Fraud Detection in Online Transactions using

Machine Learning Techniques

A SEMINAR REPORT

submitted to the Savitribai Phule Pune University, Pune

In the partial fulfilment of the requirements
for the award of the degree

BACHELOR OF ENGINEERING
(AI&DS Engineering)

BY
Roll No:

Under the guidance of

Assistant Professor

DEPARTMENT OF AI&DS ENGINEERING

Vidya Pratishthan’s Kamalnayan Bajaj Institute of Engineering
and Technology,
Vidyanagari Bhigawan Road
Baramati- 413133

2023-24
 CERTIFICATE

This is to certify that has successfully submitted her seminar report to the Depart-
ment of AI&DS Engineering, VPKBIET, Baramati, on

”Fraud Detection in Online Transactions using Machine Learning

Techniques”

during the academic year 2023-2024 in the partial fulfillment towards completion of
Third Year of Bachelor of Engineering in Artificial Intelligence & Data Science
.

Assistant Professor, Assistant Professor,

Guide, HOD,
Dept of AI&DS Engineering. Dept of AI&DS Engineering.

Dr. R.S.Bichkar
Principal
VPKBIET, Baramati.

Place: Vidya Pratishthan’s Kamalnayan Bajaj Institute of Engineering and

Technology, Baramati.
Date : 16/10/2023

i
 Acknowledgments
I am feeling happy to forward this seminar report as an image of sincere efforts.
The Successful seminar reflects my work, the effort of my guide in giving me good
information.

I would like to express my heartiest thanks to my Seminar guide for his/her

undying support which makes it possible to make this Seminar knowledgeable. She
not only provided me with the literature and guidance to study but also the platform
required for me to prepare best for this seminar.

I am also equally indebted to our seminar coordinator who has been a constant
source of inspiration and guiding star in achieving my goal.

I thank , Head of Department of AI&DS Engineering and respected Principal

for supporting and providing all facilities to complete the work. I am thankful to and
fortunate enough to get constant encouragement, support, and guidance from the entire
teaching staff of the Department of AI&DS Engineering which helped me in successfully
completing my seminar work.

ii
 Abstract

Online transactions have seamlessly integrated into modern life, offering un- par-
alleled convenience and efficiency. However, the ease of online transactions is counter-
balanced by the escalating specter of fraud, phishing, and data com- promise. With the
surge in online transaction frequency, these risks have grown significantly, potentially
undermining personal privacy and financial security. To mitigate these threats, substan-
tial investments from financial institutions and insurance companies have been dedicated
to pioneering systems that detect and prevent fraudulent transactions effectively.

These sophisticated systems harness the power of machine learning, an ad- vanced
technology that enables computers to learn from extensive datasets. These systems pro-
gressively refine their ability to identify suspicious patterns indicative of fraud. Within
this framework, both the XGBoost algorithm, a collection of decision trees, and neural
networks, computing systems inspired by the human brain, play pivotal roles.

Comparative evaluations against alternative methodologies demonstrate the remark-

able speed and accuracy of XGBoost and neural networks. This un- derscores their
importance in tackling the unique challenges posed by fraud detection. Their collective
deployment showcases the formidable capabilities of machine learning and neural net-
works, assuring the integrity of online transac- tions by robustly countering fraudulent
activities.

Keywords : Online transactions, Fraud detection, Machine learning, XGBoost algo-

rithm, Neural networks, Financial security, Data compromise, Privacy, etc.

iii
Contents

Abstract iii

List of Figures vi

1 Introduction 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Literature Survey 2

3 Motivation 5

4 System Architecture 6
4.1 Explanation of System Architecture . . . . . . . . . . . . . . . . . . . . 7

5 Requirement Specification 8
5.1 Software Requirements: . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
5.2 Hardware Requirements: . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

6 Approach for Solving the Problem 9

7 Comparison of the Approaches 11

7.1 Limitations Of Fraud detection in online transactions using Machine Learn-
ing Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
7.2 Advantages of new techniques over previous techniques . . . . . . . . . . 12
7.3 Performance Demonstration . . . . . . . . . . . . . . . . . . . . . . . . . 12

8 Advantages and Disadvantages 13

8.1 Advantages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
8.2 Disadvantages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

9 Conclusion 14

10 Future Scope 15

11 Summary 16
Bibliography 17

v
List of Figures

4.1 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Chapter 1

Introduction

1.1 Introduction
In today’s digital age, the prevalence of online transactions has surged to un-
precedented heights. With the widespread adoption of digital payment methods and
e-commerce platforms, billions of transactions occur on the web daily. However, this dig-
ital revolution has also brought forth a formidable challenge—online transaction fraud.
As online commerce continues to expand, the risk of fraudulent activities targeting per-
sonal and financial data has become increasingly significant. The implications of online
transaction fraud extend beyond individual financial losses, casting a shadow on the
health and sustainability of the digital economy itself.

This report introduces two distinctive fraud detection algorithms, each tailored to
tackle the multifaceted challenges posed by online transactions. The first, based on
a Fully Connected Neural Network, employs sophisticated data analysis techniques to
discern patterns and anomalies within transaction data. The second algorithm, rooted
in the power of XGBoost, takes advantage of hyperparameter optimization to construct
a formidable fraud detection system.

This exploration of online transaction fraud detection is not merely an academic

endeavor; it holds real-world implications. The efficacy of these algorithms transcends
theoretical discussions, as the paper also details the creation of an interactive online
transaction fraud detection system based on the XGBoost model. This system is de-
signed to empower users by automatically analyzing uploaded transaction data and
promptly delivering fraud detection results.

As we dive into the world of online transaction fraud detection, we’ll take a closer
look at how these fraud-fighting algorithms work. We’ll explore the technical details
that make them tick and see how well they perform. We’ll also consider how these tools
can be used to protect us in a world where online deception is a big concern.

1
Chapter 2

Literature Survey

1.Fraud Detection using Machine Learning in e-Commerce

Adi Saputra and Suharjito’s paper on ”Fraud Detection using Machine Learning in
e-Commerce,” pub- lished in 2019, addresses the rising volume of internet users and sub-
sequent growth in e-commerce trans- actions, noting the concurrent increase in online
fraud. As the paper aims to develop fraud prevention techniques, it evaluates various
machine learning algorithms, including Decision Tree, Na ıve Bayes, Ran- dom Forest,
and Neural Network. The study uses unbalanced data and employs Synthetic Minority
Over-sampling Technique (SMOTE) to rectify this imbalance. Evaluation through con-
fusion matrix anal- ysis reveals that the Neural Network achieves the highest accuracy
at 96 percent followed by Random Forest and Na ıve Bayes, both at 95 percent, and
Decision Tree at 91 percent. Furthermore, SMOTE sig- nificantly improves the F1-Score
and G-Mean values, underlining its efficacy in enhancing fraud detection in e-commerce
transactions. These findings contribute to the literature on fraud detection, particularly
in the e-commerce domain.

2.Credit card fraud detection using artificial neural network

Authors A. RB and S.K. KR published a paper in 2021 focusing on credit card fraud
detection using deep learning techniques and comparing them with various machine
learning algorithms. The paper addresses the growing issue of credit card fraud across
industries, particularly finance and e-commerce. It highlights the application of machine
learning and deep learning techniques, with a primary focus on artificial neural networks
(ANNs). The paper’s objectives are to detect fraudulent credit card transactions and

2
compare the performance of supervised learning and deep learning algorithms, empha-
sizing accuracy. It discusses existing credit card fraud detection systems using machine
learning algorithms like Support Vector Ma- chine, Na ıve Bayes, and k-Nearest Neigh-
bor, proposing an ANN-based system for evaluation. The research summarizes related
studies in the field, including methods like optimized light gradient boosting machines
and balanced dataset approaches. The proposed system’s architecture includes user
registration, algo- rithm selection, data preprocessing, and model training/testing. Im-
plementation methods involve dataset loading, preprocessing, and training/testing with
SVM, KNN, and ANN. The paper discusses results, em- phasizing that ANN outper-
forms SVM and KNN in accuracy for credit card fraud detection, making it a promising
solution to enhance security across industries and mitigate fraudulent transactions.

3.Fraud Detection using Machine Learning and Deep Learn-

ing
The research paper titled ”Fraud Detection using Machine Learning and Deep Learn-
ing” by Pradheepan Raghavan and Neamat El Gayar from Heriot Watt University in
Dubai, UAE, presents a comprehensive investigation into the application of various ma-
chine learning and deep learning techniques for fraud de- tection, focusing on credit card
fraud. The paper begins by highlighting the increasing challenges posed by fraudsters
in exploiting credit card information for unauthorized transactions. It emphasizes the
impor- tance of machine learning and deep learning methods in proactively identifying
and preventing fraudulent activities. The authors conduct experiments using three dis-
tinct datasets, namely the European, Aus- tralian, and German datasets, to benchmark
multiple algorithms, including K-nearest neighbor (KNN), random forest, support vec-
tor machines (SVM), autoencoders, convolutional neural networks (CNN), re- stricted
boltzmann machine (RBM), and deep belief networks (DBN). Evaluation metrics such
as the Area Under the ROC Curve (AUC), Matthews Correlation Coefficient (MCC),
and Cost of failure are employed to assess model performance. The paper concludes by
summarizing the key findings and suggesting di- rections for future research in the evolv-
ing field of fraud detection, which demands adaptive and dynamic systems to combat
ever-evolving fraud techniques.

3
Sr No Paper title and Author Methods/Algorithm Advantages Limitations/Gaps
Improved Fraud
Synthetic Minority
Detection Feature Engineer-
Over-sampling
Accuracy, ing,Model
”Fraud Detection using Technique, Decision
Effective Tuning, Handling
Machine Learning in Trees, Naıve Bayes,
Handling of Evolving Fraud
1 e-Commerce”,2019 Principal
Imbalanced Techniques,
Author: Adi Saputra, Component
Data, Flexibility Integration with
Suharjito Analysis, Genetic
of Machine Business
Algorithm, Neural
Learning Processes,etc
Network
Algorithms,etc.
Improved
Data Preprocess-
”Credit card fraud Accuracy,
ing,Supervised
detection using Automation, Interpretability,
Learning
artificial neural Reduction in Deployment
Algorithms,Support
2 network”,2021 False Positives Considerations,
Vector
Author:Asha High Accuracy Imbalanced
Machine(SVM),
RB,Suresh Kumar KR. with Deep Data,etc.
Artificial Neural
et al. Learning,etc
Network (ANN),etc

Autoencoders,
Convolutional Automated
Adversarial At-
Neural Networks Detection,
”Fraud Detection using tacks,Generalization
(CNN), Restricted Real-time
Machine Learning and Across
Boltzmann Machine Processing,
Deep Learning,2019 Institutions,
3 (RBM),Deep Belief Reduced False
Author:radheepan Cost-Benefit
Net- Positives,
Raghavan and Neamat Analysis,
works(DBN),Area Continuous
El Gayar et al. Real-world Im-
Under the ROC Monitoring, etc
plementation,etc
Curve (AUC), etc

4
Chapter 3

Motivation

1. The motivation behind developing an online transaction fraud detection system

using machine learning techniques is the increasing prevalence of fraudulent ac-
tivities in online transactions. As e-commerce and online financial transactions
continue to grow, so does the risk of fraudulent behavior. Traditional rule-based
fraud detection systems often struggle to keep up with the evolving tactics of
fraudsters, leading to financial losses and damage to consumer trust.

2. Machine learning offers a promising solution to this problem. By leveraging algo-

rithms that can learn from historical transaction data, machine learning models
can identify patterns and anomalies associated with fraudulent behavior. These
models can analyze large volumes of transactional data and make accurate predic-
tions about the likelihood of fraud in new transactions.

3. The motivation for using machine learning in online transaction fraud detection
is to improve the accuracy and efficiency of fraud detection systems. By training
models on labeled datasets, machine learning algorithms can learn to distinguish
between legitimate and fraudulent transactions based on various features such as
transaction amount, location, time, and user behavior. This enables real-time
detection and prevention of fraudulent activities, reducing financial losses and
protecting the integrity of online transactions.

4. Furthermore, machine learning models have the ability to continuously learn and
adapt to new fraud patterns, making them more effective in detecting emerging
fraud techniques. This adaptability is crucial in the dynamic and ever-changing
landscape of online transactions.

5. Overall, the motivation behind developing an online transaction fraud detection

system using machine learning techniques is to enhance the security and trust-
worthiness of online transactions, protect consumers from financial losses, and
mitigate the impact of fraudulent activities on the digital economy.

5
Chapter 4

System Architecture

Figure 4.1: System Architecture

6
4.1 Explanation of System Architecture
The system architecture of the online transaction fraud detection system based on
machine learning consists of several components that work together to detect and pre-
vent fraudulent activities in real-time

1. Data Acquisition: The system acquires raw transaction data uploaded by users.
This data includes transaction information and identity information.
2. Data Preprocessing: The acquired data undergoes preprocessing to prepare it for
analysis. This step involves feature selection and data transformation. For feature
selection, the system focuses on the transaction data set and selects relevant features.
Continuous features are transformed using logarithmic transformation and Z-score stan-
dardization, while categorical features are encoded using One-Hot Encoding.
3. Fraud Detection Algorithm: The system employs two fraud detection algorithms: the
Fully Connected Neural Network (FCNN) and XGBoost. The FCNN algorithm inte-
grates two neural network models with different loss functions to effectively mine infor-
mation from various transaction features. The XGBoost algorithm, a gradient boosting
framework, is optimized using Hyperopt to construct a high-performance fraud detec-
tion model.
4. Model Training: The selected fraud detection algorithm, in this case, the XGBoost
model, is trained using the preprocessed data. The model learns from historical trans-
action data to identify patterns and anomalies associated with fraudulent behavior.
5. Real-Time Prediction: Once the model is trained, it is deployed in the system to make
real-time predictions on new transactions. The system uses the trained XGBoost model
to predict the probability that a transaction is fraudulent. If the predicted probability
exceeds a set threshold (e.g., 0.85), a warning is issued.
6. Results Storage and User Interaction: The detection results, including the predicted
fraud probability, are saved in a database. The system also provides a web interface
where users can upload transaction data and receive real-time fraud detection results.

7
Chapter 5

Requirement Specification

5.1 Software Requirements:

1.Python: It is a dominant language in data science and machine learning. It has
libraries and frameworks.
2.Django: The system, built using the Django framework, a high-level Python web
framework, leverages Django’s tools and libraries for online transaction fraud detection.
3.Machine Learning Libraries: TensorFlow, Keras, PyTorch, and XGBoost.
4.XGBoost: XGBoost is an open-source gradient boosting library commonly used in
machine learning.
5.Database: The system likely used a database management system like MySQL or
PostgreSQL to store detection results and transaction history.

5.2 Hardware Requirements:

1.RAM: 8GB/16GB
2.Processor: Intel i7 and Greater/Amd Ryzen 7 H-series Processor.
3.Graphics: A high-performance GPU, such as an NVIDIA GeForce RTX.
4.Storage: SSD/HDD
5.Network Connectivity: A stable and high-speed internet connection for real-time
fraud detection.

8
Chapter 6

Approach for Solving the

Problem

1.Data Acquisition:

Acquire raw transaction data uploaded by users, including transaction information

and identity information.

2.Data Preprocessing:

Preprocess the acquired data by performing feature selection and data transforma-
tion. Select relevant features from the transaction data set and transform continuous
features using logarithmic transformation and Z-score standardization. Encode categor-
ical features using One-Hot Encoding

3.Fraud Detection Algorithm Selection:

Choose the fraud detection algorithms to be used. In this research, the Fully Con-
nected Neural Network (FCNN) and XGBoost algorithms were selected

4. FCNN Algorithm:

Implement the FCNN algorithm for fraud detection. This involves integrating two
neural network models with different loss functions to effectively mine information from
various transaction features

5. XGBoost Algorithm:

Implement the XGBoost algorithm for fraud detection. Optimize the XGBoost clas-
sifier using Hyperopt to construct a high-performance fraud detection model

9
 6. Model Training:

Train the selected fraud detection algorithm, in this case, the XGBoost model, using
the preprocessed data. The model learns from historical transaction data to identify
patterns and anomalies associated with fraudulent behavior

7.Real-Time Prediction:

Deploy the trained model in the system to make real-time predictions on new transac-
tions. Use the XGBoost model to predict the probability that a transaction is fraudulent.
If the predicted probability exceeds a set threshold, issue a warning

10
Chapter 7

Comparison of the Approaches

7.1 Limitations Of Fraud detection in online transactions

using Machine Learning Techniques

1.Inability to capture complex patterns: Traditional statistical and multi-

dimensional analysis techniques may not be able to capture the intricate patterns
and relationships within transaction data, limiting their effectiveness in detecting
fraud.
2.Lack of scalability: Traditional methods may struggle to handle the increas-
ing volume and complexity of online transactions. Machine learning algorithms,
on the other hand, can scale well with large datasets and adapt to changing fraud
patterns.
3.Limited feature representation: Traditional techniques may have difficulty
representing and extracting relevant features from transaction data. Machine
learning algorithms, such as the proposed Fully Connected Neural Network and
XGBoost, can automatically learn and extract important features from the data.
4.Verification based approach: Traditional methods are often verification tech-
niques that focus on identifying known fraud patterns. They may not be able to
detect emerging or unknown fraud patterns, which can limit their effectiveness in
preventing fraud.

11
7.2 Advantages of new techniques over previous techniques

1.Improved Detection Accuracy: he machine learning algorithms, specifically

the Fully Connected Neural Network and XGBoost, achieve higher accuracy in
detecting fraudulent transactions compared to traditional statistical and multi-
dimensional analysis techniques.
2.Ability to Capture Complex Patterns: Machine learning algorithms have
the capability to capture intricate patterns and relationships within transaction
data, which may not be effectively captured by traditional methods. This enables
the detection of sophisticated fraud patterns that may go unnoticed by traditional
techniques.
3.Scalability: Machine learning algorithms, such as XGBoost, can handle large
volumes of transaction data and adapt to changing fraud patterns. This scalability
allows for efficient processing of a high number of transactions in real-time, making
them suitable for online transaction fraud detection.
4.Adaptability to Emerging Fraud Patterns: Machine learning algorithms
can adapt to emerging fraud patterns by continuously learning from new data.
This flexibility allows the models to detect and adapt to evolving fraud techniques,
providing a more robust and up-to-date fraud detection system.
5.Real-Time Detection: The proposed machine learning-based fraud detection
system can make real-time predictions on new transactions, enabling immediate
identification of potential fraud. This real-time detection capability enhances the
ability to prevent fraudulent transactions and minimize financial losses.

7.3 Performance Demonstration

This paper presents two fraud detection algorithms, a Fully Connected Neural Net-
work and XGBoost, for online transaction fraud detection. The algorithms achieve
high AUC values of 0.912 and 0.969, respectively. The paper also describes the imple-
mentation of an online fraud detection system based on the XGBoost classifier, which
accurately predicts the fraud probability of network transaction behavior. The system
intercepts fraudulent behavior before the user pays, improving the performance of fraud
detection in online transactions. The XGBoost classifier is optimized using Hyperopt,
and the system is implemented using the Django framework. The research is supported
by a student innovation and entrepreneurship training program.

12
Chapter 8

Advantages and Disadvantages

8.1 Advantages
1. Improved detection accuracy: The proposed fraud detection algorithms
based on Fully Connected Neural Network and XGBoost achieve high AUC values
of 0.912 and 0.969, respectively.

2. Efficient model training: The Fully Connected Neural Network algorithm

demonstrates quick model training speed.

3. Scalability: The XGBoost algorithm can handle large volumes of transaction data
and adapt to changing fraud patterns, making it suitable for online transaction
fraud detection

4. Real-time detection: he online transaction fraud detection system based on

XGBoost provides real-time detection, allowing for immediate identification of
potential fraud.

5. Flexibility and customization: The system can be customized and fine-tuned

to specific business needs using the Django framework.

8.2 Disadvantages
1. Parameter tuning: The XGBoost algorithm requires parameter tuning, which
can be time-consuming.

2. Imbalance of sample labels: The detection accuracy of the Fully Connected

Neural Network algorithm may be affected by the imbalance of sample labels.

3. Dependency on specific data sets: The structure of the Fully Connected

Neural Network algorithm needs to be adjusted when the data set changes.

13
Chapter 9

Conclusion

In conclusion, this report presents an online transaction fraud detection system

based on machine learning algorithms, specifically the Fully Connected Neural Network
and XGBoost. The system achieves high accuracy in detecting fraudulent transactions,
with AUC values of 0.912 and 0.969 for the respective algorithms. The system allows
users to upload transaction data and receive fraud detection results in real-time. The
XGBoost algorithm performs better in terms of accuracy, while the Fully Connected
Neural Network algorithm has faster training speed. This system provides an effective
solution for detecting and preventing fraud in online transactions, contributing to the
security and trustworthiness of online transactions.

14
Chapter 10

Future Scope

The future scope of online transaction fraud detection systems based on machine
learning includes exploring ensemble methods, advanced feature engineering techniques,
deep learning models, explainability and interpretability, adaptive learning approaches,
real-time monitoring capabilities, integration with other security systems, and leverag-
ing external data sources. These directions aim to improve the accuracy, efficiency, and
effectiveness of fraud detection systems, addressing emerging challenges and evolving
fraud techniques. By incorporating these advancements, online transaction fraud detec-
tion systems can better detect and prevent fraudulent activities, minimizing financial
losses and ensuring the security of online transactions. In addition to the aforementioned
future directions, another important aspect to consider is the integration of explainabil-
ity and interpretability techniques into fraud detection systems. This would enable users
and stakeholders to understand the reasoning behind the system’s decisions and build
trust in its capabilities. By providing explanations for fraud detection outcomes, such
as highlighting the key features or patterns that contributed to a particular decision,
users can gain insights into the fraud detection process and potentially provide feedback
or suggestions for improvement. This integration of explainability and interpretabil-
ity can also help in meeting regulatory requirements and ensuring transparency in the
decision-making process.

15
Chapter 11

Summary

This research paper focuses on the development of fraud detection algorithms and an
online transaction fraud detection system. The paper introduces two algorithms, a Fully
Connected Neural Network and XGBoost, for detecting fraud in online transactions. The
algorithms achieve high AUC values of 0.912 and 0.969, respectively, indicating their
effectiveness in accurately identifying fraudulent activities.
The paper also describes the implementation of an online fraud detection system
based on the XGBoost classifier. This system allows users to upload transaction data
and receive real-time fraud detection results. The XGBoost classifier is optimized using
Hyperopt, a hyperparameter optimization library, to enhance its performance.
The online fraud detection system is designed to intercept fraudulent behavior before
the user completes the transaction, thereby minimizing financial losses. The system’s
ability to accurately predict the fraud probability of network transaction behavior con-
tributes to its effectiveness in detecting and preventing fraud.
The implementation of the system is carried out using the Django framework, which
provides a robust and scalable platform for building web applications. The research is
supported by a student innovation and entrepreneurship training program, highlighting
the practical relevance and potential impact of the work.
Overall, this research paper presents valuable insights into the development of fraud
detection algorithms and the implementation of an online transaction fraud detection
system. The algorithms and system described in the paper demonstrate high accuracy
and real-time detection capabilities, making them valuable tools for combating fraud in
online transactions.

16
Bibliography

[1] Liu, B., Chen, X. (2021). Online Transaction Fraud Detection System Based
on Machine Learning. Journal of Physics: Conference Series, 2023(1), 012054.
doi:10.1088/1742-6596/2023/1/012054

[2] Zhang, Z., Zhou, X., Zhang, X., et al. (2018). A Model Based on Convolutional Neu-
ral Network for Online Transaction Fraud Detection. Security and Communication
Networks, 10(2), 1-9.

[3] Jain, V. (2017). Perspective Analysis of Telecommunication Fraud Detection Us-

ing Data Stream Analytics and Neural Network Classification Based Data Mining.
International Journal of Information Technology, 9(3), 303-310.

[4] Kurshan, E., Shen, H. (2020). Graph Computing for Financial Crime and Fraud
Detection: Trends, Challenges and Outlook. 2020 Second International Conference
on Transdisciplinary AI (TransAI).

17