Automated Fraud Detection In Financial Services Using Hybrid Autoencoders And LSTM

Abstract

The growing sophistication of cyberattacks and transaction manipulation techniques has turned financial fraud into a big challenge in digital banking and electronic payment systems. Conventional fraud detection methods are not particularly successful in detecting new fraud patterns and sequential transaction anomalies. This research aims to present a hybrid autoencoder-LSTM model that could be used towards an automated fraud detection system for financial services based on the Kaggle Credit Card Fraud Detection Dataset. The proposed methodology is combining an autoencoder model for anomaly reconstruction analysis and sequential transaction behavior learning using a long short-term memory (LSTM) network. Firstly, the data set is subjected to preprocessing such as normalization and SMOTE (class balancing). The autoencoder identifies abnormal transaction patterns by analyzing reconstruction losses, and the LSTM network learns fraud patterns over time in sequences of transactions. The experimental results show that the proposed framework obtained an accuracy of 99.42%, a precision of 98.16%, a recall of 97.88%, an F1 score of 98.02%, and an ROC-AUC value of 99.31%. The other result of the confusion matrix analysis was that both the false positive and false negative rates were a little low, and this indicated that the proposed framework was valid in classifying frauds. The fraud detection capability and false alarm rate were also better compared to existing deep learning models. The results show that the proposed hybrid autoencoder-LSTM framework is efficient, scalable, and can be deployed in real time for intelligent financial fraud monitoring systems.