Credit Card Fraud Detection Using Deep Learning Based on Auto-Encoder

Abstract

Fraudulent activities in financial fields are continuously rising. The fraud patterns tend to vary with time, and no consistency can be observed in this regard. The incorporation of new technology by fraudsters is the reason for the execution of online fraud transactions. Given the volatility of the fraud patterns, a good fraud detection model must be able to evolve and update itself to the changing patterns. Thus we aim in this paper to analyze the fraud cases that are unable to be detected based on supervised learning or previous history, create an Auto-encoder model based on deep learning and Compare and assess the performance of the model based on data from different parts of the world and check for the demographic diversity of fraud patterns thereby inferring that the data from which part of the world the model fits the best. The proposed algorithm, deep learning based on the auto-encoder (AE) network is an unsupervised learning algorithm that utilizes backpropagation by setting the inputs and outputs identical. In this research, the Tensorflow package from Google has been employed to implement AE by using deep learning. The accuracy, precision, recall, F1 score and area under the curve(AUC) are all executed to assess the performance of the model.

Credit Card Fraud Detection Using Improved Deep Learning Models

“…A deep AE model was developed that can evolve into pattern changes [21]. The model consisted of four hidden layers, including a pair of encoders and a pair of decoders.…”

“…Firstly, the AE network is an unsupervised learning algorithm that utilizes backpropagation by setting the inputs and outputs identically [21,25]. The main goal of the AE is to approximate the distribution of the input value as accurately as possible [2].…”

“…The proposed model's performance has been evaluated by comparing its results with previous works using the hyperparameter values reported in Table 4. The proposed DL model results are compared with [19,21,23,25] in terms of accuracy, DR, and AUC as shown in Table 5. From Table 5, it is obvious that the proposed models surpass the previous works.…”

“…This is due to the inclusion of an extra set of hyperparameters in the tuning technique that was not considered in the previous work which effectively helps DL to capture the interdependence between high-dimensional features. In addition, one can see that the fraud detection performance of the proposed model is significantly better than that of [19,21,23,25] because determining the best successful combination of hyperparameter values for detecting fraud in financial transactions and the usage of binary focal loss (BFL) function improved DR results by helping the models converge faster and better and avoid noisy samples during training, it also effectively down-weighting the easy to classify normal transactions and gave more emphasis to hard-classify fraudulent transactions.…”

Credit Card Fraud Identification using Logistic Regression and Random Forest

“…KNN, Nave Bayes, and LR are tested on highly unbalanced credit card fraud data, and meta-classifiers and meta-learning strategies for handling such data are investigated [7].In certain situations, supervised learning techniques for fraud detection may not be successful. To separate out outliers from regular behavior, we build a model using a deep Auto-encoder and a constrained Boltzmann machine [8]. Credit card fraud may be detected via a hybrid technique suggested by Olena et al [9], which uses random forest and isolation forest to identify anomalous transactions.…”