Blockchain-based two-level trustable reputation framework for e-commerce platform using smart contracts

Abstract

E-commerce platforms incorporate reviews and reputation systems, allowing retailers and customers to manage and track their financial transactions. Consequently, it is crucial to design a reliable reputation system for the e-commerce environment, as it faces well-documented threats, including sybil attacks, feedback collusion, impersonation, review tampering, and whitewashing attacks. Current centralized systems are vulnerable to impersonation attacks, feedback manipulation, and lack automated verification against collusion-based reputation distortion. These unwanted ratings and reviews are highly correlated with abnormal cyber-attacks that damage both seller reputations and buyer experiences. To address these challenges, we propose a Blockchain-based Two-Level E-Commerce Trustable Reputation Framework (BTL-ETRF) utilizing deep learning-embedded transformers and redactable blockchain systems. Initially, we implement Multi-Factor Authentication for e-commerce users, utilizing three factors: PIN, OTP, and biometric fingerprint, to mitigate impersonation attacks. Only authenticated users are allowed to proceed to the reputation verification stage, where the proposed work considers five major metrics to classify user reputation using the Residual Dilated Convolution Transformer. To automate the reputation verification process, we design and employ two smart contracts, the Authentication Smart Contract and the Reputation Smart Contract which trigger automated actions based on the BTL-ETRF results. All transactions include reputation classification, and triggered actions are stored in the redactable blockchain, which can modify the stored transactions if needed. Finally, we demonstrate the performance of the proposed BTL-ETRF using Python and Ethereum Solidity, and conduct a formal analysis that shows the proposed model outperforms the compared works.

Introduction

E-commerce fraud includes deceptive activities carried out by individuals or groups to achieve unauthorized transactions¹, steal personal or financial data, or exploit e-commerce platforms for financial gain. Common examples include identity theft, phishing scams, chargeback fraud, affiliate fraud², and misleading advertising³. Therefore, detecting fraud is critical in e-commerce transactions to safeguard customers⁴ and prevent financial losses. Various techniques are employed to identify fraudulent activities, including transaction monitoring, IP address geolocation, and device fingerprinting⁵. With technological advancements, machine learning (ML) algorithms have been widely adopted to analyse transaction data and identify patterns indicative of fraudulent behaviour. By training these models on historical transaction records, suspicious transactions can be effectively detected⁶.

The rapid growth of the e-commerce sector has further amplified these concerns. Over the past decade, its share of global GDP has increased from 0.8% to over 2.7%. Simultaneously, the worldwide cost of cybercrime has risen significantly, from approximately $567 billion in 2018 to over $900 billion in 2020. Fraud and cybercrime in e-commerce include activities such as fake account creation, payment fraud, account hijacking, and fake review generation. In this context, ML techniques play an essential role in detecting and mitigating fraudulent activities within e-commerce platforms. Several large-scale platforms, including eBay, Microsoft, and LinkedIn^7–9, have integrated ML-based fraud detection mechanisms.

However, in practical deployments, ML-based fraud detection systems are often integrated into broader anti-fraud departments, where automated predictions are complemented by manual investigations conducted by fraud analysts^10,11. This hybrid approach requires incorporating fraud detection outputs into routine operational workflows. Despite extensive research in fraud detection, limited attention has been given to aligning academic solutions with the operational requirements of real-world anti-fraud departments¹². This gap between research and deployment raises concerns regarding scalability, automation, and enforcement efficiency.

Furthermore, the decentralized and dynamic nature of the Internet has motivated the adoption of blockchain-based solutions to enhance transparency and data integrity in online transactions^13,14. Blockchain enables secure and distributed storage of transactional data without reliance on centralized authorities¹⁵. Blockchain-based payment systems employ cryptographic mechanisms to ensure transaction confidentiality and integrity^16,17. Such systems improve transparency, traceability, and resistance to unauthorized data manipulation. Access to sensitive user data stored on the blockchain is restricted through cryptographic keys, thereby strengthening data protection.

Despite these advancements, reputation mechanisms in e-commerce platforms remain vulnerable to structured attacks, including Sybil attacks, ballot stuffing, bad-mouthing, whitewashing (re-entry), collusion between buyers and sellers, and impersonation during authentication. Most existing solutions focus primarily on transaction fraud detection or centralized moderation, leaving reputation manipulation insufficiently addressed. A significant limitation of current systems is their inadequate process management during trading activities^18,19, which restricts comprehensive monitoring and automated enforcement.

Trust in e-commerce platforms is largely established through user-generated reviews and reputation scores. However, these mechanisms are susceptible to coordinated fake review campaigns, artificial rating inflation, targeted reputation deflation, and unauthorized modification of stored feedback. Consequently, ensuring verifiable, tamper-resistant, and automatically validated reputation computation remains a critical research challenge. Although prior studies have explored trust modelling and long-term reputation development, limited emphasis has been placed on integrating authentication, automated behavioural verification, and secure blockchain-based enforcement mechanisms within a unified framework. Consumers²⁰ continue to express concerns regarding the reliability of recommendations and the credibility of user reviews, which directly influence purchasing decisions.

To address the above challenges, we focus specifically on reputation-system-level threats rather than general payment fraud. In this work, we consider the following adversarial behaviours such as (i) Sybil account creation, (ii) ballot stuffing and bad-mouthing through transaction-linked feedback manipulation, (iii) collusion between buyers and sellers to artificially inflate or suppress reputation, (iv) whitewashing attacks through re-entry, (v) impersonation attempts during authentication, and (vi) unauthorized modification of stored feedback records. Existing blockchain-based reputation systems ensure immutability but lack automated verification mechanisms capable of detecting such structured behavioural patterns. To overcome these limitations, we propose a Blockchain-based Two-Level E-Commerce Trustable Reputation Framework (BTL-ETRF), which integrates Multi-Factor Authentication (MFA), a Residual Dilated Convolution Transformer (RDCT) for behavioural reputation classification, automated smart contract enforcement, and a redactable blockchain mechanism to enable auditable yet flexible data modification under controlled cryptographic assumptions.

Related works

In the related work section, first, Blockchain-based Reputation Systems are discussed, followed by Blockchain-based Fraud Detection Approaches.

1. Blockchain-based reputation systems

The utilization of blockchain technology to generate a decentralized reputation system for e-commerce is presented²¹. Traditional centralized systems often experience operational issues, censorship, and single points of failure, which compromise the reliability of operator feedback. Here, the proposed blockchain-based system utilizes a decentralized, unchallengeable ledger to ensure transparent, secure, and trustworthy user reviews and ratings. Furthermore, this removes intermediaries, decreases costs, and delivers a tamper-proof environment, thereby refining the integrity and reliability of e-commerce platforms. The study²² shows the trust-based model for e-commerce applications utilizing blockchain technology. The traditional centralized security mechanisms are predisposed to cyber threats. In this study, the model utilizes blockchain to implement robust security policies with smart contracts and cryptographic techniques, confirming secure transactions and defending user credentials. This decentralized method improves data integrity and trust by creating a tamper-proof environment, thereby enhancing user confidence in e-commerce applications. The study²³ proposed a blockchain-based reputation system for e-commerce platforms that balances anonymity and verifiability. Traditional reputation systems struggle with user privacy and the authenticity of feedback. Furthermore, this system employs cryptographic protocols to anonymize user characteristics while ensuring that reviews and ratings remain verifiable. By applying zero-knowledge proofs and ring signatures, we preserve the privacy of users and the integrity of their feedback, making a safe and trustworthy reputation system that defends user anonymity without negotiating the authenticity of their interactions. The study proposed²⁴ elaborates on a blockchain-based reputation system for e-commerce that grows verifiable credentials to expand trust and accountability. Additionally, this system stores user credentials and feedback on a blockchain, confirming they are tamper-proof and simply verifiable. Furthermore, users can regulate their data and reputation, which increases transparency and reliability in the e-commerce environment. By leveraging the blockchain’s decentralized nature, this system provides a robust framework for managing reputation, fostering a more trustworthy and user-centric e-commerce experience. The study²⁵ presents a static ML-based method for assessing the usability and security of e-commerce websites. By using static analysis methods, this paper observes the website’s code and structure without execution to classify vulnerabilities and usability issues. ML algorithms measure user interface design, navigation efficiency, and security flaws. Additionally, this technique provides a quick and reliable method for enhancing e-commerce websites, enabling developers to implement actionable visions that improve usability and security, ultimately leading to enhanced user experiences and safer online transactions.

Although these blockchain-based reputation systems improve transparency and immutability, most of them primarily focus on secure storage of feedback rather than intelligent behavioural verification. They lack automated mechanisms for detecting collusion patterns, Sybil attacks, or structured reputation manipulation campaigns.

• (b)Blockchain-based fraud detection approaches

A novel fraud detection system, eFraudCom²⁶, is presented for e-commerce platforms utilizing competitive graph neural networks. Furthermore, this method enables the modeling and examination of connections between entities, thereby recognizing fraudulent behavior with increased accuracy. By developing a competitive learning environment, Efraudcom enhances its ability to detect subtle anomalies, providing a robust solution for justifying fraud in e-commerce transactions. The study in²⁷ presents a new technique for detecting fraud on the Ethereum blockchain by categorizing irregular smart contracts. By using blockchain’s transparency and immutability, we grow an anomaly detection system that utilizes ML to examine contract execution designs and recognize irregularities. Furthermore, this technique successfully pinpoints fraudulent activities, improving security and trust in decentralized applications by providing a reliable mechanism for detecting and mitigating fraudulent contracts. In the study²⁸, the author proposed a combined approach merging blockchain technology, smart contracts, and data mining to develop e-commerce operations. The blockchain ensures transparency and security, while smart contracts mechanize and enforce agreements. Although data mining techniques are used to extract valuable insights from transaction data and refine decision-making processes, this cooperative approach addresses mutual challenges in e-commerce, including fraud, inefficiencies, and data integrity issues, ultimately leading to a more secure and efficient e-commerce ecosystem. The study²⁹ presents a context-aware method for calculating transaction trust in e-commerce environments. Here, our model joins contextual information, such as transaction history, user behavior, and environmental variables, to dynamically measure trust levels. By using the Algorithms, this technique delivers a more intricate and precise trust computation, by improving security and reliability in e-commerce transactions and developing greater user confidence. The study proposed³⁰, the author develops a novel technique for detecting abnormal transactions in e-commerce via an extended data flow conformance checking. Unlike traditional methods, our approach incorporates additional transaction characteristics and behavior patterns to capture complex transaction flows. Furthermore, by categorizing deviations from expected transaction flows, this method efficiently identifies potential fraud, and our proposed system enhances the accuracy and reliability of fraud detection, thereby ensuring safer online transactions on e-commerce platforms. The study³¹ presents a combined neural network model designed to classify fake reviews on e-commerce platforms using multidomain datasets. Furthermore, the Fake reviews challenge trust and alter consumer perceptions. The model employs advanced neural network architectures to examine review patterns across multiple domains, improving its capacity to differentiate between genuine and deceptive reviews. Hereby, amalgamating diverse data sources, the model attains the highest accuracy in detecting fake reviews, thus contributing to a more trustworthy e-commerce environment. The study³² introduces a logistic regression-based framework for detecting fraud at the point of sale on e-commerce sites. But the traditional fraud detection systems often face limitations in real-time processing and accuracy. Here, our framework uses logistic regression to examine transaction data, recognizing suspicious activities proficiently. Additionally, by amalgamating the key transactional features and behavioral indicators, the model delivers a robust solution for fraud detection, enhancing the security of e-commerce transactions and protecting merchants and customers from fraudulent activities. The study proposed³³, the author of this paper has proposed an innovative technique for assessing the trustworthiness of e-commerce initiatives that suggested a high-involvement knowledge product. These products need important consumer trust due to their nature and value. Furthermore, this method combines qualitative and quantitative assessments, incorporating factors such as customer reviews, transaction history, and service quality. By retaining a complete trust evaluation model, they provided a reliable mechanism for assessing the credibility of e-commerce enterprises, thereby developing greater consumer confidence in high-stakes purchases. The study proposed³⁴, the paper shows a static ML-based technique for measuring the usability and security of e-commerce websites. Unlike dynamic analysis, this method evaluates the website’s code and construction without execution, detecting potential vulnerabilities and usability issues. Furthermore, Algorithms examine factors like user interface design, navigation efficiency, and security flaws. This system offers a fast and dependable way to enhance e-commerce websites, contributing designers’ actionable insights to improve user experience and security. The study in³⁵ introduces BCFDPS, a blockchain-based scheme considered to detect and prevent click fraud in online advertising. Furthermore, the click fraud interrupts advertising metrics and results in financial losses. BCFDPS utilizes blockchain’s transparency and immutability to record and verify click data securely. Through smart contracts and anomaly detection algorithms, it categorizes fraudulent clicks in real-time, safeguarding accurate billing and advertisers from fraud. Finally, this blockchain-enabled solution improves the integrity and trustworthiness of online advertising ecosystems. This study presents³⁶ a novel process for detecting fraud in e-commerce transactions, utilizing Specific Techniques. In addition to that, the technique employs numerous algorithms to examine transaction patterns and detect anomalies related to fraudulent behavior. By mixing a wide range of features from transactional data, the technique accomplishes high accuracy in separating genuine from fraudulent transactions. This approach improves e-commerce platform security, contributing a robust mechanism for real-time fraud detection and prevention. The study in³⁷ presents a novel approach to enhancing worker recruitment in collaborative mobile crowdsourcing. Although traditional methods struggle to measure worker trustworthiness accurately, this procedure utilizes them to model worker relationships and interactions, enabling dynamic and context-aware trust evaluation. Additionally, this method improves the selection process by identifying consistent workers, thereby refining the efficiency and quality of crowdsourcing projects overall. The Study in³⁸ explores the application of Techniques for fraud detection in e-commerce transactions. Numerous algorithms are used to examine transaction data and classify fraudulent activities. By utilizing historical transaction patterns and behavioral indicators, the proposed model efficiently differentiates between sincere and suspicious transactions. The ML-based method suggests a scalable and accurate solution for detecting fraud, improving the security and trustworthiness of e-commerce platforms. In the study³⁹, the author develops an effectual feature engineering model planned for fraud detection in e-commerce. Here, Feature engineering is vital for the Models’ performance. Furthermore, this method methodically extracts and selects the most pertinent features from transactional data, enhancing the model’s ability to detect fraudulent patterns. Ultimately, this approach enhances detection accuracy while minimizing computational complexity, making it suitable for real-time e-commerce fraud prevention applications. The Study in⁴⁰ explores how blockchain technology can enhance e-commerce supply chains, emphasizing privacy protection and sales mode selection. The blockchain’s transparency and immutability challenge key tests in supply chain management. Hence, they proposed a blockchain-based framework ensuring data privacy and allowing secure sales mode selection. Additionally, this approach improves supply chain integrity and trust, providing security, efficiency, and better decision-making for e-commerce stakeholders.

While these approaches leverage machine learning and blockchain for fraud detection, they predominantly target transactional anomalies rather than reputation-system-level manipulation. Moreover, they do not integrate multi-factor authentication, smart contract automation, and redactable blockchain mechanisms within a unified reputation verification framework.

Recent privacy-preserving reputation schemes have also been proposed in vehicular and mobile crowdsensing environments. The PPDR scheme⁴¹ introduces a dual-reputation mechanism in vehicle platoons to separate leader and follower trust scores while preserving privacy through cryptographic techniques. Similarly, the PPRU framework⁴² proposes a privacy-preserving reputation updating mechanism for cloud-assisted vehicular networks, leveraging elliptic curve cryptography and homomorphic operations to reduce Trusted Authority overhead. In addition, a lightweight privacy-preserving reputation management scheme for mobile crowdsensing in vehicular networks⁴³ focuses on minimizing computational cost while ensuring identity anonymity and secure feedback aggregation. Although these schemes provide strong privacy guarantees in vehicular contexts, their threat models and architectural assumptions differ significantly from those of large-scale e-commerce platforms. They primarily address location privacy, intermittent connectivity, and vehicular communication constraints rather than structured reputation manipulation such as collusion, ballot stuffing, and whitewashing. Moreover, they do not integrate deep learning-based behavioural analysis or redactable blockchain mechanisms for controlled data correction. In contrast, the proposed BTL-ETRF framework combines transformer-based reputation modelling, multi-factor authentication, smart contract automation, and auditable blockchain redaction tailored specifically for e-commerce ecosystems.

In summary, the existing blockchain-based reputation systems for e-commerce often suffer from centralized designs, tampering risks, and a lack of automated verification mechanisms. Many prior studies focus on blockchain integration but fail to incorporate multi-factor authentication to mitigate impersonation and unauthorized access. Furthermore, reputation verification in earlier works lacks deep learning-driven classification models that can accurately identify malicious or abnormal reviews. Another gap lies in the absence of smart contract–based automation, as most systems rely on semi-manual processes for reputation handling. Lastly, prior blockchain implementations are immutable, but they lack redactable blockchain mechanisms, limiting flexibility in securely modifying stored transactions when errors or disputes arise.

Proposed framework

The proposed Blockchain-based Two-Level E-Commerce Trustable Reputation Framework (BTL-ETRF) is designed to provide secure authentication, intelligent reputation verification, and automated enforcement within decentralized e-commerce environments. The framework integrates Multi-Factor Authentication (MFA), a transformer-based behavioural analysis module (RDCT), smart contract-driven decision enforcement, and a redactable blockchain storage mechanism. Together, these components establish a secure, tamper-resistant, and automated reputation verification pipeline. The architecture operates in two primary layers: (i) authentication-level verification to mitigate impersonation and unauthorized access, and (ii) behavioural-level reputation classification to detect structured manipulation patterns such as collusion and feedback inflation. The overall system architecture is illustrated in Fig. 1. The framework’s detailed components are described in the following subsections.

Fig. 1.

Overall architecture of the proposed BTL-ETRF Model.

System model

To delve into the proposed BTL-ETRF architecture, this supplementary section outlines the roles of the entities participating in this environment. The roles are explained as follows,

Buyers: Buyers are the ones who consume the products remotely. Their role is to purchase products based on previous ratings displayed by the application interface. They can also submit product ratings based on the product quality.

Sellers: Sellers are product distributors who deploy the product’s price, name, and pictures on e-commerce sites. They can also advertise their products through the application interface. For credibility and quality, they can also post ratings of their products.

Application interface: The role of the application interface is to ensure interactions among the smart contract, blockchain, and users. It also invokes the smart contract, collects reputation scores and product information, and distributes this information to users. Additionally, the application interface stores the product addresses in its memory.

Smart contract: A Smart contract is an executable software program that is installed in the blockchain. Smart contracts automate transactions, revocation, monetary activities, and authentication by propagating actions to the distributed applications.

Redactable blockchain: The role of blockchain is to securely store seller details and product information, buyer transactions, taken actions, and reputation scores, and provide this information as feedback to the application interface.

Business phase

In this phase, retailers and consumers perform two operations, which are mentioned as follows,

Product updating:

• Initially, retailers provide their product details, including pricing, name, pictures, and advertising content, to the application interface.

• The application interface passes the launched product description to the blockchain file system, which in turn acknowledges the product address.

• Once the acknowledged address arrives at the application interface, it then summons the smart contract.

• The smart contract provides the address to the blockchain as a block and provides feedback.

Product purchase:

• From the application interface, consumers select and review products for purchase.

• From the application interface, the consumers decide to buy the product.

• The monetary transactions are implemented and summoned by the application interface.

• The transaction actions are provided to the smart contract and blockchain, which in turn returns feedback to the application interface.

Assessment phase

Similar to the business phase, the assessment phase is also composed of two sub-phases, namely buyer and seller assessment, respectively. The steps commended in the two phases are provided as follows;

Buyer assessment:

• Once the transaction has been completed, the consumers provide ratings and post comments on behalf of retailers to the application interface.

• The posted comments and ratings are passed to the blockchain, and the application interface acknowledges it.

• From the acknowledged address, the application interface summons the smart contract. The smart contract provides monetary rewards for the ratings and comments. The reputation scores for both retailers and consumers are updated and published to the blockchain.

• The reputation scores are stored in the blockchain, and feedback assessment is provided to the application interface.

Seller assessment:

• Consumers provide ratings for retailers after completing a transaction through the application interface.

• The ratings are passed to the blockchain, and it returns the address to the application interface.

• The smart contract gets summoned once the application interface acknowledges the address.

• Similar to the buyer assessment, both the consumer and retailer’s reputation scores are updated to the blockchain. Finally, the successful feedback is provided by the blockchain through the application interface.

Threat model and security assumptions

To clearly define the scope of our proposed BTL-ETRF framework, we specify the adversarial model considered in this work.

Adversary model this research assumes that the Probabilistic Polynomial Time (PPT) adversary is capable of.

• Creating multiple fake buyer and seller accounts (Sybil behaviour)

• Performing collusion attacks between buyers and sellers

• Injecting fake positive and negative feedback (ballot stuffing and bad mouthing)

• Attempting re-entry (whitewashing) by leaving and rejoining the platform

• Launching Man in the Middle (MITM) attacks during authentication

• Attempting to tamper with store feedback records

Also, we assume that the adversary did not possess,

• The trapdoor key of the chameleon hash function

• The private keys used in the digital signatures

• Ability to break SHA-512 under standard cryptographic assumptions

Considered attack surface:

• Application layer attacks (impersonation, collusion, fake reviews)

• Authentication layer attacks (replay attacks, credential spoofing)

• Storage later attacks (blockchain manipulation, feedback tampering)

Also, we did not consider,

• 51% attacks on public blockchain consensus

• Quantum adversaries

• Physical compromise of the user device

With the above assumptions, the proposed work tends to resolve the attacks such as unauthorized modification of transaction history, collusion-based reputation inflation/deflation, feedback manipulation, and impersonation, respectively.

Two-level E-commerce reputation verification

This section provides a detailed description of the proposed Two-Level E-Commerce Reputation Verification (TLERV), which is the core component of the BTL-ETRF framework. The TLERV consists of two main processes: Multi-Factor Authentication (MFA) and transformer-based reputation verification. The detailed mathematical modelling of these processes is presented below.

Multi-factor authentication

The entities participating in the MFA include e-commerce sellers (), e-commerce buyers , a distributed e-commerce application (DE_APP), and a blockchain database. In implementing MFA, these entities are authenticated using PINs, biometric fingerprints, and OTPs, respectively. Financial transactions are primarily secured through biometric authentication, while OTPs and PINs are protected using the SHA-512 public key cryptography algorithm. The steps involved in the MFA process are detailed below.

Registration phase Before registration to the distributed e-commerce application, the and s unique identities are generated . The registration phase establishes a secure identity and credential framework for each e-commerce user prior to participating in transactions. The entities involved include the e-commerce user (buyer or seller), the Distributed E-Commerce Application (DE_APP), and the blockchain database. The DE_APP registers the and based on the upcoming steps,

Stage 1 identity submission: the DE_APP acquires the and s unique ID (), fist name (), last name (), phone number . The DE_APP collectively holds the registered information as .

Stage 2 identity verification: the generated information is provided to the DE_APP for verifying the by the and . If the entered is incorrect, the and are provided with 2 tries; else their information is stored in the database .

Stage 3 pin generation and validation: in this stage, e-commerce users must complete their registration process by entering the 6-digit and are then asked to re-enter it. If the entered and re-entered matches, it is hashed by SHA-512 and stored in the database . In contrast, if it is mismatched, the users are asked to enter the correct PIN in two attempts and revoked. Note that the origin PIN was not stored in plain text form.

Stage 4 biometric capture and key generation: once the PINs are successfully entered, the biometric fingerprint is captured by the DE_APP, in which it computes a unique public–private key pair for the users. Using public key cryptography, the gets encrypted and stored in the . In addition, the and biometric template are also encrypted and stored in the DE_APP itself.

Stage 5 blockchain registration: the DE_APP sends the present , , to the blockchain database. If the provided information already exists, then the and s are requested to enter a new as a token of new registration. Finally, the records are stored in the database and a successful registration notification is provided to the and s devices.

Authentication phase:

The stages involved in the authentication are provided below,

Stage 1: Initially, the and s are requested to enter their 6-digit to the DE_APP. They had only two attempts to provide the correct PIN.

Stage 2: Upon acquiring, the DE_APP sends , and to the for validation.

Stage 3: The , and are checked by the , if the information are matched then is generated and send to the users whereas another copy is hashed using SHA-512 and stored. If the information does not match, the session is terminated.

Stage 4: Once the arrives, the user is requested to enter the OTP within 30 s. If the stored and entered OTPs are not the same, then the session gets terminated otherwise is scanned from the users.

Stage 5: The obtained is matched with the by validating the in the . If both are the same, the and s are allowed for buying and selling; otherwise, they are revoked from the website using an authentication contract.

Transformer-based reputation verification

Once the and are authenticated, they are free to initiate transactions. For any transaction, both the retailer and consumer perform both positive and negative ratings of each other. Both ratings fall within the range [− 4, − 1]. There are several key factors in reputation assessment, including past reputation scores among consumers and sellers, transaction volume, transaction time, social trust, and user behavior. The reputation verification factors are explained as follows,

User behaviour: User behaviour encompasses several key aspects, including the number of transactions, consistency of transactions, login frequency, purchase patterns, and communication behaviour.

Social trust: Social media plays a major role in e-commerce reputation verification by assessing trustworthiness and credibility. The social trust metrics include the number of connections, the strength of connections, content sharing, mutual connections, and information consistency.

Past reputation score: The past reputation score of the retailer and consumer is determined based on the weighting factor among legitimate fair ratings and illegitimate ratings. The formulation of past reputation scores is provided below,

1

2

From the above equation, where is the Transaction time factor, which adjusts the weight based on how recent the transaction is (more recent transactions have higher influence). Transaction volume factor, representing the significance of the transaction amount (higher amounts yield higher weight). Past reputation score, reflecting historical performance of the consumer/retailer. Count of legitimate positive ratings provided by users. Count of legitimate negative ratings provided by users. and Weighted positive and negative ratings, respectively, after applying the transaction and volume factors.

Amount of transaction: The amount of the transaction is directly linked to the product they purchase. Specifically, no consumer buys more products and associates with the seller to provide unfair ratings. The amount of the transaction weighting factor can be formulated by,

3

From, and which ranges among 0 to 1. The weighting factor increases proportionally with the transaction amount.

Time of transaction: The consumer and retailer may conspire with each other on rating by consuming the products shortly after purchase. If the consumer’s transaction occurred recently, it is likely that the consumer and seller are vulnerable to collusion attacks. The formulation of the time for the transaction weighting factor can be computed by,

4

From the above equation, denotes the time interval among the present time for transaction and past time for transaction correspondingly which ranges among 0 to 1. The listed reputation verification weighting factors serve as input to the transformer named Residual Dilated Convolution Transformer (RDCT) which consists of two major blocks: the Residual Swin Transformer Block (RSTB) and the Dilated Convolution Block (DCB). The RSTB, composed of multi-head attention layers, works based on the shifted window mechanism. The RSTB consists of a Multi-Layer Perceptron (MLP), Double Layer Normalization (LN), and Multi-Head Self-Attention (MHSA). The residual connection is performed after every layer of MHSA and MLP, correspondingly. The MHSA is composed of two components: Shifted Windows MHSA (SW-MHSA) and Windows MHSA (W-MHSA). The formulation of the RSTB in a successive manner for e-commerce reputation verification is shown below,

5

6

7

8

Initially, the inputs are segmented into non-overlapping windows and reshaped to , where the number of windows is denoted as , for every window, the self-attention operation is performed. The value, key, and query are computed as,

9

From the above equation, the learnable weight matrices are denoted as , . , and respectively. The attention method is formulated as,

10

om the above equation, the key dimension is denoted as , and positional encoding is denoted as .

Since the transformers extract spatial and temporal features, some important interdependencies’ information is missed. To capture the interdependencies and global information, we adopt dilated convolution layers consecutively, which can be formulated as,

11

From the above equation, the bias function and activation function are denoted as and correspondingly. The denotes the dynamic parameter that regulates the feature shift over time. The formulation of the convolution receptive field when with an increase in size of is denoted as,

12

From the above equation, SN denotes the dilated convolution receptive field, whereas denotes the kernel size of the convolutional kernel. The output produced by the RSTB and dilated convolution provides an informative feature map with reputation verification patterns. Finally, utilizing an ML algorithm, the reputation is classified into reputed and non-reputed categories, respectively. Based on the results, the reputation smart contract invokes the action against the user and adds them to the blockchain database.

The adoption of RDCT over other deep learning algorithms was based on the practical feasibility and representational strength in large-scale e-commerce environments. Conventional models, such as CNNs, have limitations. LSTM-based models offer improved temporal modelling capabilities but struggle with extracting long-range dependencies, parallelization, and scalability in irregular user behavioral data. In addition to that, the transformer variant enables better performance in global dependency modelling but faces higher computational costs due to higher-dimensional representation signals. In contrast, the RDCT combines two complementary components named RSTB and DCB for local–global dependency extraction, and receptive field enlargement correspondingly. With this combination, the RDCT tends to model for multi-scale dependencies, long-term collusion behavior, and short-term fluctuations in reputation verification with lesser latency and higher accuracy.

To illustrate the synergy between the Residual Swin Transformer Block (RSTB) and Dilated Convolution Block (DCB), consider a collusion scenario where a group of users systematically inflates ratings over extended time intervals. Conventional CNN models capture local patterns but fail to model long-range dependencies across dispersed transactions. Standard transformers model global attention but lack explicit receptive field control for structured temporal variation.

In the proposed RDCT,

• The RSTB captures long-range dependencies across distributed user interactions through window-based multi-head self-attention.

• The DCB expands the receptive field using dilated convolutions, enabling detection of periodic rating bursts and sparse collusion patterns.

This dual mechanism allows RDCT to simultaneously capture,

• Short-term anomalous transaction bursts.

• Long-term structured collusion behaviour.

Hyperparameter tuning: The hyperparameter tuning test was conducted to ensure the reliability and fairness of the model. The RSTB consists of 8 attention heads, 4 transformer layers, with a window size of 7 and a hidden dimension of 256. The dilation rates of 2 and 4 were selected for DCB to trade off the expansion of the receptive field with better computational efficiency. With the help of the Adam optimizer, learning rate 1e-4, and a batch size of 64, the model is trained. The model was trained on the NVIDIA Tesla V100 GPU with 32 GB of memory.

Smart contract design

Our proposed work designs two smart contracts for enabling automated actions on the two-level reputation verification framework. Authentication Smart Contract (ASC) and Reputation Smart Contract (RSC) are the two prime smart contracts utilized in our work. Figure 2a, b present the architecture of the ASC and Reputation Smart Contract RSC, respectively. The design, roles, and privileges of the adopted smart contracts are provided as follows.

Fig. 2.

Architecture of ASC and RSC.

Authentication smart contract (ASC): The ASC is primarily designed to automate actions based on MFA results. In addition to that, the ASC is amalgamated with the redactable blockchain to ensure the integrity of data with essential modifications. The entities and their roles within the ASC include external entities, registered users, and contract owners.

Role of ASC.

• The main responsibilities of the contract owner include updating the authentication policy, managing user registration, and deploying smart contracts. It provided privileges to newly registered users and had rights to some of the contract parameters.

• The primary responsibility of registered users is to handle authentication requests and public key distribution. It verifies the contract identity and allows access to specific functionalities.

• The external systems, such as the application interface, are responsible for interacting with smart contracts and returning the outcomes. It could access the authentication results and had limited access to predefined authentication rules.

Authentication workflow.

The ASC performs the following functions:

• The user submits PIN, OTP, and biometric data to the DE_APP

• The DE_APP then verifies, matches stored hash, OTP validity, and biometric template match

• After verification, the DE_APP generate a nonce N

• The uses signs by using their private key

• The signed nonce is then provided to the ASC

• The ASC verifies the signature from the stored public key

• If valid, access granted; else revoked

The enforcement action is defined by

13

This separation ensures that registration, credential management, and runtime authentication remain logically independent while preserving blockchain-level verifiability.

Reputation smart contract (RSC): Similar to the ASC, the RSC is designed to automate the actions on transformer-based reputation verification. There are similar entities involved in the RSC, such as contract owners, authenticated users, and external entities. Unlike the ASC, the RSC restricts access to authenticated users, thereby reducing unwanted traffic during reputation verification.

Initially, the contract owner invokes the authentication server to provide legitimate e-commerce user details. Once received, the contract owner monitors user behavior, social trust, past reputation score, transaction amount, and transaction time, correspondingly. Based on the monitored information, the RSC can take automated action, as

14

The RCS implements when the user is reputed otherwise non-reputed .

Redactable blockchain design

Redactable blockchains are a type of blockchain that allows users to modify the blockchain without compromising security and data integrity. The primary requirement for designing a redactable blockchain is ensuring data auditability, maintaining data integrity, and preventing practical errors during block joining. There are several mechanisms involved in the redactable blockchain, such as the chameleon hash function and redactable signatures. The mathematical formulations behind the blockchain model are provided below,

Chameleon Hash Function: Different from the traditional hash function, the chameleon hash function permits a specific hash data to be modified without changing the overall block hash. The mathematical formulation of the chameleon hash function is provided as follows,

15

From the above equation, , in which denotes the hashed output, denotes the random value, and is the actual data. Furthermore, the chameleon hash function had the trapdoor function that permits modification of without completely changing the hash function as,

16

Redactable Signature: The validity of the signature is maintained by enabling the redactable signature to be modified in a specific hashed part. The redactable signature is denoted as for the message as,

17

From the chameleon hash function, the updation of the redactable signature can be denoted as,

18

The above equation ensures that the original message parts get valid even though it gets redacted .

Scalability and Security for Redactable Blockchain: The scalability of the proposed redactable blockchain mechanism is evaluated in terms of computational overhead introduced by the chameleon hash function and redactable signature operations. Unlike conventional immutable blockchains, which require global recomputation when block content changes, the proposed design localizes modifications using trapdoor-based collision control. From a computational perspective, let denote the size of the feedback message and represent the chameleon hash function. The update complexity of a redacted block is , as only the modified message portion requires recomputation rather than reprocessing the entire chain. This localized update mechanism avoids full-chain recalculation and reduces modification latency compared to traditional immutable blockchain architectures.

To empirically evaluate performance, we measured block modification latency under varying feedback message sizes ranging from 1 to 770 KB. The observed modification time increases approximately linearly with message size, ranging from 182 to 1500 ms under the specified experimental setup (Ubuntu 18.04, Solidity v0.5.0, NVIDIA Tesla V100 GPU). This indicates predictable scaling behaviour under controlled workload conditions.

The additional computational overhead introduced by the chameleon hash function was measured relative to a conventional SHA-based hash computation. Experimental observations indicate an overhead in the range of approximately 12–15% under the evaluated test conditions. However, this overhead is offset by the elimination of global block recomputation, which improves practical modification efficiency in scenarios requiring controlled redaction.

From a security perspective, the redactable blockchain relies on:

• Collision resistance of the chameleon hash function under trapdoor secrecy.

• Unforgeability of redactable signatures under chosen-message attack assumptions.

• Secure management of trapdoor keys via threshold cryptographic techniques.

Under the Random Oracle Model (ROM), an adversary without access to the trapdoor key has a negligible probability of generating two distinct messages that produce the same chameleon hash value. Furthermore, the use of distributed threshold key management mitigates single-point compromise risks during authorized redaction operations. It is important to note that the proposed scalability and security claims are evaluated within the defined experimental and adversarial assumptions and do not address consensus-level attacks such as 51% attacks or network-layer denial-of-service attacks.

Large scale deployment considerations: While the experimental evaluation was conducted under controlled workload conditions, large-scale e-commerce platforms may process millions of transactions daily. In such scenarios, potential bottlenecks include,

• Trapdoor key coordination latency.

• Smart contract execution gas costs.

• Blockchain write throughput limits.

Assuming transactions per second and redaction rate , the expected overhead introduced by threshold key reconstruction is denoted by in which is the threshold reconstruction size. Since redaction events are rare compared to transaction frequency, threshold cryptographic operations do not significantly affect overall throughput. However, in extremely high-frequency modification scenarios, batching or off-chain processing mechanisms may be required.

Experimental results

Experimental setup

In this subsection, we designed the experimental setup for the proposed TLERV framework. The experimental configuration of the proposed TLERV is described as follows. IPFS and Web3.js are JavaScript-based tools used in the implementation, where IPFS is employed for decentralized storage and Web3.js serves as the JavaScript API library to interact with the redactable blockchain and smart contracts. In addition, Web3.js provides JavaScript objects and functions that enable users to communicate with the blockchain and deployed smart contracts. We implemented and evaluated the proposed TLERV framework on the Ubuntu 18.04 operating system, utilizing Visual Studio Code and Solidity v0.5.0.

Dataset description

The experimental evaluation of the proposed BTL-ETRF framework was conducted using the publicly available Amazon Product Review Dataset (Stanford SNAP Repository)⁴⁴. This dataset contains large-scale user-product interaction records, including review text, rating score, timestamp, product identifier, and user identifier. For this study, we utilized a filtered subset of the Electronics category, which contains 192,403 users, 63,001 products, and 1,689,188 review entries. Every record includes the user ID, product ID, rating score (1–5 scale), review timestamp, and review text.

The dataset was partitioned using stratified sampling, in which 70% used for the training set, 15% used for the validation set, and used for 15% testing set. To ensure robustness, experiments were repeated five times with different random seeds, and the average results were reported. Class imbalance was handled using weighted cross-entropy loss during training.

Latency analysis

This subsection examines latency using two methods: first, it determines the time cost of adding blocks, and second, it calculates the proposed model’s delay during reputation verification.

Blockchain latency analysis

In our study, we examine the blockchain’s time cost across three activities: product purchase, product upload, and an increase in block count. Table 1 represents verified feedback tokens, adds the feedback to the blockchain, and evaluates the reputation score, and Fig. 3 illustrates two key observations:

• The time required to complete each of these three activities increases as the number of blocks on the blockchain grows. Each operation consists of two main time-consuming components: accessing and storing data in the Redactable blockchain and executing the functionalities of the smart contract. Consequently, as the number of blocks increases, the time costs associated with data access and storage also rise. Additionally, low-complexity blockchain addresses must upload product details, transaction actions, and updated reputation ratings to perform smart contract functionalities.

• The assessment phase takes more time than the product purchase and upload phases. This is primarily because the review process involves determining, updating, and storing user reputation ratings in the blockchain. Table 2 and Fig. 4 demonstrate the running time in each stage.

Table 1.

Verify the feedback token, add the feedback to the blockchain, and evaluate the reputation score.

Size of feedback message	Response time	Block size generated in blockchain
130B	2470 ms	6 K
252B	2520 ms	6 K
500B	2522 ms	7 K
1 KB	2576 ms	8 KB
2 KB	2506 ms	11 KB
4 KB	2508 ms	15 KB
8 KB	2542 ms	22 KB
16 KB	2650 ms	36 KB
24 KB	2546 ms	54 KB
32 KB	2598 ms	70 KB

Fig. 3.

The number of ratings from other users increases users’ reputation scores.

Table 2.

Performance analysis of running time.

Stage	Registration		Rating Creation		Reputation Updating		Verifying Rating
	Provider	Platform	Customer	Provider	Platform	Provider	Verifier
Runtime	31	0.97	87	44	0.26	0.015	0.04

Fig. 4.

Average latency in individual stage.

In contrast, the product purchase and upload operations only require uploading product data or completing transaction activities. Additionally, a graphic compares our blockchain to other existing efforts, including RepChain²¹, BC-DRS²², and TPB-HIHF²³. Due to the effective use of redactable signatures and chameleon hash functions, our proposed redactable blockchain achieves lower latency than the compared baselines (Table 3) and facilitates blockchain edits without compromising security or data integrity. These techniques allow specific sections of the blockchain to be updated quickly and accurately without the need for laborious recalculations and verifications, which often cause delays. Using chameleon hash functions, blockchain entries can be selectively modified without impacting the entire chain, significantly reducing computational costs. Redactable signatures further enhance this efficiency by allowing individual signatures to be changed without breaking the overall chain, enabling rapid implementation of changes. In contrast, existing systems like RepChain, BC-DRS, and TPB-HIHF experience higher latency due to their complex and multidimensional architectures:

Table 3.

Comparative analysis of proposed BTL-ETRF with existing systems.

Model	Average Latency (ms)	Accuracy (%)	ROC-AUC
BTL-ETRF	~ 181–2887 ms (low)	97.8%	0.98
RepChain	~ 3500–5000 ms (high)	89.5%	0.84
BC-DRS	~ 3100–4700 ms (high)	87.6%	0.82
TPB-HIHF	~ 2800–4000 ms (moderate-high)	90.2%	0.85
ML-USA	~ 2600–4000 ms (moderate-high)	91.1%	0.86
TM-ECA	~ 2400–3800 ms (moderate)	92.3%	0.88

RepChain: This system focuses on reputation management while protecting privacy. To maintain integrity and privacy, it employs complex cryptographic processes. These procedures require extensive computations and verifications, leading to increased delays.

BC-DRS²²: The decentralized reputation verification process in BC-DRS adds layers of complexity and increases latency due to the substantial computation and communication required among nodes.

TPB-HIHF²³: Using Hyperledger Fabric and Hyperledger Indy in a multi-permissioned architecture introduces extra coordination and verification processes. The necessity for cross-system interactions inherently causes delays due to the integration of multiple systems and authorization levels.

Overall, our proposed redactable blockchain reduces delay through practical usability and efficient redaction procedures by minimizing computational overhead and preventing frequent errors during block addition. This approach results in a much faster system compared to RepChain, BC-DRS, and TPB-HIHF, which are impeded by their complex and resource-intensive processes. Figure 5 shows the overall performance analysis of the proposed work. Furthermore, the overall performance comparison of the proposed work and existing works is outlined in Table 3.

Fig. 5.

Performance analysis.

Latency analysis of reputation verification

Automation and advanced technology significantly enhance the latency performance of our Blockchain-based BTL-ETRF. Only authorized users who have completed the multi-factor authentication (MFA) process using their fingerprint, PIN, and OTP can proceed to the reputation verification phase. This ensures that only valid requests are processed, reducing unnecessary load and increasing efficiency. Our methodology classifies user reputations using RDCT, based on five key variables. RDCT’s faster and more accurate data processing allows for quick and precise reputation evaluations, thereby reducing verification time. Additionally, we utilize two smart contracts: the Reputation Smart Contract (RSC) and the Authentication Smart Contract (ASC). These smart contracts automate the reputation verification process, eliminating the need for human intervention and ensuring rapid implementation of reputation changes. The ASC quickly responds to user authentication requests, further enhancing the system’s overall efficiency. While the RSC effectively manages reputation data, the ASC swiftly processes user authentication.

Conversely, the lack of automation and less advanced technology in current systems like RepChain, BC-DRS, TPB-HIHF, ML-USA, and TM-ECA results in higher latency. For example, RepChain processes data more slowly because it lacks the deep learning and smart contract features present in our framework. BC-DRS’s lack of powerful deep learning models and a redactable blockchain leads to less effective verification. TPB-HIHF and ML-USA experience higher latency due to their reliance on less automated, simpler techniques. Despite emphasizing security and trust, TM-ECA does not integrate redactable blockchain, smart contracts, and MFA as comprehensively as our system. In conclusion, our BTL-ETRF significantly reduces verification time compared to existing systems by achieving excellent latency performance through the use of RDCT for quick and accurate reputation categorization, and ASC and RSC for automation. Table 4 outlines the response time while receiving feedback.

Table 4.

Receive feedback from seller.

Feedback in size (KB)	Response time (ms)
1	182
2	248
3	216
4	275
8	226
16	240
28	272
34	304
68	218
134	442
178	390
268	562
440	812
770	1500
1034	2108

Security analysis

Our proposed TLERV model is evaluated in terms of security analysis in two aspects. The model is initially examined in terms of a few aspects of information security. Second, our proposed TLERV is evaluated based on its ability to protect against a few acknowledged reputation system attacks.

Aspect of information security

Here, information security is analyzed in terms of identity privacy, credential likability, and the integrity of stored feedback data.

Identity Privacy We analyse identity privacy under the assumption of a probabilistic polynomial-time (PPT) adversary attempting to infer the real-world identity of a user from blockchain records or feedback tokens. Let each registered user be associated with,

• A public key ,

• A unique identifier ,

• Encrypted biometric template .

The adversary is assumed to have full read access to blockchain records but does not possess:

• The private key ,

• The trapdoor key of the chameleon hash function,

• Access to off-chain biometric databases.

Privacy goal: The adversary should not be able to distinguish between two users and with non-negligible advantage based solely on publicly visible blockchain data.

Formally, identity privacy is preserved if:

19

From the above equation, the is negligible.

Since,

• Biometric templates are encrypted under public key cryptography,

• Only hashed PIN values are stored,

• No plaintext personal identifiers are written on-chain; the adversary’s advantage reduces to breaking standard encryption or hash assumptions. Under the hardness of SHA-512 preimage resistance and public-key encryption security, identity linkage is computationally infeasible.

Data integrity We analyse data integrity against an adversary attempting to modify stored feedback records without authorization.

Let each block contain,

20

From the above equation, the is the chameleon hash, and the is the redactable signature.

Integrity goal: An adversary without trapdoor access should not be able to modify while preserving blockchain validity. Suppose the adversary attempts to produce,

21

Such that and blockchain verification still succeeds. This would require either:

• Generating a collision in the chameleon hash without trapdoor access, or

• Forging a valid redactable signature.

Under the Random Oracle Model,

• The probability of producing a collision without trapdoor knowledge is negligible.

• The probability of forging a signature under a chosen-message attack is negligible.

  22

From the above equation, the is negligible in the security parameter. Authorized redaction is permitted only through threshold-controlled trapdoor key usage, ensuring accountability and auditability (Fig. 6).

Fig. 6.

Data integrity.

Defense against known attacks

Our proposed TLERV framework offers countermeasures for multiple acknowledged attacks against the reputation system.

Re-entry/whitewashing: Re-entry and whitewashing attacks can be mitigated by ensuring the seller’s record accurately reflects the individual’s real-world identity. Demanding digital recognition from legal authorities confirms that sellers have legal entity identifiers associated with their records in the system. Henceforth, leaving the system and re-joining does not affect prior feedback or reputation scores.

Slandering/bad-mouthing: Slandering and bad-mouthing attacks aimed at minimizing the competitor’s reputation. Several studies suggest that requiring users to provide feedback only when they encounter complications during the transaction process can result in slanderous or negative comments. An attacker would need to make purchases from a competitor in order to leave negative feedback. Our model requires reviewers to have feedback tokens in the method, enabling them to send feedback. Additionally, social trust, reputation score, transaction amount, and several credentials are considered for reputation verification based on a DL-based approach. Henceforth, feedback is directly linked to specific transactions, which helps reduce negative comments and slanderous attacks.

Self-promotion/ballot stuffing: Restricting feedback suggestions to only users involved in transactions, similar to slandering and bad-mouthing attacks, also mitigates self-promoting and ballot-stuffing attacks. For sellers with stores in marketplaces, it is not possible to inject feedback for themselves without participating in transactions, as marketplaces issue feedback tokens on behalf of sellers. This mechanism helps reduce ballot-stuffing and self-promoting attacks. However, sellers with their e-commerce platforms cannot be prevented from issuing fake tokens to submit positive feedback for themselves. Disclosing reviewer information may lead to an unfairly high number of positive reviews. The privacy-preserving feature of our model helps reduce dishonest positive feedback.

Man-in-the-middle (MITM): MITM attacks occur when an attacker captures and potentially alters the interaction between two parties (i.e., an e-commerce website and a user), without their knowledge. As a result, sensitive information, including credit card details, personal data, and login credentials, is exposed. To mitigate this attack, we have employed a redactable blockchain, where both authentication and reputation smart contracts store user data and their addresses in our proposed blockchain.

Discriminatory attacks: In the context of e-commerce, a discriminatory assault targets particular user groups according to attributes like geography, IP address, behaviour, past purchases, or even perceived socioeconomic position. The goal can be anything from the unlawful modification of user data to price manipulation and service denial. These assaults have the potential to compromise the integrity of the e-commerce platform and treat users unfairly. In our work, we perform MFA by considering OTP, PIN, and fingerprint, which allows legitimate users into the website, thereby alleviating discriminatory attacks. In Table 5, the comparative analysis of security techniques against attacks is presented.

Table 5.

Comparative security mechanisms against attacks.

Attack type	Proposed BTL-ETRF	RepChain	BC-DRS	TPB-HIHF	ML-USA	TM-EC
Re-Entry/Whitewashing	Legal entity identifiers + Redactable blockchain	Cryptographic privacy only, no binding to real identity	Node-based consensus only	Permissioned blockchain, but no real-world ID binding	ML-based scoring, no entity binding	Trust emphasis, but no strong legal identity
Slandering/Bad-Mouthing	Feedback tokens + DL verification of transaction-linked feedback	Cryptographic privacy, no DL analysis	Distributed verification, no feedback linkage	Transaction integrity is ensured, but no review validation is performed	Static ML, vulnerable to fake reviews	Security policies, no deep learning validation
Self-Promotion/Ballot Stuffing	Feedback restricted to verified transactions + marketplace-issued tokens	Basic privacy, but vulnerable to fake reviews	No prevention against seller-issued fake feedback	Hyperledger consensus only, no review validation	ML-only, no blockchain feedback control	Reputation rules, but no transaction-linked tokenization
MITM Attack	Redactable blockchain + ASC/RSC smart contracts securing addresses	Privacy focus, not MITM specific	Vulnerable (no blockchain-level mitigation)	Hyperledger consensus only, vulnerable to MITM	No blockchain security, vulnerable	Security focus, but no blockchain-level authentication
Discriminatory Attacks	Multi-Factor Authentication (OTP, PIN, Fingerprint) ensures fairness	No MFA	No MFA	No MFA	No MFA	No MFA

Formal security proof

To provide format guarantees, this research analyzes the proposed framework under the University Composability (UC) and Random Oracle Model (ROM) frameworks. The redactable blockchain is based on the chameleon hash function that has a collision-resistant capability under trapdoor secrecy in ROM. An adversary without the trapdoor key has a negligible probability which tends to produce two distinct message , so that . In similar manner, the signature are made as unforgeable under the chosen message attacks in the ROM that guarantees only the authorized entities can modify the signed set of blocks. The SHA-512 was utilized for protecting the OTP and PIN values during authentication as a random oracle model, which also guarantees collision resistance. For composability, both the Reputation Smart Contract (RSC) and Authentication Smart Contract (ASC) are shown to be UC as they check the authenticity and reputation updates, respectively. On the whole, under the standard hardness conditions, the signature schemes and cryptographic hash function of the proposed work are secure under tampering, replay, and forgery attacks, correspondingly.

Analysis of reputation verification performance

In reputation verification, our proposed BTL-ETRF outperforms the compared baselines, particularly in terms of accuracy and the ROC (Receiver Operating Characteristic) curve. This enhanced performance is achieved through a carefully designed process that includes the following key components:

Reputation verification authentication users: Unlike other systems, BTL-ETRF ensures that only verified users participate in the reputation verification phase. This enhances the reliability and security of the process, reducing the likelihood of fraudulent or malicious activities.

Metrics for classifying user reputation: BTL-ETRF uses five primary metrics to classify user reputation: social trust, prior reputation score, transaction amount, transaction duration, and user behavior. These metrics provide a comprehensive and multifaceted view of user reputation, enabling more precise and dependable verification.

RDCT integration: Our framework integrates RDCT, enabling sophisticated processing and analysis of reputation data. RDCT identifies intricate patterns and correlations within the data, leading to more precise classification and prediction of user reputation.

Automated reputation verification with smart contracts: BTL-ETRF utilizes two smart contracts to automate and expedite the reputation verification process.

ASC (authentication smart contract): Ensures that only authorized and valid users can access the reputation verification step.

RSC (reputation smart contract): Manages the assessment of reputation metrics and initiates automatic actions based on BTL-ETRF outcomes.

By automating critical processes and reducing human involvement, these smart contracts enhance the efficiency, transparency, and security of the verification process. Table 6 outlines the verifiable credentials considered for message exchange. Current systems, including RepChain, BC-DRS, TPB-HIHF, ML-USA, and TM-ECA, encounter several challenges that impact their accuracy and ROC curve performance. Notably, RepChain and BC-DRS lack sophisticated ML models, such as RDCT, which results in less accurate reputation categorization. TPB-HIHF uses static credentials, potentially leading to incomplete reflection of user activities and real-time transactions.ML-USA relies on static ML models that are not well-suited for dynamic and evolving e-commerce environments.TM-ECA emphasizes security and trust, but it does not integrate extensive metrics and sophisticated models for reputation verification. In contrast, BTL-ETRF’s use of RDCT and smart contracts like ASC and RSC ensures more accurate and reliable reputation verification, enhancing overall system performance in terms of accuracy and ROC curve metrics. A comparative analysis of proposed and existing works is presented in Figs. 7 and 8, which show the ROC curve and accuracy, sequentially, demonstrating that our proposed work achieves superior performance. Tables 7 and 8 define the stage metrics utilized during seller registration and feedback receipt.

Table 6.

Verifiable credentials taken for message exchange.

Type of message exchange	Situations	Average latency
Credential of issue	Identity of digital Token of discount Token of feedback	171 ms
Request proof	Identity of digital Token of discount Token of feedback	114 ms
Send proof	Identity of digital Token of discount Token of feedback	164 ms

Fig. 7.

ROC Curve.

Fig. 8.

Accuracy.

Table 7.

Stage metrics of seller registration.

Phase	Latency in average
Issue a digital identity	171 ms
Digital identity request proof	114 ms
Verify the digital identity and add the seller to the blockchain	2436 ms
Send Proof	166 ms
Total	2887 ms

Table 8.

Stage metrics of feedback.

Phase	Latency in average
Request verification of a feedback token	114 ms
Send Proof	163 ms
Verify the feedback token, record the feedback on the blockchain, and evaluate the reputation score	2542 ms
Token of issue discount	170 ms
Total	2989 ms

Reliability analysis

Our proposed BTL-ETRF takes a holistic approach to addressing critical security and integrity concerns, making it more reliable than other solutions for e-commerce platforms. Reviews and reputation management systems are crucial for handling and monitoring financial transactions on these platforms. Thus, maintaining user satisfaction and trust hinges on a robust reputation system. Many current reputation systems suffer from centralized architectures and tampering issues, leading to unwarranted reviews and ratings that cyberattacks can exploit to damage vendors’ reputations and consumers’ experiences. BTL-ETRF addresses these issues through three key components: the Redactable Blockchain System ensures data integrity while allowing necessary modifications. Multi-Factor Authentication (MFA) protects user identities and prevents impersonation attempts by using three factors: PIN, OTP, and biometric fingerprint. Deep Learning-Embedded Transformers enhance the accuracy and reliability of reputation assessments. This integrated framework demonstrates improved performance compared to baseline systems under the evaluated metrics. By utilizing a redactable blockchain, we guarantee data integrity and prevent tampering. This technology allows for secure, targeted modifications without compromising the overall integrity of the blockchain. Here, Fig. 9 shows that users’ reputation scores increase in proportion to the number of ratings they receive from other users.

Fig. 9.

Users’ reputation scores rise in proportion to the quantity of other users’ ratings.

Our methodology utilizes RDCT to categorize user reputations based on five key indicators for reputation verification. This advanced deep learning technology provides a more precise and reliable evaluation of user reputations compared to conventional approaches. We implement the ASC and RSC smart contracts to automate the process. These contracts reduce the risk of human error and manipulation by automating actions triggered by BTL-ETRF results. Figure 10 illustrates that users receive a higher average monetary compensation as the number of ratings they provide increases. This automation, combined with robust authentication and sophisticated verification methods, ensures a trustworthy and secure reputation system. Conversely, existing methodologies like RepChain, BC-DRS, TPB-HIHF, ML-USA, and TM-ECA exhibit shortcomings across various domains. RepChain offers anonymity and verifiability, but it lacks sophisticated deep learning integration and extensive multi-factor authentication (MFA), which makes it less reliable against impersonation attacks and less accurate in reputation assessment. BC-DRS lacks deep learning models and redactable blockchain technology, which are crucial for enhanced tamper resistance and reliability. TPB-HIHF and ML-USA rely on static ML and valid credentials for security analysis, whereas RDCT and smart contracts in our BTL-ETRF enable dynamic and automated verification. TM-ECA emphasizes security and trust but lacks the comprehensive feature set that BTL-ETRF provides, reducing its dependability and resilience. In contrast, our BTL-ETRF integrates advanced deep learning algorithms, multi-factor authentication (MFA), and smart contract automation, offering a robust solution for dependable reputation management in e-commerce environments. This approach addresses the shortcomings of existing methodologies by enhancing security, reliability, and accuracy in reputation verification.

Fig. 10.

Users receive a higher average monetary compensation the more ratings they have provided.

Regulatory compliance considerations

The redactable blockchain design supports regulatory compliance with data protection laws such as the General Data Protection Regulation (GDPR).

• Right to rectification: authorized redaction allows correction of erroneous user data without compromising blockchain consistency.

• Right to erasure (Article 17 GDPR): sensitive personal information is stored off-chain in encrypted form, enabling controlled deletion while preserving hashed audit references on-chain.

• Data minimization: only hashed identifiers and public keys are stored on-chain, avoiding storage of plaintext personal data.

• Auditability: Redactable signatures ensure that all modifications are traceable and verifiable.

These properties allow the framework to balance immutability with regulatory flexibility.

Conclusion

In this research, we design a BTL-ETRF: a blockchain-based two-level reputation verification framework for the e-commerce environment. The first stage of our reputation verification process involves MFA, which authenticates e-commerce buyers and sellers by utilizing three factors: PIN, OTP, and biometric fingerprint. For the second stage of reputation verification, we adopted RDCT for reputation classification into two stages, as reputed and non-reputed, by considering five factors, such as transaction time, transaction amount, past reputation scores, social trust, and user behaviour. Based on the results from BTL-ETRF, the automated actions are triggered sequentially by ASC and RSC. ASC triggers an action based on the MFA results, whereas RSC triggers an action based on the RDCT results. The results obtained are stored in the redactable blockchain, where the blockchain authority can redact the blockchain data without affecting the originality of the data, which firmly ensures the blockchain’s reliability. The implementation is carried out using Python for transformer-based reputation verification and Ethereum with Solidity for blockchain implementation. The results were analyzed using several performance metrics, which show that the proposed work outperforms the compared works. Our proposed BTL-ETRF framework achieves a latency reduction of approximately 35–40% compared to RepChain, BC-DRS, and TPB-HIHF, while also enhancing reputation verification by close to 10–15%, thereby validating the reliability and effectiveness of our proposed work contributions.

Abbreviations

BTL-ETRF

Blockchain based two level E-commerce trustable reputation framework

TLERV

Two-level E-commerce reputation verification

MFA

Multi-factor authentication

RDCT

Residual dilated convolution transformer

RSTB

Residual swin transformer block

DCB

Dilated convolution block

ASC

Authentication smart contract

RSC

Reputation smart contract

DE_APP

Distributed E-commerce application

ROM

Random oracle model

UC

Universal composability

Author contributions

Conceptualization, KSK, RCD, CA, DE, SA, WP, HMR, SK; methodology, KSK, RCD, CA, DE, SA, WP, HMR, SK; writing—original draft preparation, KSK, RCD, CA, DE, SA, WP, HMR, SK; writing—review and editing, KSK, RCD, CA, DE, SA, WP, HMR, SK; visualization, KSK, RCD, CA, DE, SA, WP, HMR, SK. All authors have read and agreed to the published version of the manuscript.

Funding

There is no external funding.

Data availability

The data used to support the study’s results can be acquired by contacting the corresponding author.

Declarations

Competing interests

KSK, RCD, CA, DE, SA, WP, HMR, and SK declare that they have no competing interests.