See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/331396279

An Efficient Real Time Model For Credit Card Fraud Detection Based On Deep
Learning

Conference Paper · October 2018

DOI: 10.1145/3289402.3289530

CITATIONS READS
50 6,032

3 authors:

Youness Abakarim Mohamed Lahby

École normale supérieure de Casablanca, Morocco, Casablanca Université Hassan II de Casablanca
6 PUBLICATIONS 84 CITATIONS 65 PUBLICATIONS 799 CITATIONS

SEE PROFILE SEE PROFILE

Abdelbaki Attioui
ENS- Université Hassan II de Casablanca Morocco
18 PUBLICATIONS 113 CITATIONS

SEE PROFILE

All content following this page was uploaded by Mohamed Lahby on 16 July 2019.

The user has requested enhancement of the downloaded file.

 An Efficient Real Time Model For Credit Card Fraud Detection
Based On Deep Learning
Youness Abakarim, Mohamed Lahby, Abdelbaki Attioui
Laboratory of Mathematics and Applications, University Hassan II of Casablanca, Ecole Normale Superieure (ENS)
Casablanca, Morocco
[email protected],[email protected],[email protected]

ABSTRACT institutions are faced with the challenge to build effective and
In the last decades Machine Learning achieved notable results in proactive fraud detection systems. Machine learning represents a
various areas of data processing and classification, which made the promising solution to deal with this problem, and this by using the
creation of real-time interactive and intelligent systems possible. gathered historical customers data and their real-time transaction
The accuracy and precision of those systems depends not only on details [3].
the correctness of the data, logically and chronologically, but also In banking and financial sectors, machine learning is used ac-
on the time the feed-backs are produced. This paper focuses on tively today for different applications, notably in portfolio man-
one of these systems which is a fraud detection system. In order agement, trading, risk analysis, prevention and fraud detection . In
to have a more accurate and precise fraud detection system, banks the financial landscape, for example, Machine Learning is used to
and financial institutions are investing more and more today in build Chatbots, an artificial intelligence software that can interact
perfecting the algorithms and data analysis technologies used to with the customers and respond to there queries. In trading, Deci-
identify and combat fraud. Therefore, many solutions and algo- sion Trading Support Systems or Algorithmic Trading, is used to
rithms using machine learning have been proposed in literature to make extremely fast decisions [4]. Moreover, one of the primary
deal with this issue. However, comparison studies exploring Deep use of machine learning in the banking industry is the protection
learning paradigms are scarce, and to our knowledge, the proposed against fraud. With the help of ML algorithms, detecting suspicious
works don’t consider the importance of a Real-time approach for activities became and easier task. Based on the transactions history,
this type of problems. Thus, to cope with this problem we propose machine learning showed promising new methods to analyze the
a live credit card fraud detection system based on a deep neural net- behavior of users and detect if there is a fraud or not [12] [14] [15].
work technology. Our proposed model is based on an auto-encoder In this work, we are aiming our attention on one aspects of the
and it permits to classify, in real-time, credit card transactions as latter : Credit Fraud detection.
legitimate or fraudulent. To test the effectiveness of our model, four In a McKinsey article published in 2017 [5] Deep Learning is
different binary classification models are used as a comparison. The presented as a very promising solution to deal with fraud in fi-
Benchmark shows promising results for our proposed model than nancial transactions, making to best use of banks big-data. Deep
existing solutions in terms of accuracy, recall and precision. learning is a generic term that refers to machine learning using
deep multilayer artificial neural network (ANN). It is a biologically
KEYWORDS inspired model of human neurons; composed of multilevel hidden
layers of nonlinear processing units, where each neuron is able to
Deep Learning, Real-Time Data, Binary Classification, Fraud Detec-
send data to a connected neuron within hidden layers [7]. Deep
tion.
neural networks attracted much attention in the field of machine
ACM Reference Format: learning. It’s currently providing the best precision and accuracy to
Youness Abakarim, Mohamed Lahby, Abdelbaki Attioui. 2018. An Efficient
many problems; providing promising results in many field, notably
Real Time Model For Credit Card Fraud Detection Based On Deep Learning.
in binary classification.
In Proceedings of October 2018 (SITA’18). ACM, New York, NY, USA, 7 pages.
https://doi.org/10.1145/nnnnnnn.nnnnnnn In a data analysis view, Fraud detection is a binary classifica-
tion problem, the transactions data is analyzed and classified in
1 INTRODUCTION "legitimate" or "fraudulent". Binary classification is a simple case of
classification, where a collection of data is classified into two classes,
In the recent years the number of bank transactions via credit cards
based on some features. It is mainly used in situation where we
raised drastically and with it the number of frauds and card theft.
want to predict a specific outcome, that can only take two distinct
The 2018 Association for Financial Professionals Payments Fraud
values. Some typical examples include medical diagnosis, spam
Survey, underwritten by J.P. Morgan [1], reported a new increase
detection, or in our case : fraud detection. Although fairly simple,
in payments fraud. An unprecedented record of 78 percent of all
binary classification is a very basic problem. There are numerous
organizations experimented payments fraud last year, a total of 700
paradigms used for learning binary classifiers, such as: Decision
treasury and finance professionals according to the survey. With
Trees, Neural Networks, Bayesian Classification, Support Vector
the rise of digital payment, finance institutions has lost billions
Machines, Logistic regression, K-nearest neighborhood, ect. [8]
due to credit card fraud [2]. Due to this issue banks and financial
In the other hand, Real-time data processing has become an
SITA’18, Rabat, Morocco, important field of research. We speak of Real-time data when the
2018. ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00 information is delivered directly after it has been gathered, making
https://doi.org/10.1145/nnnnnnn.nnnnnnn
SITA’18, Rabat, Morocco,
Youness Abakarim, Mohamed Lahby, Abdelbaki Attioui

the data accuracy time dependent. A Real-time system has been there have been many researchers around the use of SVMs in bi-
described by John A. Stankovic(1988) as : " One which controls an nary classification problems, in particular for image classification
environment by receiving data, processing them, and returning the problems [25], and in the financial related problems [26]. In [27],
results sufficiently quick to affect the environment at that time" [11]. compared SVM with various other paradigms in solving credit card
The novelty of our work, is the Real-time approach to credit card fraud, the authors focused on creating new inputs by aggregating
fraud detection, using state of the art machine learning technology. common transactional variables.
By using Deep neural network based on an auto-encoder, our model Artificial neural network (ANN) is a computing system that
classifies live feed of consumer transactions, and gives a real-time consists of networks interconnected elementary computing units,
decision - are they legitimate or fraudulent. designed to imitate the functioning of neurons in the human brain.
To test the effectiveness of our model, different binary classifica- Ghosh and Reilly [16] used data from a credit card issuer, one of
tion models are used for benchmark. Although the criteria used by the neural network based fraud detection system was trained on a
banks, differs for one institution to another, the use of linear models large sample of labeled credit card account transactions and tested
is well known in the banking sector [17]. Therefor, in the present on a holdout data set that consisted of all account activity over
work, we have selected logistic regression as a comparison, as well a subsequent two-month period of time. The method gave good
as linear SVM regression. Support vector machines (SVM) were results with high fraud account detection with fewer false positives
introduced for the first time by Vapnik [6], as a state-of-the art tech- by a factor of 20 compared with traditional methods. Cardwatch
nique to solve binary classification problems. SVMs have drawn a [19] developed in 1997 is a data mining system based on an artifi-
lot of attention in recent studies due to their major performance as cial neural network, trained with historical data of customers. By
classifiers. However, in some cases using non linear regression ANN processing the transaction of a specific customer it detect possible
has given good results in comparison with other linear models [18]. anomalies. Dorrronso et al. [20] presented an online system for
Hence, we have also selected a non linear artificial neural network fraud detection of credit card operations based on a neural classifier.
algorithms as a benchmark. As for the data sample, in this exercise The system was installed in a transactional hub and relies on the
we have used, the publicly available, European cards transaction of information of the operation and its previous history. A nonlinear
September 2012. version of Fisher’s discriminant analysis was used to overcome
This paper presents two key point. First, a real-time deep learn- imbalance of the rate of normal and fraudulent operations. In 2014
ing approach to the credit card fraud detection problem, based on [21] the authors proposed a credit card fraud detection model using
an auto-encoders. Second, a comparison of different binary clas- frequent item set mining. in [22] the authors used a feed forward
sification methods applied to this financial problem. We observe ANN for detecting fraudulent transactions, the method is found to
that our model based on deep machine learning with auto-encoder be effective.
gives promising results on this binary classification problem. Recently, and with the hype surrounding deep learning, some
The rest of the paper is organized as follows: In Section 2 we give works have shown that deep neural network, such as recurrent
a short review our problem’s related works. In section 3, we describe neural networks, have promising results in this field. However,
the architecture of our deep neural network prediction model. In since deep learning is a new approach in machine learning, the use
section 4, we give a description of our testing environment, then and analysis of various deep learning paradigms remains scarcely
we show and analyze the test results. Finally, in section 5 we draw explored. In [29] the authors used a convolutional neural network, a
some conclusions and indicate the future works. feed forward neural network inspired form the animal visual cortex,
to classify a set of card transaction in being fraudulent or not. In [30]
the authors used a Hidden Markov Model in a eCommerce fraud
2 RELATED WORK
detection model, the method is found to detect 80% of fraudulent
Over the past few years, many solution have been proposed to transaction. In [31] the authors found that deep learning method
cope with the problem of credit card fraud. The major approaches based on auto-encoder performed better than gradient boosted trees
are either statistical (a survey of Boltan and Hand in 2002 [9]) or for fraud detection. However, the proposed works don’t consider
based on artificial intelligence [10]. Various paradigms have been the importance of a Real-time approach for this type of problems,
used, some notable examples are : Regression models, Support and to our knowledge, comparison studies exploring Deep learning
Vector Machines, Restricted Boltzmann Machines, Artificial Neural paradigms are scarce. To cope with this situation we introduce a live
network, ect. binary classification method based on a deep learning auto-encoder.
Statistical models are increasingly applied to financial data min- In the second contribution, we propose an exhaustive comparison
ing task, including logistic regression, regression analysis, multiple with some typical binary classification algorithms. Our motivation
discriminant analysis, and Probit method, ect [23]. In the literature about the benchmark algorithms is shown in Tab. 3.
Logistic regression is widely used for binary classification problems
[24]. In [15] the author compared various traditional model for 3 OUR REAL-TIME CLASSIFICATION
fraud detection, Logistic regression is found to be the most accurate
of the traditional methods. SVMs have drawn a lot of attention in
APPROACH
recent studies due to its major performance as a classifier. Unlike 3.1 Prediction Model
ANNs which minimize empirical risk, SVMs are based on structural As a first approach we propose a classification method based on two
risk minimization. They use a nonlinear mapping to transform the stages: a periodical offline training of the historical data, by which
input data into a multidimensional feature space. In recent years we build our machine learning models. This stage first consists of a
 SITA’18, Rabat, Morocco,
An Efficient Real Time Model For Credit Card Fraud Detection Based On Deep Learning

Figure 1: Architecture of the Real-Time classification model

Figure 2: Implementation model of the live data classification technologies used.

feature engineering process, to transform the transaction data into math library. Tensorflow is the main used here for building and
features and labels for our machine learning classification. The data training our ML algorithms.
is then split into test and training sets. Afterward, Our models are Keras is an open source neural network library written in Python
build with the training features and labels. We test the models with [36]. Due of it’s capability to run over Tensorflow, it’s a good choice
the test features to get our first predictions and then compare the in our study.
test prediction with our test labels. The process is redone multiple Kafka is an Apache distributed streaming platform. It is used
times until we are satisfied with the models accuracy. In the second to build real-time streaming application that react and transforms
stage the models are used for prediction on a live stream of new the streams of Data. In our study, Kafka is used for building the
data. Fig. 1 shows the methodology followed to produce the results. real-time streaming data pipes lines.
To build the mentioned model we used the following technolo- Memsql is a distributed SQL data base. We have chosen this
gies : distribution for it’s wide use with Apache Kafka, and the amount
• Tensorflow : For building our machine learning model of documentation available on its implementation.
• Kafka : For building our RealTime streaming data pipeline.
• Memsql : For data pipelines. 3.2 Deep Auto-Encoder
Fig. 2 shows how the different technologies are implemented to For deep Learning we use an Auto-Encoder. Auto-Encoders are
build our live classification model. neural networks with an equal input and output. There architecture
For experiments, implementation and analysis we are using the is basically a two stacked Restricted Boltzmann Machines parallel
following environment: Python 3.5, Keras over Tensorflow Back- to each other, see Fig. 3. An auto encoder consists on two part, the
end, GPU GTX 660, Memory 8 GB. encoder and the decoder :
TensorFlow is an open-source software library for machine learn- • Encoder : compressing of the input into a fewer number
ing across a range of tasks [34]. In our study, we used it’s symbolic of bits. The part of the network with fewer bit is called
SITA’18, Rabat, Morocco,
Youness Abakarim, Mohamed Lahby, Abdelbaki Attioui

Figure 4: Neural network design: Each hidden layer is noted

by L1, L2... followed by the number of neurons in each layers

Table 1: Dataset features

Figure 3: Auto-encoder
Data set Characteristics Multivariate
Attribute characteristics Categorical, Integer
bottleneck or the "maximum point of compression" since Associated Task Classification
at this point the input is compressed the maximum. These Number of instances 284807
compressed bits that represent the original input are together Number of Attributes 30
called an “encoding” of the input. Missing Values N/A
• Decoder : here the input is reconstructed using the encoding
of the input. A successful encoding is when the decoder is
able to reconstruct the input exactly as it was fed to the
encoder.
In this work we use an hyperbolic tangent function "tanh" for the
encoding and decoding of the input to the output. The equations
are as follows :
Encoder
h(x) = tanh(Wx ) (1)
Decoder
â = tanh(W ∗ h(x)) (2)
The error reconstruction is done by back-propagation; the error
signal is computed and propagated backward in the network. The Figure 5: Data set distribution Over-Sampled, the fraud class
errors form desired and actual output values is used as condition. counts for 3,1% of all transactions
Parameter gradients is used for the back-propagation realization.
Time and Amount have not been transformed and all the other
3.3 Deep Neural network features are represented by V0, V1, . . . V26 values. See Tab. 2.
For our Auto encoder deep neural network we used 3 encoders and
3 decoders for a total of 6 hidden layers. The composition of the 4.2 Performance metrics
neural network is shown in Fig. 4. Due to the high metrics of the
This Data-set classifies transactions by being fraudulent or not.
"Tanh" activation function, it has been used in every hidden layer
We have 492 frauds out of 284807, which is highly unbalanced
for the Auto-encoder neural network.
0.173%. To solve this class unbalance, Random over-Sampling is
used. Fig. reffig:datasetOverSampling shows the distribution of the
4 EXPERIMENTAL RESULTS Data-set After Over-Sampling .
4.1 Dataset The Dataset is spliced into training and test sets. For a Pre-trained
The Data-set used in this work contains the transactions made in model performance check, we split the data into two separate train-
two days by European cards in September 2012, gathered and ana- ing sets and one independent test set for final model comparison.
lyzed during a research collaboration of Worldline and the Machine See Tab. 4.
Learning Group of ULB on big data mining and fraud detection. It
is freely available on Kaggle. See Tab. 1. 4.3 Paradigms
The data contains only numerical values. Due to confidentiality In order to measure the efficiency of deep learning in this case
the values where changed by PCA transformation. The features of study, three learning classification methods were chosen for
 SITA’18, Rabat, Morocco,
An Efficient Real Time Model For Credit Card Fraud Detection Based On Deep Learning

Table 2: Dataset attributes

Variable name Description Type

V0, V1, ..., V26 Transaction features after PCA transformation Integer
Time Seconds elapsed between each and the first transaction Integer
Amount Amount of transaction Integer
Class Non fraudulent or fraudulent 0 or 1

Table 3: Comparison Algorithms

Support Vector Machine Linear SVM Regression SVM has been proven to surpasses traditional neural
network models for solving complex non-linear prob-
lem, whch makes SVM a good choice for solving the
complex changeable data structure problems such as
fraud detection [32] [33].
Regression Logistic Regression As mentioned in section 2, in [15] the author used lo-
gistic regression to perform various credit card fraud
experiments, the latter is found to have promising re-
sults in comparison of the traditional methods
Classical ANN Non Linear Auto regression NN Using non linear regression ANN has given good results
in comparison with other linear models [18]
Deep Learning Deep Neural network (DNN) with auto encoders Beside being data-specific, Auto-encoders don’t need
explicit labels to train on, the labels are self generated
withing the model. Due to the diversity of the fraudulent
transaction, we believe that auto-encoders could be the
most fitted for this task.

Table 4: Instances distribution

Number of Instances 284807

Split ratio for Pre-training 0.2
Split ratio for Training 0.4
Independent test set 0.4

Figure 6: confusion matrix illustration

comparison, totaling 4 typical algorithms. The methods used for

comparison were chosen by their redundancy in various binary The precision is the measure of the exactness of the model, it is
classification research papers and their known good results. See the accuracy on transaction predicted as fraud defined as the num-
Tab. 3 for more details. ber of positive predictions divided by the total number of positive
For the metrics we used a combination of the confusion matrix class predicted by the model, as follow:
and a column chart, as well as the precision and recall. A confu-
sion matrix of a binary classifier is a table that shows the number Precision = T P/T P + F P . (3)
of instances classified correctly/incorrectly in each class. Fig. 6
illustrates the confusion matrix of a binary classifier. The recall is the measure of the completeness of the model, it
In our case, positive represents the fraudulent transactions and is the accuracy on fraud transactions defined as the number of
negative represents the legitimate ones. True positive (TP) repre- positive predictions divided by the number of positive class values
sents the fraudulent transactions classified as fraud. True negative of the test data, as follow:
(TN) represents legitimate transactions classified as legitimate. False
positive (FP) represents the misclassified legitimate transactions as Recall = T P/T P + F N . (4)
fraud. False negative (FN) represents the misclassified fraudulent We use F1 score to convey the balance between precision and
transactions as legitimate. the recall defined as follow:
SITA’18, Rabat, Morocco,
Youness Abakarim, Mohamed Lahby, Abdelbaki Attioui

Figure 7: Results metrics comparison

Table 5: Confusion Matrix results

F 1Score = 2((precision ∗ recall)/(precision + recall)). (5)

TP FP TN FN
4.4 Results analysis Linear SVM Regression 341 2111 115173 234
In card fraud detection systems, when a transaction raise a fraud Logistic Regression 302 1290 116002 270
flag, the transaction is refused, and the user must go through a
NN Based Classification 384 2229 115049 198
verification process to verify if it’s the case of a false flag or a real
fraud. This verification processes vary from a phone call to a series Non Linear Auto Regression 430 3871 113408 151
of verification forms. The cost of a false flag is then equivalent to Deep NN Auto encoders 358 1457 115793 257
the cost of these processes, which is significantly lower than the
cost of a fraud case. However, when the number of false flags is
important, purchases get blocked more frequently by error, making
one’s use of a credit card tedious and time consuming, moreover it to its dependencies (accuracy paradox[35] ). Therefor, Recall and
can generate big loses for either sides of the transaction. thus, our precision are shown in Fig. 7.
model should have a balanced mount of fraudulent transactions The results, in Fig. 7, show that in our model Non linear auto
caught and false flags raised. regression has the best Recall, but at the cost of precision. Deep NN
Tab. 5 shows the experimental results of our implemented algo- auto encoder, in the other hand, has the best precision, with results
rithms. Each algorithm was tested four times and for each algorithm close to logistic regression. We can’t draw conclusion from recall
we created a confusion matrix, and represented the TP FN FP TN an precision only, our model should have a well balanced values
in the table. The non linear auto regression has caught the most between the two metrics. Let’s see the F1 scores Tab. 6. In this case of
amount of the fraudulent transactions, but at cost of the false flags. study, Deep NN Auto encoder has the overall best F1 score followed
Logistic regression raised the least amount of false flags but is by logistic Regression, which confirms that Deep NN auto-encoder
not the best at catching the fraudulent transactions. Deep Learn- is the best fitted algorithm amongst the tested ones. Moreover, the
ing based on the auto encoder has balanced results, with a good deep learning algorithm used in this study is very simple; with
amount of frauds caught and a fair amount of false flags. First re- more tuning of the parameter (Hyper-parameter Tuning with Grid
sults seems promising for the Deep neural network model. Let’s Search) we may have better results. This provides good insight in
see the accuracy of our models. Fig. 8 shows the accuracy of each which algorithm should be chosen in building our prediction model,
algorithm. and we choose Deep Neural network with auto-encoder.
The accuracy is defined as correct predictions made divided
by the total number of predictions made. The data shows that 5 CONCLUSION
Logistic regression followed by the Auto-encoder have the best In this paper we proposed a Real-time model for credit card fraud
accuracy and the typical neural network classification method has detection, for a real-life data set of Credit Card transactions, us-
the worst. Because our data is not well balanced, we shouldn’t draw ing Deep Learning. After the tests and the comparison study, on
conclusions from accuracy only. Accuracy can be misleading due the performance of some typical real-time binary classifiers versus
 SITA’18, Rabat, Morocco,
An Efficient Real Time Model For Credit Card Fraud Detection Based On Deep Learning

[14] Masoumeh Zareapoor, Seeja.K.R, and M.Afshar.Alam.: Analysis of Credit Card

Fraud Detection Techniques: based on Certain Design Criteria. In: International
Journal of Computer Applications (0975 – 8887) Volume 52– No.3, August (2012)
[15] Bhattacharyya, Siddhartha, et al. "Data mining for credit card fraud: A compara-
tive study." Decision Support Systems 50.3 (2011).
[16] Ghosh, Sushmito, and Douglas L. Reilly. "Credit card fraud detection with a
neural-network." System Sciences, 1994. Proceedings of the Twenty-Seventh
Hawaii International Conference on. Vol. 3. IEEE, (1994).
[17] Bakar, Nor Mazlina Abu, and Izah Mohd Tahir. "Applying multiple linear regres-
sion and neural network to predict bank performance." International Business
Research 2.4 (2009): 176.
[18] Landi, Alberto, et al. "Artificial neural networks for nonlinear regression and
classification." Intelligent Systems Design and Applications (ISDA), 2010 10th
International Conference on. IEEE, (2010).
[19] Aleskerov, Emin, Bernd Freisleben, and Bharat Rao : Cardwatch: A neural network
based database mining system for credit card fraud detection. Computational In-
Figure 8: Accuracy of each algorithm telligence for Financial Engineering (CIFEr), 1997., Proceedings of the IEEE/IAFE
1997. IEEE, 1997.
Table 6: F1 scores results [20] Dorronsoro, Jose R., et al. : Neural fraud detection in credit card operations. IEEE
transactions on neural networks 8.4 (1997)
[21] Seeja, K. R., and Masoumeh Zareapoor. "FraudMiner: A novel credit card fraud
detection model based on frequent itemset mining." The Scientific World Journal
Classifier F1 score 2014 (2014).
[22] Hassibi, Khosrow. "Detecting payment card fraud with neural networks." World
Linear SVM Regression 0,225 Scientific Book Chapters (2000): 141-157.
Logistic Regression 0,279 [23] ALTMAN, Edward I., MARCO, Giancarlo, et VARETTO, Franco.: Corporate
NN Based Classification 0,240 distress diagnosis: Comparisons using linear discriminant analysis and neural
networks (the Italian experience). Journal of banking & finance, vol. 18, no 3, p.
Non Linear Auto Regression 0,176 505-529.(1994)
Deep NN Auto encoders 0,294 [24] D.W. Hosmer, S. Lemeshow : Applied Logistic Regression, 2nd Ed, Wiley-
Interscience, (2000)
[25] GOH, King-Shy, CHANG, Edward, et CHENG, Kwang-Ting.: SVM binary classifier
ensembles for image classification. In : Proceedings of the tenth international
Deep Neural Network with Auto-encoder, the benchmark experi- conference on Information and knowledge management. ACM. p. 395-402.(2001)
ments show that Deep NN Auto encoder has very promising results, [26] Shin, K.-S., Lee, T. S., & Kim, H.-J.: An application of support vector ma-
chines in bankruptcy prediction model. Expert Systems With Applications, 28(1),
with the best F1 score. The experiment confirms that although the 127–135.(2005)
well known performance of logistic regression Deep learning out [27] Whitrow, Christopher, et al. "Transaction aggregation as a strategy for credit
card fraud detection." Data Mining and Knowledge Discovery 18.1 (2009): 30-55.
performs it. Therefore, future studies will focus mainly on state [28] Y. Bengio.: Learning deep architectures for ai, Found. Trends Mach. Learn., vol.
of the art deep learning paradigms for this type of Real-time Data 2, no. 1, pp. 1–127, Jan. (2009)
Classification problems. The proposed Framework can be used by [29] FU, Kang, CHENG, Dawei, TU, Yi, et al.: Credit Card Fraud Detection Using Con-
volutional Neural Networks. In : International Conference on Neural Information
Credit Card Providers To monitor any unusual behavior and to Processing. Springer International Publishing. p. 483-490. (2016)
detect possible fraud attempts. [30] Srivastava, Abhinav, et al. "Credit card fraud detection using hidden Markov
model." IEEE Transactions on dependable and secure computing 5.1 (2008): 37-48.
[31] Rushin, Gabriel, et al. "Horse race analysis in credit card fraud—deep learn-
REFERENCES ing, logistic regression, and Gradient Boosted Tree." Systems and Information
[1] J.P.Morgan. "Payments Fraud and Control Survey", PNC Financial Services Group, Engineering Design Symposium (SIEDS), 2017. IEEE, (2017)
2018. [32] Patel, S., and S. Gond. "Supervised Machine (SVM) learning for credit card fraud
[2] Boston Consulting Group. (2017). Global Payments 2017 – Deepening The Cus- detection." International Journal of Engineering Trends and Technology (2014).
tomer Relationship. [33] Mareeswari, V., and G. Gunasekaran. "Prevention of credit card fraud detection
[3] Bose, Indranil, and Radha K. Mahapatra. "Business data mining—a machine based on HSVM." Information Communication and Embedded Systems (ICICES),
learning perspective." Information and management 39.3 (2001): 211-225. 2016 International Conference on. IEEE, (2016).
[4] TUNG, Hui-Hsuan, CHENG, Chiao-Chun, CHEN, Yu-Ying, et al. Binary Classifi- [34] Abadi, Martín, et al. "Tensorflow: a system for large-scale machine learning."
cation and Data Analysis for Modeling Calendar Anomalies in Financial Markets. OSDI. Vol. 16. (2016).
In : Cloud Computing and Big Data (CCBD), 2016 7th International Conference [35] A. Ng.: CS 229 machine learning course materials, Stanford University, (2016)
on. IEEE, (2016). p. 116-121. [36] F. Chollet.: Keras: deep Learning library for Theano and TensorFlow, Github,
[5] Jacomo Corbo, Carlo Giovine, and Chris Wigley(2017, April). Applying analytics https://github.com/fchollet/keras (2015)
in financial institutions fight against fraud. McKinsey Analytics. Retrieved from [37] Jason Brownlee, Machine Learning Mastery with python (2016)
https://www.mckinsey.com. [38] John, S. N., et al. : Realtime Fraud Detection in the Banking Sector Using Data
[6] Vapnik, V.: Statistical learning theory. Wiley, New York.(1998) Mining Techniques/Algorithm. Computational Science and Computational Intel-
[7] I. Goodfellow, Y. Bengio, A. Courville.: Deep learning, Cambridge, Massachusetts, ligence (CSCI), (2016)
The MIT Press, (2016)
[8] Phyu, Thair Nu. "Survey of classification techniques in data mining." Proceedings
of the International MultiConference of Engineers and Computer Scientists. Vol.
1. (2009).
[9] Bolton, Richard J., and David J. Hand. "Statistical fraud detection: A review."
Statistical science (2002): 235-249.
[10] Zhou, Xun, et al. "A state of the art survey of data mining-based fraud detection
and credit scoring." MATEC Web of Conferences. Vol. 189. EDP Sciences, 2018.
[11] John A. Stankovic. :Misconceptions about real-time computing. IEEE Computer,
21(10), 10-19 (1988).
[12] S. Benson Edwin Raj, A. Annie Portia.: Analysis on Credit Card Fraud Detec-
tion Methods. In: International Conference on Computer, Communication and
Electrical Technology – ICCCET2011, 18th & 19th March, (2011)
[13] Martin, James.: Programming Real-time Computer Systems. Englewood Cliffs,
NJ: Prentice-Hall Inc. p. 4. ISBN 0-13-730507-9. )(1965)

View publication stats