A MODEL FOR FRAUD DETECTION IN CREDIT CARD TRANSACTION USING AUTOENCODER

ABSTRACT In response to the escalating volume of credit card-based financial transactions, there is an increasing need to effectively identify and prevent fraudulent activities. This study proposes a two-stage model designed to enhance the trustworthiness of fraud detection systems by minimising both missed fraud identifications and false alarms. The central challenge lies in comprehending and learning the intricate associations among transaction attributes. In the initial stage of the proposed model, an autoencoder is employed to transform transaction attributes into a lower-dimensional feature vector. This feature vector serves as the input for a classifier in the second stage. The objective is to mitigate the complexity of transaction attribute associations, thereby facilitating more accurate fraud detection. Experimental validation is conducted on a benchmarked dataset. The results indicated that, particularly in terms of F1-measure, the proposed two-stage model outperforms systems relying solely on classifiers and other autoencoder-based approaches. This suggests that the combination of an autoencoder for feature transformation in the first stage, followed by a classifier in the second stage, yields superior performance in identifying fraudulent transactions compared to alternative methods.