DEEP NEURAL NETWORKS FOR DETECTION OF CREDIT CARD FRAUD
Authors
Keywords
credit card fraud, deep neural network, GAN.
Summary
This publication presents research on the possibility of using Deep Neural Net-works (DNNs) for detection of credit card fraud. The research is based on an open da-taset and aims to determine how the varying number of hidden layers and other archi-tectural parameters (width of each layer, activation functions) affects training and predictive potential of the network. It also aims to determine how to use generative-adversarial approach efficiently in order to overcome the problem with the imbalance between legitimate and fraudulent transactions, which is crucial characteristic of the datasets related to detection of credit card fraud.
The publication presents the results of various experiments, conducted to achieve the stated goals. A number of experiments have been carried out combining network training of various depth of the network and with different parameters of the learning process, for example, activation function and dropout values. The experiments focus on using generative-adversarial approach as a method for oversampling, in par-ticular.
The results indicate that deep neural networks with moderate hidden layers, combined with methods for oversampling are performing much better in identifying fraudulent transactions than the classical neural networks, consisting of only one hid-den layer. The results also indicate that using generative-adversarial approach as a method for oversampling is comparable in efficiency to other widely used techniques.
Pages: 24
Price: 3 Points
