Hybrid BiLSTM-Transformer Model for Identifying Fraudulent Transactions in Financial Systems

Abstract

This paper proposes a credit card fraud detection method based on the combined model of BiLSTM and Transformer. With the popularity of electronic payment and online transactions, the problem of credit card fraud has become more and more serious, and traditional fraud detection methods have limited effectiveness in dealing with complex transactions. Therefore, this study combines the temporal data processing capability of BiLSTM and the global feature modeling capability of Transformers to build a new fraud detection model. The experimental results show that the proposed BiLSTM+Transformer model outperforms traditional machine learning models on several evaluation indicators. Specifically, the model showed significant advantages in metrics such as accuracy, recall, AUC, and F1-score, especially when dealing with unbalanced data sets and complex transaction patterns to better identify potential fraud. In addition, the effective combination of BiLSTM and Transformer and their complementarity in feature extraction and sequence modeling are verified through ablation experiments. The results of this study provide a new solution for credit card fraud detection and also provide a reference for risk management and security protection in the financial field. Future studies can further explore the optimization and extension of the model to improve its application capability in large-scale data and more complex scenarios.