CREDIT CARD FRAUD DETECTION BASED ON MACHINE

LEARNING AND DEEP LEARNING

A PROJECT REPORT

Submitted by

DHARSHINI M (960121243013)
MONISHA M (960121243030)
SAJITHA C (960121243039)
SALINI I (960121243040)

in partial fulfillment for the award of the degree of

BACHELOR OF TECHNOLOGY

IN

ARTIFICIAL INTELLIGENCE AND DATA SCIENCE

ANNAI VAILANKANNI COLLEGE OF ENGINEERING

ANNA UNIVERSITY: CHENNAI 600 025

MAY 2025
 ANNA UNIVERSITY: CHENNAI 600 025

BONAFIDE CERTIFICATE

Certified that this Report titled “CREDIT CARD FRAUD DETECTION

BASED ON MACHINE LEARNING AND DEEP LEARNING” is the
bonafide work of “DHARSHINI M (960121243013), MONISHA M
(960121243030),SAJITHA C(960121243039), SALINI I (960121243040)”
who carried out the work under my supervision.

SIGNATURE SIGNATURE

Mrs. B. JENEFA, M.Tech Mrs. B. JENEFA, M.Tech

HEAD OF THE DEPARTMENT SUPERVISOR

Assistant Professor and Head, Assistant Professor and Head,

Department of AI & DS, Department of AI & DS,

Annai velankanni college of Annai velankanni college of

Engineering, pottalkulam, Engineering, pottalkulam,

Azhagappapuram, Azhagappapuram,

Kanyakumari – 629 401 Kanyakumari -629 401

Submitted to project and viva Examination held on ________________

INTERNAL EXAMINER EXTERNAL EXAMINER

ii
 ABSTRACT

With the rapid evolution of the technology, the world is turning to use
credit cards instead of cash in their daily life, which opens the door to many
new ways for fraudulent people to use these cards in a fraudulent way. The
Federal Trade Commission estimates that 10 million people are victimized by
credit card theft each year. Credit card companies lose close to $50 billion
dollars per year because of fraud. This is a very relevant problem that demands
the attention of communities such as machine learning and data science where
the solution to this problem can be automated. The main objective of this paper
is to predict the chances of a fraudulent activity, to improve the prediction
accuracy, and to accomplish self-learning ability. In this article, we focus on
obtaining deep feature representations of legal and fraud transactions from the
aspect of the loss function of a deep neural network by using Squirrel
Optimization and Advance DLMNN classifier. The purpose of this paper is to
obtain better separability and discrimination of features so that it can improve
the performance of our fraud detection model and keep its stability.

iii
 ACKNOWLEDGEMENT

First and foremost, we acknowledge the abiding presence and the

abounding grace of our almighty god for his unseen hand yet tangible guidance
all through the formation of this project. Also, we express our cordial thanks to
our beloved parents

We express our sincere gratitude to our chairman Dr. D PETER

JESUDHAS, for providing us all supports.

We are extremely grateful to our principal for his

inspiration to us for preceding this project Our heartfelt thanks to our
principal Dr. R ANGELINE PRABHAVATHY, Ph.D PRINCIPAL, For
her constant help affectionate support suggestions and valuable ideas in
conducting my study.

We would like to extend our heartfelt thanks to

Mrs. B. JENEFA, M. Tech Head of artificial intelligence and data science
Department, for granting us permission and for providing a good working
environment to complete this project work

We express our heartiest thanks to all staff of the artificial intelligence

and data science department who have helped us with the successful
completion of the project.

iv
 TABLE OF CONTENTS

CHAPTER TITLE PAGE

NO
ABSTRACT iii
LIST OF FIGURES vii
ABBREVATION viii

1 INTRODUCTION 1

1.1 OVERVIEW 1
1.2 EXISTING CHALLENGES 2
1.3 SIGNIFICIANCE RESAERCH 3
1.4 APPLICATIONS 4
1.5 OBJECTIVE 5

2 LITERATURE REVIEW 6
2.1 Existing system 12
2.2 Drawback 14
2.3 Proposed system 15

3 SYSTEM SPECIFICATION 18
3.1 HARDWARE REQUIREMENTS 18
3.2 SOFTWARE REQUIREMENTS 18
3.3 OPERATING SYSTEM 18
3.4 PYTHON 20
3.5 JUPYTER NOTEBOOK 22

4 SYSTEM DESIGN 24

v
 4.1 ARCHITECTURE DESIGN 24
5 PROJECT DESCRIPTION 26
5.1 DATA COLLECTION AND PRE PROCESSING 26
5.2 FEATURE SELECTION 26
5.3 FRAUD CLASSIFICATION 26
5.4 MODEL TRAINING AND OPTIMIZATION 27
5.5 REAL TIME FRAUD DETECTION SYSTEM 27
5.6PERFORMANCEEVALUATIONANDMONITORING 27
5.7 USER INTERFACE AND REPORTING 27

6 SYSTEM TESTING 29
6.1 SYSTEM TESTING 29

7 SYSTEM IMPLEMENTATION 32
7.1 DATASET 32
7.2 FEATURE ENGINEERING 33
7.3 DATA BALANCING 34
7.4 MODEL TRAINING 34
7.5 MODEL EVALUATION 36
8 FUTURE ENHANCEMENT 38
9 CONCLUSION 43
10 BIBILOGRAPHY 45
10.1 JOURNAL REFERENCES 45

vi
 LIST OF FIGURES

FIGURE NO DESCRIPTION PAGE NO

2.1 THE WHOLE FRAMEWORK OF THE METHOD 14

4.1 SYSTEM ARCHITECTURE 25

vii
 LIST OF ABBREVATIONS

Abbreviation Full Form

ML Machine Learning
DL Deep Learning
AI Artificial Intelligence
ANN Artificial Neural Network
CNN Convolutional Neural Network
RNN Recurrent Neural Network
LSTM Long Short-Term Memory
SVM Support Vector Machine
RF Random Forest
DT Decision Tree
KNN K-Nearest Neighbors
PCA Principal Component Analysis
ROC Receiver Operating Characteristic
AUC Area Under the Curve
TPR True Positive Rate

viii
 CHAPTER 1
INTRODUCTION

1.1 OVERVIEW

In this paper, the aim is to build a credit card fraud detection model based
on deep representation learning methods that can learn effective representations
of transaction behaviors. Simultaneously, we hope that our model can have
good stability. For the class imbalance problem, there are many methods to
handle it. This article pays more attention to a better learning representation
that can both enhance the performance of fraud detection and keep the stability
of performance. As mentioned in the literature, a representation learning
method is to learn representations of the data that can easily extract useful
information when building classifiers or other predictors. Representation
learning has been applied widely such as person reidentification and face
recognition.

Our paper suggests an aspect of latest machine learning algorithms to

detect fraudulent or anomalous events commonly called outliers. We have used
a dataset provided by Kaggle; this dataset comprises the transaction records of
the European card holders in the year 2013. Inside the dataset there are 31
columns out of which 30 are used as features and the remaining 1 column is
used as class. Our features include Time, Amount and Number of transactions.

Detection of credit card fraud is an intent part of testing for the

researchers over a long time and will be an interesting part of testing in the
coming time. We are introducing a fraud detection system for credit-cards by
applying three different algorithms and training our machine using these
algorithms with the transaction records we have. The model that we built helps
the authorities to get notified of the fraud in credit-cards and take the further
necessary steps over the transaction and label the transaction as fraud or
1
legitimate transaction. These algorithms show us that the given transaction
tends to be a type of fraud or not, these algorithms were selected using
experimentation, discussion and feature importance techniques as shown in
methodology. It is a real-time transaction data from European credit-card
holders which explains the skewness of data.

1.2 EXISITNG CHALLENGES

Despite significant advancements in fraud detection techniques, several

challenges persist in accurately identifying fraudulent transactions while
minimizing false positives. One of the primary challenges is the imbalance in
datasets, where the number of fraudulent transactions is significantly lower
than legitimate ones. This imbalance leads to biased machine learning models
that tend to favor non-fraudulent cases, making it difficult to detect rare fraud
occurrences effectively. Traditional models often struggle to learn meaningful
patterns from such skewed data, resulting in poor generalization and limited
real-world applicability.

Another major challenge is the evolving nature of fraud. Fraudsters

continuously adapt their techniques, making rule-based and even some machine
learning-based approaches less effective over time. New fraud schemes such as
identity theft, synthetic fraud, and account takeovers require fraud detection
systems to be highly adaptive and capable of learning from emerging threats.
Static models often fail to capture these evolving fraud patterns, necessitating
continuous model updates and retraining.

Additionally, real-time fraud detection poses a critical challenge. Credit

card transactions occur within seconds, requiring fraud detection models to
process vast amounts of data at high speed without compromising accuracy.
Many traditional models are computationally expensive and struggle to provide

2
instant fraud classification, leading to potential delays or missed fraudulent
activities.

Another concern is feature selection and representation. Fraudulent

transactions do not always follow a consistent pattern, making it difficult to
extract discriminative features that clearly differentiate between fraud and
legitimate transactions. Poor feature representation can lead to
misclassification, increasing both false positives (flagging legitimate
transactions as fraud) and false negatives (failing to detect actual fraud).

Lastly, privacy and security concerns also pose challenges in fraud

detection. Financial transaction data is highly sensitive, and ensuring secure
data handling while maintaining compliance with regulatory policies is
essential. Many organizations face difficulties in sharing fraud-related data due
to privacy regulations, which limits the availability of comprehensive datasets
for training robust models. Addressing these challenges requires advanced
methodologies, such as deep learning-based approaches, that can improve fraud
detection accuracy while adapting to new fraud patterns and ensuring real-time
processing.

1.3 SIGNIFICANCE OF RESEARCH

The increasing adoption of credit cards in daily transactions has made

financial operations more convenient, but it has also given rise to significant
security threats in the form of credit card fraud. As cybercriminals employ more
sophisticated techniques, traditional rule-based fraud detection methods have
proven insufficient in effectively distinguishing between legitimate and
fraudulent transactions. The financial impact of credit card fraud is staggering,
with billions of dollars lost annually, making it imperative to develop more
advanced, automated, and intelligent fraud detection systems. This project
leverages machine learning and deep learning techniques to enhance fraud

3
detection by analyzing transaction patterns and identifying fraudulent activities
with higher accuracy. By incorporating Squirrel Optimization and the
Advanced DLMNN classifier, the model aims to improve feature
representation and achieve superior discrimination between fraudulent and
legitimate transactions. This advancement not only enhances the prediction
accuracy but also ensures adaptability and self-learning capabilities, allowing
the system to dynamically evolve with emerging fraud patterns. Furthermore,
an efficient fraud detection system helps financial institutions reduce financial
losses, improve customer trust, and strengthen overall security in the digital
payment ecosystem. The broader significance extends beyond just economic
benefits; it contributes to cybersecurity advancements, minimizes identity theft,
and protects consumers from the distress caused by fraudulent transactions.
Ultimately, this project provides a cutting-edge approach to fraud detection,
leveraging deep learning methodologies to create a more robust and reliable
system for safeguarding financial transactions in an increasingly digital world.

1.4 APPLICATIONS

• Banking and Financial Services

• E-Commerce and Online Retail
• Insurance Fraud Detection
• Cybersecurity and Identity Theft Prevention
• Healthcare Payments
• Government and Regulatory Compliance
• Cryptocurrency Transactions
• Point-of-Sale (POS) Systems
• Travel and Hospitality
• Telecommunication Industry

4
1.5 OBJECTIVE

The primary objective of this project is to develop an efficient and

intelligent credit card fraud detection system using machine learning and deep
learning techniques. The model aims to accurately predict fraudulent
transactions, minimize false positives, and enhance detection accuracy by
leveraging advanced algorithms such as Squirrel Optimization and the
Advanced DLMNN classifier. By improving feature representation and
ensuring adaptability, the system will be capable of identifying emerging fraud
patterns while maintaining high-speed real-time detection. Ultimately, this
project seeks to provide a robust, scalable, and automated solution to safeguard
financial transactions and reduce financial losses caused by credit card fraud.

5
 CHAPTER 2
LITERATURE REVIEW

System analysis is the detailed study of the colourful operations

performed by the system and their connections within and outside the system.
The crucial question in this phase is what all the problems there in the present
system and what must be done to break those problems. The success of the
system depends largely on how easily the problem is defined, completely
delved and duly carried out. The analysis should give the medium of problem
understanding and a framework for its result. Then the being system was
studied completely by the collection of applicable data about the system.

2.1. Transaction fraud detection based on total order relation and behavior
diversity, Zheng, G. Liu, C. Yan, and C. Jiang [2018]:

Proposed a Markov chain models are unsuitable for the representation of

these behaviors. In this paper, they propose logical graph of BP (LGBP) which
is a total order-based model to represent the logical relation of attributes of
transaction records. Based on LGBP and users' transaction records, they can
compute a path-based transition probability from an attribute to another one. At
the same time, they define an information entropy-based diversity coefficient
in order to characterize the diversity of transaction behaviors of a user. In
addition, they define a state transition probability matrix to capture temporal
features of transactions of a user. Consequently, they can construct a BP for
each user and then use it to verify if an incoming transaction is a fraud or not.
Their experiments over a real data set illustrate that their method is better than
three state-of-the-art oneness.

6
2.2 Credit card fraud detection: A realistic modeling and a novel learning
strategy, A. D. Pozzolo, G. Boracchi, O. Caelen, C. Alippi, and G.
Bontempi [2018]:

Achieved three major contributions. First, they propose, with the help of
their industrial partner, a formalization of the fraud-detection problem that
realistically describes the operating conditions of FDSs that everyday analyze
massive streams of credit card transactions. They also illustrate the most
appropriate performance measures to be used for fraud-detection purposes.
Second, they design and assess a novel learning strategy that effectively
addresses class imbalance, concept drift, and verification latency. Third, in their
experiments, they demonstrate the impact of class unbalance and concept drift
in a real-world data stream containing more than 75 million transactions,
authorized over a time window of three years.

2.3. Cost-sensitive learning of deep feature representations from

imbalanced data, S. H. Khan, M. Hayat, M. Bennamoun, F. A. Sohel, and
R. Togneri [2018]:

Proposed an approach applicable to both binary and multi-class

problems without any modification. Moreover, as opposed to data level
approaches, they do not alter the original data distribution which results in a
lower computational cost during the training process. They report the results of
their experiments on six major image classification datasets and show that this
approach significantly outperforms the baseline algorithms. Comparisons with
popular data sampling techniques and cost sensitive classifiers demonstrate the
superior performance of their proposed method.

7
2.4. Representation learning: A review and new perspectives, Y. Bengio,
A. Courville, and P. Vincent [2013]:

Reviewed recent work in the area of unsupervised feature learning and

deep learning, covering advances in probabilistic models, autoencoders,
manifold learning, and deep networks. This motivates longer term unanswered
questions about the appropriate objectives for learning good representations,
for computing representations (i.e., inference), and the geometrical connections
between representation learning, density estimation, and manifold learning.

2.5. Deep representation learning with part loss for person re-
identification, H. Yao, S. Zhang, R. Hong, Y. Zhang, Q. Tian [2019]:

Proposed the discriminative power gaining on unseen person images,

they propose a deep representation learning procedure named part loss network,
to minimize both the empirical classification risk on training person images and
the representation learning risk on unseen person images. The representation
learning risk is evaluated by this part loss, which automatically detects human
body parts and computes the person classification loss on each part separately.
Compared with traditional global classification loss, simultaneously
considering part loss enforces the deep network to learn representations for
different body parts and gain the discriminative power on unseen persons.
Experimental results on three person ReID datasets, i.e., Market1501, CUHK03
and VIPeR, show that their representation outperforms existing deep
representations.

2.6. A light CNN for deep face representation with noisy labels, X. Wu, R.
He, Z. Sun, and T. Tan [2018]:

Introduced a variation of maxout activation, called Max-Feature-Map

(MFM), into each convolutional layer of CNN. Different from maxout
activation that uses many feature maps to linearly approximate an arbitrary
8
convex activation function, MFM does so via a competitive relationship. MFM
can not only separate noisy and informative signals but also play the role of
feature selection between two feature maps. Second, three networks are
carefully designed to obtain better performance meanwhile reducing the
number of parameters and computational costs. Lastly, a semantic
bootstrapping method is proposed to make the prediction of the networks more
consistent with noisy labels.

2.7. A discriminative feature learning approach for deep face recognition,

Y. Wen, K. Zhang, Z. Li, and Y. Qiao [2016]:

Proposed in order to enhance the discriminative power of the deeply

learned features, this paper proposes a new supervision signal, called center
loss, for face recognition task. Specifically, the center loss simultaneously
learns a center for deep features of each class and penalizes the distances
between the deep features and their corresponding class centers. More
importantly, they prove that this center loss function is trainable and easy to
optimize in the CNNs. With the joint supervision of softmax loss and center
loss, they can train a robust CNNs to obtain the deep features with the two key
learning objectives, inter-class dispension and intra-class compactness as much
as possible, which are very essential to face recognition. It is encouraging to
see that their CNNs (with such joint supervision) achieve the state-of-the-art
accuracy on several important face recognition benchmarks.

2.8. Neural fraud detection in credit card operations, Pattern Recognition,

J. Dorronsoro, F. Ginel, C. Sgnchez, and C. Cruz, [2007]:

Presented an online system for fraud detection of credit card operations

based on a neural classifier. Since it is installed in a transactional hub for
operation distribution, and not on a card-issuing institution, it acts solely on the
information of the operation to be rated and of its immediate previous history,

9
and not on historic databases of past cardholder activities. Among the main
characteristics of credit card traffic are the great imbalance between proper and
fraudulent operations, and a great degree of mixing between both. To ensure
proper model construction, a nonlinear version of Fisher's discriminant
analysis, which adequately separates a good proportion of fraudulent operations
away from other closer to normal traffic, has been used.

2.9. Detection of credit card fraud transactions using machine learning

algorithms and neural networks: A comparative study, D. Dighe, S. Patil,
and S. Kokate [2018]:

Proposed the use of online transactions in day-to-day life has been

increasing since last decade due to advancements in technology and network
connectivity. Due to ease, simplicity and user friendliness of the online
transaction system, new users are constantly joining the vast population
benefitting from such system. Credit card fraud resulting from misuse of the
system is defined as theft or misuse of one's credit card information which is
used for personal gains without the permission of the card holder. To detect
such frauds, it is important to check the usage patterns of a user over the past
transactions. Comparing the usage pattern and current transaction, they can
classify it as either fraud or a legitimate transaction. In this paper, the
techniques used are KNN, Naïve Bayes, Logistic Regression, Chebyshev
Functional Link Artificial Neural Network (CFLANN), Multi-Layer
Perceptron and Decision Trees which are evaluated on basis of their result
evaluated in terms of various accuracy metrics.

2.10. Detecting credit card fraud using selected machine learning

algorithms, M. Puh and L. Brkic, [2019]:

Described that due to the immense growth of e-commerce and increased

online based payment possibilities, credit card fraud has become deeply

10
relevant global issue. Recently, there has been major interest for applying
machine learning algorithms as data mining technique for credit card fraud
detection. However, number of challenges appear, such as lack of publicly
available data sets, highly imbalanced class sizes, variant fraudulent behavior
etc. In this paper they compare performance of three machine learning
algorithms: Random Forest, Support Vector Machine and Logistic Regression
in detecting fraud on real-life data containing credit card transactions. To
mitigate imbalanced class sizes, they use SMOTE sampling method. The
problem of ever-changing fraud patterns is considered with employing
incremental learning of selected ML algorithms in experiments. The
performance of the techniques is evaluated based on commonly accepted
metric: precision and recall.

11
 2.1EXISTING SYSTEM

2.1.1 Introduction

In the existing method there are some challenging issues for supervised
learning and unsupervised learning in fraud detection. On the other hand,
machine learning (ML) techniques were employed to predict the suspicious and
non-suspicious transactions automatically by using classifiers. Therefore, the
combination of machine learning and data mining techniques were able to
identify the genuine and non-genuine transactions by learning the patterns of
the data in accurate labeled dataset. The most commonly techniques used fraud
detection method is K-Nearest Neighbor (KNN) Algorithm. This technique can
be used alone or in collaboration using ensemble or meta-learning techniques
to build classifiers.

Credit card fraud detection is a typical classification problem where the

objective is to classify a transaction as either fraudulent or legitimate. KNN
algorithm can be used for this purpose by training a model on a dataset of past
transactions that have been labeled as either fraudulent or legitimate. Then, the
model can be used to predict the label of new transactions based on their
features such as the amount, location, and time of the transaction.

2.1.2Theoretical Background

K-Nearest Neighbor (KNN) Algorithm:

The K-Nearest Neighbor (KNN) algorithm is a machine learning

algorithm that can be used for classification and regression problems. It is based
on the idea that similar instances tend to belong to the same class, and thus, it
determines the class of a new instance by finding the K closest instances to it
and taking a majority vote of their classes.

12
2.1.3 Methodology

Prepare the data:

The first step is to prepare the data by cleaning it, normalizing it, and
splitting it into training and testing datasets. The training dataset is used to train
the KNN model, while the testing dataset is used to evaluate its performance.

Choose the value of K:

The next step is to choose the value of K, which determines the number
of neighbors that are considered in the classification. A small value of K can
lead to overfitting, while a large value can lead to underfitting. A common
approach is to use cross-validation to choose the optimal value of K that
maximizes the accuracy of the model.

Compute distances:

For each new transaction in the testing dataset, the distances to all
transactions in the training dataset are computed using a distance metric such
as Euclidean distance or Manhattan distance.

Select K nearest neighbors:

The K transactions with the smallest distances are selected as the nearest
neighbors of the new transaction.

Classify the transaction:

The label of the new transaction is determined by taking a majority vote

of the labels of its K nearest neighbors. If the majority of the neighbors are
labeled as fraudulent, the new transaction is classified as fraudulent, otherwise,
it is classified as legitimate.

13
Evaluate the performance:

The performance of the KNN algorithm is evaluated by computing

metrics such as accuracy, precision, recall, and F1 score on the testing dataset.
These metrics can be used to compare the performance of different values of K
or to compare the KNN algorithm with other machine learning algorithms.

The advantages of using KNN algorithm for credit card fraud detection
are: It is a simple and intuitive algorithm that does not require extensive
training or parameter tuning. It can be effective in detecting fraud patterns that
are similar to past frauds and have been previously labeled. It can handle
imbalanced datasets where the number of fraudulent transactions is much
smaller than the number of legitimate transactions.

Figure 2.1: The whole framework of the method

2.3 DRAWBACK

This system does not perform very well when the data set has more noise
i.e., target classes are overlapping.

14
 This method may require a large amount of historical transaction data to
learn the user behavior and detect abnormal patterns, which can be a challenge
in situations where data is scarce or incomplete.

The performance of this method may be affected by the choice of

parameters and the specific implementation details, which can require
extensive tuning and experimentation to achieve optimal results.

Another limitation of the system is its potential difficulty in adapting to

evolving fraud tactics. Fraudsters continually change their methods to evade
detection, which means the model must be frequently retrained with updated
data to remain effective. Without ongoing updates, the system may fail to
recognize new patterns of fraudulent behavior, leading to decreased detection
accuracy over time.

Moreover, computational complexity can be a concern, especially when

dealing with large-scale transaction datasets. The need for extensive parameter
tuning and processing vast amounts of data can result in high training times and
resource consumption. This may limit the system’s applicability in real-time or
resource-constrained environments where quick decision-making is crucial.

2.4 PROPOSED SYSTEM

In addition, most companies and institutions now tend to move their

business toward online services due to the rapid increase of using modern
technology in all fields. Thus, online transactions (e.g., booking a hotel) require
a customer to have a credit card to access the services and complete the
transaction in such an efficient way that it might be hard and time-consuming
to perform while using cash payment. However, a credit card is susceptible to
cybercriminals causing credit card fraud. The fraudsters perform fraudulent
activities by making unauthorized access to credit card information and such

15
activities cause a financial loss for both company and customer. Thus, the
challenges of fraudulent activities increased the demand for systems to detect
credit card fraud. The researchers try to build fraud detection systems using
machine learning, deep learning, and data mining techniques to detect the
transaction whether it is fraudulent transactions or genuine based on datasets
that include information about the transactions. However, credit card fraud
detection is becoming more complex since the fraudulent transactions for the
cards are more and more like legal ones.

This paper is proposed to tackle the problem of credit card fraud using
machine learning and deep learning models performed on the Fraud Detection
dataset provided by Kaggle. The main contribution of this work is to develop a
fraud detection model using deep learning modified neural network (DLMNN)
classifier. Finally, evaluate the performance the proposed system with existing
methods. Model Evaluation Metrics like Precision, Recall, F1-Score are used
for performance evaluation.

The proposed techniques focus to detect the Credit Card Fraudulent on

transactions whether it is a genuine or a fraud transaction and the approaches
used to separate fraud and non-fraud are Squirrel Optimization and adaptive
DLMNN classifier and finally we will determine which approach is best for
detecting credit card frauds. The proposed DLMNN classifier algorithm detects
the fraud in credit card system. To evaluate the algorithms, 80% of the dataset
is used for training and 20% is used for testing. Accuracy, F1-score, precision,
and recall score are used to evaluate the performance this approach.

For predicting these transactions banks make use of DLMNN Classifier,

past data has been collected and new features are been used for enhancing the
predictive power and Squirrel optimization technique to rank the features based
on the feature importance. The performance of fraud detecting in credit card

16
transactions are greatly affected by the sampling approach on data-set, selection
of variables and detection techniques used. Dataset of credit card transactions
is collected from kaggle and it contains a total of 2,84,807 credit card
transactions of a European bank data set. It considers fraud transactions as the
“positive class” and genuine ones as the “negative class”.

17
 CHAPTER 3
SYSTEM SPECIFICATIONS

3.1HARDWARE REQUIREMENTS

System : Intel i3 3.5 GHz to Latest Version.

Hard Disk : 40 GB.

RAM : 1 GB
Monitor : 14’ Color Monitor.
Mouse : Optical Mouse.
Keyboard : 101 Keys

3.2 SOFTWARE REQUIREMENTS

Operating system : Windows 10

IDE : Jupyter Notebook
Tools used : Python Programming

3.3 OPERATING SYSTEM

Windows 11:

Windows 11 is the rearmost major release of Microsoft's Windows NT

operating system, released in October 2021. It's a free upgrade to its precursor,
Windows 10 ( 2015), and is available for any Windows 10 bias that meet the
new Windows 11 system conditions.

Windows 11 features major changes to the Windows shell told by the

cancelled Windows 10X, including a redesigned launch menu, the relief of its"
live penstocks" with a separate" contraptions" panel on the taskbar, the
capability to produce tiled sets of windows that can be minimized and restored

18
from the taskbar as a group, and new gaming technologies inherited from Xbox
Series X and Series S similar as Auto HDR and Direct Storage on compatible
tackle. Internet Discoverer ( IE) has been replaced by the Chromium- grounded
Microsoft Edge as the dereliction web cyber surfer, like its precursor,
Windows 10, and Microsoft brigades is integrated into the Windows shell.
Microsoft also blazoned plans to allow further inflexibility in software that can
be distributed via the Microsoft Store and to support Android apps on Windows
11(including a cooperation with Amazon to make its app store available for the
function).

Citing security considerations, the system conditions for Windows 11

were increased over Windows 10. Microsoft only officially supports the
operating system on bias using an eighth- generation Intel Core CPU or newer
(with some minor exceptions), a alternate- generation AMD Ryzen CPU or
newer, or a Qualcomm Snapdragon 850 ARM system- on- chip or newer, with
UEFI secure charge and Trusted Platform Module( TPM)2.0 supported and
enabled( although Microsoft may give exceptions to the TPM2.0 demand for
OEMs). While the OS can be installed on unsubstantiated processors, Microsoft
doesn't guarantee the vacuity of updates. Windows 11 removed support for
32- bit x86 CPUs and bias that use BIOS firmware.

Windows 11 entered a mixed event at launch. Pre-release content of the

operating system concentrated on its stricter tackle conditions, with
conversations over whether they were legitimately intended to ameliorate the
security of Windows or as a ploy to upsell guests to newer bias and over thee-
waste associated with the changes. Upon release, it was praised for its bettered
visual design, window operation, and stronger focus on security, but was
blamed for colourful variations to aspects of its stoner interface that were seen
as worse than its precursor, as an attempt to inhibit druggies from switching
to contending operations.
19
Windows 11 security

Windows 11 comes with cutting-edge features that help protect you from
malware. While staying vigilant is the most important protective measure you
can take, security features in Windows 11 also help provide real-time detection
and protection.

Its’ve innovated solutions that redefine log-in credentials. Windows 11

validates your credentials using either a device-specific PIN code, fingerprint
or facial recognition protecting you from phishing and other network attacks,
including password leaks.

Windows 11 protects your most valuable information in multiple ways.

In addition to the built-in protections that help keep you from downloading
suspicious or potentially unwanted apps, Windows 11 also comes with a suite
of Microsoft-developed apps that help keep you protected, both online and off.

3.4 PYTHON

Python is a general purpose, dynamic, high position and interpreted

programming language. It supports an Object acquainted programming
approach to develop operations. It's simple and easy to learn and provides lots
of high- position data structures.

Python is easy to learn yet important and protean scripting language

which makes it seductive for Application Development.

Python's syntax and dynamic codifying with its interpreted nature, makes
it an ideal language for scripting and rapid-fire operation development.

Python supports multiple programming pattern, including object

acquainted, imperative and functional or procedural programming styles.

20
 Python isn't intended to work on special areas similar as web
programming. That's why it's known as multipurpose because it can be used
with web, enterprise, 3D CAD etc.

We do not need to use data types to declare variable because it's stoutly
compartmented so it can be written as a = 10 to assign an integer value in an
integer variable.

Python makes the development and debugging fast because there's no

compendium step included in python development and edit- test- debug cycle
is veritably fast.

Python Features

Python provides lots of features that are listed below.

1) Easy to Learn and Use

Python is easy to learn and use. It is developer-friendly and high-level

programming language.

2) Expressive Language

Python language is more expressive means that it is more understandable

and readable.

3) Interpreted Language

Python is an interpreted language i.e.; interpreter executes the code line

by line at a time. This makes debugging easy and thus suitable for beginners.

4) Cross-platform Language

Python can run equally on different platforms such as Windows, Linux,

Unix and Macintosh etc. So, we can say that Python is a portable language.

21
5) Free and Open Source

Python language is freely available at official web address. The source-

code is also available. Therefore, it is open source.

6) Object-Oriented Language

Python supports object-oriented language and concepts of classes and

objects come into existence.

7) Extensible

It implies that other languages such as C/C++ can be used to compile the
code and thus it can be used further in our python code.

8) Large Standard Library

Python has a large and broad library and provides rich set of module and
functions for rapid application development.

9) GUI Programming Support

Graphical user interfaces can be developed using Python.

10) Integrated

It can be easily integrated with languages like C, C++, and JAVA etc.

3.5 JUPYTER NOTEBOOK

Project Jupyter is a design to develop open- source software, open norms,

and services for interactive computing across multiple programming languages.
It was spun off from IPython in 2014 by Fernando Pérez and Brian Granger.
Project Jupyter's name is a reference to the three core programming languages
supported by Jupyter, which are Julia, Python and R. Its name and totem are an
homage to Galileo's discovery of the moons of Jupiter, as proved in scrapbooks
attributed to Galileo. Project Jupyter has developed and supported the

22
interactive computing products Jupyter Notebook, JupyterHub, and
JupyterLab. Jupyter is financially patronized by NumFOCUS.

The first interpretation of Notebooks for IPython was released in 2011

by a platoon including Fernando Pérez, Brian Granger, and Min Ragan- Kelley.
In 2014, Pérez blazoned a spin- off design from IPython called Project Jupyter.
IPython continues to live as a Python shell and a kernel for Jupyter, while the
tablet and other language- agnostic corridor of IPython moved under the
Jupyter name. Jupyter supports prosecution surroundings (called “kernels") in
several dozen languages, including Julia, R, Haskell, Ruby, and Python(via the
IPython kernel).

In 2015, about 2, 00,000 Jupyter notebooks were available on GitHub.

By 2018, about2.5 million were available. In January 2021, nearly 10 million
were available, including notebooks about the first observation of gravitational
swells and about the 2019 discovery of a supermassive black hole.

Major cloud computing providers have espoused the Jupyter Tablet or

outgrowth tools as a frontend interface for pall druggies. Exemplifications
include Amazon SageMaker Notebooks, Google's Colaboratory, and
Microsoft's Azure Tablet.

Visual Studio Code supports original development of Jupyter notebooks.

As of July 2022, the Jupyter extension for VS Code has been downloaded over
40 million times, making it the second-most popular extension in the VS Code
Marketplace.

The Atlantic published a composition entitled" The Scientific Paper Is

Obsolete" in 2018, agitating the part of Jupyter Notebook and the Mathematica
tablet in the future of scientific publishing.

23
 CHAPTER 4

SYSTEM DESIGN

4.1 ARCHITECTURAL DESIGN

4.1.1 Introduction

Architectural design is about putrefying the system into interacting

factors. It's expressed as a block illustration defining an overview of the system
structure, features of the factors, and how these factors communicate with each
other to partake data. It identifies the factors that are necessary for developing
a computer- grounded system and communication between them i.e.,
relationship between these factors. It defines the structure and parcels of the
factors that are involved in the system and also the non intercourses between
these factors. The architectural design process is about relating the factors i.e.
subsystems that makeup the system and structure of the sub-system and they’re
interaction. It's an early stage of the system design phase. It acts as a link
between specification conditions and the design process.

Architectural design is crucial because it lays the foundation for the

entire software development process. A well-planned architecture helps in
managing system complexity by dividing it into smaller, manageable
components. It ensures that the system fulfills both functional and non-
functional requirements such as performance, scalability, security, and
maintainability. By organizing the system into clear modules with defined
responsibilities, it becomes easier to understand, test, and extend in the future.

24
 Min-max
Scalar
Data Data Transformation Squirrel
Correlation Optimization
collection Cleaning
Data
Normalization
Training Feature Selection
Phase
Pre processing

DLMNN
Based
Prediction
Model
Testing Training
Phase
Min-max Scalar
Data Transformation
Input Data
normalization

Pre processing

Fig 4.1 System Architecture

There are different architectural styles that can be applied depending on

the system's goals and constraints. These include layered architecture, client-
server architecture, microservices, service-oriented architecture (SOA), and
event-driven models. Each of these has specific benefits and trade-offs. For
example, layered architecture supports modular development and easier testing,
while microservices allow for independent deployment and scaling of
components, making them ideal for large-scale or dynamic systems.

Architectural design also supports effective collaboration among

development teams. By defining clear interfaces and responsibilities for each
component, teams can work in parallel with minimal interference. Additionally,
the architecture serves as a vital documentation artifact that guides developers
throughout the project lifecycle. It provides a reference for design decisions,
helps onboard new team members, and ensures that changes to the system can
be implemented in a structured and consistent manner
25
 CHAPTER 5
PROJECT DESCRIPTION

5.1 DATA COLLECTION AND PREPROCESSING

This module is responsible for gathering transactional data from

financial institutions, credit card providers, or publicly available datasets. Since
credit card fraud detection involves highly imbalanced data, this module
applies techniques such as oversampling, undersampling, or Synthetic Minority
Over-sampling Technique (SMOTE) to balance the dataset. Additionally, it
performs data cleaning, handling missing values, and normalizing transaction
features to prepare the data for model training.

5.2 FEATURE SELECTION

Feature selection plays a crucial role in improving fraud detection

accuracy. The Squirrel Optimization Algorithm (SOA) is employed to identify
the most relevant features that contribute to distinguishing fraudulent
transactions from legitimate ones. By reducing redundant and irrelevant
features, this module enhances the efficiency of the model, reduces
computation time, and prevents overfitting.

5.3 FRAUD CLASSIFICATION

This module leverages the Advanced Deep Learning Metric Neural

Network (DLMNN) classifier to categorize transactions as fraudulent or
legitimate. DLMNN focuses on extracting deep feature representations from
transaction data, allowing for better discrimination between different classes. It
utilizes a specialized loss function that improves the model’s ability to learn
patterns and relationships in high-dimensional data, resulting in enhanced
accuracy and lower false-positive rates.

26
 5.4 MODEL TRAINING AND OPTIMIZATION

Once the feature selection and classification model are defined, the
system undergoes extensive training using historical transactional data. This
module optimizes the model using techniques such as backpropagation,
dropout regularization, and hyperparameter tuning. Additionally, the model is
continuously updated with new fraud patterns to improve adaptability and
maintain high performance over time.

5.5 REAL-TIME FRAUD DETECTION SYSTEM

This module is responsible for deploying the trained model in a real-

world setting to monitor transactions in real time. It processes incoming
transactions, applies the trained classification model, and instantly flags
suspicious activities. The system is integrated with financial platforms to
provide real-time alerts, enabling swift action to prevent fraudulent
transactions.

5.6 PERFORMANCE EVALUATION AND MONITORING

To ensure the effectiveness of the fraud detection system, this module

evaluates the model using various performance metrics such as accuracy,
precision, recall, F1-score, and Area Under the Curve (AUC). The system also
incorporates monitoring mechanisms to track false positives and false
negatives, allowing for further refinement of the model as fraud techniques
evolve.

5.7 USER INTERFACE AND REPORTING

A user-friendly interface is developed to provide financial analysts, fraud

investigators, and businesses with insights into detected fraudulent
transactions. This module includes visualization tools, detailed transaction
27
reports, and automated alerts to enhance decision-making and fraud prevention
strategies.

By integrating these modules, the proposed credit card fraud detection

system aims to provide a comprehensive, automated, and highly accurate
solution to combat financial fraud while ensuring real-time processing and
adaptability to emerging fraud patterns.

28
 CHAPTER 6

SYSTEM TESTING

INTRODUCTION

The purpose of testing is to discover errors. Testing is the process of

trying to discover every conceivable fault or weakness in a work product. It
provides a way to check the functionality of factors, subassemblies, assemblies
and or a finished product it’s the process of exercising software with the intent
of icing that the Software system meets its conditions and stoner prospects and
doesn't fail in an inferior manner. There are colourful types of tests. Each test
type addresses a specific testing demand.

6.1 SYSTEM TESTING

In this section I perform analysis and experiments to validate the

contribution of each component of the network. It is critical software quality
assurance method for discover errors. For my intelligent machine fault
diagnose system, system testing will involve testing the entire system as a
whole, rather than testing individual modules separately.

In this section, I perform the system testing for each class in the dataset
acquired from Kaggle. This testing can be done by providing input dataset
containing two-day transactions made on 09/2013 by European cardholders.
The dataset contains 492 frauds out of 284,807 transactions. Thus, it is highly
unbalanced, with the positive (frauds) accounting for only 0.17%.

This testing can be done by providing validation split:

Split train test

1. Nominal - 5%

2. Outer Race - 5%
29
 3. Inner Race - 5%

Validation split

25% of Train set.

Here, I tested the proposed approach by varying different parameters. I

altered the various parameters of the system that outweighed the verification.
Such measurement shows the performance of the proposed approach in
different scenarios. The effect of these various factors on the proposed
approach in terms of their accuracy.

Cross-ValidationTesting
To ensure the model's performance is not biased due to the specific train-
test split, k-fold cross-validation was applied. This technique divides the dataset
into k subsets and iteratively trains and tests the model on different
combinations, providing a more reliable estimate of generalization accuracy.

HyperparameterSensitivityTesting
Different hyperparameters such as learning rate, number of layers, number
of neurons, and activation functions were varied to observe their effects on
model accuracy. This testing helped identify optimal hyperparameter
configurations that enhance model performance without overfitting.

NoiseRobustnessTesting
To simulate real-world conditions, random noise was introduced into the
dataset. This test evaluated how well the proposed approach could maintain
accuracy when presented with noisy or imperfect input data, indicating its
reliability in practical applications.

ScalabilityTesting
The system was tested using datasets of increasing size to assess how well

30
it scales. Performance metrics like training time, memory usage, and accuracy
were monitored to understand how the model behaves with larger volumes of
data.

ClassImbalanceTesting
To determine the system's robustness in handling imbalanced datasets,
experiments were conducted where certain classes had significantly fewer
samples than others. Techniques such as oversampling, undersampling, and
weighted loss functions were used to mitigate imbalance and evaluate the
system’s fairness and precision.

31
 CHAPTER 7

SYSTEM IMPLEMENTATION

This project is implemented using python software. Anaconda

software helps you create an environment for many different versions of Python
and package versions. Anaconda Navigator is a GUI tool that is included in the
Anaconda distribution and makes it easy to configure, install, and launch tools
such as Jupyter Notebook. A Conda Python environment is an isolated
environment. It allows you to install packages without modifying your system's
Python installation. Anaconda software helps you create an environment for
many different versions of Python and package versions. Anaconda is also used
to install, remove, and upgrade packages in the project environments.

The implementation of the credit card fraud detection system involves a

structured pipeline, starting with data collection from financial transaction
records. The data is preprocessed to handle missing values, normalize features,
and balance class distribution using oversampling or undersampling
techniques. Feature selection is performed using the Squirrel Optimization
Algorithm (SOA) to extract the most relevant features for fraud classification,
reducing redundancy and improving model efficiency. The optimized feature
set is then fed into the Advanced Deep Learning Metric Neural Network
(DLMNN) classifier, which learns deep feature representations to differentiate
between fraudulent and legitimate transactions. The model undergoes training
and optimization using historical transaction data, applying techniques such as
hyperparameter tuning and regularization to enhance accuracy. Once trained,
the model is deployed in a real-time fraud detection system that continuously
monitors incoming transactions, flagging suspicious activities and generating
alerts for financial institutions. A user-friendly interface is developed for

32
visualizing fraud reports, monitoring system performance, and assisting fraud
analysts in decision-making. The system is further evaluated using key
performance metrics, ensuring its reliability, adaptability to evolving fraud
patterns, and effectiveness in minimizing financial losses due to fraud.

7.1 DATASET

The dataset is the Kaggle Credit Card Fraud Detection dataset here. It
contains two-day transactions made on 09/2013 by European cardholders. The
dataset contains 492 frauds out of 284,807 transactions. Thus, it is highly
unbalanced, with the positive (frauds) accounting for only 0.17%.

Due to the extreme imbalance in the dataset, special attention must be

paid to evaluation metrics and validation strategies. Traditional accuracy can
be misleading because the model might simply predict the majority class (non-
fraud) and achieve high accuracy without effectively detecting fraud.
Therefore, metrics such as precision, recall, F1-score, and the area under the
ROC curve (AUC-ROC) are more informative for assessing model
performance. Additionally, techniques like stratified sampling during data
splits and the use of synthetic data generation methods (e.g., SMOTE) help
ensure that the model is properly trained to recognize the minority class. This
careful approach improves the reliability and robustness of fraud detection in
real-world applications.

7.2 FEATURE ENGINEERING

The first step is feature engineering that is aiming at extracting

informative features of users’ transaction behaviors. The raw features, such as
transaction time/date and transaction amount, cannot well characterize the
transaction behaviors of cardholders and fraudsters. One of the commonly used
methods is to derive some new features using the transaction aggregation
strategy. The aggregation features are derived through grouping the
33
transactions according to a selected time interval, card number, transaction
type, and merchant code. Then, the number of transactions and the total amount
spent on those transactions are calculated. After the process of transaction
aggregation strategy, a single transaction with raw features is transformed into
a feature matrix with more informative aggregation features.

7.3 DATA BALANCING

After feature engineering, a classifier can be trained as a binary

classification task. However, if the class imbalance problem is not considered,
the learned classifier will tend to identify most of the fraud transactions as
genuine ones. The reason is that almost all classifiers have a default assumption
of a balanced data set, and thus, the learned decision boundary tends to bias
toward the class with more samples. Hence, dealing with the class imbalance
problem has become an indispensable step before training a fraud detection
model. The most commonly used method of handling the class imbalance
problem is data sampling. Especially, the under-sampling method can reduce
the redundancy of genuine transactions and speed up the model training.
Randomly under sampling is one of the most famous under sampling methods
due to its simplicity and effectiveness. However, these sampling methods do
not consider the spatial distribution of instances from different classes. The
Gaussian mixture under sampling method can be applied to sample more
informative instances and, thus, improve the performance of classifiers.

The undersampling method removes majority class observations until we

have the same number of instances in the two classes.

7.4 MODEL TRAINING

A fraud transaction detection model, as a binary classifier, can be trained

with a relatively balanced data set after handling the class unbalance problem.
The Deep Learning Modified Neural Network (DLMNN) Classifier algorithm
34
is successfully used to detect fraud transactions. Almost it belongs to
representation learning. It aims at discovering better representations of inputs
by learning transformations of data that disentangle factors of variation in data
and retain most of the information. Especially, the deep representation learning
with deep neural networks has achieved remarkable success in many domains
in recent years due to some advanced structures. The latest neural network
architectures make a deep representation learning model not only deeper with
much more layers but also easier for model training. These advanced
architectures significantly enhance the ability of complex nonlinear mapping
of a deep representation learning model. On the other hand, the ingeniously
designed loss functions can supervise the process of deep representation so that
the final model can obtain an ideal result.

Basically, an DLMNN is a mathematical model of a process developed

empirically rather than using mass and energy balances around the process. A
neural network consists of a network of partially connected processing
elements or nodes, arranged in layers. They further include interconnections
between the nodes of successive layers. A schematic configuration of the basic
structure of a single neuron or a node within a neural network. with inputs, an
activation function and a single output. The connections between nodes are
calculated values called weights. The weights represent the “strength” of
connection between neurons, Yi is the output.

Each neuron in the hidden layer receives weighted input plus bias from
each neuron in the previous layer.

35
 j −1
Where XK denotes the input from k-th node in j-th layer, Wk,j is the
weight of the link between node k and all the nodes in the previous layers, and
bi is the bias to the node ,Nj-1 is the number of nodes in the layer j-1.

This sum is passed along to an activation function, to produce the output

of the node, calculate as: Yi=f(Zi). The function is the most commonly used
activation defined as:

The activation function serves to model nonlinear behaviour.

7.5 MODEL EVALUATION

To measure how the proposed model performs, we used different

metrics. Since the dataset used in this paper was highly imbalanced, using the
accuracy metric alone will not be accurate to measure the performance of the
model. Different evaluation metrics were used to evaluate our work in this
dataset like Confusion Matrix, Area Under the Receiver Operating
Characteristic Curve (AUC) and F1-Measure.

These evaluation metrics collectively provide a deeper understanding of

the model’s performance, especially in the context of fraud detection where
correctly identifying rare positive cases is critical. The Confusion Matrix helps
identify the types of errors the model makes, such as false negatives, which
represent missed fraud cases and can have serious consequences. The AUC-
ROC metric evaluates the model’s discriminative power regardless of the
classification threshold, making it useful for comparing different models.
Meanwhile, the F1-Measure offers a balanced view by considering both
precision (how many detected frauds are actually frauds) and recall (how many
actual frauds were detected), which is essential when the cost of false negatives

36
and false positives differ significantly. Using these metrics together ensures that the
proposed approach is not only accurate but also reliable and effective in real-world scenarios.

For instance, examining how performance changes when the threshold

for classification is adjusted can help fine-tune the balance between detecting
frauds and minimizing false alarms. Moreover, testing the model on different
subsets of the data, such as during peak transaction times or on new unseen
data, can reveal its generalization capabilities. This comprehensive evaluation
approach ensures that the proposed model maintains high performance not only
in controlled experiments but also in practical, real-world environments where
fraud patterns may evolve over time.

37
 CHAPTER 8
FUTURE EHANCEMENT

Future enhancements in credit card fraud detection can further improve

the efficacy and efficiency of fraud prevention measures, leveraging
advancements in technology, data analytics, and cybersecurity. Integration of
more advanced machine learning techniques such as deep learning, ensemble
methods, and anomaly detection algorithms can enhance the accuracy and
robustness of fraud detection models. Deep learning architectures can
effectively capture complex patterns and temporal dependencies in transaction
data, leading to more precise fraud detection. Real-time Transaction
Monitoring: Enhancing real-time transaction monitoring capabilities by
leveraging streaming analytics and in-memory processing technologies.
Implementing predictive analytics models that continuously analyze
transaction data as it flows through the system, enabling immediate detection
and prevention of fraudulent activities.

The integration of behavioral biometrics offers a promising enhancement

to credit card fraud detection by analyzing unique user patterns such as typing
speed, mouse movements, touchscreen behavior, and device usage habits.
These behavioral indicators enable continuous user authentication, providing
an added layer of security beyond traditional credentials. When combined with
contextual data like geolocation, device fingerprinting, and IP addresses,
systems can detect subtle anomalies that might indicate fraudulent behavior.
This multi-dimensional approach improves the precision of fraud detection
models, especially in scenarios where conventional methods may fail to
identify sophisticated attacks.

38
 Another advancement lies in the application of blockchain technology,
which can offer tamper-proof and transparent records of transactions through
decentralized ledgers. This can enhance the integrity of transaction data and
reduce the chances of manipulation by malicious actors. Additionally, the
adoption of Explainable AI (XAI) allows fraud detection systems to produce
interpretable decisions, fostering greater trust and facilitating regulatory
compliance. As data privacy laws become more stringent, implementing
privacy-preserving methods such as differential privacy and federated learning
will be essential. These technologies enable collaborative model training across
institutions without exposing sensitive customer data, thereby strengthening
fraud prevention while maintaining user privacy.

39
OUTPUT:

40
41
42
 CHAPTER 9
CONCLUSION

Credit card fraud is one of the most important problems that financial
institutions are currently facing. In this paper, a deep representation learning
model is proposed for credit card fraud detection that has the advantage to
achieve a good and stable performance. This paper uses a deep learning model
to detect whether an online transaction is legitimate or fraud. This paper uses
deep learning techniques like, Squirrel Optimization and Advance DLMNN
classifier for the detection of credit card fraudulent activities. For this work
Jupyter Notebook tool is used. The Kaggle dataset for credit card transactions
is used in this paper. This work achieves maximum accuracy of 99.66%.
Although the proposed method obtains good results on small set data, there are
still some problems such as imbalanced data. The future work will focus on
solving these problems and improving the algorithm.

The integration of the Squirrel Optimization Algorithm with the

DLMNN classifier plays a crucial role in enhancing the model’s learning
efficiency and classification accuracy. The Squirrel Optimization technique is
inspired by the foraging behavior of flying squirrels and is effective in
optimizing complex nonlinear functions, which is particularly beneficial for
high-dimensional data like credit card transactions. By embedding this
optimization strategy into the training phase of the neural network, the model
not only converges faster but also avoids common pitfalls such as local minima,
leading to more reliable fraud detection results.

In comparison with traditional machine learning algorithms such as

decision trees, support vector machines (SVM), and logistic regression, the
proposed deep learning-based method demonstrates superior performance in

43
terms of precision, recall, and overall accuracy. However, to enhance its
practicality, future work should explore the deployment of the model in real-
time transaction environments and test it against dynamic, evolving fraud
patterns. Incorporating real-time data streams, developing adaptive learning
mechanisms, and employing more extensive datasets with class-balancing
strategies will further improve the model’s applicability and resilience in real-
world scenarios.

44
 CHAPTER 10
BIBLIOGRAPHY

Bibliography is usually a physical academic study of books, cultural

facilities; in this sense it is also known as a bibliography. Carter and Barker
describe bibliography as a two-fold scientific discipline—an ordered list of
books and a systematic description of books as objects.

10.1 JOURNAL REFERENCES

1. L. Zheng, G. Liu, C. Yan, and C. Jiang, “Transaction fraud detection

based on total order relation and behavior diversity,” IEEE Trans.
Comput. Soc. Syst., vol. 5, no. 3, pp. 796–806, Sep. 2018.
2. A. D. Pozzolo, G. Boracchi, O. Caelen, C. Alippi, and G. Bontempi,
“Credit card fraud detection: A realistic modeling and a novel learning
strategy,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 8, pp.
3784–3797, Sep. 2018.
3. S. H. Khan, M. Hayat, M. Bennamoun, F. A. Sohel, and R. Togneri,
“Cost-sensitive learning of deep feature representations from imbalanced
data,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 8, pp. 3573–
3587, Aug. 2018.
4. Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A
review and new perspectives,” IEEE Trans. Pattern Anal. Mach. Intell.,
vol. 35, no. 8, pp. 1798–1828, Aug. 2013.
5. H. Yao, S. Zhang, R. Hong, Y. Zhang, C. Xu, and Q. Tian, “Deep
representation learning with part loss for person re-identification,” IEEE
Trans. Image Process., vol. 28, no. 6, pp. 2860–2871, Jun. 2019.
6. X. Wu, R. He, Z. Sun, and T. Tan, “A light CNN for deep face
representation with noisy labels,” IEEE Trans. Inf. Forensics Security,
vol. 13, no. 11, pp. 2884–2896, Nov. 2018.

45
7. Y. Wen, K. Zhang, Z. Li, and Y. Qiao, “A discriminative feature learning
approach for deep face recognition,” in Proc. Eur. Conf. Comput. Vis.
(ECCV). Cham, Switzerland: Springer, 2016, pp. 499–515.
8. J. Dorronsoro, F. Ginel, C. Sgnchez, and C. Cruz, “Neural fraud
detection in credit card operations,” IEEE Trans. Neural Netw., vol. 8,
no. 4, pp. 827–834, Jul. 1997.
9. D. Dighe, S. Patil, and S. Kokate, “Detection of credit card fraud
transactions using machine learning algorithms and neural networks: A
comparative study,” in 2018 Fourth International Conference on
Computing Communication Control and Automation (ICCUBEA).
IEEE, 2018, pp. 1–6.
10.M. Puh and L. Brkic´, “Detecting credit card fraud using selected
machine learning algorithms,” in 2019 42nd International Convention on
Information and Communication Technology, Electronics and
Microelectronics (MIPRO). IEEE, 2019, pp. 1250–1255.

46