Credit card fraud detection is widely powered by machine learning models that achieve extremely high offline accuracy, often reporting AUC-ROC scores close to 1.0. However, these evaluations usually ignore real-world deployment constraints, especially the strict latency requirements of payment gateways, which typically allow less than 50 ms per transaction for authorization. This creates a significant gap between model performance in research settings and their practicality in production systems. In this work, we propose KD-Fraud, a framework that systematically studies this gap and applies knowledge distillation to enable real-time fraud detection. We first analyze the latency of six representative model families and show that even highly accurate ensemble methods may fail deployment constraints; for example, a stacked tree-based model reaches an AUC-ROC of 99.99 but requires 50.78 ms per transaction, exceeding typical limits. To solve this, KD-Fraud compresses an ensemble of XGBoost, LightGBM, and Random Forest models into a lightweight three-layer MLP using a combined cross-entropy and KL-divergence loss. On a dataset of 568,630 transactions with 29 PCA-transformed features, the distilled model achieves an AUC-ROC of 99.99, an F1-score of 99.95, and perfect recall on 56,863 fraud cases, while reducing inference time to 0.152 ms per transaction. This results in a 333.5× speedup over the teacher ensemble and a 541× reduction in model size. We also find that when teacher predictions are highly confident, temperature scaling has minimal effect, and performance gains come mainly from architectural compression.

credit card fraud detection,knowledge distillation,inference latency,model compression,payment gateway,tabular distillation