Fraud detection system using neural network is totally based on the human brain working principal. As human brain it learns through past experience and uses its knowledge or experience [2] in making the decision about fraudulent or non-fraudulent. It can be configured by supervised and unsupervised learning technique [3] . Supervised learning algorithms are defined as the desired output is known for the input. Back propagation neural network is mostly used supervised technique [4] . Initially the Supervised learning algorithm uses a supervised training data. The training data consists of training examples such as the last one or two year data of particular consumer (occupation, income, large amount of purchased placed, frequencies of large Purchase, location) and the various ATM card fraud face by a particular bank previously. Then it analyzes the training dataset and produces a classifier. When test data is given to the input it Compare the test data with training data. If test data matches with fraudulent record than it will be fraud. Otherwise transaction will be done securely. The layer of neural network in ATM card fraud detection system [2] is shown in Figure 1.

The unsupervised techniques do not need the previous knowledge of fraudulent and Normal records. These methods raise alarm for those transactions that are most dissimilar from the Normal one. Self-organizing map (SOM) is one of the most popular unsupervised neural networks learning. It operates in two phase: training and mapping [5] . Initially the map is built and weights of the neurons are updated iteratively, based on input samples. Test data is classified automatically into normal and fraudulent classes through the procedure of mapping. After training the SOM, new unseen transactions are compared to normal and fraud clusters. If a new incoming transaction is similar to all previous transactions from genuine set, and then it is considered genuine otherwise it is fraudulent.

The Genetic Algorithms are evolutionary algorithms whose main objective is to obtain the better solution to the problem. It detects the fraud in real time and also minimizes the number of false alerts. The fraud that is detected is based on the customers’ behavior. The customer confidential information (such as the credit card usage frequency count, credit card usage location, credit card overdraft, current bank balance, average daily spending etc.) is stored in the data warehouse that is exposed to the rule engine which consists of the fraud rule set. The filter and the priority module set’s the priority of the information and then send it to the genetic algorithm which performs its function and generates the output. The process [6] is shown in Figure 2 and Figure 3.

A hidden Markov model (HMM) is a statistical model in which the system being

modeled is assumed to be a Markov process with unobserved state [7] . It works on the user spending profiles to detect frauds which can be divided into three types, such as: 1) Lower profile; 2) Middle profile; and 3) Higher profile [2] . The training phase and detection and prevention phase [8] [9] process is given in Figure 4 and Figure 5.

In this model at first start the bank sever and HMM server. When client initiates transaction, HMM starts observing and comparing the operation. It traps the transaction if identified fraud and is blocked .User reply with password on mobile using Bluetooth is same bank ATM else using SMS. Password is verified for authentication and transaction is allowed. Transaction is totally blocked after three failed attempts.

Bayesian networks are statistical techniques in data mining. The goal of Bayesian network is to correctly predict the value of a designated discrete class variable given a vector of predictors or attributes [10] . For the purpose of fraud detection, two Bayesian networks are constructed that describe the behavior of user. First, a Bayesian network is constructed to model behavior under the assumption that the user is fraudulent (F) and another model under the assumption that the user is a legitimate (NF). The “fraud net” is set up by using expert knowledge. The “user net” is set up by using data from non fraudulent users. During operations, user net is adapted to a specific user based on emerging data. By inserting evidence to these networks, the result of any transaction has been classified as fraudulent or non fraudulent behavior. Probability of fraud = P (F) then P (NF)

= 1 − P (F) in general and by applying Bayes rule, it gives the probability of fraud for any incoming transaction. The fraud probability that has obtained of training can be used as an alarm level.

A decision tree is one widely used machine learning technique that has been effective for classification or regression. Its usefulness results from the ability to compensate for missing values and having a highly flexible hypothesis space [11] . First, training sets are generated by selecting a set of measurements from a single smart meter within a particular time period for a particular customer. After the selection of the training dataset, this dataset is used to generate the decision rules representative of the normal energy consumption behavior model for the customer in question. Here, Prediction is achieved by using the generated decision rules to predict the expected energy consumption values based on the feature set (year, day of the week, time) of the validation set. For calculation the Root Mean Squared Error (RMSE) is a widely used statistical method. This calculation is used as an indicator of deviation between the predicted and the actual value in the validation dataset and is calculated [11] as Equation (1).

R M S E = ∑ i = 1 n ( y ′ t − y t ) 2 n (1)

Clustering is a process of arranging data into groups of similar objects. Different grouping results are obtained from various clustering methods available to group the dataset. The choice of a particular method will depend on the desired output [12] . K-Means clustering is a simple and efficient method to cluster the data. The system architecture [13] of the clustering based approach is given in Figure 6.

First, the variables used in this program are declared such as transaction amount, credit card number, new transaction, transaction date, merchant category id, and transaction type id and transaction country. Then validation process shows the validity of the details required for the transaction. Now the data table that is generated before is entered into the database. Then the data which is removing from the table is now entered to get transaction data. Then an array is used so that the transaction detail will produce row wise. After that clusters are named/labeled as low cluster, high cluster, and middle risky cluster. To detect fraud or genuine transaction now the current transaction details were taken by using k-means clustering algorithm. If it is fraud then the message will display “fraud transaction” or else it will display “legitimate transaction”.

The Support Vector Machines (SVM) is statistical learning techniques and has successful application in a range of problems. The basic idea of SVM classification algorithm is to construct a hyper plane as the decision plane which making the distance between the positive and negative mode maximum [10] . SVMs are a popular machine learning method for classification, regression, & other learning tasks. LIBSVM is a library for Support Vector Machines (SVMs). A typical use of LIBSVM involves two steps: first, training a data set to obtain a model & second, using the model to predict information of a testing data set. The main functions carried out by SVM are as 1) At first set up the training data for model creation. 2) Then set up SVM’s parameters for the dataset that is created and then send them for SVM training. 3) SVM Trainer, which trains each & every individual data from the large dataset. 4) Once the dataset is trained completely and the SVM Predictor does prediction of that trained data.

Artificial immune systems (AIS) represent an important strategy inspired by biological systems. The main developments within AIS have focused on three main immunological theories: clonal selection, immune networks and negative selection. The immune system can distinguish between self and non-self [14] . In the concept of credit card fraud detection, self (S) represents all patterns in a finite space that is legitimate and non-self (Ŝ) represents all patterns that are not in self. The AIS consists of artificial lymphocytes (ALCs) that able to classify any pattern as self or non-self by detecting only non-self patterns. For training AIS system only needs positive examples but it can identify items as non-self. The system arbitrarily generates an ALC, test it against the set of self patterns and if it doesn’t match any of the self patterns, it is included in the set of mature ALCs. When an ALC does match with any of the self patterns, it is replaced by a new randomly generated ALC which then needs to be tested as well. The ALC becomes mature or adult, by training it with the known self patterns. This training method is known as negative selection. High level model of AIS has been applied for credit card fraud detection. The aim of the AIS system is to have a high anomalous transaction detection rate and a low false positive rate [10] .