towards dependable and ex- 2018, 17.
plainable machine learning using automated reasoning, in: International Grafberger, A., Chadha, M., Jindal, A., Gu, J., Gerndt, M., 2021. Fedless:
Conference on Formal Engineering Methods, Springer. pp. 412–416. Secure and scalable federated learning using serverless computing, in:
Bücker, M., Szepannek, G., Gosiewska, A., Biecek, P., 2022a. Trans- 2021 IEEE International Conference on Big Data (Big Data), pp. 164–
parency, auditability, and explainability of machine learning models in 173. doi:10.1109/BigData52589.2021.9672067.
credit scoring. Journal of the Operational Research Society 73, 70–90. Guo, Y., Liang, C., 2016. Blockchain application and outlook in the banking
Bücker, M., Szepannek, G., Gosiewska, A., Biecek, P., 2022b. Trans- industry. Financial innovation 2, 1–12.
parency, auditability, and explainability of machine learning models in Hara, S., Hayashi, K., 2018. Making tree ensembles interpretable: A
credit scoring. Journal of the Operational Research Society 73, 70–90. bayesian model selection approach, in: International conference on
Calvaresi, D., Mualla, Y., Najjar, A., Galland, S., Schumacher, M., 2019. artificial intelligence and statistics, PMLR. pp. 77–85.
Explainable multi-agent systems through blockchain technology, in: In- Harris, T., 2015. Credit scoring using the clustered support vector machine.
ternational Workshop on Explainable, Transparent Autonomous Agents Expert Systems with Applications 42, 741–750.
and Multi-Agent Systems, Springer. pp. 41–58. Hassija, V., Bansal, G., Chamola, V., Kumar, N., Guizani, M., 2020. Se-
Casino, F., Dasaklis, T.K., Patsakis, C., 2019. A systematic literature review cure lending: Blockchain and prospect theory-based decentralized credit
of blockchain-based applications: Current status, classification and open scoring model. IEEE Transactions on Network Science and Engineering
issues. Telematics and informatics 36, 55–81. 7, 2566–2575.

Jovanovic et al.: Preprint submitted to Elsevier Page 18 of 20

 Blockchain, XAI and Credit Scoring

Hieu, N.Q., Anh, T.T., Luong, N.C., Niyato, D., Kim, D.I., Elmroth, Mao, D., Wang, F., Hao, Z., Li, H., 2018. Credit evaluation system
E., 2020. Resource management for blockchain-enabled federated based on blockchain for multiple stakeholders in the food supply chain.
learning: A deep reinforcement learning approach. arXiv preprint International journal of environmental research and public health 15,
arXiv:2004.04104 . 1627.
Hossain, M.S., Muhammad, G., Guizani, N., 2020. Explainable ai and mass Martinez, I., Francis, S., Hafid, A.S., 2019. Record and reward federated
surveillance system-based healthcare framework to combat covid-i9 like learning contributions with blockchain, in: 2019 International confer-
pandemics. IEEE Network 34, 126–132. ence on cyber-enabled distributed computing and knowledge discovery
Hurley, M., Adebayo, J., 2016. Credit scoring in the era of big data. Yale (CyberC), IEEE. pp. 50–57.
JL & Tech. 18, 148. McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A., 2017.
Hurlin, C., Pérignon, C., Saurin, S., 2022. The fairness of credit scoring Communication-efficient learning of deep networks from decentralized
models. arXiv preprint arXiv:2205.10200 . data, in: Artificial intelligence and statistics, PMLR. pp. 1273–1282.
Imteaj, A., Amini, M.H., 2022. Leveraging asynchronous federated learning Moinet, A., Darties, B., Baril, J.L., 2017. Blockchain based trust
to predict customers financial distress. Intelligent Systems with Appli- & authentication for decentralized sensor networks. arXiv preprint
cations 14, 200064. arXiv:1706.01730 .
Islam, M.R., Ahmed, M.U., Barua, S., Begum, S., 2022. A systematic Moscato, V., Picariello, A., Sperlí, G., 2021. A benchmark of machine
review of explainable artificial intelligence in terms of different appli- learning approaches for credit score prediction. Expert Systems with
cation domains and tasks. Applied Sciences 12, 1353. Applications 165, 113986.
Issa, W., Moustafa, N., Turnbull, B., Sohrabi, N., Tari, Z., 2023. Nakamoto, S., Bitcoin, A., 2008. A peer-to-peer electronic cash system.
Blockchain-based federated learning for securing internet of things: A Bitcoin.–URL: https://bitcoin. org/bitcoin. pdf 4, 15.
comprehensive survey. ACM Computing Surveys 55, 1–43. Nassar, M., Salah, K., ur Rehman, M.H., Svetinovic, D., 2020. Blockchain
Karandikar, N., Abhishek, R., Saurabh, N., Zhao, Z., Lercher, A., Marina, for explainable and trustworthy artificial intelligence. Wiley Interdisci-
N., Prodan, R., Rong, C., Chakravorty, A., 2021. Blockchain-based pro- plinary Reviews: Data Mining and Knowledge Discovery 10, e1340.
sumer incentivization for peak mitigation through temporal aggregation Nguyen, D.C., Ding, M., Pham, Q.V., Pathirana, P.N., Le, L.B., Seneviratne,
and contextual clustering. Blockchain: Research and Applications 2. A., Li, J., Niyato, D., Poor, H.V., 2021. Federated learning meets
Khan, I.A., Moustafa, N., Pi, D., Sallam, K.M., Zomaya, A.Y., Li, B., 2021. blockchain in edge computing: Opportunities and challenges. IEEE
A new explainable deep learning framework for cyber threat discovery Internet of Things Journal 8, 12806–12825.
in industrial iot networks. IEEE Internet of Things Journal . Nie, G., Rowe, W., Zhang, L., Tian, Y., Shi, Y., 2011. Credit card churn
Konečnỳ, J., McMahan, H.B., Yu, F.X., Richtárik, P., Suresh, A.T., Bacon, forecasting by logistic regression and decision tree. Expert Systems with
D., 2016. Federated learning: Strategies for improving communication Applications 38, 15273–15285.
efficiency. arXiv preprint arXiv:1610.05492 . Nyaletey, E., Parizi, R.M., Zhang, Q., Choo, K.K.R., 2019. Blockipfs-
Lampathaki, F., Agostinho, C., Glikman, Y., Sesana, M., 2021. Moving blockchain-enabled interplanetary file system for forensic and trusted
from ‘black box’to ‘glass box’artificial intelligence in manufacturing data traceability, in: 2019 IEEE International Conference on Blockchain
with xmanai, in: 2021 IEEE International Conference on Engineering, (Blockchain), IEEE. pp. 18–25.
Technology and Innovation (ICE/ITMC), IEEE. pp. 1–6. Oseni, A., Moustafa, N., Creech, G., Sohrabi, N., Strelzoff, A., Tari, Z.,
de Lange, P.E., Melsom, B., Vennerød, C.B., Westgaard, S., 2022. Explain- Linkov, I., 2022. An explainable deep learning framework for resilient
able ai for credit assessment in banks. Journal of Risk and Financial intrusion detection in iot-enabled transportation networks. IEEE Trans-
Management 15, 556. actions on Intelligent Transportation Systems .
Li, D., Han, D., Weng, T.H., Zheng, Z., Li, H., Liu, H., Castiglione, A., Li, Óskarsdóttir, M., Bravo, C., Sarraute, C., Vanthienen, J., Baesens, B., 2019.
K.C., 2022a. Blockchain for federated learning toward secure distributed The value of big data for credit scoring: Enhancing financial inclusion
machine learning systems: a systemic survey. Soft Computing 26, 4423– using mobile phone data and social network analytics. Applied Soft
4440. Computing 74, 26–39.
Li, D., Luo, Z., Cao, B., 2022b. Blockchain-based federated learning Patel, S.B., Bhattacharya, P., Tanwar, S., Kumar, N., 2020. Kirti: A
methodologies in smart environments. Cluster Computing 25, 2585– blockchain-based credit recommender system for financial institutions.
2599. IEEE Transactions on Network Science and Engineering 8, 1044–1054.
Li, H., Wu, J., Xu, H., Li, G., Guizani, M., 2021. Explainable intelligence- Polyviou, A., Velanas, P., Soldatos, J., 2019. Blockchain technology:
driven defense mechanism against advanced persistent threats: A joint financial sector applications beyond cryptocurrencies. Multidisciplinary
edge game and ai approach. IEEE Transactions on Dependable and Digital Publishing Institute Proceedings 28, 7.
Secure Computing 19, 757–775. Qiao, Y., Lan, Q., Wang, Y., Jia, S., Kuang, X., Yang, Z., Ma, C., 2023a.
Li, Y., Chen, C., Liu, N., Huang, H., Zheng, Z., Yan, Q., 2020. A Pevachain: Privacy-preserving ridge regression-based credit evaluation
blockchain-based decentralized federated learning framework with com- system using hyperledger fabric blockchain. Expert Systems with
mittee consensus. IEEE Network 35, 234–241. Applications , 119844.
Lo, S.K., Staples, M., Xu, X., 2021. Modelling schemes for multi-party Qiao, Y., Lan, Q., Wang, Y., Jia, S., Kuang, X., Yang, Z., Ma, C., 2023b.
blockchain-based systems to support integrity analysis. Blockchain: Pevachain: Privacy-preserving ridge regression-based credit evaluation
Research and Applications 2, 100024. system using hyperledger fabric blockchain. Expert Systems with
Lundberg, S.M., Lee, S.I., 2017. A unified approach to interpreting model Applications 223, 119844.
predictions. Advances in neural information processing systems 30. Qiao, Y., Lan, Q., Zhou, Z., Ma, C., 2022. Privacy-preserving credit eval-
Ma, S., Shi, J., Huang, Y., Qin, S., Hou, Z., 2022. Muc-driven feature uation system based on blockchain. Expert Systems with Applications
importance measurement and adversarial analysis for random forest. 188, 115989.
arXiv preprint arXiv:2202.12512 . Qu, Y., Uddin, M.P., Gan, C., Xiang, Y., Gao, L., Yearwood, J., 2022.
Malhotra, D., Srivastava, S., Saini, P., Singh, A.K., 2021. Blockchain based Blockchain-enabled federated learning: A survey. ACM Computing
audit trailing of xai decisions: Storing on ipfs and ethereum blockchain, Surveys 55, 1–35.
in: 2021 International Conference on COMmunication Systems and Radha, S.K., Taylor, I., Nabrzyski, J., Barclay, I., 2021. Verifiable badging
NETworkS (COMSNETS), pp. 1–5. doi:10.1109/COMSNETS51098.2021. system for scientific data reproducibility. Blockchain: Research and
9352908. Applications 2, 100015.
Mallah, R.A., Lopez, D., Marfo, G.B., Farooq, B., 2021. Untargeted Rajbahadur, G.K., Wang, S., Oliva, G.A., Kamei, Y., Hassan, A.E., 2022.
poisoning attack detection in federated learning via behavior attestation. The impact of feature importance methods on the interpretation of defect
arXiv preprint arXiv:2101.10904 . classifiers. IEEE Transactions on Software Engineering 48, 2245–2261.

Jovanovic et al.: Preprint submitted to Elsevier Page 19 of 20

 Blockchain, XAI and Credit Scoring

Ribeiro, M.T., Singh, S., Guestrin, C., 2016. " why should i trust you?" for secure storage of human readable justifications for credit risk as-
explaining the predictions of any classifier, in: Proceedings of the 22nd sessment, in: Garg, D., Wong, K., Sarangapani, J., Gupta, S.K. (Eds.),
ACM SIGKDD international conference on knowledge discovery and International Advanced Computing Conference, Springer, Singapore.
data mining, pp. 1135–1144. pp. 55–72. doi:https://doi.org/10.1007/978-981-16-0404-1_5.
Rjoub, G., Bentahar, J., Wahab, O.A., 2022. Explainable ai-based feder- Wamba, S.F., Queiroz, M.M., 2020. Blockchain in the operations and
ated deep reinforcement learning for trusted autonomous driving, in: supply chain management: Benefits, challenges and future research
2022 International Wireless Communications and Mobile Computing opportunities. International Journal of Information Management 52,
(IWCMC), IEEE. pp. 318–323. 102064. doi:https://doi.org/10.1016/j.ijinfomgt.2019.102064.
Sachan, S., Yang, J.B., Xu, D.L., Benavides, D.E., Li, Y., 2020. An Wan, A., Dunlap, L., Ho, D., Yin, J., Lee, S., Jin, H., Petryk, S., Bargal,
explainable ai decision-support-system to automate loan underwriting. S.A., Gonzalez, J.E., 2020. Nbdt: neural-backed decision trees. arXiv
Expert Systems with Applications 144, 113100. preprint arXiv:2004.00221 URL: https://arxiv.org/abs/2004.00221.
Sahakyan, M., Aung, Z., Rahwan, T., 2021. Explainable artificial intelli- Wang, Z., Hu, Q., 2021. Blockchain-based federated learning: A compre-
gence for tabular data: A survey. IEEE Access 9, 135392–135422. hensive survey. arXiv preprint arXiv:2110.02182 URL: https://arxiv.
Salah, K., Rehman, M.H.U., Nizamuddin, N., Al-Fuqaha, A., 2019. org/abs/2110.02182.
Blockchain for ai: Review and open research challenges. IEEE Access West, D., 2000. Neural network credit scoring models. Computers
7, 10127–10149. & Operations Research 27, 1131–1152. doi:https://doi.org/10.1016/
Sattler, F., Wiedemann, S., Müller, K.R., Samek, W., 2019. Robust and S0305-0548(99)00149-5.
communication-efficient federated learning from non-iid data. IEEE Wu, L., Ruan, W., Hu, J., He, Y., 2023. A survey on blockchain-based
transactions on neural networks and learning systems 31, 3400–3413. federated learning. Future Internet 15, 400.
Schwalbe, G., Finzel, B., 2021. A comprehensive taxonomy for explain- Xia, Y., Liu, C., Li, Y., Liu, N., 2017. A boosted decision tree approach
able artificial intelligence: A systematic survey of surveys on meth- using bayesian hyper-parameter optimization for credit scoring. Expert
ods and concepts. arXiv e-prints , arXiv–2105doi:doi.org/10.1007/ Systems with Applications 78, 225–241. doi:https://doi.org/10.1016/
s10618-022-00867-8. j.eswa.2017.02.017.
Siddiqi, N., 2012. Credit risk scorecards: developing and implementing Yang, F., Abedin, M.Z., Hajek, P., 2023. An explainable federated learning
intelligent credit scoring. volume 3. John Wiley & Sons. and blockchain-based secure credit modeling method. European Journal
Siddiqi, N., 2017. Intelligent credit scoring: Building and implementing of Operational Research .
better credit risk scorecards. John Wiley & Sons. Yang, F., Qiao, Y., Abedin, M.Z., Huang, C., 2022a. Privacy-preserved
Srinivasan, R., Chander, A., Pezeshkpour, P., 2019. Generating user- credit data sharing integrating blockchain and federated learning for
friendly explanations for loan denials using gans. arXiv preprint industrial 4.0. IEEE Transactions on Industrial Informatics 18. doi:10.
arXiv:1906.10244 doi:https://doi.org/10.48550/arXiv.1906.10244. 1109/TII.2022.3151917.
Stiehle, F., Weber, I., 2022. Blockchain for business process enactment: Yang, F., Qiao, Y., Qi, Y., Bo, J., Wang, X., 2022b. Bacs: blockchain
A taxonomy and systematic literature review, in: Marrella, A., Matule- and automl-based technology for efficient credit scoring classifica-
vičius, R., Gabryelczyk, R., Axmann, B., Bosilj Vukšić, V., Gaaloul, W., tion. Annals of Operations Research , 1–21doi:https://doi.org/10.
Indihar Štemberger, M., Kő, A., Lu, Q. (Eds.), International Conference 1007/s10479-022-04531-8.
on Business Process Management, Springer. pp. 5–20. doi:doi.org/10. Yin, X., Zhu, Y., Hu, J., 2021. A comprehensive survey of privacy-
1007/978-3-031-16168-1_1. preserving federated learning: A taxonomy, review, and future direc-
Stockburger, L., Kokosioulis, G., Mukkamala, A., Mukkamala, R.R., Avi- tions. ACM Computing Surveys (CSUR) 54, 1–36. doi:https://doi.
tal, M., 2021. Blockchain-enabled decentralized identity management: org/10.1145/3460427.
The case of self-sovereign identity in public transportation. Blockchain: Yobas, M.B., Crook, J.N., Ross, P., 2000. Credit scoring using neural and
Research and Applications 2, 100014. doi:doi.org/10.1016/j.bcra.2021. evolutionary techniques. IMA Journal of Management Mathematics 11,
100014. 111–125. doi:https://doi.org/10.1093/imaman/11.2.111.
Tanwar, S., 2022. Blockchain Technology: From Theory to Practice. Zeng, X., Hao, N., Zheng, J., Xu, X., 2019. A consortium blockchain
Springer Nature. paradigm on hyperledger-based peer-to-peer lending system. China
Thantharate, P., Thantharate, A., 2023. Zerotrustblock: Enhancing security, Communications 16, 38–50. doi:10.23919/JCC.2019.08.004.
privacy, and interoperability of sensitive data through zerotrust permis- Zhang, C., Xie, Y., Bai, H., Yu, B., Li, W., Gao, Y., 2021a. A survey on
sioned blockchain. Big Data and Cognitive Computing 7, 165. federated learning. Knowledge-Based Systems 216, 106775. doi:https:
Thomas, L., Crook, J., Edelman, D., 2017. Credit scoring and its applica- //doi.org/10.1016/j.knosys.2021.106775.
tions. SIAM. Zhang, G., Hou, Z., Huang, Y., Shi, J., Bride, H., Dong, J.S., Gao, Y.,
Thomas, L.C., 2000. A survey of credit and behavioural scoring: forecasting 2021b. Extracting optimal explanations for ensemble trees via logical
financial risk of lending to consumers. International Journal of Forecast- reasoning. arXiv preprint arXiv:2103.02191 URL: https://arxiv.org/
ing 16, 149–172. doi:doi.org/10.1016/S0169-2070(00)00034-0. abs/2103.02191.
Turkson, R.E., Baagyere, E.Y., Wenya, G.E., 2016. A machine learning Zhang, J., Lyu, T., Li, R., 2019. A study on smie credit evaluation model
approach for predicting bank credit worthiness, in: 2016 Third Inter- based on blockchain technology. Procedia CIRP 83, 616–623. doi:https:
national Conference on Artificial Intelligence and Pattern Recognition //doi.org/10.1016/j.procir.2019.05.003.
(AIPR), IEEE. pp. 1–7. doi:10.1109/ICAIPR.2016.7585216. Zhang, J., Tan, R., Su, C., Si, W., 2020a. Design and application of
Uddin, M.A., Stranieri, A., Gondal, I., Balasubramanian, V., 2021. A a personal credit information sharing platform based on consortium
survey on the adoption of blockchain in iot: challenges and solutions. blockchain. Journal of Information Security and Applications 55,
Blockchain: Research and Applications 2, 100006. doi:doi.org/10.1016/ 102659. doi:https://doi.org/10.1016/j.jisa.2020.102659.
j.bcra.2021.100006. Zhang, W., Xu, W., Hao, H., Zhu, D., 2020b. Cost-sensitive multiple-
Verma, A., Bhattacharya, P., Madhani, N., Trivedi, C., Bhushan, B., Tanwar, instance learning method with dynamic transactional data for personal
S., Sharma, G., Bokoro, P.N., Sharma, R., 2022. Blockchain for industry credit scoring. Expert Systems with Applications 157, 113489.
5.0: Vision, opportunities, key enablers, and future directions. IEEE Zhang, Z., Niu, K., Liu, Y., 2020c. A deep learning based online credit
Access doi:10.1109/ACCESS.2022.3186892. scoring model for p2p lending. IEEE Access 8, 177307–177317. doi:10.
Vilone, G., Longo, L., 2020. Explainable artificial intelligence: a systematic 1109/ACCESS.2020.3027337.
review. arXiv preprint arXiv:2006.00093 doi:doi.org/10.48550/arXiv: Zhu, X., 2020. Blockchain-based identity authentication and intelligent
2006.00093. credit reporting. Journal of Physics: Conference Series 1437, 012086.
Walambe, R., Kolhatkar, A., Ojha, M., Kademani, A., Pandya, M., Kathote, doi:10.1088/1742-6596/1437/1/012086.
S., Kotecha, K., 2020. Integration of explainable ai and blockchain

Jovanovic et al.: Preprint submitted to Elsevier Page 20 of 20

View publication stats