Edge-Enabled Real-Time Fraud Detection for Network Lending Terminals under Low-Latency Constraints

This study proposes an adaptive edge–cloud collaborative framework for real-time fraud detection in network lending terminals, addressing the challenges posed by extreme class imbalance and latency constraints. Using the publicly available Credit Card Fraud Detection dataset, nine machine learning algorithms were evaluated in combination with four oversampling techniques (SMOTE, Borderline-SMOTE, SVMSMOTE, and ADASYN). The results demonstrate that ensemble tree-based methods—particularly Random Forest, LightGBM, and XGBoost combined with SMOTE—achieve the best trade-off between accuracy, fraud recall, and false-positive rate. The framework integrates edge-level pre-scoring with cloud-based model refinement, reducing end-to-end latency by up to 35% while preserving detection accuracy. These findings underscore the potential of hierarchical, cost-sensitive learning pipelines to strengthen financial transaction security in real-time environments.