Sei sulla pagina 1di 9

International Journal of Computer Systems (ISSN: 2394-1065), Volume 04– Issue 07, July, 2017

Available at http://www.ijcsonline.com/

Credit Card Fraud Identification Using Artificial Neural Networks

Chandrahas MishraA, Dharmendra Lal GuptaB, Raghuraj SinghC

A
Department of Computer Science & Engineering, Kamla Nehru Institute of Technology, Sultanpur, UP. INDIA
B
Department of Computer Science & Engineering, Kamla Nehru Institute of Technology, Sultanpur, UP. INDIA
C
Computer Science and Engineering Department, HBTI Kanpur, UP. INDIA.

Abstract

In the present time, internet has become more popular in almost every domain of life. E-commerce, online shopping and
transactions are increasing day by day. Fraudulent activities have also increased in online payment system. Credit card
fraud is more frequently occurring in present scenario. Many credit card fraud identification systems have been
introduced till now. This paper presents the performance analysis and a comparative study of the results of various
techniques used for credit card fraud identification. In addition, we have also proposed a model for credit card fraud
identification using artificial neural network. Accuracy of BR, GDA, and LM techniques for the proposed artificial
neural network model applied on two different datasets taken from UCI Repository has been compared.

Keywords: Fraud, Credit Card Fraud, Fraud Identification Systems (FISs), Machine learning (ML), Neural Nets Models,
Artificial Neural Network (ANN).

2. Second fraud is done when submitting the application


I. INTRODUCTION
form to the bank for issuing the credit card with the
As credit card is the most prominent method of fake information.
payment for online and day to day purchases nowadays, 3. Third fraud is done online when purchasing any item
the cases of fraud associated with it are also growing.
Fraud can be defined as the illegal usage on any system or by submitting the card information of any credit card
goods. Credit card fraud is done through the use of the without letting known to the owner.
credit card details, with an intension of obtaining goods 4. Fourth fraud is the stealing of someone’s credit card
without actually paying for it. The financial losses due to and using it by showing as an owner of this card until
credit card frauds have become so high that it seems to be that card become blocked by bank.
a menace to the economy.
Two types of frauds named as physical fraud and Credit card fraud results the financial losses of
virtual fraud are more prevalent. Physical fraud is the organizations and also affects the individual user, if the
fraud that is done by stealing the card and using it for the credit card details get steal. To address this issue, financial
payment or purchasing, whereas virtual fraud is being institutions apply many fraud prevention strategies such as
conducted by misusing someone’s card details through credit card authorization, address verification system
internet. Various methods of credit card fraud are shown (AVS), rule-based detection, etc. However, fraudsters are
in figure 1. [1] adaptive and come up with several tricks and ideas to
break the existing prevention mechanisms. Once the fraud
prevention techniques are failed, there is the requirement
of developing an efficient fraud identification system (FIS)
credit card for identifying credit card frauds and maintaining the
fraud
viability of the payment system.
The idea of Artificial Neural Network was
presented in late 1943 by Walter Pitts and Warren
Bankruptcy Application Behavioral Theft/count
fraud fraud fraud erfiet fraud
S.McCulloch as a data processing unit for classification or
prediction problems.
According to E. Don Box et al. [2] neural networks
Figure 1. Classification of credit card fraud. model can be used for demand forecasting in untemper
environment. Neural network are designed and trained on
the premise of total requests of the gatherings of
1. In the first kind of fraud, the user uses its own credit overviewed clients of various classes,
card knowing that he has no money in his credit card The most as often as possible experienced basic
and bank has to pay by sending bill to the address. decision making tasks of human activity is classification.
Classification problem occurs when an article should be

151 | International Journal of Computer Systems, ISSN-(2394-1065), Vol. 04, Issue 07, July, 2017
Chandrahas Mishra et al Credit Card Fraud Identification Using Artificial Neural Networks

referred as a predefined class taking into account number Ghosh and Reilly [12] have performed a feasibility
of observed attributes related to that object. Many study to check the effectiveness of the neural network for
problems in business, science, industry, and medicine can the identification of credit card fraud for Mellon bank and
be treated as classification problems. Examples includes they have achieved the reduction of around 20%-40% in
bankruptcy prediction, credit scoring, medical diagnosis, the overall loss.
quality control, handwritten character recognition, and A new approach to this problem is given by Aleskerov
speech recognition described by peter et al.[3]. et al. [13] as CARDWATCH, acredit card fraud detection
Neural Network and Genetic Algorithms are used for (CCFD system based on neural network and data-mining
web mining. Sankar K. Pal et.al [4] described the soft system for).
computing framework based web mining. Quah and Sriganesh [14] proposed a framework that
Artificial Neural Networks (ANN) technique of Data can be applied in real-time scenario, where first an outlier
Mining and Artificial Intelligence are often applied in analysis is done separately for each customer using self-
fraud identification system [5] [6] [7] [8] [9] [10]. organizing map (SOM) and classification algorithm is
In this paper, we focuson machine learning techniques. used to classify a transaction as fraudulent or genuine.
The paper provides a platform which is intended to A fraud identification system consists of four
classify a current transaction intofraud or non-fraud. We component that areconnected serially was proposed by
consider credit card fraud identification problem as a Panigrahi et al. [15]. The main concept is to determine the
classification problem and use two techniques (supervised suspicious transactions by applying Bayesian learning to
and unsupervised) for credit card fraud identification. identify the fraudulent activities.
These are the machine learning techniques, in which the Srivastava et al. [16] have proposed a new model
first one uses training data to build the model, which have which uses the Hidden Markov Model (HMM) that is
all the attributes including class label i.e. it already trained with normal behaviour of the card holders and
contains the attribute which tells whether previous based on it the classification is done.
transaction is fraud or not. In the second technique, A self- organised credit card fraud detection system by
training data do not contain the class label i.e. this applying Artificial Neural Network (ANN) and Bayesian
technique is class less. Belief Network (BBN) has been implemented by Maes et
al. [17]
This paper proposes a credit card fraud To distinguish fraudulent transactions via credit card,
identification system using artificial Neural Networks. there was a requirement for classification techniques ready
Neural Network is a supervised machine learning to recognize deceitful exchanges taking into account
technique. ArtificialNeural Network works similar to the historical data. To that end a Multilayer Perceptron (MLP)
human mind. Human cerebrum comprise of number of Artificial Neural Network classifier was chosen [18] [19]
neurons associated with each other. Similarly ANN [20] [21].
comprises of artificial neurons, called nodes in network, ANN comes in many forms like Recurrent NN,
connected with each other. Associative NN, etc. Simulated Annealing method used to
train the Feed-Forward Neural Network for real time
This paper is organized as follows. In section 1, credit card fraud detection was proposed by Azeem Ush
we have presented the introduction of the problem along Shan Khan et al. [22].
with its proposed solution. In section 2, we provide the Combination of clustering and neural network used for
related work previously done for credit card fraud credit card fraud identification was proposed by Tanmay k
identification system. In section 3, we have described the Behera et al. [23]
proposed methodology and model for credit card fraud
identification system. Section 4 deals with the details of III. PROPOSED METHODOLOGY
experimental setup i.e. tool and information about dataset. This section describe the methodology of proposed work
In section 5, performance analysis and comparative study i.e. what process is required to get the result. Steps
of all the techniques on the basis of result obtained from required to implement the process to design the network
experiment is done. In section 6, conclusions of the study model are mentioned below.
and results are described.
1. Collect the public dataset.
II. RELATED WORKS 2. Create the network model.
For the first time, Dorronsoro et al. in 1997 developed a 3. Configure the network.
system to detect credit card fraud using Neural Network. 4. Initialize the weights and biases matrices.
ANN have been successfully applied in business failure 5. Train the network.
prediction, stock price prediction, credit fraud 6. Validate the network.
identification and many more area till now. The solution 7. Simulate the network.
for credit card fraud identification using genetic algorithm 8. Use the network.
and scatter search was proposed by Duman and Ozcelik
[11]. Some scores are associated with every transactions, Above mentioned 8 steps should be followed to get the
and on the basis of scores every transactions are classified desired network.
as fraudulent or genuine.

152 | International Journal of Computer Systems, ISSN-(2394-1065), Vol. 04, Issue 07, July, 2017
Chandrahas Mishra et al Credit Card Fraud Identification Using Artificial Neural Networks

layer both. Formula for sigmoid activation function is


given below:

1
𝑆 𝑡 =
1 + 𝑒 −𝑡

Incoming Transactions

User authentication and card validation phase


Figure 2. Feed Forward Back Propagation Neural Network Enter login & credit card details
architecture.

Figure 2 is the proposed architecture of customized feed


forward neural network for credit card fraud identification Fail
Validatio
system. Configuration of neural network is as follows: n
check?
Type of network = feed forward back propagation
No of inputs=1
No of outputs=1
No of neurons in input layer = 20 Pass Reject transaction
No of neurons in output Layer =1
No of hidden layers=2 Neural Network training phase
No of neuron in hidden layer1=10
Training data
No of neuron in hidden layer2 =20
Input Output
Training algorithm=Bayesian Regularization.
Transfer function= transig.
Maximum no of iteration = 1000.
Neural
Accuracy is calculated by comparison of actual output Network
given in dataset and predicted output of the network
model after performing the simulation. Formula for Desired output Predicted output
calculating the accuracy is given below.
Change in weights
Accuracy=100-Percentage Error. +/-
Error
Percentage Error=100*mod (average Predicted Training
Output - average Actual Output)/average Actual Algo Objective
Output. Function

Transactio
Reject
n allowed
(Fraudulent)
(Genuine)
Figure 4. Proposed Implementation Model

Implementation model proposed for credit card fraud


identification system is depicted in figure 4. There are two
phase in the model, first one is user authentication and
validation phase and the second one is neural network
training phase. In first phase login credential and credit
card details are checked. If validation check is pass then
transaction is passed to second phase otherwise transaction
Figure 3. Sigmoid Activation Function Graph. will be rejected. In second phase neural network training
based learning is performed. Transaction is fraudulent or
Sigmoid activation function depicted in figure 3 alludes genuine is decided by second phase. If transaction is
to one of the instance of logistic function. It works for fraudulent then rejected and if genuine then allowed by
real input values and returns just positive result [24]. It is second phase.
used as the activation function at hidden layer and output

153 | International Journal of Computer Systems, ISSN-(2394-1065), Vol. 04, Issue 07, July, 2017
Chandrahas Mishra et al Credit Card Fraud Identification Using Artificial Neural Networks

Simulation is performed on MATLAB 2015. Neural second one is Australian credit dataset. Exact data is not
Network Train (nntrain) tool is used to perform the provided in the dataset due to security issue. Some integer
training of the network on 75% of the dataset and 15% are assigned to each value in the dataset. Availability of
of dataset is used for testing and 10 % is used for real world dataset is very big challenge for credit card
validation process. Customized multi-layer feed forward fraud identification system. Desired output is either 1 or 0
back-propagation neural network is created, architecture in the given dataset.
of customized neural network is shown in the figure 2.
Details of the simulation process is described in section
4.

IV. EXPERIMENTAL SETUP AND SIMULATION


In this section, we present the experimental setup to find
out the result, which is based on real world dataset. The
dataset consists of 1000 instances is taken from UCI
repository. This dataset contains useful data about
transactions. This dataset contains 20 attributes and value
relating to every row is changed over into typical or
numerical structure due to some protection policy. But the
types of attributes are shown in the table I.

TABLE 1.GERMAN (STATLOG)CREDIT DATASET.


Sl
Attribute Name
.No
1 Status of existing checking Account
2 Duration in month
3 History of credit taken
4 Purpose of transaction
5 Credit amount
6 Saving account/bonds
7 Present employment since
8 Instalment rate in %
9 Personal status & sex
10 Guarantors
11 Present address since
12 Property
13 Age
14 Other instalment plans
15 Housing
16 Existing credits at this bank
17 Job
18 Number of people liable to provide
maintenance for
19 Telephone
Figure 5. Training process of neural network on nntrain tool.
20 Foreign worker(Yes or No)
Above figure 5 is the snapshot of training process of the
experiment. Neural Network is trained by BR technique.
In this experiment we have used two different dataset Best validation performance is also shown in the figure .
taken from UCI Repository [25]. First dataset is German
(Statlog) credit dataset mentioned in above table 1 and

154 | International Journal of Computer Systems, ISSN-(2394-1065), Vol. 04, Issue 07, July, 2017
Chandrahas Mishra et al Credit Card Fraud Identification Using Artificial Neural Networks

15 1 1.078 1.741034008 1.497512706


16 2 1.9986 1.521666412 1.210227784
17 1 1 1.024937115 1.034604852
18 1 1.0001 1.131244711 1.308728917
19 2 1.9501 1.2120957 1.137590726
20 1 1 1.070347306 1.031234724
21 1 1 1.038522871 1.122649941
22 1 1.0613 1.082266085 1.056088436
23 1 1 1.143370296 1.037903509
24 1 1 1.039790856 1.023726286
25 1 1 1.067849418 1.02912808
26 1 1 1.172262969 1.132648721
27 1 1.0619 1.061526353 1.039581446
28 1 1.0002 1.229570321 1.412475921
29 1 1 1.041843763 1.098130185
30 2 1.9999 1.707175134 1.185515638
31 1 1.0158 1.112508745 1.211409236
32 1 1.0708 1.567117416 1.338320058
33 1 1.057 1.298462117 1.676886695
34 1 1 1.024876818 1.027073827
35 1 1.0566 1.200725326 1.115587195
36 2 2 1.25595547 1.374574434
37 1 1.0001 1.229837523 1.171760753
38 2 1.9155 1.459903487 1.221956388
39 1 1 1.107884687 1.056488858
Figure 6. Regression chart on training of neural network.
40 1 1.0266 1.105410033 1.087017312
Validation performance and regression plot of neural 41 1 1 1.288382022 1.415455228
network after performing training testing and validation 42 1 1 1.281036534 1.080351514
check is depicted in figure 6. 43 1 1 1.084194604 1.067192815
44 1 1 1.189726422 1.058276597
TABLE 2. RESULT OF SIMULATION ON GERMAN 45 2 2 1.619310378 1.340606729
(STATLOG)CREDIT DATASET. 46 1 1 1.065225563 1.04640388
S . N O . Desired Predicted Output Predicted Predicted 47 1 1 1.057183393 1.025597673
Output Using BR Output Using Output Using 48 1 1 1.180454678 1.161966011
Technique GDA Technique LM Technique 49 1 1 1.041670678 1.06949032
1 1 1 1.047638778 1.032920792 50 1 1.0557 1.118502078 1.103677247
2 2 2 1.542768409 1.419541553 51 1 1.0484 1.109697008 1.113767098
3 1 1 1.017609411 1.016688575 52 1 1.0518 1.145080899 1.045512008
4 1 1.7321 1.098926107 1.264753276 53 1 1.1737 1.081085538 1.125618244
5 2 1.9562 1.424087928 1.548892156 54 1 1.0114 1.026963645 1.075093057
6 1 1 1.077017816 1.093028745 55 2 1.9247 1.349029922 1.251524064
7 1 1 1.027496592 1.024500469 56 1 1 1.158356539 1.229146983
8 1 1.0001 1.36144352 1.116357442 57 2 1.9991 1.208326235 1.147574084
9 1 1 1.035692649 1.046175544 58 1 1.0001 1.162475924 1.139279383
10 2 2 1.30622961 1.700113329 59 1 1.012 1.433291204 1.160497772
11 2 2 1.828641333 1.721826268 60 2 1.9463 1.689524114 1.383668768
12 2 2 1.87180088 1.728723793 61 1 1 1.191731034 1.163684108
13 1 1.7282 1.392841394 1.162822019 62 1 1 1.030666736 1.053153584
14 2 1.9512 1.156135281 1.044354737 63 2 2 1.749325264 1.095491458

155 | International Journal of Computer Systems, ISSN-(2394-1065), Vol. 04, Issue 07, July, 2017
Chandrahas Mishra et al Credit Card Fraud Identification Using Artificial Neural Networks

64 2 2 1.828311627 1.790042683 General formula, for finding the percentage error and
65 1 1 1.170017209 1.1011867 percentage accuracy of BR Algorithm, GDA algorithm
66
and LM algorithm are given below:
1 1.5368 1.079270005 1.270777663
Percentage Error = | average predicted outputs – average
67 1 1.0549 1.218556118 1.155691961 actual outputs | / average actual outputs *100 [26].
68 1 1 1.196594179 1.263808444
69 2 1.9845 1.230265841 1.151366221 Result 1: German (Statlog) Credit(100 samples)
70
(i) Accuracy of BR Algorithm:
1 1 1.096312785 1.184800802
Percentage Error = |1.265778- 1.25| / 1.25 * 100 = 1.26%
71 1 1.0001 1.411114368 1.171060863 Percentage Accuracy = 100 - 1.26 % = 98.74%
72 1 1 1.03161638 1.147513342 (ii) Accuracy of GDA algorithm:
73 1 1.015 1.247527991 1.223354795 Percentage Error = |1.23304- 1.25| / 1.25 * 100 = 1.37%
74 Percentage Accuracy = 100 - 1.37 % = 98.63%
1 1.0639 1.320294277 1.509414194
(iii) Accuracy of LM Algorithm:
75 2 1.016 1.309747118 1.090837544 Percentage Error = |1.194613- 1.25| / 1.25 * 100 = 4.43%
76 1 1 1.09840862 1.022306366 Percentage Accuracy = 100 - 4.43 % = 98.63% =95.57
77 2 1.9782 1.792446386 1.540253235 From the above calculation it is found that accuracy of
78 BR, GDA and LM on German (Statlog) Credit dataset is
1 1 1.06982001 1.061662392
98.74, 98.63 and 95.57 respectively.
79 1 1 1.116311494 1.275661241
80 1 1.017 1.393477528 1.282265909 Result 2: German (Statlog) Credit(1000 samples)
81 2 1.9219 1.117472606 1.274384359 (i) Accuracy of BR Algorithm:
82 Percentage Error = |1.3019258 - 1.3| / 1.3 * 100 =
1 1 1.041182375 1.022763057
0.1481%
83 1 1 1.138719775 1.076702358 Percentage Accuracy = 100 – 0.1481% = 99.85%
84 1 1.0008 1.155789092 1.023513808 (ii) Accuracy of GDA algorithm:
85 1 1.0019 1.109793706 1.028101449 Percentage Error = |1.243893918 - 1.3| / 1.3 * 100 =
86 4.3158%
1 1 1.040405987 1.034950339
Percentage Accuracy = 100 – 4.3158 % = 95.68%
87 1 1.0644 1.153431744 1.359950001 (iii) Accuracy of LM Algorithm:
88 2 2 1.408263423 1.378957161 Percentage Error = |1.221545331 - 1.3| / 1.3 * 100 =
89 1 1 1.300505009 1.562948448 6.0439%
90 2 1.9993 1.751692307 1.497327047
Percentage Accuracy = 100 – 6.00439 % = 93.96%
Above experiment is also performed on Australian credit
91 1 1 1.031891115 1.017125441 dataset on 690 samples and obtained the result as follows:
92 1 1 1.062935718 1.03697635
93 2 1.998 1.137041979 1.030115845 Result 3: Australian Credit Approval (690 samples)
94 1 1 1.043863035 1.139127501
(i) Accuracy of BR Algorithm:
Percentage Error = |0.4493-0.449| / 0.449 * 100 =
95 1 1 1.038027544 1.076816913
0.9810%
96 2 1.9927 1.84452728 1.446735268 Percentage Accuracy =100 - 0.9810% = 99.019% =
97 1 1 1.017513801 1.087322099 99.02%
98 1 1 1.164480757 1.126354948
(ii) Accuracy of GDA Algorithm:
99
Percentage Error = |0.4614- 0.4449| / 0.4449 * 100 =
1 1.0001 1.064395574 1.063488267
3.71%
100 1 1.0488 1.100594373 1.023050351 Percentage Accuracy = 100 – 3.71 % = 96.29%
Average Average = Average = Average = (iii) Accuracy of LM Algorithm:
101 =1.25 1.265778 1.234304 1.194613 Percentage Error = |0.4176- 0.4449| / 0.449 * 100 = 6.14%
Percentage Accuracy = 100 – 6.14 % = 93.86%
From above calculations performed in Result 2, it is
described the accuracy of BR, GDA, LM technique on
From the above table 2, it has been established that the Australian Credit dataset is 99.02%, 96.29%, 93.86%
result of the predicted output of BR, GDA and LM is respectively.
calculated by using testing data from the German (Statlog) V. RESULT AND ANALYSIS
credit dataset. Desired output is that output which is given
in the German (Statlog) credit dataset. As a result, the Credit card fraud identification system can be
average of actual outputs is 1.25 and average Predicted implemented using several techniques i.e. clustering,
decision tree, support vector machine (SVM) etc. but the
O/P using BR, average Predicted O/P using GDA and
most efficient technique is artificial neural network. Neural
average Predicted O/P using LM are 1.265778, 1.23304
network is supervised machine learning technique used for
and 1.194613 respectively. credit card fraud identification. Various learning techniques

156 | International Journal of Computer Systems, ISSN-(2394-1065), Vol. 04, Issue 07, July, 2017
Chandrahas Mishra et al Credit Card Fraud Identification Using Artificial Neural Networks

are available to train the neural network. In this paper we Figure 8 shows the performance of Gradient Descent
have discussed three learning techniques i.e. BR, GDA and Adaptive learning (GDA) technique used for training the
LM to train the neural network for credit card fraud neural network. It gives good accuracy for credit card
identification system. On the basis of Result 1 and Result 3 fraud identification using artificial neural network but less
shown in section 4, it is found that the accuracy of BR in comparison to BR technique. From the chart shown in
technique is 98.74% and GDA is 98.63% as well as LM figure 8, it is clear that the deviation of predicted output
technique is 95.57% on German (Statlog) dataset, while from actual output is more in GDA technique in
accuracy of BR technique is 99.02% and GDA is 96.29% comparison Bayesian Regularization (BR) technique.
as well as LM technique is 93.86% on Australian credit
dataset. Result 3 shows the accuracy of BR technique,
GDA technique and LM technique are 99.02%, 96.29%
and 93.86 respectively for German (Statlog) Credit dataset Performance of LM Technique
of 1000 samples.
Actual output Predicted Output

2.5
Performance Chart of BR technique
2

Output Values
Actual output Predicted output
1.5
2.5
1
2
otput values

0.5
1.5
0
1 1
8
15
22
29
36
43
50
57
64
71
78
85
92
99
0.5 # of samples

0
Figure 9. Performance of LM technique on German credit
92
1
8
15
22
29
36
43
50
57
64
71
78
85

99

dataset.
# of samples

Figure 7. Performance of BR technique on German credit Figure 9 shows the performance of LM technique used for
dataset. training the neural network. It gives average accuracy for
credit card fraud identification using neural network which
In Figure 7, Bayesian Regularization (BR) technique is is less in comparison to BR technique and GDA technique
used for training the neural network. It gives good both
accuracy for credit card fraud identification using neural
network.
ACCURACY FOR CREDIT
CARD FRAUD DETECTION
Performance of GDA Technique
SYSTEM
Actual output Predicted output 98.74 98.63
99
2.5 98
Accuracy in %

2 97
Output values

96 95.57
1.5
95
1
94
0.5 93
BR Technique GDA Technique LM Technique
0
Type of Technique
22

43
1
8
15

29
36

50
57
64
71
78
85
92
99

# of samples
Figure 10. Accuracy chart of techniques on German credit
dataset.
Figure 8. Performance of GDA technique on German credit
dataset.

157 | International Journal of Computer Systems, ISSN-(2394-1065), Vol. 04, Issue 07, July, 2017
Chandrahas Mishra et al Credit Card Fraud Identification Using Artificial Neural Networks

Accuracy of BR technique, GDA technique and LM The performance of neural network trained by LM
technique is shown in figure 10. Accuracy of BR technique for Australian credit dataset is shown in Figure
technique is 98.74 which is more than GDA (98.63) and 13.
LM (95.57).

Figure 11. Accuracy chart of techniques on German credit Performance of GDA technique
dataset (1000 samples)
actual output predicted output
Above figure 11 shows the accuracy of BR technique,
GDA technique, LM technique is 98.65, 95.68, 93.96 1.2
respectively for German (Statlog) dataset for 1000
1
samples

Output Values
0.8

Again same experiment is performed on different data 0.6


set i.e. Australian credit data set. BR, GDA and LM 0.4
techniques are used to train the network for Australian
credit dataset. Results are shown below. 0.2
0

31

73
1
7
13
19
25

37
43
49
55
61
67

79
85
91
97
# of samples
Performance of BR Technique
actual output predicted output
Figure 14. Performance of GDA technique on Australian credit
1.2 dataset.
1
Output Values

Figure 14depicts the plot of actual output and predicted


0.8 output of artificial neural network which is trained by
0.6 gradient descent adaptive technique for Australian credit
0.4
dataset.
0.2 The Bar Chart of accuracy for all the three
0 techniques used in this experiment is shown in Figure 15.
1
8
15
22
29
36
43
50
57
64
71
78
85
92
99

It can also be seen from the chart that accuracy of BR


# of samples technique is better than other two techniques i.e. GDA and
LM on Australian credit dataset.

Figure 12.Performance of BR technique on Australian credit


dataset.
ACCURACY ON AUSTRALIAN
Figure 12shows the performance of Bayesian C R E D I T A P P R O VA L D ATA S E T
Regularization technique on Australian credit dataset. 100
Chart shows the plot of actual output and predicted output. 99.02
99
98
97 96.29
Accuracy %

Performance of LM Technique 96
95
actual output predicted output 93.86
94
1.5
93
Output Values

1 92
91
0.5 BR Technique GDA Technique LM Technique
Techniques
0
1
8
15
22
29
36
43
50
57
64
71
78
85
92
99

# of samples Figure 15. Accuracy chart of techniques on Australian credit


dataset.
Figure 13. Performance of LM technique on Australian credit
dataset

158 | International Journal of Computer Systems, ISSN-(2394-1065), Vol. 04, Issue 07, July, 2017
Chandrahas Mishra et al Credit Card Fraud Identification Using Artificial Neural Networks

VI. CONCLUSION [9] Nutan Suman, “Review Paper on Credit Card Fraud Detection,”
International Journal of Computer Trends and Technology (IJCTT)
In this paper a deep discussion about the methods of V.4 (7):2207-2215 July Issue 2013.
training the neural network and their implementations for [10] C. Khyati and M. Bhawna, “Exploration of data mining techniques
credit card fraud identification system are discussed. It is in fraud detection: credit-card,” International Journal of Electronics
clearly shown that different methods provides different andComputer Science Engineering 1(3), 1765-1771, 2012.
performances. Accuracy of the system is also depending [11] E. Duman, and M.H. Ozcelik, “Detecting credit card fraud by
upon the architecture and learning methods of the genetic algorithm and scatter search,” Expert Systems with
Applications, 38, pp. 13057–13063, 2011.
systems.In this paper we have used three learning methods
[12] S. Ghosh and D.L.Reilly, “Credit card fraud detection with a neural
i.e. Gradient descent, Bayesian Regularization and LM network,” in: Proceedings of the Annual International Conference
technique. All the learning techniques are applied on same on System Science, pp. 621–630, 1994.
network and same configuration but we found different [13] E. Aleskerov, B. Freisleben & B. Rao, “CARDWATCH: a neural
accuracy i.e. 98.74%, 98.63% and 95.57% using BR, GDA network based database mining system for credit card fraud
and LM techniques respectively on German (Statlog) credit detection,” in: Proceedings of the Computational Intelligence for
dataset and 99.01%, 96.28% and 93.86% using BR, GDA Financial Engineering, pp. 220–226, 1997.
and LM respectively on Australian credit dataset . Now [14] J. T. S. Quah and M. Srinagesh, “Real-time credit fraud detection
using computational intelligence,” Expert Systems with
from above experiment it is clear that the accuracy of result Applications, 35, pp. 1721–1732, 2008.
increased when network is trained by Bayesian
[15] S. Panigrahi, A. Kundu, S. Sural, & A. Majumdar, “Credit card
Regularization Technique on both datasets i.e. German fraud detection a fusion approach using Dempster Shafer theory and
(Statlog) credit and Australian credit dataset. It is also bayesian learning,”InformationFusion.pp.354-363,2009A.
observed that when sample size is increased then the [16] Shrivastava, A. Kundu, S. Sural and A.K. Majumdar, “Credit Card
performance of BR technique is also increased while Fraud Detcetion using Hidden Markov Model,” IEEE transactions
performance of GDA and LM is decreased. Performance of on dependable and secure computing, Vol. 5, pp.37-48, 2008..
BR technique is better than other two technique i.e. GDA [17] S. Maes, K. Tuyls, B. Vanschoenwinkel and B. Manderick, “Credit
and LM for larger size dataset. Therefore Bayesian card fraud detection using Bayesian and neural networks,” in:
Proceedings of the First International NAISO Congress on Neuro
Regularization Technique is better approach for training Fuzzy Technologies, 2002.
the multi-layer feed forward back propagation neural [18] Ahmad Waqas, Abdul Rehman Gilal, Zeeshan Bhatti and Abdul
network for credit card fraud identification system. Waheed Mahessar, “Investigating ANNs and Applications,” Indian
Availability of the real world dataset is the key challenge Journal of Automation and Artificial Intelligence Vol: 1 Issue: 2
for credit card fraud identification system because financial February 2013.
institution don’t allow to share the historical data of their [19] K. Nirusha and R.V.Krishnaiah, “An effective approach towards
customers due to privacy policy. This approach can also be creditcard deception discovery methods,” International Journal of
Computerand Electronics Research [Volume 2, Issue 3, June 2013.
used in another application which requires classification for
[20] Nikita Gupta, Archana Dhankar, Karan Singh Ranawat and
further study. ShivyaNautiyal, “Credit Card Fraud Detection using Rough Sets
and Artificial Neural Network,” Journal of Computer Applications
REFERENCES ISSN: 0974 –1925, Volume-5, Issue EICA2012-1, February 10,
[1] Linda Delamaire (UK), Hussein Abdou (UK), John Pointon 2012.
(UK),”Credit card fraud and detection techniques: areview”, Banks [21] Sheela Thiruvadi and Sandip C. Patel, “Survey of Data-mining
and Bank Systems, Volume 4, Issue 2, 2009. Techniques used in Fraud Detection and Prevention,” Information
[2] BoxE. Don,Charytoniuk Wiktor, Lee Wei-Jen, Mo-Shing, Kotas Technology Journal, 10: 710-716, 2011.
Chen Paul and Olinda Peter Van, “Neural-Network-Based Demand [22] Azeem Ush Shan Khan, Nadeem Akhtar and Mohammad Naved
Forecasting in Deregulated Environment,” IEEE Trans. On Industry Qureshi, ”Real-Time Credit-Card Fraud Detection using Artificial
Applications, Vol. 36, No. 3, May/June 2000. Neural Network Tuned by Simulated Annealing Algorithm,”
[3] Zhang Guoqiang Peter, “Neural Networks for Classification: A Association of Computer Electronics and Electrical Engineers,
Survey,” IEEE Trans. On Systems, Man, And Cybernatics—Part C: 2014.
Applications and reviews, Vol. 30, No. 4, November2000. [23] Tanmay K. Behera, suvasini Panigrahi, “Credit Card Fraud
[4] Sankar K. Pal, Varun Talwar, and Pabitra Mitra, “Web Mining in Detection: A Hybrid ApproachUsing Fuzzy Clustering & Neural
Soft Computing Framework: Relevance, State of the Art and Future Network,” IEEE ICACCE, volume 33, 2015.
Directions,” IEEE Trans. On Neural Networks, Vol. 13, NO. 5, [24] Han, Jun; Morag, Claudio, ”The influence of the sigmoid function
September 2002. parameters on the speed of Back propagation learning," In Mira,
[5] Arash Bahrammirzaee “A comparative survey of artificial José, Sandoval, Francisco, From Natural to Artificial Neural
intelligence applications in finance: artificial neural networks, Computation. pp. 195–201, 1995.
expert system and hybrid intelligent systems,” Neural Computing [25] http://archive.ics.uci.edu/ml/datasets.html, UCI Machine
and Applications November 2010, Volume 19, Issue 8, pp 1165-
1195. Learning Repository.
[26] RohitMahajan, Sunil Kumar Gupta, Rajeev Kumar Bedi, “Design
[6] Hetvi Modi, Shivangi Lakhani, Nimesh Patel and Vaishali Patel, Of Software Fault Prediction Model Using BRTechnique,” ICICT
“Fraud Detection in Credit Card System Using Web Mining,” 2014, Procedia Computer Science Volume 46, 849 – 858, 2015.
International Journal of Innovative Research in Computer and
Communication Engineering,, 2013. 1(2): p. 175-179.
[7] S. Benson, Edwin Raj, A. Annie Portia, “Analysis on Credit Card
Fraud Detection Methods,” International Conference on Computer,
Communication and Electrical Technology – ICCCET2011, 18th &
19th March, 2011.
[8] Sowjanya Chintalapati and G. Jyotsna, “Application of Data
Mining Techniques for Financial Accounting Fraud Detection
Scheme,” International Journal of Advanced Research in Computer
Science and Software Engineering Volume 3, Issue 11, November
2013.

159 | International Journal of Computer Systems, ISSN-(2394-1065), Vol. 04, Issue 07, July, 2017

Potrebbero piacerti anche