In this era of digitization, credit card frauds shake down the spirits of not only the customers but also the merchants, which incurs a loss of billions of dollars globally. To combat such frauds, a robust and responsive system is needed that can flag the fraudulent transaction instantly before it happens. The existing systems are great at detecting and battling with fraud after it has happened but slouch in case of prevention of such crimes. They aren't good¬¬ at optimization and also struggle in terms of response time. The ineficiency of existing systems is attributed to either working on a single machine learning technique, or just combining two of them. We present a Trio-Hybrid of K-means, Genetic algorithm, and artificial neural network approaches to deal with the aforementioned problems. The K-means algorithm helps in reducing the training time of neural networks and the genetic algorithm helps in feature selection to prevent the neural network from being over-trained, thereby reducing the cost of the system. We leverage the benefits provided by these three techniques and put them together into a trio for the first time and achieve an accuracy of 99.94% in detecting the fraudulent credit card transactions.
The unauthorized and ill-intended use of credit cards to commit a crime and causing monetary
harm to its owner is defined as Credit Card fraud [
Applications and Use Cases of a Credit Card Fraud Detection System 1. Any terminal that provides the electronic hardware used to swipe cards such as the Point of Service terminals at retail stores. 2. Assistance at the merchants’ site for example merchants like insurance companies will be able to identify a fraud client easily. 3. Check at the bank’s end. Banks may take a step on detecting fraud taking place associated with an account in their bank.
Challenges Involved in modelling a Fraud Detection System 1. Inaccessible data or lack of availability of data. The data associated with a customer’s credit card and account information is highly confidential hence no bank or company easily avails the dataset of their customers. As a result, the data is not readily available. 2. Imbalance in the data i.e. the number of fraud transactions is very less as compared to the genuine ones. 3. The behavior of a fraudulent profile keeps changing and its nature is dynamic. 4. The time taken by the system to decide whether a transaction is fraud or not must be very less. 5. Overlapping of transactions i.e. a genuine transaction’s nature is very similar to a fraud one. 6. Features and parameters to be processed are very large in number. 7. Selecting the optimal parameters or features is a demanding and challenging task.
The pattern followed by all the fraudulent transactions is generally very similar and we can categorize some transaction as fraud using any of the following pattern recognition systems as K-Nearest Neighbour (KNN), Artificial Neural Networks, Fuzzy Logic Based System, Artificial Immune System, Naïve Bayesian Network, Hidden Markov Model, Support Vector Machine (SVM), Decision Trees, Ensemble Classifier and Logistic Regression. We should know their advantages and disadvantages in order to leverage the benefits, when some technique is chosen to be applied. Table 1 shows the certain advantages and disadvantages of various techniques for credit card fraud detection.
This work presents a novel trio-hybrid of artificial neural network, k-means clustering, and genetic algorithm (GA) that can precisely and accurately detect and prevent fraudulent credit card transactions while they are in transit. The major contributions and the objectives of the proposed work are (i) to minimize the time required to train the neural network system by using k-means clustering, (ii) utilizing a genetic algorithm to prevent the system from being over-trained, thereby reducing the cost of the system. We have provided an algorithm for the trio-hybrid of the three mentioned approaches and have found an Accuracy of 99.94% and Loss Value of 0.561%. The paper has the following sections: Section II is the Literature Review of the noteworthy existing works that are parallel to the proposed solution and a comparative analysis between them followed by some of the benchmark systems in credit card fraud detection. Section III contains an explanation of the methodologies used in the proposed model followed by algorithms used. Section IV is the operational analysis which contains the architecture and lfowchart used in the system and the combined algorithm is also explained. Section V explains the metrics on which the system is evaluated and summarizes various operational results and ifndings. Section VI has the conclusion along with the future scope of the proposed solution.
A number of researchers are doing experiments to achieve better accuracy and reduce time for detecting the credit card frauds [6, 7, 8, 9, 10]. [11] have provided a review on the diferent credit card fraud detection practices. [12] proposed a hybrid of Bayesian Networks and Artificial Neural Networks as a technique to detect frauds in credit cards. Their paper consists of a discussion which states the speed of Bayesian networks was accelerated by ANN and after a short training period, they gave good results. [13] put forward a hybrid technique of artificial neural networks and decision trees. Firstly, the classification results obtained by the Decision tree and Multilayer perceptron generated a new dataset. This new dataset is then fed into Multilayer perceptron for the final classification of the data. High reliability is obtained by this model as a result of a very low false detection rate. [14] in their paper have devised a hybrid of K-means clustering combined with the Hidden Markov Model (HMM) and Multilayer Perceptron (MLP). The dataset is fed to K-means and then its output is given to HMM and MLP for their training which classifies the incoming transaction. As seen in the observations, the combination of “MLP with K-means clustering” gives more accurate results or higher accuracy. However, when used with 10-fold cross-validation the result is reversed. [15] in their paper Credit Card Fraud Detection Using Autoencoder Neural Network have proposed a de-noising autoencoder neural network (DEA) algorithm to handle the imbalanced nature of Credit Card datasets along
Advantages Disadvantages • ANN has ability to learn from the past. It • High processing time in case of
does not need to be reprogrammed. large neural networks. • ANN is capable of detecting the fraudu- • Excessive training required. It is diflent activity during the transaction. ficult to set up and operate. • High detection and processing speed.
• Excessive training needed. • High accuracy • Expensive • Sensitivity to data format. • Poor at processing large datasets • Expensive. It has a low speed of
detection. • Medium accuracy. It lacks trans
parency • Low speed of detection • Highly expensive • Cannot detect fraud during the
transaction. • The algorithm is complex. Even a small change in data can distract the structure. • Deliver a unique solution, by choosing an
appropriate generalization code • Robust • Very fast in detection/accurate • High maintainability • High flexibility. Explainable • Easy to understand and implement • Can handle nonlinear data as well • Capable of detecting frauds at the time • Cannot detect fraud in the initial
of the transaction few transactions. • Reduces the false positive • Not scalable to large size datasets • Expensive • Predictive model is not required before • The method accuracy depends on classification the measure of distance • Cannot detect fraud during the transaction. with SMOTE (Synthetic Minority Oversampling Technique) and SoftMax function in neural network classification to model the system. [ 16] in their paper have trained their model using Artificial Neural Network and three diferent learning mechanisms Gradient Descent Adaptive Learning, Bayesian Regularization (BR) and LM algorithm, and found that BR gave the best results. Table 2 shows the various systems parallel to the proposed system. [12] [13]
84% detec- imbalance and noise in the minority class of the tion rate dataset could be removed using the autoencoder method. r9da5et.t5ee7c%tion tsTBouhaleyatsrepespwialaiytrnheoRntvheatgerhiuboeleuansrstiesttaarwcatcioionurnirkna.TgcIeyttcechwconahmisnqpiufqaoeruuapentsridovaevtvlihydaa.eitldatbhrleee
r9da9et.t5ee8c%tion ftuMehsreeientnnhgtopPddrreaiaptnanrcosdiepctMeasslistnCae–kdoMemunapsfxiornonsgmceanAlttarhbrAeiantKlragaaloyrgyrsgiitlsAeh(smrPseiCpgfooAnrsm)iNtaeononrdrytmalization.
Neural Networks have the powerful capability of identifying patterns and the correlations and diferences among those patterns. K-means Algorithm can detect outliers even in overlapping pattern set and Genetic Algorithm uses powerful concepts of evolution to generate an optimized system. Together they make a very fast, precise, and powerful fraud detection system. The advantages provided by these three techniques are leveraged in this tri-hybrid approach. In this section, the working of each of them is explained in detail. 3.1. K-means Clustering Fuzzy Clustering also called soft clustering or K-means clustering is a clustering technique that separates data points into diferent clusters based on how much similar are they to each other and how much they difer from other data points [ 18]. In fuzzy clustering, one data point can belong to more than one cluster. Clusters are diferentiated based on similarity measures such as intensity, connectivity, and distance. Fuzzy Clustering makes our system have higher accuracy and Lower False Alarm Rates. Given n data points 1, . . . , K-means aim to find k no. of centers 1, . . . , and assignments 1, . . . ., of the data points to the centers such that sum of distances is minimized.
(1......, 1....) = ∑︁ ∑︁ * || − ()|| *
=1 =1
The first center c1 is selected at random from the data points x1,. . . ,xn, and then the distance
between this center and all points xic1p is calculated. The second center c2 is then selected
from the data points with their probability proportional to the distance. Using the minimum
distance to the centers collected so far; repeat the procedure to obtain other centers. Each data
point is assigned to the cluster from which it has a minimum distance. To calculate the distance
of each data point from each of the centroids Euclidian Distance has been used. Other methods
that can be used are Cosine Distance (cosine of the angle between the data points), Manhattan
Distance (absolute diference between coordinates of the two data points), and Minkowski
Distance (average or generalized distance). Once each data point has been assigned a cluster,
we re-calculate the centroid as a mean of all its constituting data points. Then a cluster of all
the data points is re-calculated and this process repeats till no data point shifts between clusters.
The output of clustering is used to find out the transactions that are the outlier. An outlier is an
object that is inconsistent concerning our data. They do not confer to the normal data and hence
need to be evaluated. In the case of credit card fraud detection, we have a highly unbalanced
nature of data i.e. the number of data of frauds in comparison to a genuine transaction is very
minuscule. Hence finding transactions that are the outlier and sending only those to the neural
network for classification not only makes our system better trained but also faster. Outlier
detection is done by calculating the distance between each transaction in a cluster with the
center of the cluster. All the transactions that fall above a threshold value that is calculated as
the average of all the distances are assigned as the outlier.
3.2. Genetic Algorithm
A genetic algorithm is an evolutionary optimization technique [
The aforementioned parameters are chosen over the others because they are the most crucial parameters which play an important role in the correct classification. Hence, these parameters need to be made stronger than the others so that they have a greater say in output. If the number of layers is less than this could lead to weak computing and faulty processing while a large number of layers would cause the neural network to slow down. Hence, it is optimized using a genetic algorithm. Generally, how many neurons would comprise the input layer is decided by the number of variables in the input dataset which is being processed. The calculation of the number of neurons in the hidden layer is tricky but by the rule of the thumb, they should be smaller than or equal to the minimum number of neurons in the input or output layer approx. 23 the size of the input layer plus the output layer. This is more accurately computed by feeding data on synapses to the genetic algorithm. A genetic algorithm helps the neural network by discarding the unnecessary as well as the insignificant neurons; thus, speeding up the learning. The dense layer activation function plays a major role in deciding which neuron will be activated which in turn has a significant impact on a correct classification. Network optimizers play a significant role in minimizing the loss function thereby contributing towards the error-free output. Hence, it needs to be chosen for optimization for genetic algorithms. 3.3. Neural Network A neural network is a network that comprises several nodes (neurons) present in each layer. Each node of a layer is connected to every other node in the next layer and each edge connecting them has a weight assigned to them [19]. Activation of each node in the next layer depends on the sigmoidal function which computes data in each node in the previous layer by using the weights of the edges connecting them and some bias to give the activation of the next node. The nodes in layer one activate the nodes in layer two and further the process continues until ifnally a node in the final layer is activated, which is considered as the output. The total weight on a neuron Y with three input neurons will be
1 (1 + ) − And activation of neuron Y will be given by
= = 11 + 22 + 33
Feed Forward Back Propagation Learning Algorithm In this fraud detection system, a five-layer feed-forward back propagation neural network is being used with supervised learning which is one of the most powerful learning algorithms [20]. The feed-forward will diagnose the transactions while the back propagation will calculate errors generated and accordingly correct the weights on the edges as they play the main role in the activation of nodes and thus in the output. Supervised learning means that we will be provided with the pair of input and output data values and will compare the generated output with the desired output to calculate the percentage of error. ℎ+ = * * (1 − ) The feed-forward back propagation algorithm carries the data in one direction and does not allow loops either in a forward direction or in a backward direction.
The dataset applied in the present study is taken from the Kaggle repository, a subsidiary of Google LLC and is available at https://www.kaggle.com/mlg-ulb/creditcardfraud/home. It contains transactions made by credit cards in September 2013 by European cardholders over two days, having 492 frauds out of 284,807 transactions. The dataset is highly unbalanced with the positive class (frauds) accounting for 0.172% of all transactions. It has in total 31 features out of which 28 correspond to attributes of a customer like name, age, occupation, location, account balance, type of card, etc. For security purposes, they have been PCA transformed. The other three features are time, amount, and class. 4.1. System Architecture 4.2. Flowchart Figure 2 depicts the flowchart of the working of the system. The dataset used is partitioned into two sets, Train Dataset and Test Dataset in the ratio 4:1. The training dataset then goes into the K-means clustering algorithm where clusters are formed and the transactions are then categorized to be of either low risk or high risk. If the transaction is of low risk then the bias send to the Neural network is lower otherwise it is higher. This bias of each transaction is input to a genetic algorithm which transfers this information to the neural network. A genetic algorithm (GA) also optimizes the parameters and the initial weight matrix that is input to a neural network for training. GA first randomly generates a generation, evaluates its fitness using fitness scores, and generates further generation using selection criteria of crossover and mutation. Once the stopping condition is met the dataset with the information of bias and parameter selection (weight matrix) is send as input to Neural Network to get trained by using the feed-forward back propagation algorithm. After training has been completed, a trained system is achieved on which test dataset runs to find the accuracy of the system. 4.3. Algorithm Development The dataset is firstly pre-processed by splitting into train dataset and test dataset. Split the dataset into 70:30 ratio for training and validation (testing) datasets i.e. 70% of the dataset rows are used as train dataset and 30% as the test dataset. The training dataset contains the rows and columns which would be used for training the network. The test data is then used against the trained network to measure the accuracy and other evaluation metrics. The last column “class” of the dataset is dropped as it is not needed during the training part. Removal of non-constant features is useful for better training of the model. After forming the clusters in the dataset, the distance of each point is calculated from its center, this is the K-means application. These data points are nothing but diferent column names in the dataset like be the location of the transaction, amount of transaction, etc. The closest cluster to each point is predicted iteratively until there is no shift between the clusters. This process optimally reduces the dataset for the genetic algorithm as the assignment of high and low bias has been included in each data point’s information. The genetic algorithm now plays its part on these data points for the best feature selection. These features or parameters decide the structure of the neural network. The genetic algorithm forms the initial generation by creating a random combination of features. For example, for one neural network no. of layers might be 4 and for another, it might be 5. For each of these combinations of genes, the genetic algorithm performs mutation on some, crossover on some, and uses some (which have the highest score) directly for the next generation. As a result, the best suitable structure for the network is decided and then this network is formed using the Sequential() method which is a predefined method in the Sequential class of the Keras library. As the neural network is formed using the parameters given by the genetic algorithm, the test dataset comes into play. This data is then given to the neural network and the evaluation metrics are measured as the network shows which transaction is fraud and which is genuine to raise the alarm.
Here are the steps involved in the algorithm explained above. The input and the final output are mentioned before the algorithm begins. The input is a CSV file named creditcard.csv which contains the PCA reduced values of various transactions. The column names denote the features which will be clustered using the K-means clustering and the rows in the dataset are values corresponding to diferent transactions.
Algorithm: Trio-Hybrid Algorithm for Detection of Fraudulent Credit Card Transactions Input: credicard.csv dataset Output: Scalar consisting of loss and the values of the metrics. 1. Pre-process the dataset • Split the dataset into train dataset D and the test dataset. • Drop the column “class” of the dataset.
• Scale the train dataset D for standardization i.e. remove non-constant features. 2. Apply the techniques i.e. K-means clustering, genetic algorithm, and neural network. 2a. K-Means Clustering • Form clusters and divide the sample data D among all clusters. • Calculate the distance of each point in sample space from all the clusters using K-means algorithm. • Predict the closest cluster to which each sample in D belongs. • If the sample point is already closest to its own cluster
– Stop Else
– Move the sample point to the cluster it has least distance with. • Repeat the above step until no sample shift between the clusters. • The dataset has now been optimally reduced after the data points have been grouped accordingly into the clusters having similar data points. 2b. Genetic Algorithm • The data points with their cluster information i.e. high risk or low risk are then used by the genetic algorithm to form a random combination of parameters such as number of layers in the network, number of neurons in the layer etc. to be used in the formation of most suitable neural network. This random combination forms the first generation for the genetic algorithm. • Store this random combination of parameters into a variable GENES. • Calculate the fitness score upon each of the generated combination or GENES by mean squared error fitness () function: = 1 ∑︀ =1( − )2 Lesser the score, fitter the GENES. • Sort the population of these combinations into increasing order of fitness score. • If fitness score is found as 0
– Stop Else – 10% of the current fittest population is taken to be included in the next generation – 50% are mutated to produce ofspring i.e. take two combinations and permute the combination of parameters i.e. perform crossover – For the rest of the GENES, insert a random parameter into their combination i.e.
perform mutation.
– Repeat the above steps until a best score is achieved for a set of parameters. • The most suitable parameters for the neural network are thus obtained after best feature selection by the genetic algorithm. grouped accordingly into the clusters having similar data points. 2c. Artificial Neural Network • Create object of Sequential() • Use this object to add input and output layers to the model (making of the network). • Fit the data into the network i.e. train the network using Epoch=200, batch size=500 • This network is then tested against the test dataset by considering the output produced which is a scalar consisting of loss value and metrics value to indicate the fraud and genuine transactions.
End (a) with k=10 (b) with k=20 (c) with k=15 (d) Accuracy and FAR at diferent values
The results obtained by experiments and the metrics on which they are evaluated are explained below in the two subsections. The results are analyzed and discussed through graphs and figures obtained through experiments on diferent parametric values. No proposal can be modelled into a system without some experiments to support it. The results and outputs included have been produced by this system under various inputs and parameters. To get the optimal value of K in the K-means algorithm we varied the number of clusters from 10 to 20. Figure 3a shows clustering when K=10 and Figure 3b shows clustering when K=20. It depicts that with an increase in the number of clusters there is a higher chance of forming better clusters. While running the K-means clustering algorithm on diferent K values it was observed that the best accuracy is given when K was 15. Figure 3c shows the output produced by the K-means clustering algorithm when K=15. In figure 3d the accuracy of the algorithm at diferent K values is shown. Fig. 4(a) and Fig. 4(b) shows the cross-validation score of best genes and average genes in each generation at seed value 108 and 10000 respectively. The relationship between initial seed values given to genetic algorithm and mean square error of cross-validation score generated for that seed has been depicted. It can be observed from Fig.4(c), that when the seed is around 5000 the mean square error is minimum. It increases both decreasing the seed value as well as increasing the seed value. It is shown that before feature selection of the most important features, the mean square error was 37.13% which was reduced to 28.92%. Although the number of generations taken for the average and best features in graphs Fig.4(a) and Fig.4(b) to reduce the fitness score (lesser the score, better the generation) is a little lesser compared to Fig.4(c) i.e. on a seed value of 5008, it is compensated by the reduced mean square error. This score hugely depends on the initial seed value. Fig.5a shows the time required by the neural network to train at diferent Epoch values. Fig.5b depicts that when we increase the batch size the time taken to train the neural network gets reduced. With the increase in batch size, the loss value decreases up to a certain batch size which is shown in Fig.5c. We get the least loss value at a batch size equal to 500 and after that increasing the batch size though decrease the time taken by the neural network to get trained but it also increases the loss value. After getting a trained neural network we run the test data set on the system and the observations made on the output are mentioned below. Accuracy = 99.94% Loss Value = 0.561%
The system described here is faster than the present systems as the training time of neural networks is reduced by the use of the k-means algorithm and the eficiency is increased by the use genetic algorithm. The genetic algorithm helped in the best parameter selection and removed the redundant parameters. 100 epochs are ideal for this fraction of the dataset since lesser than this will cause under-fitting of the system and if more epochs are used then the system will be over-trained for that particular dataset. Henceforth, the results indicate that the hybrid of these three techniques gave a faster and optimized system which is the need of the (a) Time vs. Epoch (b) Time vs. batch size (c) Loss value vs. Batch size present global scenario.
The loophole in the existing systems is that they aren’t able to adapt themselves quickly in the changing environment which is compensated by the use of k-means clustering The only setback of the aforementioned system as of now is the cost of its implementation since its complex to implement because of its hybrid nature. Fraud detection is a field that will never be dormant as there are always new strategies that can be found to commit fraud. Various other hybrid techniques could be experimented as future scope for better results. [4] R. Patidar, L. Sharma, et al., Credit card fraud detection using neural network, International
Journal of Soft Computing and Engineering (IJSCE) 1 (2011). [5] M. Zareapoor, K. Seeja, M. A. Alam, Analysis on credit card fraud detection techniques: based on certain design criteria, International journal of computer applications 52 (2012). [6] M. Seera, C. P. Lim, A. Kumar, L. Dhamotharan, K. H. Tan, An intelligent payment card fraud detection system, Annals of operations research (2021) 1–23. [7] N. K. Trivedi, S. Simaiya, U. K. Lilhore, S. K. Sharma, An eficient credit card fraud detection model based on machine learning methods, International Journal of Advanced Science and Technology 29 (2020) 3414–3424. [8] O. Owolafe, O. B. Ogunrinde, A. F.-B. Thompson, A long short term memory model for credit card fraud detection, in: Artificial Intelligence for Cyber Security: Methods, Issues and Possible Horizons or Opportunities, Springer, 2021, pp. 369–391. [9] N. Boutaher, A. Elomri, N. Abghour, K. Moussaid, M. Rida, A review of credit card fraud detection using machine learning techniques, in: 2020 5th International Conference on Cloud Computing and Artificial Intelligence: Technologies and Applications (CloudTech), IEEE, 2020, pp. 1–5. [10] R. Dash, R. Rautray, R. Dash, A legendre neural network for credit card fraud detection, in: Intelligent and Cloud Computing, Springer, 2021, pp. 411–418. [11] Y. Jain, N. Tiwari, S. Dubey, S. Jain, A comparative analysis of various credit card fraud detection techniques, Int J Recent Technol Eng 7 (2019) 402–407. [12] S. Maes, K. Tuyls, B. Vanschoenwinkel, B. Manderick, Credit card fraud detection using bayesian and neural networks, in: Proceedings of the 1st international naiso congress on neuro fuzzy technologies, volume 261, 2002, p. 270. [13] J. Esmaily, R. Moradinezhad, J. Ghasemi, Intrusion detection system based on multi-layer perceptron neural networks and decision tree, in: 2015 7th Conference on Information and Knowledge Technology (IKT), IEEE, 2015, pp. 1–5. [14] S. G. Fashoto, O. Owolabi, O. Adeleye, J. Wandera, Hybrid methods for credit card fraud detection using k-means clustering with hidden markov model and multilayer perceptron algorithm (2016). [15] J. Zou, J. Zhang, P. Jiang, Credit card fraud detection using autoencoder neural network, arXiv preprint arXiv:1908.11553 (2019). [16] C. Mishra, B. Gupta, R. Singh, Credit card fraud identification using artificial neural networks, International Journal of Computer Systems 4 (2017) 151–159. [17] T. K. Behera, S. Panigrahi, Credit card fraud detection: a hybrid approach using fuzzy clustering & neural network, in: 2015 second international conference on advances in computing and communication engineering, IEEE, 2015, pp. 494–499. [18] P. Chougule, A. Thakare, P. Kale, M. Gole, P. Nanekar, Genetic k-means algorithm for credit card fraud detection, International Journal of Computer Science and Information Technologies (IJCSIT) 6 (2015) 1724–1727. [19] T. Razooqi, P. Khurana, K. Raahemifar, A. Abhari, Credit card fraud detection using fuzzy logic and neural network, in: Proceedings of the 19th Communications & Networking Symposium, 2016, pp. 1–5. [20] F. Amato, N. Mazzocca, F. Moscato, E. Vivenzio, Multilayer perceptron: an intelligent model for classification and intrusion detection, in: 2017 31st International Conference on Advanced Information Networking and Applications Workshops (WAINA), IEEE, 2017, pp. 686–691.