Real-Time Edge-Enabled Cyber Credit Card Fraud Detection Using an Interpretable Multi-Relation Spiking Imperialist Competitive Bidirectional Encoder Graphormer

Abstract

Credit card fraud significantly burdens financially secure environments, with most of the traditional detection systems lacking in accuracy and interpretability. Current designs regarding the credit card fraud detection system often fail to encapsulate the intricate and hidden fraud patterns. This study investigates the role of deep learning networks in fraud detection with emphasis on class imbalance. A hybrid bidirectional spiking Graphormer module is implemented for spotting credit card remittance scams, and its performance is estimated on recall, accuracy, F1-score, precision, and Area under Curve (AUC). The comparative performance of the module is further tested over various cross validated class imbalance methods. The dataset amounts to 1,284,020 remittance records, of which 96,210 are deemed fraudulent, while 1,187,810 are legitimate transactions. Given the presence of multiple outliers, the local outlier factor is employed during the pre-processing stage to ensure the detection and removal of anomalous data points. The proposed module exhibits high predictive accuracy, achieving a performance rate of 99.25% when validated on external datasets. Results show that the combination of deep-learning-based scam spotting approaches, incorporating Explainable Artificial Intelligence (XAI) against the challenging domain of fraud detection. The Residual Network (ResNet) in an ensemble setting showed excellent performance, reaching a cross-validation score of 0.9912 against several other models. As a result, fraud detection becomes more competent than before for anomalous credit card remittances.