Financial Fraud Detection Using Hybrid Convolutional and Recurrent Neural Networks: An Analysis of Unstructured Data in Banking

Financial fraud detection is a critical domain in which the continuous evolution of fraudulent tactics necessitates advanced and adaptive detection mechanisms. This paper addresses the challenges of fraud detection within bank payments by proposing a novel approach that integrates Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). Leveraging the synthetically generated Banksim dataset from Kaggle, the study employs intelligent systems such as rule engines and machine learning to combat fraudulent activities. While rule engines execute predefined business rules, machine learning, implemented using Python software, offers a precise approach by learning evolving fraud patterns. The proposed model's methodology involves preprocessing the dataset using Min-Max Normalization to scale numerical features, ensuring equitable contributions during model training. Feature extraction employs CNNs to capture hierarchical patterns in unstructured textual data, followed by sequential modelling with RNNs to understand temporal dependencies in transaction logs. The proposed CNN-RNN hybrid model achieves remarkable accuracy, with a reported accuracy of 99.2%, surpassing existing methods by 4%. The integration of spatial and temporal modelling techniques enhances the model's precision and reliability in detecting fraudulent transactions. The ROC analysis further substantiates the proposed CNN-RNN hybrid model's discriminative power, with an impressive ROC value of 0.9.