Audit Trail Reconstruction Using Mobile Money Transaction Fingerprinting: A Machine Learning Approach with Tamper Constraint Theory

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background Mobile money platforms in low‑trust environments face audit trail forgery risks. Existing methods rely on server logs that can be altered by insiders. Methods We propose transaction fingerprinting using temporal, network, device, and spatial features, grounded in Tamper Constraint Theory (TCT)a formal framework positing that forgeries violate temporal continuity, spatial feasibility, device persistence, or behavioural entropy constraints. We evaluate on a Kenyan dataset (2.5M transactions, 12 months) with adversarially constrained simulated forgeries (temporal shift, agent swap, device spoofing) filtered via expert realism scoring (n=3 fraud analysts), plus 212 real fraud cases for external validation. A random forest classifier is benchmarked against rule‑based, logistic regression, and isolation forest models. We report PR‑AUC, calibration, group‑aware validation (leave‑device‑out, leave‑agent‑out), cost‑sensitive performance, and adversarial robustness. Results The fingerprinting method achieves AUC = 0.97 (95% CI: 0.96–0.98), PR‑AUC = 0.94 (0.92–0.96), sensitivity = 94% at 3% FPR, Brier score = 0.04. Temporal entropy (STi=0.41) and device consistency (STi=0.33) dominate. Group‑aware validation shows minimal overfitting (AUC drop ≤0.012). Under adaptive adversarial attacks, AUC degrades to 0.85–0.91. Expected monetary loss reduction is 82%, ROI = 480% for a regulator. Conclusion TCT‑grounded transaction fingerprinting provides a robust, server‑independent audit trail. Regulators can deploy it cost‑effectively.