Abstract

Mobile banking fraud presents a rapidly evolving challenge as transaction volumes scale globally. This paper introduces a hybrid machine learning framework that fuses behavioral biometrics (typing rhythm, swipe velocity, device orientation) with transactional metadata to achieve real-time anomaly detection with high precision.

Key Contributions

  1. A novel feature fusion pipeline combining time-series behavioral signals with categorical transaction features
  2. An online learning component that adapts to user-specific behavior drift
  3. A lightweight deployment architecture suitable for edge inference on mobile devices

Results

The framework achieves 97.3% AUC on held-out test sets, with false positive rates below 0.8%, making it viable for production deployment without significantly disrupting legitimate users.

Citation

Journal: SCIENTIFIC CULTURE
DOI: 10.5281/zenodo.18681515
Year: 2026