Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
Available online at www.sciencedirect.com

ScienceDirect
Procedia Computer Science 00 (2019) 000–000
Procedia
Procedia Computer
Computer Science
Science 16700 (2019)
(2020) 000–000
254–262 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia

International Conference on Computational Intelligence and Data Science (ICCIDS 2019)

International Conference on Computational Intelligence and Data Science (ICCIDS 2019)
An
An Autoencoder
Autoencoder Based
Based Model
Model for
for Detecting
Detecting Fraudulent
Fraudulent Credit
Credit Card
Card
Transaction
Transaction
Sumit Misraaa , Soumyadeep Thakuraa , Manosij Ghoshaa , Sanjoy Kumar Sahaaa
Sumit Misra
a
, Soumyadeep Thakur , Manosij Ghosh , Sanjoy Kumar Saha
Department of Computer Science and Engineering, Jadavpur University, Kolkata, India
a Department of Computer Science and Engineering, Jadavpur University, Kolkata, India

Abstract
Abstract
With the rapid growth in credit card based financial transactions, it has become important to identify the fraudulent ones. In
With the rapid
this work, a twogrowth in credit
stage model card based
is proposed financialsuch
to identify transactions,
fraudulentittransactions.
has becomeTo important to identify
make a fraud the system
detection fraudulent ones. In
trustworthy,
this
bothwork,
miss ina two stage
fraud modeland
detection is proposed to identify
false alarms such fraudulent
are to minimized. transactions.
Understanding andTolearning
make athefraud detection
complex system trustworthy,
associations among the
both miss inattributes
transaction fraud detection andproblem.
is a major false alarms are to minimized.
To address Understanding
this issue, at the first stageand learning
of the the complex
proposed model anassociations
autoencoderamong
is usedthe
to
transaction
transform the transaction attributes to a feature vector of lower dimension. The feature vector thus obtained is used as theis input
attributes is a major problem. To address this issue, at the first stage of the proposed model an autoencoder used to
to
transform
a classifierthe transaction
at the attributes
second stage. to a feature
Experiment vector
is done on aofbenchmarked
lower dimension. TheItfeature
dataset. vector
is observed thus
that obtained
in terms is used as theproposed
of F1-measure, input to
atwo
classifier at the performs
stage model second stage.
betterExperiment is donerelying
than the systems on a benchmarked dataset.
on only classifier andIt other
is observed that inbased
autoencoder terms systems.
of F1-measure, proposed
two stage model performs better than the systems relying on only classifier and other autoencoder based systems.

©c 2020
2019TheTheAuthors.
Author(s). Published
Published byby Elsevier
Elsevier B.V.
B.V.
c 2019
This is an
anThe Author(s).
open Published
access article
article underbythe
Elsevier
the B.V.
CC BY-NC-ND
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
This is open access under CC license
This is an
Peer-reviewopen access
under article under
responsibility of the
theCC BY-NC-ND
scientific license
committee
Peer-review under responsibility of the scientific committee of(http://creativecommons.org/licenses/by-nc-nd/4.0/)
of the the International
International Conference
Conference on Computational
on Computational Intelligence
Intelligence and
and Data
Peer-review
Science
Data under
(ICCIDS
Science responsibility
2019).
(ICCIDS 2019). of the scientific committee of the International Conference on Computational Intelligence and
Data Science (ICCIDS 2019).
Keywords: Fraud Detection; Financial Transaction; Autoencoder
Keywords: Fraud Detection; Financial Transaction; Autoencoder

1. Introduction
1. Introduction
Over the past few decades, there has been a massive increase in the use of e-commerce by various organizations,
Over theand
companies pastgovernment
few decades, there has
agencies. been
This hasaimproved
massive increase in the
productivity inuse of e-commerce
sectors by various
such as banking, organizations,
telecommunication,
companies and government agencies. This has improved productivity in sectors such as banking, telecommunication,
retail stores, health insurance, automobile insurance and online auction system [7, 3, 21]. The increasing popularity
retail stores,
of these health insurance,
technologies also createautomobile insurance
opportunities and online
for fraudsters auctionhavoc.
to wreak systemAs[7, 3, 21].financial
a result, The increasing
fraud haspopularity
become
of these technologies also create opportunities for fraudsters to wreak havoc. As a result, financial fraud
a menace with far reaching consequences in the financial and corporate sectors, as well as in government agencies. has become
aSuch
menace with far reaching consequences in the financial and corporate sectors, as well as in government
frauds can be defined as a criminal deception with the primary purpose of acquiring financial gains by illegal agencies.
Such
means.frauds
Highcan be defined
dependence onasinternet
a criminal deception
is seeing with credit
increased the primary purpose of acquiring
card transactions. Since thefinancial gainsused
most widely by illegal
mode
means. High dependence on internet is seeing increased credit card transactions. Since the most widely used mode

∗ Corresponding Author: Soumyadeep Thakur. Tel.: +91-8420281793.

∗ Corresponding
E-mail address:Author: Soumyadeep Thakur. Tel.: +91-8420281793.
[email protected]
E-mail address: [email protected]

1877-0509  c 2019 The Author(s). Published by Elsevier B.V.

1877-0509 
1877-0509 © 2020 The
c 2019 The Authors.
Author(s).Published bybyElsevier
ElsevierB.V.
This is an open access article underPublished
the CC BY-NC-ND B.V. (http://creativecommons.org/licenses/by-nc-nd/4.0/)
license
This isisan
This anopen
openaccess
access article
article under
under the the BY-NC-ND
CC CC BY-NC-ND licenselicense (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review
Peer-review under
underresponsibility
responsibilityofofthethe
scientific committee
scientific committee of(http://creativecommons.org/licenses/by-nc-nd/4.0/)
the International
of the Conference
International on Computational
Conference Intelligence
on Computational and Data
Intelligence Science
and Data
Peer-review
(ICCIDS under responsibility of the scientific committee of the International Conference on Computational Intelligence and Data Science
2019).
Science (ICCIDS 2019).
(ICCIDS 2019).
10.1016/j.procs.2020.03.219
 Sumit Misra et al. / Procedia Computer Science 167 (2020) 254–262 255
2 Author / Procedia Computer Science 00 (2019) 000–000

of payment, both online and offline, is through credit cards, cases of credit card fraud are also on the rise. This has a
dramatic impact on the economy, law as well as human moral values [2].
Credit card fraud can be categorized as inner card fraud or external card fraud [4]. Inner card fraud occurs as a
result of consent between cardholders and bank by using false identity to commit fraud. On the other hand external
card fraud involves the use of stolen credit card to get cash through dubious means. Inner credit card fraud, as a result,
is hard to detect as these do not follow any predictable patterns. A lot of research has been devoted to detection of
external card fraud which accounts for majority of the credit card frauds.
To prevent loss through cybercrimes, fraud detection systems have become essential for all credit card issuing
banks to minimize their losses. However, the sheer scale of electronic commerce renders human based detection al-
most ineffective and costly. Therefore, use of scalable machine learning algorithms to evaluate such transactions is
preferable. The most commonly used fraud detection methods [7, 1], are based on Neural Network (NN), association
rules, fuzzy system, decision trees, Support Vector Machines (SVM), Artificial Immune System (AIS), genetic algo-
rithms, K-Nearest Neighbour algorithms. These techniques can be used alone or in collaboration using ensemble or
meta-learning techniques to build classifiers.
A number of challenges are associated with credit card fraud detection. First, the dynamic nature of fraudulent
behaviour profile is an issue. Fraudulent transactions tend to look like legitimate ones to evade detection. Moreover,
credit card transaction data sets are heavily imbalanced (or skewed). Optimal feature (variables) selection for the
models and suitable metric to evaluate the performance of techniques on skewed credit card fraud data [4] are a few
more issues. This makes classification error-prone due to class bias. Keeping these in mind, this work aims to build a
methodology that learns the meaningful transaction attributes using an autoencoder which are subsequently utilized
in classification. The methodology can also be tuned so that it can be used for stream data. In first stage features are
extracted using autoencoders and the data, thus obtained, is then classified by a classifier. A brief survey is presented
in Section 2. Proposed methodology is elaborated in Section 3. Results and concluding remarks are placed in Section 4
and 5 respectively.

2. Related Works

Credit card fraud detection is the process of identifying whether a given transaction is fraudulent or legitimate. As
credit card becomes the most general mode of payment (for both online transactions and regular purchase), fraud rate
also tends to increase along with it. Detecting fraudulent transactions using traditional methods of manual detection
are time consuming and inaccurate. Moreover, it is impossible to detect in real time. With the advent of data mining
and machine learning techniques, it is possible to get rid of manual detection system.
From a statistical point of view, fraud detection methods can be broadly classified into supervised and unsupervised
methods. In supervised approach, detecting fraud is primarily a binary classification problem, where the objective is
to classify a transaction as legitimate (negative class) or fraudulent (positive class). In unsupervised fraud detection,
the problem can be thought of as an outlier detection system, assuming outlier transactions as potential instances
of fraudulent transactions. A detailed survey on supervised and unsupervised techniques in fraud detection is found
in [17]. Any kind of fraud detection system would be prone to error such as falsely identifying a legitimate transaction
as fraud or vice versa. It is necessary to strike a balance to minimize both. High number of missed fraud can incur
huge loss to people and corporations. On the other hand a high cases of stating a legitimate one as fraud would cause
people to lose trust in the organization with such system. Hence, the problem becomes quite challenging.
Credit card fraud detection has been studied for long. In the early days, Ghosh et al. [14] used neural network
for detecting fraud. In this work, data from a credit card issuer was used. The fraud detection system was based on
training a neural network on a large sample of labelled credit card transactions and testing on a validation data set
that consisted of all account activity over a subsequent two-month period of time. The neural network was trained on
examples of fraud caused by several factors like lost cards, stolen cards, application fraud, counterfeit fraud, mail-
order fraud and NRI (non-received issue) fraud. Brause et al. [8] combined advanced data mining techniques like
association rules mining, with neural networks to obtain a low false alarm rate .
Recently, a lot of machine learning and optimization techniques have been applied to identify credit card fraud.
Supervised learning techniques such as Artificial Neural Networks [20] and Support Vector Machines [27] have been
used for detecting fraud. A comparative analysis of the performance of machine learning classifiers including Naive
256 Sumit Misra et al. / Procedia Computer Science 167 (2020) 254–262
Author / Procedia Computer Science 00 (2019) 000–000 3

Bayes, K-nearest Neighbours, and Logistic Regression combined with data sampling techniques has been done in [4].
Ensemble of classifiers has been used successfully to identify credit card fraud. AdaBoost and Majority Voting ensem-
ble methods have shown to provide better results than single classifiers [24]. Mishra et al. [19] have used Chebyshev
Function Link Artificial Neural Network (CFANN) to identify credit card fraud. CFANN has two parts, functional
expansion and learning. In this work the authors made a comparative study between CFANN, Multi-Layer Perceptron
(MLP), and the Decision Tree algorithms.
Apart from machine learning techniques, data mining methods like frequent itemset mining was used by Seeja et
al. [26]. Genetic Algorithm was used in [23], where a set of parameter values were optimized using genetic algorithm
which were then used in certain rules. The rules were used to decide whether a transaction is fraudulent or not.
Migrating birds optimization was used in [11] to detect fraud.
Fu et al. [12] explores how to learn hidden intrinsic patterns associated with fraudulent behavior using a Convolu-
tional Neural Network (CNN). This paper also discusses a cost based sampling method to overcome the problem of
the data being imbalanced in favour of legitimate transactions. Autoencoders are a category of neural networks that
learn efficient data encodings, and their reconstructions, and hence can be used as a classifier based on their recon-
struction error. Energy based probabilistic models like Restricted Boltzmann Machines (RBMs) can be used to learn
the distribution of data. Autoencoders and RBMs have been used to detect fraud in [22].
From the survey it can be concluded that although learning algorithms have been largely used for fraud detection.
However, little attempt has been made to extract features from the transaction attributes. This extraction allows the
learning model to learn the distributions of fraud more effective unencumbered by irrelevant features. The model
proposed in this work uses an autoencoder to extract important features from the transaction data. Autoencoders have
the ability to detect complex non-linear correlations among features of the data. This ensures that the encoded data
thus obtained from the autoencoder is devoid of correlated and irrelevant features The compressed data thus obtained
is used to train a classifier.

3. Proposed Method

Fraud detection models have to deal with biased data as well as presence of irrelevant features in the input. These
two factors hinder the ability of a classifier to properly learn from the large amount of data. To address these issues,
a two-stage approach has been proposed where in the first stage a lower dimension of features are extracted from the
input and in the subsequent stage a model decides whether the transaction is fraud or not.
In this work, a fraud detection model is proposed that uses Autoencoders [5] for extracting essential features from
the input data, followed by a classification algorithm. For the purpose of detecting credit card fraud, our method
would predominantly work on credit card transactions. A given transaction can have a lot of features (attributes),
including the time and amount of the transaction, mode of transaction (deposit or withdrawal), the customers’ account
number, their age, location of the ATM used, etc. Having unnecessary features may cause classification algorithms to
perform poorly. Also, since real transaction data can have a lot of attributes, and hence very high dimensions, dealing
with such data becomes very expensive as far as time complexity is concerned. It is very difficult to identify and
focus on predetermined attributes and to determine their interrelationship to make the judgment. Hence the primary
objective is to find only the meaningful attributes and thereby reduce the number of attributes which are to be used for
classification. For feature selection, mutual information score [6] of a feature and the class is a statistical filter method.
But, it disregards the intricate relationships that multiple features may have among each other. Wrapper based feature
selection techniques [13] on the other hand are computationally expensive to run on a large dataset. Moreover, no
wrapper methods can guarantee generation of the optimal result.
To address these problems, this work uses Autoencoders, which can efficiently handle create a lower dimensional
representation of the input data, while being able to discover non-linearly correlated features. The details of an Au-
toencoder network are discussed in the following section.

3.1. Autoencoders

Autoencoders [5] are a specific category of feed-forward neural networks that are used to learn efficient encodings
of the training data. An autoencoder network has the same input and output dimensions; it transforms the input to a
 Sumit Misra et al. / Procedia Computer Science 167 (2020) 254–262 257
4 Author / Procedia Computer Science 00 (2019) 000–000

hidden representation, having a different dimension than the input (and output) dimension, and then reconstruct the
input from this hidden representation. It tries to learn the function fθ (X) = X for an input X, where θ denotes the
function parameters to be learned. In other words, it tries to approximate the identity function, which can be done
trivially, but by placing constraints on the network, such as by limiting the number of hidden units, the trivial solution
can be eliminated.

Fig. 1. Architecture of an undercomplete autoencoder with a single encoding layer and a single decoding layer

The most common type of an autoencoder is the undercomplete autoencoder [5] where the hidden dimension is
less than the input dimension. The architecture of such an autoencoder is shown in Figure 1. It shows an autoencoder
with input dimension 6 and hidden dimension of 3. An autoencoder has two parts - an Encoder and Decoder. For
an undercomplete autoencoder dim(Z) ≤ dim(X) always holds, where dim(X) denotes the dimensional of X. Also,
since X  is a reconstruction of X, dim(X  ) = dim(X) holds. The encoders and decoders are modelled by deep neural
networks.

• Encoder:
The encoder maps the input data X into hidden form Z. Let Wφ and Bφ be the weights and biases for the encoder
layer, then the hidden form Z can be represented as:

Z = fE (Wφ × X + Bφ ) (1)

fE is the Encoder activation function

• Decoder:
The decoder transforms Z to the reconstruction X  of the original data X. Let Wθ and Bθ be the weights and
biases for the encoder layer, then the hidden form Z can be represented as:

X  = fD (Wθ × Z + Bθ ) (2)
258 Sumit Misra et al. / Procedia Computer Science 167 (2020) 254–262
Author / Procedia Computer Science 00 (2019) 000–000 5

fD is the Decoder activation function

The activation functions fE (.) and fD (.) can be non-linear, and hence an autoencoder is able to detect non-linearly
correlated features in the input, which are otherwise hard to detect.
The goal of the autoencoder network is to provide a transformed feature vector Z such that reconstructed data X 
is close to the original data X. The reconstruction error ∆ measures the distance between the reconstructed input from
the autoencoder and the original input. In this work, Euclidean distance between X and X  was considered as the
reconstruction error, i.e.

∆(X, X  ) = ||X − X  ||2 (3)

Like any good supervised learning models, the ideal autoencoder should be sensitive to the inputs to accurately
build the reconstructions, but not to an extent so that the model overfits. The problem of over-fitting can be solved
by adding a regularizer, which works by slightly tuning the objective function of the learning algorithm. One popular
regularization technique used in case of autoencoders was introduced by Rifai et al. [25]. It penalizes large derivatives
of the hidden data with respect to the input.

3.2. Classification

In the first stage of the proposed model an autoencoder is trained using the transaction attributes. The autoencoder
is therefore able to produce a transformed (encoded) representation of the attributes, Z which can be used to retrieve
the original features. The representative features have a smaller dimension than the original features which makes
learning of the classifier in the second stage easier. For the transformation of transaction attributes only the encoder
network of the autoencoder is used. In the second stage, a classifier is trained with the labelled transactions where
each transaction is represented by Z, the features generated by the autoencoder. For testing, the transaction attribute
vector passes through autoencoder (only encoder) and corresponding transformed vector is fed to trained network for
classification. The model proposed here is a general one and any classifier can be used in the second stage depending
on the user requirements. Our model has been tested using three different classifiers to prove the generality of our
model. The classifiers used are Multi-Layer Perceptron, K-Nearest Neighbour and Logistic Regression.

4. Experimental Results

4.1. Dataset

In this experiment, credit card transaction dataset from ULB Machine Learning Group [18] was used, which was
downloaded from https://www.kaggle.com/ntnu-testimon/paysim1/downloads/paysim1.zip/2. It con-
tains credit card transactions made by cardholders in Europe in 2013. The dataset has a total of 284, 807 transactions,
and the fraudulent ones make up only a meagre 0.172% of the data with 492 such transactions. So, the data is highly
imbalanced towards the fraudulent class. It contains only numeric input variables which are as a result of a Principal
Component Analysis (PCA) resulting in 28 principal components. Apart from these, the amount of money involved
in the transaction and its time of occurrence are included as features, so a total of 30 features are present. The feature
by the name T ime denotes the number of seconds that have elapsed since the first transaction. The the amount of
money involved in the transaction is denoted by Amount. The fraudulent transactions are considered as belonging to
the positive class while the non-frauds are considered as belonging to the negative class in our experiment.

4.2. Model Creation

The data is pre-processed before training. First, the T ime attribute is modified to denote the hour of the day. The
data is then normalized to [-1,1], and split into training and test sets, with the training sets having 70% of the total
number of transactions. The autoencoder is trained with all our training data and it learns the distribution of the data.
 Sumit Misra et al. / Procedia Computer Science 167 (2020) 254–262 259
6 Author / Procedia Computer Science 00 (2019) 000–000

In our experiment an autoencoder with a 2-layer encoder network and a 2-layer decoder network is used. The
dimension of the hidden data is 15, so a 30 dimensional data is encoded using only half the number of dimensions. The
reconstruction error is calculated using the Euclidean distance between the input and its reconstruction. The classifiers
subsequently used are Multi Layered Perceptron (MLP), K-nearest Neighbor (KNN) and Logistic Regression (LR).
An MLP with 2 hidden layers having dimensions 13 and 7 is used. It has an adaptive learning rate starting with 0.0001
and uses Adam Optimization Algorithm [15] for reducing the classification error. The KNN classifier is trained using
3 closest training examples. In case of Logistic Regression, L2-regularization [10] is used for regularization, and
Limited-memory Broyden–Fletcher–Goldfarb–Shanno (LM-BFGS) [9] is used as optimizer.

4.3. Performance Evaluation

The proposed model is tested using the test set described in the previous section. In this experiment, the positive
class is the fraudulent class. A classification outcome has four cases, True Positive (TP), False Positive (FP), True
Negative (TN) and False Negative (FN). To measure the performance, the metrics used are Accuracy, Precision,
Recall and F1-score, which are defined as follows.

TP + TN
Accuracy = (4)
T P + FP + T N + FN

TP
Precsion(P) = (5)
T P + FP

TP
Recall(R) = (6)
T P + FN

2×P×R
F1 − S core = (7)
P+R

Table 1. Description of models used in comparison of our proposed model

Name Description
M1 Proposed model
M2 Use of classifier alone
M3 Autoencoder used to extract features and undersampling of negative class is used while training
classifier
M4 Classifier used along side under-sampling of negative class

To study the performance, different approaches are considered as mentioned in Table 1. Proposed one (M1) is
autoencoder based, with extracted features fed to the classifier. M2 stands for directly using the transaction attribute
260 Sumit Misra et al. / Procedia Computer Science 167 (2020) 254–262
Author / Procedia Computer Science 00 (2019) 000–000 7

based vectors to classifier. M3 considers encoded feature but negative (non-fraud) class is under-sampled for training
the classifier. M4 is similar to M1 but the classifier is trained with data with the non-frauds (negative class) under-
sampled. For all the four approaches experiment is carried out with three different classifiers and results are tabulated
in Table 2-4. The precision of proposed model (M1) is far higher than other models (M2, M3 and M4). For MLP and
KNN classifier, Proposed model (M1) performs better than the rest in terms of accuracy, precision and F1-measure
(as shown in Table 2 and 3). In case of LR classifier, the precision of the proposed model better than rest but suffers
in recall and hereby in F1-measure (as shown in Table 4). This may be attributed to the fact that performance of LR
classifier is limited by its weakness in dealing with non-linearity class features.
As F1-measure takes care of both miss in fraud detection and false alarm rate, focus is put on this metric and in
this regard, an autoencoder followed by an MLP performs best. This combination has been considered for subsequent
study. The proposed model has been further compared with other autoencoder based model for comparison of their
performance. Pumsirirat et al. [22] have presented a system with an autoencoder trained only on non-fraud data and
used the same for classification. Variational Autoencoders (VAEs) [16, 28] are quite popular for their use as generative
models in addition to learning encodings. Performances of these two systems and the proposed model are shown in
Table 5. It may be noted that both the systems have reasonable recall but precision is very low. It indicates that huge
amount of false alarms are generated for those systems. But the proposed method is a balanced one and in terms of
F1-measure it outperforms both by large extent.

Table 2. Comparison of performance of different models in identifying fraud using MLP classifier

Model Accuracy Precision Recall F1-Score

M1 0.9994 0.8534 0.8015 0.8265
M2 0.9993 0.7794 0.7794 0.7794
M3 0.9986 0.5385 0.8750 0.6667
M4 0.9964 0.2896 0.8603 0.4333

Table 3. Comparison of performance of different models in identifying fraud using KNN classifier

Model Accuracy Precision Recall F1-Score

M1 0.9995 0.9340 0.7279 0.8182
M2 0.9995 0.9100 0.7426 0.8178
M3 0.9973 0.3517 0.8382 0.4957
M4 0.9970 0.3324 0.8603 0.4795

Table 4. Comparison of performance of different models in identifying fraud using LR classifier

Model Accuracy Precision Recall F1-Score

M1 0.9992 0.8571 0.5735 0.6872
M2 0.9991 0.8452 0.5221 0.6455
M3 0.9985 0.5174 0.7647 0.6172
M4 0.9993 0.7863 0.7574 0.7715

The proposed model can be established through offline training and thereafter it can be well utilized in handling
the transaction streams. During fraud detection, a transaction passes through encoder stage to generate the feature
vector and thereafter classification proceeds. Both these tasks can be accomplished in real time (in the order of 17.56
microseconds per transaction on a machine with a CPU clock frequency of 2.3 GHz and 4 Gigabytes of RAM using
 Sumit Misra et al. / Procedia Computer Science 167 (2020) 254–262 261
8 Author / Procedia Computer Science 00 (2019) 000–000

Table 5. Comparison of proposed models with other contemporary models

Model Accuracy Precision Recall F1-Score

Method of Pumsirirat et 0.9705 0.0470 0.8367 0.0890
al. [22]
Variational Autoencoder 0.9890 0.1049 0.7868 0.1851
Proposed Method 0.9994 0.8534 0.8015 0.8265

Multilayered Perceptron as classifier). However, once the model is trained it assumes the transaction profile follows
the similar trend. In reality it may change with time. To cope up with this the model has to be retrained.

5. Conclusion

In this work, a two stage model has been proposed to detect the fraudulent ones in credit card transactions. Relation-
ships among the transaction attributes are quite complex. Proper understanding of the same can help in classification.
At the first stage of proposed model autoencoder focuses on this aspect by transforming the transaction attributes to a
lower dimensional feature vector. Such feature vectors is then fed to a classifier at second stage. Experimental results
show that proposed methodology maintains a good balance between precision and recall in detecting the frauds. It also
outperforms the systems based on either different classifiers or variants of autoencoder. It establishes the efficiency of
proposed two stage model. In future, the proposed two stage model can be tuned to handle stream data. The model
can be trained on a batch of transactions, and the trained model can be utilized in predicting the future transactions.
However, to cope up with the changes in the pattern of fraud, periodic retraining of the model will be an important
challenge.

References

[1] Abdallah, A., Maarof, M.A., Zainal, A., (2016). Fraud detection system: A survey. Journal of Network and Computer Applications 68, 90–113.
[2] Alexopoulos, P., Kafentzis, K., Benetou, X., Tagaris, T., Georgolios, P., (2007). Towards a generic fraud ontology in e-government., in:
Proceedings of the ICE-B, Barcelona, Spain. pp. 269–276.
[3] Allan, T., Zhan, J., (2010). Towards fraud detection methodologies, in: Proceedings of the 5th International Conference on Future Information
Technology, IEEE, Changsha, China. pp. 1–6.
[4] Awoyemi, J.O., Adetunmbi, A.O., Oluwadare, S.A., (2017). Credit card fraud detection using machine learning techniques: A comparative
analysis, in: Proceedings of the 2017 International Conference on Computing Networking and Informatics (ICCNI), IEEE, Lagos, Nigeria. pp.
1–9.
[5] Baldi, P., (2012). Autoencoders, unsupervised learning, and deep architectures, in: Proceedings of the ICML Workshop on Unsupervised and
Transfer Learning, pp. 37–49.
[6] Battiti, R., (1994). Using mutual information for selecting features in supervised neural net learning. IEEE Transactions on Neural Networks
5, 537–550.
[7] Bolton, R.J., Hand, D.J., (2002). Statistical fraud detection: A review. Statistical Science , 235–249.
[8] Brause, R., Langsdorf, T., Hepp, M., (1999). Neural data mining for credit card fraud detection, in: Proceedings of the 11th International
Conference on Tools with Artificial Intelligence, IEEE, Chicago, IL, USA. pp. 103–106.
[9] Byrd, R.H., Lu, P., Nocedal, J., Zhu, C., (1995). A limited memory algorithm for bound constrained optimization. SIAM Journal on Scientific
Computing 16, 1190–1208.
[10] Cortes, C., Mohri, M., Rostamizadeh, A., (2009). L 2 regularization for learning kernels, in: Proceedings of the Twenty-Fifth Conference on
Uncertainty in Artificial Intelligence, AUAI Press, Montreal, QC, Canada. pp. 109–116.
[11] Duman, E., Elikucuk, I., (2013). Solving credit card fraud detection problem by the new metaheuristics migrating birds optimization, in:
Proceedings of the 2013 International Work-Conference on Artificial Neural Networks, Springer, Tenerife, Spain. pp. 62–71.
[12] Fu, K., Cheng, D., Tu, Y., Zhang, L., (2016). Credit card fraud detection using convolutional neural networks, in: Proceedings of the 2016
International Conference on Neural Information Processing, Springer, Kyoto, Japan. pp. 483–490.
[13] Ghosh, M., Guha, R., Mondal, R., Singh, P.K., Sarkar, R., Nasipuri, M., (2018). Feature selection using histogram-based multi-objective
GA for handwritten devanagari numeral recognition, in: Proceedings of the 2018 Intelligent Engineering Informatics. Springer, Bhubaneswar,
India, pp. 471–479.
[14] Ghosh, S., Reilly, D.L., (1994). Credit card fraud detection with a neural-network, in: Proceedings of the Twenty-Seventh Hawaii International
Conference on System Sciences, IEEE, Wailea, HI, USA. pp. 621–630.
 262 Sumit Misra et al. / Procedia Computer Science 167 (2020) 254–262
Author / Procedia Computer Science 00 (2019) 000–000 9

[15] Kingma, D.P., Ba, J., (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 .
[16] Kingma, D.P., Welling, M., (2013). Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 .
[17] Kou, Y., Lu, C.T., Sirwongwattana, S., Huang, Y.P., (2004). Survey of fraud detection techniques, in: Proceedings of the 2004 International
Conference on Networking, Sensing and Control, IEEE, Taipei, Taiwan. pp. 749–754.
[18] Lopez-Rojas, E., Elmir, A., Axelsson, S., (2016). Paysim: A financial mobile money simulator for fraud detection, in: Proccedings of the 28th
European Modeling and Simulation Symposium, (EMSS), Dime University of Genoa, Larnaca, Cyprus. pp. 249–255.
[19] Mishra, M.K., Dash, R., (2014). A comparative study of chebyshev functional link artificial neural network, multi-layer perceptron and decision
tree for credit card fraud detection, in: Proceedings of the 2014 International Conference on Information Technology, IEEE, Bhubaneshwar,
India. pp. 228–233.
[20] Ogwueleka, F.N., (2011). Data mining application in credit card fraud detection system. Journal of Engineering Science and Technology 6,
311–322.
[21] Pejic-Bach, M., (2010). Profiling intelligent systems applications in fraud detection and prevention: survey of research articles, in: Proceedings
of the 2010 International Conference on Intelligent Systems, Modelling and Simulation, IEEE, Liverpool, UK. pp. 80–85.
[22] Pumsirirat, A., Yan, L., (2018). Credit card fraud detection using deep learning based on auto-encoder and restricted boltzmann machine.
International Journal of Advanced Computer Science and Applications 9, 18–25.
[23] RamaKalyani, K., UmaDevi, D., (2012). Fraud detection of credit card payment system by genetic algorithm. International Journal of Scientific
& Engineering Research 3, 1–6.
[24] Randhawa, K., Loo, C.K., Seera, M., Lim, C.P., Nandi, A.K., (2018). Credit card fraud detection using adaboost and majority voting. IEEE
Access 6, 14277–14284.
[25] Rifai, S., Vincent, P., Muller, X., Glorot, X., Bengio, Y., (2011). Contractive auto-encoders: Explicit invariance during feature extraction, in:
Proceedings of the 28th International Conference on International Conference on Machine Learning, Omnipress, Bellevue, WA, USA. pp.
833–840.
[26] Seeja, K., Zareapoor, M., (2014). Fraudminer: A novel credit card fraud detection model based on frequent itemset mining. The Scientific
World Journal 2014, 1–10.
[27] Singh, G., Gupta, R., Rastogi, A., Chandel, M.D., Riyaz, A., (2012). A machine learning approach for detection of fraud based on svm.
International Journal of Scientific Engineering and Technology 1, 194–198.
[28] Xu, H., Chen, W., Zhao, N., Li, Z., Bu, J., Li, Z., Liu, Y., Zhao, Y., Pei, D., Feng, Y., et al., (2018). Unsupervised anomaly detection via
variational auto-encoder for seasonal kpis in web applications, in: Proceedings of the 2018 World Wide Web Conference, International World
Wide Web Conferences Steering Committee, Lyon, France. pp. 187–196.