Enhancing Fraud Detection Performance on Imbalanced Datasets: A ResNet-Based Approach Integrating Multi-Attention Techniques and SMOTE-ENN

Background: The detection of financial fraud has become increasingly significant because electronic transactions and advanced frauds have been growing rapidly, leading to huge losses for the economy. Yet, significant class imbalance, data heterogeneity, evolving fraud trends, and a small number of labelled samples remain as challenges for building effective fraud detection systems.Objective: This article assesses an integrated approach to a fraud-detection framework based on residual networks (ResNet), different attention mechanisms, and class-imbalance handling techniques across heterogeneous financial datasets.Methods: The proposed experimental framework combined ResNet with self-attention, channel attention, gated attention and multihead attention. Class imbalance was addressed using the synthetic minority oversampling technique (SMOTE) and edited nearest neighbours (ENN). The model was evaluated on the European Credit Card Fraud dataset, the UCI Taiwan dataset, and the combined dataset to assess its performance across both transactional and demographic feature spaces.Results: On the European Credit Card Fraud dataset, the ResNet multihead attention model outperformed all other models, while on the UCI Taiwan dataset, the ResNet gated-attention model had the best F1 performance. The channel-attention model performed the best in terms of accuracy and AUC for the merged data. AUC was not always the best choice for the various experimental setups, but the most consistent improvements were in F1 score and ranking stability.Conclusion: The integrated ResNet-attention framework achieved competitive performance on fraud detection in imbalanced and heterogeneous financial datasets, particularly in terms of F1 score. The results show that the utility of single attention mechanisms varies across datasets and does not indicate that any single architecture is superior. The framework should be tested on other real-world financial datasets, and further optimisation should be explored to enhance its generalisability and robustness.