Towards a Feed-Forward Neural Network for Financial Fraud Detection

This paper presents the development of a feed-forward neural network model for the early detection and mitigation of fraudulent transactions in banking systems. Using synthetically generated data that mimics real-world class imbalance between legitimate and fraudulent activities, we build a binary classifier capable of distinguishing between the two with high accuracy. The model achieved an overall accuracy of 98%, with a recall of 84% for fraudulent transactions. Results indicate that simple yet well-structured deep learning architectures can effectively reduce false negatives—a critical factor in fraud detection. Visualization techniques including Principal Component Analysis (PCA) projection and confusion matrices were used to evaluate classification performance and feature space separation. The approach provides a scalable foundation for integrating AI into financial security pipelines and can serve as a core component of the Advanced Financial Risk Analytics and Management System (AFRAMES).