A SMOTE enhanced multi stream deep learning ensemble for credit card fraud detection

Credit card fraud detection is highly imbalanced, making minority-class metrics and deployment efficiency more informative than accuracy alone. This study proposes SMOTE-EnsembleNet (SENet), combining CNN, BiLSTM, and Transformer streams through learnable cross-attention fusion. Among 100%, 200%, and 500% SMOTE ratios with focal loss, 200% SMOTE achieved the best overall balance. SENet obtained AUC-ROC values of 0.9967 (95% CI 0.9955–0.9979), 0.9567 (95% CI 0.9542–0.9592), and 0.9923 (95% CI 0.9911–0.9935) on ULB, IEEE-CIS, and PaySim, respectively. Compared with no SMOTE, AUC-ROC improved by 1.55, 1.55, and 1.34 percentage points, while cross-attention improved AUC-ROC by 0.22 percentage points over concatenation. Fold-level significance testing ((p < 0.05)) was conducted against all tuned baselines. SENet achieved 12.4 ms inference latency per transaction on an NVIDIA A100 GPU while supporting interpretable error and feature-importance analyses. AUC-ROC and AUC-PR are threshold-independent, whereas precision, recall, and F1-score are reported separately in the experiments. The 12.4 ms latency applies only to the stated NVIDIA A100 configuration. Cross-attention additionally improved AUC-PR by 1.3 percentage points over concatenation, and the small differences between 200% and 500% SMOTE are interpreted together with fold-level statistical tests.