Credit Card Fraud Detection Based On Autoencoder And Generative Adversarial Network

With the fast development of Internet,online payment has become the most common payment approach in people's daily life,and credit card payment is one of the most popular online payment methods.The means of credit card frauds have been gradually turning to more diverse and complicated as well as the reasons of credit card frauds are also different.Some people want to steal funds from credit card accounts of other people,while others want to obtain goods for free by using credit cards of other people.Credit card frauds cause hundreds of millions of dollars in financial losses every year,therefore,designing an effective credit card fraud detection approach can help to reduce the financial risk during the use of credit cards.Previous fraud detection algorithms typically involve data mining and machine learning technologies such as Bayesian algorithm,logistic regression algorithm,decision tree algorithm,neural network algorithm and so on.These all algorithms can be considered as binary classification algorithms in this case.As binary classification algorithms,they often have relatively strict constraints on training dataset,which require that there are normal transactions and fraudulent transactions in the dataset and the number of both can not be too different from each other.However,there is a problem of data imbalance in credit card fraud detection in real-life cases,since the number of normal transactions are much greater than the number of fraudulent transactions.Imbalanced dataset can reduce the performance of common classification algorithms heavily,since the model only need to identify those normal transactions correctly and ignore those fraudulent transactions to obtain a high precision.However,the trained model will lose the ability to detect the fraudulent transactions from the entire dataset.The work of this paper is encouraged by a competition of credit card fraud detection on Kaggle.In the dataset of the competition,the number of fraudulent transactions is only 0.17% of the total number of transactions.Therefore,this paper extend this case to one-class classification scenario,in which we need to obtain an effective credit card fraud detection model when there are no fraudulent transactions in dataset.Aiming at this problem,this paper try to apply some current deep learning methods to solve it,by proposing a new model combining autoencoder and generative adversarial network to distinguish fraudulent transactions from normal ones.This technique can be considered as a one-class classification method which belongs to unsupervised learning.It only needs normal transactions to construct the credit card fraud detection model.This is suitable for the situation that there are few or even no fraudulent transactions.The main work of our paper is as follows:(1)Proposing a feature extraction method based on autoencoder.The experimental results have shown that the key features extracted by sparse autoencoder can help the model learn the boundary between normal transactions and fraudulent transactions better.(2)Proposing a credit card fraud detection model based on generative adversarial network.The key features extracted by sparse autoencoder are taken as the input of generative adversarial network to train it.During the training process,it will optimize the discriminator to get the ability of identifying whether a sample is fraudulent or not.Finally,combining the trained sparse autoencoder and the discriminator of trained generative adversarial network to obtain the ultimate credit card fraud detection model.The experimental results have shown that the proposed model outperforms the other state-of-the-art one-class methods.(3)Proposing an improved credit card fraud detection model.During those experiments,we find that the model will take long time to achieve convergence when using basic generative adversarial network and the result is not stable.Therefore,this paper helps the discriminator to learn the boundary between normal transactions and fraudulent transactions better,by shifting the goal of generator from generating normal transactions to generating fraudulent transactions.The experimental results have shown that the new model are more stable than the original one and take less time to achieve convergence.(4)Extending the proposed algorithm to news spam detection scenario.The experimental results on the news dataset proves the applicability of the proposed algorithm on similar problems.