Online payment fraud detection is a critical problem in the financial sector due to the increasing volume of

digital transactions. This paper proposes a machine learning-based fraud detection system using CatBoost,

XGBoost, and a soft voting ensemble model. Principal Component Analysis (PCA) is applied for dimensionality

reduction, and SMOTE is used to address class imbalance. The models are evaluated using precision, recall,

Credit cards have transformed how we make purchases and manage our finances. A Online payment is a plastic

payment card provided by a financial organization, usually a bank that allows the cardholder to borrow money

to purchase goods and services. Unlike debit cards, which remove cash straight from the cardholder's bank

account, credit cards provide a line of credit that must be returned at a later date, typically with interest. The

cardholder may use the Online payment for a variety of purchases, both in person and online, making it a

convenient and generally recognized payment method across the world [1] [2]. Credit cards have become an

essential aspect of modern consumer culture, giving people more purchasing power and the freedom to manage

their finances properly. When a consumer applies for a Online payment and is authorized, the financial institution

sets them a credit limit, which is the most money they may borrow with the card. The cardholder may use the

Online payment to make purchases up to the limit. Each month, the Online payment firm delivers a billing

transactions across several domains. As technology progresses and financial transactions move to digital

platforms, the danger of fraudulent activity increases, making fraud detection more important than ever [8].

As the finance sector continues to undergo digital transformation, online payment systems are now ubiquitous

in modern business, giving users unprecedented speed and convenience. But the growth of this convenience has

simultaneously provided the perfect conditions for sophisticated fraudulent behaviour which causes a critical

risk that can result in significant global financial losses for individuals and organizations. The development of

Online Payment Fraud Detection Systems has consequently become a top priority to protect the integrity and

security of the digital economy. Unlike traditional, inflexible rulebased solutions that can't keep up with the rapid

pace of change of techniques by cybercriminals, contemporary solutions such as this one use contemporary

machine learning (ML) algorithms to analyse complex transactional data. [6]ML models are capable to detect

reputations, and impaired consumer trust.

Fraud can have serious consequences, including financial losses, damaged reputations, and compromised

customer trust. Fraud detection systems play an important role in avoiding such repercussions by proactively

identifying and flagging questionable transactions. Effective fraud detection benefits financial organizations by

not just protecting their assets but also ensuring regulatory compliance and increasing client confidence.

Maryam Habibpour [1] and colleagues propose in this work Numerous research have used deep neural networks

(DNNs) to identify Online payment fraud, with the goal of improving point prediction accuracy and avoiding

unwanted biases through the development of various network architectures or learning models. It is critical to

measure uncertainty in conjunction with point estimate because it lowers model unfairness and allows

In this paper, Asma Cherif [2] et al. Online payment fraud is becoming a serious and growing problem as new

technologies and communication channels emerge, such as contactless payment. This article gives a

comprehensive examination of current research on detecting and forecasting fraudulent Online payment

transactions from 2015 to 2021. The 40 papers picked for consideration are analysed and classified based on the

topics they cover (class imbalance problem, feature engineering, etc.) and the machine learning approach they

use (conventional and deep learning modelling). Our analysis reveals that deep learning has received little

research, implying that more research is needed to address the difficulties in detecting Online payment fraud

using cutting-edge technologies such as big data analytics, large-scale machine learning, and cloud computing.

Our study is a significant resource for academic and industrial researchers in analysing financial fraud detection

systems and designing trustworthy solutions by highlighting existing research challenges and future research

CCF detection models. The imbalanced CCF dataset is then used to build detection algorithms with DRL.

Jacobo Chaquet-Ulldemolins [4] et al. proposed this system. Artificial intelligence (AI) has lately gained

popularity in the global economy due to its exceptional ability to analyse and model data in a variety of

disciplines. As a result of this condition, society is rapidly becoming more automated, and these new approaches

may be combined to form a beneficial tool for addressing the difficult challenge of credit fraud detection.

However, severe restrictions make it impossible for financial institutions to comply with them while employing

modern procedures. From a methodological approach, auto encoders have demonstrated effectiveness in finding

nonlinear features in a range of problem domains. However, auto encoders are opaque and often referred to as

"black boxes." In this study, we provide an interpretable and impartial CFD approach.

Esra Faisal Malik [5], for example. As mentioned in this article financial crimes have progressively harmed

financial institutions. Various single and hybrid machine learning algorithms have been used to detect crimes

phase was used to create the hybrid approaches. The hybrid models created were separated into two phases. Our

results revealed that the hybrid model Adaboost + LGBM is the best model due to its superior performance.

In their article, Ibtissam Benchajiet [6] discusses a new system for Online payment fraud detection, making

improvements to current testing methods through the use of sequential modelling in traditional machine learning.

dataset, allowing for techniques that can be developed to optimize values that may not be present or available

within the dataset. The methodology for achieving this will be strengthened the combination of three

costeffective probabilistic dimensionality reduction feature cross-validator selection (LSTM). As Online

payments are increasingly common, it should not be surprising that fraudulent activities are occurring. To

properly combat and address fraud, financial institutions must take steps to enhance their monitoring systems to

decrease significant losses. The proposed model being summarized should attempt to address scam fraud

activities.

In this paper, Ebenezer Esenogho [7] and colleagues have suggested, "Recent improvements in e-commerce and

communication systems have played a major role in the increased usage of credit cards both for online purchases,

as well as traditional shopping. Nonetheless, the rate of fraudulent activity in online payment transactions has

purchasing habits of online payment customers.

According to Samaneh Sorournejad [8] et.al. in this article, Online payment plays a detrimental and significant

role in today's economy, as it has become an unavoidable element of household, business, and global activities.

Although credit cards can have tremendous advantages when used properly with restraint, significant damage to

credit and finances can be created through fraud. Numerous methods have been proposed to resolve the growth

of Online payment fraud, but all have the same purpose of detecting Online payment fraud. Each of these

techniques has certain drawbacks, advantages, and features. The focus of this paper is to address the challenges

of Online payment fraud detection and to review the state of the art in Online payment fraud detection techniques,

and datasets, and evaluation criteria. The advancements and drawbacks of the fraud detection techniques will be

presented and compared. In addition, a taxonomy of the aforementioned techniques is presented, distinguishing

learning discriminative representations from transactions. Since fraudsters often commit fraud by mimicking

cardholders' behavior, the performance of existing fraud detection models is poor. Here we present an Adaptive

by financial systems. Several studies have developed predictive models for Online Payment Fraud detection

using different machine learning methods. In this work, we propose an adaptive framework that enhances Online

the type of transaction and the clients profile.

In today's digital economy, credit cards are indispensable, and as their use has recently increased significantly,

so has Online payment theft. Algorithms for machine learning (ML) have been used to detect Online payment

fraud. However, it has proven challenging for ML classifiers to function at their best due to Online payment

holders' dynamic shopping habits and the issue of class imbalance. This paper presents a robust deep-learning

method to address this issue, which combines a multilayer Perceptron (MLP) as the meta-learner with long short-

term memory (LSTM) and gated recurrent unit (GRU) neural networks as base learners in a stacking ensemble

architecture. To balance the class distribution in the dataset, the hybrid synthetic minority oversampling

methodology and edited nearest neighbor (SMOTE-ENN) method are used. According to the experimental

The Online Payments Fraud Detection System follows a methodical approach to reliably identify fraudulent

transactions. First, transactional datasets are gathered and uploaded, which include features such as transaction

type, amount, account balances, and target fraud label. The feature names can either be uploaded from a file, or

taken from the dataset to maintain consistency. The data must also be cleaned for completion and correctness in

order to be prepared for modeling. Principal Component Analysis (PCA) is used to mitigate computational

complexity, while improving model performance. PCA will convert the original high-dimensional feature space

into a lowerdimensional representation that retains 95% of the variance. This step assists in minimizing the

potential redundancies and allows the models to focus on the more informative components that signify fraud

patterns. In the predictive modeling section, three approaches to model fitting are utilized - CatBoost with PCA,

deploy a web application based on Streamlit supporting transactions’ single and batch predictions. The