Deep Learning Framework for Financial Fraud Detection: Systematic Feature Engineering and Comparative Evaluation of Neural Architectures

ABSTRACT

Financial fraud detection presents a significant challenge due to the extreme class imbalance inherent in transaction datasets. This paper presents a comprehensive evaluation of deep learning architectures for fraud detection in structured transaction data, with a focus on model performance, interpretability, and operational relevance. Specifically, we compare three deep tabular models, namely, an advanced multilayer perceptron (AdvancedMLP), an attention‐based residual network (AttentionFraudNet), and an advanced residual network (AdvancedResNet), against three traditional machine learning baselines, including Random Forest, Gradient Boosting, and Logistic Regression. The experiments are conducted on the BankSim Simulator dataset, comprising 594,643 transactions with a fraud rate of approximately 1.21%. We evaluate models using multiple metrics, including PR‐AUC, ROC‐AUC, recall at low false positive rates, and precision at various operating thresholds. Our results show that AdvancedMLP achieves the highest PR‐AUC (0.8921 ± 0.0020), while Gradient Boosting attains the best ROC‐AUC (0.9987 ± 0.0001) and the highest recall at 1% FPR (0.9498 ± 0.0087). AttentionFraudNet provides interpretable attention weights that highlight transaction‐level features contributing to fraud predictions, enabling fraud analysts to understand and validate model decisions. Business impact analysis reveals that the proposed models can reduce fraud losses while maintaining false positive rates below 1%, translating to estimated annual savings of approximately 49.2 M$ for a mid‐sized financial institution. All models achieve inference times below 1 ms per transaction on an NVIDIA RTX 3090 GPU, demonstrating computational feasibility for near‐real‐time screening. Our findings highlight the effectiveness of deep tabular models for fraud detection and provide practical guidance for selecting models based on operational requirements and performance trade‐offs.