UPI Fraud Detection Using Convolutional Neural

Networks(CNN)
MELAM NAGARAJU

Seshadri Rao Gudlavalleru Engineering College

Yarramreddy Chandrasena Reddy Dept. of IT
Seshadri Rao Gudlavalleru Engineering College
Polavarapu Nagendra Babu
Seshadri Rao Gudlavalleru Engineering College
Venkata Sai Pavan Ravipati
Seshadri Rao Gudlavalleru Engineering College
Velpula Chaitanya
Seshadri Rao Gudlavalleru Engineering College

Research Article

Keywords: Convolutional Neural Networks, fraud detection, online banking, machine learning, imbalanced
datasets, feature engineering

Posted Date: March 14th, 2024

DOI: https://doi.org/10.21203/rs.3.rs-4088962/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License

Additional Declarations: The authors declare no competing interests.

Page 1/16
Abstract
In response to the escalating threat of online banking fraud, exacerbated by the COVID-19 pandemic, this
paper introduces a novel approach utilizing Convolutional Neural Networks (CNNs) for fraud detection.
The study focuses on developing machine learning models tailored for recognizing fraudulent
transactions and addresses challenges such as imbalanced datasets, feature transformation, and
engineering. The proposed CNN-based model exhibits superior accuracy, particularly in handling
imbalanced datasets, offering a promising solution compared to traditional algorithms. The research
emphasizes the adaptability of CNNs to unconventional data types, such as banking transactions, and
showcases their ability to capture intricate fraud patterns. Evaluation metrics include precision, recall, F1-
Score, ROC Curve, and AUC, providing a comprehensive assessment of the model's effectiveness.

II. INTRODUCTION
In the ever-evolving landscape of financial transactions, the advent of online banking has ushered in
unprecedented convenience and efficiency. However, this digital transformation has brought about new
challenges, particularly the escalating threat of fraudulent activities accompanying the surge in online
transactions. The global COVID-19 pandemic has further accelerated this shift towards online operations,
creating an environment conducive to malicious actors seeking to exploit vulnerabilities.

With the growing dependence of both financial institutions and users on remote transactions, there is a
heightened demand for advanced fraud detection mechanisms. The accelerated proliferation of digital
transactions, further exacerbated by uncertainties introduced by the pandemic, emphasizes the critical
importance of establishing resilient strategies to effectively prevent and detect fraudulent activities within
the domain of online banking. As financial interactions increasingly shift towards digital platforms, the
necessity for sophisticated security measures becomes more evident. The dynamic nature of online
transactions, combined with the unique challenges posed by external factors such as the pandemic,
highlights the pressing need for the development of robust and adaptive strategies to safeguard the
integrity of online banking systems. In light of these evolving circumstances, the urgency to stay ahead of
potential threats and fortify the security infrastructure of digital financial operations

becomes a paramount concern for both financial institutions and individual users.

In light of the escalating challenges posed by the evolving landscape of online transactions, especially
within the Unified Payments Interface (UPI), this research aims to address the critical issue of fraud
detection. Departing from conventional methodologies commonly employed in this domain, the study
takes a novel approach by incorporating Convolutional Neural Networks (CNNs) as a sophisticated tool
for identifying fraudulent banking transactions. This departure from traditional methods signifies a
recognition of the inherent limitations in conventional approaches and an acknowledgment of the
distinctive capabilities that CNNs bring to the field.

Page 2/16
The shift away from conventional methods is prompted by the need for a more adaptive and effective
approach to tackle the complexities of transaction data. Convolutional Neural Networks, known for their
ability to autonomously learn hierarchical features, offer a promising avenue for enhancing the detection
accuracy of fraudulent activities within the dynamic landscape of online transactions, particularly in the
context of the UPI. This research underscores the importance of embracing innovative technologies and
methodologies to address the evolving challenges in the realm of fraud detection, emphasizing the
unique strengths that CNNs bring to the analysis of transactional data.

Convolutional Neural Networks (CNNs), initially crafted for image analysis, have emerged as a
revolutionary choice for bolstering fraud detection, leveraging their inherent capacity to autonomously
acquire hierarchical features. In the realm of banking transactions, this transformative capability
translates into a robust mechanism capable of discerning intricate patterns across both local and global
levels within transaction sequences. In the context of the existing framework, this research advocates for
the adoption of a CNN-based model, positioning it as a computationally efficient and adept alternative.

The ensuing sections of this study delve into the nuanced aspects of CNN architecture, its tailored
application to banking transaction data, and the meticulous selection of a specific dataset for rigorous
evaluation. The primary objective of this research extends beyond merely enhancing the accuracy of
fraudulent transaction detection; it also aspires to make a meaningful contribution to the broader
discourse surrounding the security of digital financial transactions. This is particularly relevant in an era
dominated by online operations, where the need for robust safeguards against fraudulent activities is
paramount. Through a comprehensive exploration of CNN-based models, this study endeavors to fortify
the foundations of secure digital financial transactions, addressing the evolving challenges posed by the
dynamic landscape of online operations.

III. PURPOSE OF THE PAPER

The primary objective of this study is to address the evolving challenges in the landscape of banking
transactions, particularly in the context of Unified Payments Interface (UPI), by proposing and evaluating
a comprehensive scam detection system. The rapid growth in prevalence of online banking and the surge
in digital transactions, the paper focuses on leveraging advanced ML techniques, including CNN, Decision
Trees, Naive Bayes, and Logistic Regression with L1 and L2 regularization, for the accurate identification
of fraudulent UPI transactions. Against the backdrop of the COVID-19 pandemic, which has accelerated
the transition to online financial operations, the study recognizes the heightened risk of fraud and
emphasizes the urgency of robust fraud prevention strategies. The significance of this research lies in the
development and application of innovative machine learning models tailored to the unique challenges
posed by UPI transactions, where the nature of fraud patterns may be intricate and subtle. The specific
objectives of the paper include harnessing the power of CNNs to adaptively learn hierarchical features
from UPI transaction data, utilizing Decision Trees for their interpretability and efficiency, incorporating
Naive Bayes for its simplicity and efficiency in handling feature independence assumptions, and

Page 3/16
employing Logistic Regression with L1 and L2 regularization to enhance interpretability, streamline the
model, and address multicollinearity concerns.

Through the analysis of a dataset containing historical banking transaction records, including both
fraudulent and genuine transactions, the paper aims to evaluate the proposed fraud detection system's
performance using metrics such as precision, recall, F1-Score, ROC Curve, and AUC. Ultimately, the goal is
to provide power of the proposed models in safeguarding users and financial institutions from potential
losses and damages caused by fraudulent UPI transactions in the digital era.

IV. LITERATURE REVIEW

The work by Aleskerov, Freisleben,et al. [1] and Rao (1997) represents a pivotal contribution in the domain
of fraud detection, specifically focusing on credit card transactions. In this study, the authors introduce
the CardWatch system, which utilizes neural network-based database mining techniques to enhance the
detection of credit card fraud. This marks a significant departure from conventional approaches,
highlighting the application of computational intelligence in the intricate task of identifying fraudulent
activities within financial transactions.

In the pioneering work by Dorronsoro et al. [14], on fraud detection in credit card operations, marking a
significant advancement in the utilization of artificial intelligence in financial security. Their research
focuses on leveraging the adaptive learning capabilities of neural networks to discern patterns indicative
of fraudulent activities. The study provides knowledge about neural-based approaches. The neural
network paradigm presented by Dorronsoro et al. serves as a pivotal reference for the current literature
review, highlighting the historical progression and ongoing relevance of artificial intelligence in enhancing
security measures within credit card transactions.

Abdallah, Maarof, and Zainal's survey et al. [3] is a comprehensive exploration of fraud detection
systems. This survey critically examines various approaches, techniques, and technologies employed in
fraud detection across diverse domains. By offering a systematic overview, the authors contribute to the
understanding of current challenges and advancements in the field. The survey's coverage of a broad
spectrum of fraud detection methodologies makes it a valuable resource for researchers, providing
insights that can inform the development of effective and adaptive fraud detection systems. The
inclusion of this survey in a journal paper contributes to the scholarly discourse, guiding future research
endeavors in the dynamic landscape of fraud prevention.

The edited volume authored by Andrews and Peterson (1990) et al. [4] stands as a comprehensive
resource delving into the domain of criminal intelligence analysis. While its primary focus is not on
financial fraud exclusively, the compilation serves as a foundational piece that enriches the
understanding of analytical techniques applicable in a broader context, including fraud detection. By
presenting a holistic view of criminal intelligence analysis, the volume contributes theoretical and
methodological insights that can be adapted and applied to various domains, including the intricacies of
financial fraud detection.
Page 4/16
Artís, Ayuso, and Guillén's 1999 [5] work on insurance fraud modeling provides insights applicable to
financial fraud scenarios. While not directly focused on financial fraud, the study's findings offer valuable
insights into nuanced fraudulent patterns, enhancing fraud detection methodologies. Summers and
Sweeney's et al. [10] research presents a rigorous examination of the relationship between fraudulently
misstated financial statements and insider trading activities. By leveraging a robust dataset and
employing empirical methodologies, the authors contribute substantively to the literature on financial
fraud detection. The findings of this study offer implications for regulatory practices and underscore the
importance of considering insider trading as a potential signal of fraudulent financial reporting. This
research enriches the understanding of the complex interplay between financial misstatements and
insider trading, making it a noteworthy addition to the scholarly discourse in the field of accounting and
finance. The inclusion of this study in a journal paper would contribute to the ongoing dialogue on
effective mechanisms for detecting and preventing financial fraud.

Sambra et al.'s [12] paper introduces an innovative framework that addresses the challenges of
centralized social applications. The authors present the Solid platform, grounded in linked data principles,
offering a decentralized alternative for social applications. The research advocates for user-centric
control over personal data and privacy. The adoption of linked data enhances interoperability and
facilitates seamless information exchange among decentralized applications. Sambra et al.'s work
contributes to the evolving landscape of decentralized social applications and serves as a valuable
reference for scholars and practitioners exploring linked data architectures in social networking
platforms.

Sahin's (2017) thesis et al. [2] about Telephony Fraud, conducted at École Doctorale Informatique,
Télécommunication et Électronique, Paris, contributes significantly to telecommunications and fraud
prevention. The research explores the intricacies of telephony fraud, aiming to unveil patterns and
vulnerabilities to inform more effective countermeasures. Sahin's findings are expected to impact the
telecommunications industry, law enforcement, and cybersecurity, offering insights for the development
of proactive measures and technologies. In a journal paper, Sahin's work stands as a valuable reference
for researchers and professionals involved in telecommunications security, providing a synthesis of
evidence, theoretical frameworks, and practical recommendations.

In the study by Becker et al. [13], the authors delve into the history and lessons learned from fraud
detection in the telecommunications sector, providing valuable insights that can inform contemporary
research in the field. The examination of fraud detection methodologies and their evolution over time
offers a comprehensive understanding of the challenges and advancements in telecommunications
security. This historical perspective serves as a foundation for the current study, contributing to the
context and theoretical framework for the exploration of fraud detection methodologies in the specific
domain of the project. The authors emphasize the importance of continuous learning from past
experiences to enhance the effectiveness of fraud detection systems, a principle that resonates with the
ongoing pursuit of innovation and improvement in the field.

Page 5/16
Hand et al. [15] present a critical examination of performance criteria offering a nuanced perspective on
the evaluation metrics essential for assessing the effectiveness of fraud detection systems. The study
underscores the importance of considering various performance indicators to comprehensively evaluate
the accuracy and efficiency of fraud detection models. By addressing the nuanced challenges specific to
plastic card transactions, the authors contribute essential insights into refining evaluation methodologies.
This pivotal work establishes a benchmark for performance assessment in fraud detection, providing a
valuable reference for the ongoing exploration of optimal criteria in the present thesis focused on UPI
fraud identification. The integration of Hand et al.'s performance criteria framework enriches the literature
review by incorporating a robust evaluation perspective tailored to the unique challenges posed by plastic
card fraud.

Barao and Tawn, along with their collaborators [6], have made a significant contribution to statistical
analysis by addressing the challenges posed by short series with outliers. Their paper specifically
explores extremal analysis, with a focus on applications in sea levels and athletics records. Through this
work, they provide valuable insights into the intricate task of handling extreme events within the realm of
statistical analysis. In a parallel vein, Blunt and Hand, along with their colleagues [7], have produced a
comprehensive technical report centered on the UK credit card market. This report stands out as a
relevant and insightful contribution to the fields of applied statistics and financial markets. By delving
deeply into the nuances of the credit card market, their work serves as a valuable resource for
understanding and navigating the complexities of this financial domain.

Bolton and Hand, in their seminal work [8], introduced pioneering unsupervised profiling methods aimed
at enhancing fraud detection. Their research, unveiled at the Conference on Credit Scoring and Credit
Control, showcases innovative approaches that have significantly contributed to the field. Phua et al.'s [9]
(2010) seminal work serves as a foundational resource in the field. systematically reviews and
consolidates advancements in data mining techniques for fraud detection. The research critically
analyzes existing methodologies, identifies trends, and offers valuable insights into the evolving
landscape of fraud detection. This comprehensive survey, rich with references, provides an essential
knowledge base for researchers and practitioners in data mining, cybersecurity, and fraud prevention. As
a citable source, it enhances the scholarly discourse on fraud detection, guiding further research and the
development of robust strategies in the ongoing battle against fraudulent activities.

Brockett, Xia, and Derrig (1998) et al., [11] employed Kohonen's self-organizing feature map (SOFM) to
reveal fraudulent automobile bodily injury claims. Their study, published in The Journal of Risk and
Insurance, focused on detecting fraudulent activities within insurance claims using advanced data
analysis techniques. This research signifies a pioneering effort in leveraging SOFM for fraud detection in
the insurance sector, showcasing the potential of sophisticated machine learning algorithms in
combating fraudulent behaviors.

V. ABOUT DATASET

Page 6/16
Within this dataset, a comprehensive set of 31 attributes is meticulously curated, offering crucial insights
into financial transactions. The temporal dimension is encapsulated in the "Time" attribute, which
delineates the chronological sequence of transactions, providing a temporal context for analysis. This
attribute allows for the exploration of patterns and trends over time, enriching the dataset with a temporal
perspective. The vectors V1-V28 delve into the intricate details of each transaction, constituting a
multifaceted set of attributes. These vectors have undergone a transformation into z-scores, a statistical
standardization process. This transformation ensures that the values within each vector are
standardized, facilitating meaningful comparisons and analyses across the entire dataset. By converting
the transaction details into z-scores, the dataset is rendered more amenable to the application of machine
learning algorithms, as it mitigates the influence of varying scales and magnitudes in the original data.
The standardization of these vectors is particularly valuable in enhancing the interpretability and
effectiveness of machine learning models. It enables algorithms to discern subtle patterns and anomalies
within the transaction details, contributing to the accurate identification of noteworthy trends related to
fraud detection. This dataset is a nuanced compilation of attributes, with the "Time" attribute providing a
temporal context, and vectors V1-V28 furnishing detailed transaction information in a standardized form
through the utilization of z-scores. This thoughtful preprocessing enhances the dataset's suitability for
sophisticated analyses and machine learning applications, particularly in the realm of fraud detection
within financial transactions.

The attribute labeled "Amount" in this dataset serves as a critical indicator, representing the monetary
value associated with each individual transaction. This attribute is pivotal in providing a quantitative
dimension to the dataset, offering insights into the financial magnitude of the transactions under
scrutiny. By capturing the monetary aspect of each transaction, the "Amount" attribute becomes a key
factor in understanding the financial dynamics within the dataset. The inclusion of the "Amount" attribute
allows for a nuanced exploration of potential patterns and trends related to transaction amounts.
Analyzing the distribution and characteristics of transaction amounts can unveil distinct patterns
associated with fraudulent activities. Unusual spikes, irregularities, or patterns in transaction amounts
may indicate anomalous behavior that could be indicative of fraudulent transactions. Moreover, the
"Amount" attribute enables researchers and data analysts to assess the financial impact of fraudulent
activities. Understanding the financial magnitude associated with fraudulent transactions is essential for
evaluating the severity of potential risks and losses incurred by financial institutions or users. The
"Amount" attribute adds a valuable layer of granularity to the dataset, allowing for a more comprehensive
exploration of potential patterns and anomalies related to transaction amounts. This attribute plays a
crucial role in enhancing the dataset's utility for fraud detection purposes, providing a quantitative
perspective that contributes to a more nuanced understanding of the financial dynamics within the realm
of the analyzed transactions.

The focal point of the dataset is the "Class" attribute, serving as the target variable. It classifies
transactions into two categories: "0" signifies transactions with no fraud, while "1" designates
transactions involving fraudulent activities. This binary classification facilitates the development of
models aimed at accurately distinguishing between legitimate and fraudulent transactions. The
Page 7/16
incorporation of z-scores in the preprocessing stage for the vectors V1-V28 is a strategic choice aimed at
optimizing the effectiveness of machine learning algorithms in analyzing transaction-related information.
Z-scores, also known as standard scores, are used to standardize the values within each vector. This
standardization process is integral for creating a uniform scale across the entire dataset, ensuring that
the variables are comparable and eliminating potential biases introduced by differing scales. By
standardizing these vectors, the machine learning algorithms can more effectively discern patterns and
anomalies within the dataset. Standardization enhances the model to identify subtle variations and
irregularities in the transaction-related features, as it eliminates the influence of varying scales among the
original attributes. This is crucial for the accurate detection of patterns associated with fraudulent
transactions, as anomalies may manifest in different vectors, and standardization allows for a fair
comparison across these diverse attributes. The standardized vectors provide a common ground for
analysis, facilitating the identification of meaningful relationships and trends. This preprocessing step is
particularly beneficial when employing machine learning models that rely on distance metrics or
comparative measures, as it ensures that the algorithm is not unduly influenced by the magnitude of
values in different vectors. The use of z-scores in preprocessing V1-V28 vectors is a methodical approach
to enhance the analytical capabilities of machine learning algorithms. Standardization creates a level
playing field for these vectors, allowing the algorithms to effectively uncover subtle patterns and
anomalies within the dataset, ultimately contributing to more accurate and robust fraud detection.

VI. PROPOSED METHODOLOGY

1. Data Collection and Preprocessing:

In the initial phase of our methodology, we focus on acquiring the "FraudDetectionDataset" from internal
transaction data. This dataset encompasses a substantial volume of 284,807 transactions, each
characterized by 30 features. These features include essential transaction information such as
transaction amount, time, and anonymized PCA components, offering a comprehensive representation of
the transactional landscape. One notable challenge in this dataset is the imbalanced class distribution,
where only 492 out of the 284,807 transactions are identified as fraud cases. To address this issue and
ensure robust model training, we implement strategic preprocessing steps.

Firstly, to standardize the data and facilitate optimal model performance, we employ feature scaling
using standardization. This process involves transforming the features to have a mean of 0 and a
standard deviation of 1. By doing so, we mitigate the impact of varying scales among different features,
providing a more consistent and effective input for our algorithms.

Secondly, to handle any missing data within the dataset, a meticulous approach is undertaken.
Fortunately, in this dataset, there are no missing values. However, a robust preprocessing pipeline should
be equipped to address such instances if they were present. This ensures the dataset's completeness and
enhances the reliability of subsequent analyses. By systematically collecting and pre-processing the
"FraudDetectionDataset," we establish a solid foundation for the subsequent phases of our fraud

Page 8/16
detection project. This meticulous preparation is crucial in dealing with the challenges posed by
imbalanced class distribution and ensuring the dataset's suitability for training and evaluating our
machine learning models.

2. Algorithm Selection and Implementation:

In the subsequent stage of our fraud detection project, we adopt a comprehensive approach by deploying
a diverse ensemble of sophisticated algorithms. This ensemble includes the Feedforward Neural Network
(FNN), Convolutional Neural Network (CNN), Decision Tree, Naive Bayes, Logistic Regression with both L1
and L2 regularization, and K-Nearest Neighbors (KNN). Each algorithm in this varied set brings unique
strengths and characteristics, contributing to a robust and well-rounded fraud detection system.

To initiate the model training process, we utilize the preprocessed dataset obtained in the earlier stage.
The preprocessing steps, encompassing feature scaling through standardization and meticulous
handling of missing data, ensure that the data is appropriately prepared for training. This curated dataset,
with its standardized features and complete information, serves as the foundation for training and
evaluating the performance of our selected algorithms. An 80 − 20 split ratio is employed for dividing the
dataset into training and testing sets. This ratio allocates 80% of the data for training the models,
allowing them to learn patterns and relationships within the dataset, and reserves the remaining 20% for
testing the models' generalization capabilities. This partitioning strategy aims to strike a balance between
providing sufficient data for training and preserving a separate set for evaluating the models'
performance on unseen data, thus gauging their real-world applicability.

Each algorithm undergoes a tailored training process, leveraging the training set to learn the intricacies of
the dataset and develop a robust understanding of fraudulent and non-fraudulent patterns. This diverse
set of algorithms ensures that our fraud detection system is capable of capturing a wide range of
patterns and anomalies, enhancing its effectiveness in identifying potential fraudulent transactions. By
employing this ensemble of algorithms and adopting a systematic training approach, we aim to create a
well-performing and versatile fraud detection system that can address the challenges posed by varying
patterns in fraudulent activities. The subsequent evaluation will shed light on the strengths and
weaknesses of each algorithm, enabling us to make informed decisions about their real-world
deployment.

3. Hyperparameter Tuning:

In the pursuit of enhancing the performance of our fraud detection algorithms, a critical step involves the
application of hyperparameter tuning techniques. Two prominent methods employed for this purpose are
grid search and cross-validation.

Grid search involves systematically testing a predefined set of hyperparameter values to identify the
combination that yields the best model performance. This exhaustive search across the hyperparameter
space ensures that the optimal configuration is discovered, maximizing the algorithm's ability to discern

Page 9/16
fraudulent transactions accurately. The hyperparameters chosen for grid search are specific to each
algorithm and are carefully selected based on their potential impact on the model's performance.
Simultaneously, cross-validation is implemented to assess the algorithm's performance across different
subsets of the data. This technique involves splitting the dataset into multiple folds and training the
model on different combinations of training and validation sets. Cross-validation provides a more
comprehensive evaluation of the algorithm's robustness and generalization capabilities, thereby reducing
the risk of overfitting to a specific dataset.

The hyperparameter tuning process is highly customized, considering the unique characteristics of each
algorithm. For instance, the learning rate and the number of hidden layers and neurons may be crucial for
fine-tuning a Feedforward Neural Network (FNN), while the kernel size and stride might be pivotal for
optimizing a Convolutional Neural Network (CNN). Logistic Regression, with its L1 and L2 regularization,
requires careful adjustment of regularization strength, and K-Nearest Neighbors (KNN) necessitates
optimizing the number of neighbors. This tailored approach to hyperparameter tuning ensures that the
unique features and requirements of each algorithm are duly accounted for, optimizing their performance
in the context of fraud detection. The final result is a set of finely tuned algorithms, each configured to its
optimal state, ready to be evaluated for their efficacy in accurately identifying fraudulent transactions
within the given dataset.

4. Performance Evaluation and Results

Upon completing the training phase, the next crucial step involves assessing the performance of our
fraud detection models using key metrics. Two primary metrics employed for this evaluation are test
accuracy and average precision score. Test accuracy measures the proportion of correctly classified
transactions in the test set, providing an overall indicator of the model's correctness. On the other hand,
the average precision score assesses the precision-recall trade-off, particularly crucial for imbalanced
datasets like ours.

A detailed comparison of results is conducted for each algorithm, unveiling their respective strengths and
weaknesses. The Feedforward Neural Network (FNN) and Convolutional Neural Network (CNN) showcase
high accuracy but exhibit relatively low average precision scores. This suggests a potential high false
positive rate, a factor that is critical in fraud detection scenarios. In contrast, machine learning models,
including Decision Tree, Naive Bayes, Logistic Regression (L1 and L2 regularization), and K-Nearest
Neighbors (KNN), demonstrate strong overall performance. Specifically, Decision Tree, Naive Bayes, and
KNN exhibit perfect average precision scores, indicating exceptional accuracy in identifying fraud cases
with minimal false positives. Logistic Regression with L1 regularization also performs well in terms of
average precision. However, Logistic Regression with L2 regularization shows a comparatively lower
precision score, suggesting a higher rate of false positives compared to its counterparts.

An insightful observation arises when considering the trade-off between accuracy and precision. While
FNN and CNN achieve high accuracy, their lower average precision scores imply a potential challenge in
accurately identifying fraud cases without triggering false positives. On the other hand, machine learning
Page 10/16
models strike a balance, achieving high accuracy while maintaining excellent precision. In practical fraud
detection scenarios, this trade-off has significant implications. The decision to prioritize accuracy over
precision or vice versa depends on the specific requirements of the application. For instance, in financial
settings, minimizing false positives (precision) might be prioritized to avoid inconveniencing legitimate
customers, even at the expense of overall accuracy. These nuanced insights guide the selection and fine-
tuning of models for deployment in real-world fraud detection systems.

Results:

The evaluation of our implemented models has provided valuable insights into their performance.
Notably, the Feedforward Neural Network (FNN) and Convolutional Neural Network (CNN) have exhibited
commendable accuracy in test sets, showcasing their ability to accurately classify transactions. However,
the relatively low average precision scores for both models indicate potential challenges with false
positive rates. This emphasizes the significance of considering precision-recall trade-offs, especially in
the context of imbalanced datasets like ours.

In contrast, the machine learning models, including Decision Tree, Naive Bayes, Logistic Regression with
L1 and L2 regularization, and K-Nearest Neighbors (KNN), have demonstrated robust overall performance
with varying precision scores. Particularly noteworthy is Logistic Regression with L1 regularization (LR1),
which stands out by achieving the best overall accuracy among all algorithms. This remarkable accuracy
underscores LR1's effectiveness in correctly identifying fraud cases with minimal false positives. Such
precision is of paramount importance in fraud detection, where minimizing false positives is crucial to
avoid inconveniencing legitimate users. These results provide actionable insights for refining fraud
detection strategies in real-world applications. For instance, the outstanding precision achieved by LR1,
along with the perfect precision scores attained by Decision Tree, Naive Bayes, and KNN models,
suggests their suitability for deployment in scenarios where false positives must be minimized. On the
other hand, while FNN and CNN models demonstrate high accuracy, further optimization or consideration
of additional factors may be necessary to address potential challenges with false positives.

Moreover, in the context of user interface design, presenting these insights in a comprehensible manner
becomes essential. A user-friendly interface should provide clear visualizations and summaries of the
model performance, emphasizing the trade-offs between accuracy and precision. This aids decision-
makers in selecting models aligned with the specific requirements of the application. Additionally,
incorporating user feedback mechanisms into the interface can enhance the adaptability of the system,
allowing for iterative improvements based on real-world usage and evolving fraud patterns. Overall, the
integration of insightful model evaluations and a user-friendly interface forms a cohesive strategy for
refining and deploying effective fraud detection systems in practical applications.

VII. CONCLUSION
In summary, this research has addressed the pressing challenge of fraud detection in Unified Payments
Interface (UPI) transactions, particularly in the context of the evolving landscape of digital banking and
Page 11/16
exacerbated by the COVID-19 pandemic. By presenting and evaluating a comprehensive fraud detection
system utilizing advanced ML techniques such as CNN, Decision Trees, Naive Bayes, and Logistic
Regression with L1 and L2 regularization, the study contributes valuable insights to the realm of financial
security.

The research is significant given the increasing reliance on online financial transactions, necessitating
robust fraud prevention measures. The adaptability of machine learning models showcased, especially
the innovative use of CNNs, Decision Trees, and Naive Bayes, enhances the ability to capture complex
patterns within UPI transaction data. Logistic Regression with L1 regularization improves interpretability
and aids in feature selection, while L2 regularization addresses multicollinearity concerns, collectively
streamlining the fraud detection process. Evaluation metrics, including precision, recall, F1-Score, ROC
Curve, and AUC, provide a comprehensive assessment of the models' performance in accurately
identifying and preventing fraudulent UPI transactions. These findings contribute to ongoing efforts to
strengthen fraud prevention strategies in the financial sector. The adaptability demonstrated by machine
learning models in this study opens avenues for further refinement, ensuring the resilience of fraud
detection mechanisms in the dynamic landscape of online banking and digital transactions.

Declarations
Funding:

The authors declare that no funds, grants, or other support were received during the preparation of this
manuscript.

Competing Interests:

The authors have no relevant financial or non-financial interests to disclose

Author Contributions:

All authors contributed to the study conception and design. Material preparation, data collection and
analysis were performed by Melam Nagaraju and Yarramreddy Chandrasena Reddy. The first draft of the
manuscript was written by Melam Nagaraju and all authors commented on previous versions of the
manuscript. All authors read and approved the final manuscript.

Ethics approval:

This is an observational study. The SRGEC Research Ethics Committee has confirmed that no ethical
approval is required.

Consent to participate:

Informed consent was obtained from all individual participants included in the study.

Page 12/16
Consent to publish:

We consent to the research staff collecting and processing our information.

Data Availability Statement:

The dataset generated during and/or analyzed related to this publication are available from the
corresponding author on reasonable request

Author information:

Authors and Affiliations

1) Melam Naga Raju, Assistant professor, Dept. of IT, SR Gudlavalleru Engineering College, Gudlavalleru,
AndhraPradesh–521356,India, [email protected],

2) Yarramreddy Chandrasena Reddy, Dept. of IT, SR Gudlavalleru Engineering College, Gudlavalleru,

Andhra Pradesh–521356, India [email protected]

3) Polavarapu Nagendra Babu, Dept. of IT, SR Gudlavalleru Engineering College, Gudlavalleru, Andhra
Pradesh–521356, India, [email protected]

4) Venkata Sai Pavan Ravipati, Dept. of IT, SR Gudlavalleru Engineering College, Gudlavalleru, Andhra
Pradesh–521356, India, [email protected]

5) Velpula Chaitanya, Dept. of IT, SR Gudlavalleru Engineering College, Gudlavalleru, Andhra Pradesh–
521356, India, [email protected]

References
1. ALESKEROV E, FREISLEBEN, B., and, RAO B (1997) CARDWATCH: A neural network-based database
mining system for credit card fraud detection. In Conference (pp. 220–226). IEEE, Piscataway, NJ
2. Sahin M (2017) Understanding Telephony Fraud as an Essential Step to Better Fight it [Thesis]. École
Doctorale Informatique, Télécommunication et Électronique, Paris
3. Abdallah A, Maarof MA, Zainal A (2016) Fraud detection system: A survey. J Netw Comput Appl
68:90–113
4. ANDREWS PP, PETERSON MB (eds) (1990) Criminal Intelligence Analysis. Palmer Enterprises,
Loomis, CA
5. ARTÍS M, AyUSO M, GUILLÉN M (1999) Modeling different types of automobile insurance fraud
behavior in the Spanish market. Insurance Math Econ 24:67–81
6. BARAO MI, TAWN JA (1999) Extremal analysis of short series with outliers: Sea-levels and athletics
records. Appl Stat 48:469–487

Page 13/16
 7. BLUNT G, HAND DJ (2000) The UK credit card market. Technical report, Department of Mathematics,
Imperial College, London
8. BOLTON RJ, HAND DJ (2001) Unsupervised profiling methods for fraud detection. In Conference on
Credit Scoring and Credit Control 7, Edinburgh, UK, 5–7 Sept
9. Phua C, Lee V, Smith K, Gayler R (2010) A comprehensive survey of data mining-based fraud
detection research. https://doi.org/10.48550/ARXIV.1009.6119
10. Summers SL, Sweeney JT (1998) Fraudulently misstated financial statements and insider trading:
An empirical analysis. 73(1):131–146https://www.jstor.org/stable/248345
11. BROCKETT PL, XIA X, DERRIG RA (1998) Using Kohonen's self-organizing feature map to unveil
automobile bodily injury claims fraud. J Risk Insur 65:245–274
12. Sambra AV, Mansour E, Hawke S, Zereba M, Greco N, Ghanem A, Zagidulin D, Aboulnaga A, Berners-
Lee T (2016) Solid:a platform for decentralized social applications based on linked data
13. Becker RA, Volinsky C, Wilks AR (2010) Fraud Detect Telecommunications 52(1):20–33
14. Dorronsoro JR, Ginel F, Sanchez C, Santa Cruz C (1997) Neural fraud detection in credit card
operations. IEEE 8:827–834
15. Hand C, Whitrow DJ, Adams C, Juszczak NM, Weston P D (2008) Performance criteria for plastic
card fraud detection. JORS 59(7):956–962

Figures

Page 14/16
Figure 1

comparison of accuracies of algorithms

Figure 2

Main System Interface for transaction uploading

Page 15/16
Figure 3

Verifying the transaction and displaying the result

Page 16/16