SAE-IF: A hybrid sparse autoencoder–isolation forest framework for real-time credit-card fraud detection and sustainable digital-economy protection

Abstract

The rapid expansion of online payments and e-commerce has increased exposure to credit-card fraud and related financial crimes, threatening the security and stability of digital economies. Conventional supervised detectors require labelled fraud data, which are scarce and highly imbalanced. This study presents SAE-IF, a hybrid semi-supervised framework that combines a sparse autoencoder (SAE) with an isolation forest (IF) for anomaly detection in credit-card transactions. The SAE learns latent representations from transaction features without using class labels in its reconstruction objective, while labels are used for resampling, validation, fusion-weight optimisation, and threshold selection. A leakage-aware protocol employs stratified data splitting, robust feature scaling, and Bayesian hyperparameter optimisation with Optuna. The framework is evaluated on the ULB/Kaggle credit-card dataset, the IEEE-CIS Fraud Detection dataset, and the PaySim mobile-money dataset. SAE-IF achieved precision--recall area under the curve (PR-AUC) values of 0.832, 0.801, and 0.845, respectively, while maintaining precision above 0.90 and low false-positive counts. On a standard central processing unit, mean offline inference latency was 1.41 ± 0.23 ms per transaction. These results indicate that combining representation learning with anomaly detection can provide accurate and computationally feasible fraud screening in label-scarce financial environments, subject to validation on production infrastructure.