【美联储】CardSim：支付卡欺诈检测研究的贝叶斯模拟器-2025_42页_1mb

报告摘要

Payment card fraud is increasing, with criminals leveraging generative AI tools, necessitating innovative fraud detection research.
Publicly available payment transaction data is scarce due to privacy and security concerns, limiting research pace and reproducibility.
This paper introduces CardSim, a flexible, scalable Bayesian simulator for generating synthetic payment transaction fraud data without compromising data privacy.

Calibration to Real-World Data: Uses publicly available surveys (e.g., Diary of Consumer Payment Choice, Federal Reserve Payments Study) to parameterize transaction characteristics.
Bayesian Approach: Employs Bayes' theorem to associate transaction features with fraud propensity, allowing probabilistic updates based on new evidence.
Modular and Scalable: Designed for easy adaptation to evolving payment trends and fraud tactics through adjustable parameters.
Public Availability: Accompanied by a Python software package for researcher use.

Simulates payment card transactions, incorporating payer and payee characteristics (e.g., transaction frequency, amount, location).
Uses Bayesian inference to generate fraud flags based on continuous and discrete features, with class imbalance addressed via ranking.
Operationalizes simulations via an agent-based model-like structure but assumes static behavior, avoiding complex dynamics.

Supports testing and evaluation of machine learning workflows, such as fraud detection classification, using real-world proxies.
Demonstrates performance with models like logistic regression, XGBoost, and neural networks on simulated data.
Enables interpretability frameworks like SHAP for understanding model predictions.

Simulation produces high-fidelity data reflecting real payment behaviors and fraud patterns.
Test results show effective fraud detection with metrics like precision (76-81%), recall (57-65%), and AUPRC, improving with tuning.
Computational efficiency is high, with up to 3.16 million records generated in 10 seconds, but runtime grows nonlinearly with increased complexity.

Focuses on simplistic fraud relationships and excludes typologies like spending sprees or unsupervised detection.
Assumptions about payer to payee ratios and geographic limitations in current design.
Scope restricted to payment cards; future extensions should target other payment systems and unsupervised anomaly detection.

CardSim advances fraud detection research by providing a publicly accessible tool for rigorous experimentation, addressing data gaps while balancing privacy.

展开完整摘要