Data Science: Concern for Credit Card Scam with Artificial Intelligence

Abstract

In nowadays time we are on the specific education to a cashless society. Similarly, the use of credit card in developing international locations has grown to be a not unusual scenario. The world goes to be digital by means of making the availability of services to peoples together with credit score playing cards, ATM, Internet banking, and mobile banking services. With the rapid increase of the e-trade or online business the use of credit card is so common as it makes the payment easier and saves a time. There are lots of blessings of using the credit card system such as ease of purchase, preserve purchaser credit score history, protection of Purchases, and so on. As we recognize each coin has two faces so the usage of credit card has hazards too inclusive of credit score card fraud in which the hacker theft the credit card info and the misuse the credit score card by means of stealing the details of the card together with credit score card range, CVV, expiry date, etc. Fraud is intentional hypocrisy for the motive of obtaining economic profits or inflicting losses by means of implicit or express deception. However, Fraud is the violation of public law wherein the scammer profits an unauthorized advantage or causes unlawful damage.

1 Introduction

As of now, Unlawful use of a cc and obtaining its private information without the owner's knowledge is considered cc fraud [2]. Credit card fraud is increasing day by day with the development of technology leading to the loss of billions of dollars of customers around the world every year. Hence, Fraud detection involves identifying fraudulent activity scattered across many legal transactions as quickly as possible. This is a widespread event problem called outlier analysis, anomaly detection, exception retrieval, rare class retrieval, unbalanced data retrieval, and so on the amount of fraud transitions is much less than the total number of transitions, so the accurate detection of the fraud transition is very difficult and questionable for this, we must use very efficient methods and algorithms. Therefore, in this article we try to collect and integrate the whole series of research in the literature and to analyze it from various aspects. According to the World Payment Report, in 2016 total non-cash transactions increased 10.1% from 2015 for a total of 482.62 billion transactions. Actually, the credit card fraud techniques are mainly categorized into two category application fraud and the behavioral fraud. Application fraud occurs when scammer request new cards from the bank or issuing companies using the wrong information or obtain the wrong information from the other [3]. However, multiple requests can be made by a single user with the same user details or by a different user with identical details. On the other hand, behavioral fraud has four main types: mail theft, fake card, stolen/lost card, and the cardholder has no fraud [5]. Although Due to the fraudment of the credit card the flip side fraudment rise we move to the EVM smart chip-based card for security purposes (Figs. 1, 2 and 3).

Fig. 1

The data of credit card that needs to be confidential by the owner of the credit card in the credit card to protect it from the hacker [6]

Fig. 2

The steady growth of the non-cash transactions [2]

Fig. 3

Due to the fraudment of the credit card the flip side fraudment rise we move to the EVM smart chip-based card for security purposes [5]

1.1 Categories of Credit–Card Fraud Finding

There are two kinds of main findings:

• a. Finding of misuse

• b. Finding of Anomaly.

Extortion of misuse identification is a procedure that manages directed grouping movement at the exchange level. In these strategies, exchanges are hailed as extortion or ordinarily dependent on verifiable information from the client's past progress model [4]. This dataset is utilized to make the grouping model that can foresee the state (typical or false) of new records. There are numerous models for making techniques for the common two-class grouping task, for example, rule enlistment, choice trees, and neural systems. This methodology has demonstrated to be the dependable identification of the majority of the misrepresentation recommendations that have been seen previously. It is otherwise called misuse location. Utilization conduct examination misrepresentation (abnormality recognition) manages solo location techniques dependent on account conduct [3]. In this strategy, an exchange is identified as extortion in the event that it clashes with typical client conduct. This is on the grounds that we don't anticipate that con artists should act similarly as the record holder or to know about the proprietor's example of conduct [5]. To do this, we have to remove the real client conduct model, for example, the client profile for each record and afterward recognize false exercises dependent on it by contrasting new practices with agreeing with this model, various exercises that are adequately extraordinary are viewed as misrepresentation. Profiles may contain data about record movement, for example, trader types, sum, spot, and season of exchanges. This strategy is otherwise called irregularity detection [5]. In the table beneath, a few strategies are quickly introduced speaking to some current extortion discovery methods that are applied to the assignments of the charge card misrepresentation recognition framework. It additionally speaks to the favorable position and detriment of each approach [5, 5] (Fig. 4).

Fig. 4

Due to the fraudment of the credit card the flip side fraudment rise we move to the EVM smart chip-based card for security purposes

2 Methodology

User comes and selects the transaction methods and after that the process is matched with the stored datasets and is seen that is there any issues in the process of the transaction if the datasets find any issues then the transitions are rejected and the fraud is confirmed [6]. For further if there is any issues like the user isn’t fraud then security questions are checked and if the security question can’t be answered then it is a complete fraud and the system exits the user from the transaction. The flow chart diagram of the Credit Card Fraud detection technique that is used by us is of the user behavioral model that is shown in the figure below. In this the new transition is matched with the past transition pattern such as amount of transition pattern, location of the transition and the type of purchasing the transition if the system found it fraud than it raises the security question if it answered than transition further proceed otherwise the system will abort the transition and report to the holder about the fraudment of the transition [10] (Fig. 5).

Fig. 5

Block diagram of credit card fraud detection system

2.1 Stages of CC Fraud Detection Classification Methods

The Credit card Fraud Detection system is an 8 step process in which we will use different machine learning algorithms and use them to analyze data import dataset plot different graphs use ANN that is artificial neural network to recognize patterns will see different curves.

• 1. Importing Datasets

• 2. Data Exploration

• 3. Data manipulation

• 4. Data Modeling

• 5. Fitting Logistic Regression Model

• 6. Fitting Decision Model

• 7. Artificial Neural Network

• 8. Gradient Boosting.

In the very first step import the library and different datasets and then when it is successful then move to our second step and which is data exploration. After that there is a need to explore the imported data that will manipulate it with step called data manipulation and after that Model the data. The Data Modelling is an important part as it will compare with many states and transactions with this modeled data and if there is any issue in modeled data then the system won’t work properly [].

Now from the next step start using different ML algorithms, at first do fitting logistic Regression Model. Then apply fitting Decision Model after that apply ANN artificial neural networks and at the end do gradient boosting [].

• a. Importing Datasets: Here import the datasets that contain exchanges made by Visas or check card holders. We have utilized Kaggle.com which has the pre-made informational collections which contained 285,500 lines of information and 32 sections. Out of the apparent multitude of segments, the main ones that are the most sense were Time, Amount, and Class (misrepresentation or not extortion). The other 29 sections were changed by utilizing what is by all accounts a PCA dimensionality decrease so as to ensure the client personality [11].

• b. Data Exploration: In this we will investigate the information that take a gander at various fields’ tables and different subtleties like kind of charge card utilized generally the territory which has more number of fakes. We investigate the information that is contained in the MasterCard information outline. We will continue by indicating the charge card utilizing the head () and the tail () work.

  Now the data that use to run through a few initial comparisons between the three columns that Time, Amount, and Class (Fig. 6).

  Fig. 6

  

  Comparison train data of time, amount and class

  Now we have the data that we want to run through a few initial comparisons between the three columns that Time, Amount, and Class.

  Figure 7 is depicting that, while most of the big deals are very small, this distribution is also planned. Most of the time, day-to-day transactions aren't overly expensive (most are <$50), but they're probably the most fraudulent transactions to happen too.

  Fig. 7

  

  Distribution of amount of credit card data

  Figure 8 is depicting that the visible distribution is valid data for two days. It is for regular consumers for most purchases made during the day and when people leave work/school and go home, purchases decrease overnight. Figure 9 is illustrating the Class (Fraud/Not Fraud).

  Fig. 8

  

  Time distribution of credit card data

  Fig. 9

  

  Fraudulent and Non-Fraudulent distribution of credit card data

  In the dataset, there is only 493 fraud transactions, i.e., Only the approx. 0.179% of all of the transactions in this dataset.

• c. Data Manipulation: The information control is applied to the sum part of Visa informational collection. In this we do Scaling. Scaling is otherwise called the exceptional normalization by the assistance of which the scaling of information is organized as indicated by a particular range. When we've normalized our whole dataset, we'll split our dataset into a preparation set and a test set with a split proportion of 0.83. This implies 83% of our information is given to preparing information while 17% of information is given to test information [12] (Fig. 10).

  Fig. 10

  

  Divison of test data and train data

• d. Data Modeling: Later we have normalized our whole dataset, we will partition our dataset into preparing set just as test set with the split proportion of 0.83 which implies that 83% of the information is ascribed to the preparation information though 17% of the information is given to the test information [9, 9].

  Figure 11 is depicting that, here not added any numbers because it would be very, very difficult for the reader to see. So look for anything that shows strong correlation.

  Fig. 11

  

  Correlation among time and amount of credit card fraud

• e. Logistic Regression and Random Forests: Calculated relapse is a factual model that endeavors to decrease the expense of an awful gauge. The Random Forests calculation is a gathering of choice trees that all things considered anticipate whether the change is a misrepresentation or not. In the event that progressing is a fake, I run my information through these two models and get extraordinary outcomes. The complete score for calculated relapse and arbitrary woods models is exceptionally encouraging for our dataset. Each model has a high evident positive rate and takes into consideration a bogus positive rate which we need here [14].

• f. Fitting Logistic Regression Model: In this credit card fraud detection system project, we will adapt to our standard model. So let's start with logistic regression. In our case, we are trying to find out if it is a fraud/non-fraud transition. [6, 6] (Fig. 12).

  Fig. 12

  

  Logistic regression of credit card fraud/non-fraud transition

• g. Decision TREE: In the choice tree—Regression, the choice tree makes relapse or characterization models as a tree structure. We partition the dataset into littler and littler subsets and simultaneously build up a related choice tree. The final product is a tree with choice hubs and leaf hubs. The choice tree can contain both numeric and clear cut information (Fig. 13).

  Fig. 13

  

  Decision tree—regression model for credit card fraud/non-fraud transition

• h. ANN Artificial Neural Network: ANN's are the sort of AI calculation (ml) methodology displayed based on the human sensory system. ANN models picked up utilizing authentic information and rank dependent on input information. The information imported from the neural system bundle permits us to actualize our ANN. At that point we can plot it utilizing the plot work. Presently, on account of fake neural systems, there is a scope of qualities somewhere in the range of 1 and 0. We characterize limit esteem. 0.5 Otherwise the worth more prominent than 0.5 will be 1 and the rest is 0 [9] (Fig. 14).

  Fig. 14

  

  Artificial neural networks for credit card fraud/non-fraud transition

3 Proposed Algorithm

Some common methods are used for making the system. Algorithm used: HMM (Hidden Markov Model (HMM)); Semi HMM; Multiple HMM; Optimized HMM, and Advanced HMM.

The comparison between different algorithms on their accuracy here the optimized HMM is most accurate in the functioning whereas the HMM is lead Functioning. Whereas the dataset trading using R language is also a good accurate algorithm this functionally performs equally to the optimized HMM. Among all the methods of finding the fraud detection technique system build the Hidden Markov Model (HMM) is the one of the best and accurate method for the detection of the fraud transition [7]. HMM method has the very high accuracy among all the methods as shown in the figure below among all the methods of building the fraud detection system techniques (Fig. 15).

Fig. 15

Accuracy comparison of different methods

3.1 ROC Curve of the Algorithms

A Receiver Operator Characteristics curve (ROC) is a graphical diagram used to show the diagnostic capability of binary classifiers. It was first used in the theory of signal detection but it is now used in many other field such as medicine, radiology, natural hazards, and machine learning [8] (Fig. 16).

Fig. 16

Receiver operator characteristics curve of various methods

4 Issues: CC Fraud—Detection

Fraud detection system is subject to a lot of trouble and has a lot of problems mentioned below. Effective fraud detection techniques that most should have the ability to eliminate these difficulties to obtain and achieve the best performance.

• a. Data Misbalancing: credit card data fraud detection is skewed, which means that a very small percentage of all credit card transactions are fraud. This makes detection difficult and inaccurate fraudulent transactions.

• b. Misclassification of data: In a fraud detection system tasks, different misclassifications have different interests. One classifies normal transactions such as fraud harmless as normal detection fraudulent transactions. Because in the first case of misclassification are identified by further investigation [].

5 Conclusion

To find the fraud transactions with credit cards are actually important specifically in digital society. So detection of credit card fraud is very multifarious and important matter that entails the large number of data analysis and a very precise system which will detect fraud in maximum number of cases and also able to handle error and some other inputs. It can be considered as part of machine learning and data science and this type of system ensures the customer and organization that the money transfer is always safe and there is no loss of the money or anything. Future work will focus on more about the fake bin detection which authors covered in the introduction part. This is one of the most important things to look on as there is much fraud using this bin system that authors analyzed. Another thing is the POS hacking or hijacking which is also an important task to add in it. The POS hacking is also a big fraud and we should always look on that to and try our system to tackle this problem also.