2025-03-09-美联储-CardSim_支付卡欺诈检测研究的贝叶斯模拟器（英）_41页_1mb

CardSim: A Bayesian Simulator for Payment Card Fraud Detection Research Summary

Background

Payment card fraud has increased in the US due to factors like the COVID-19 pandemic and accessible generative AI tools. Generative AI enhances fraud capabilities, while financial institutions use AI for detection, creating a need for innovative research. Publicly available payment data are scarce due to privacy and economic constraints, hampering research reproducibility and innovation. CardSim is introduced as a simulation tool to address these data gaps by generating synthetic payment transaction data without compromising security.

Methodology

CardSim employs a Bayesian approach to simulate payment card transactions and fraud patterns. It is modular, scalable, and calibrated to publicly available data. The simulation process involves three key phases:

1. Developing payer and payee characteristics based on surveys like the Diary of Consumer Payment Choice.
2. Running a core transaction simulator to generate features like card type, amount, location, distance, and time.
3. Using Bayes' theorem to assign fraud labels based on statistical distributions.
   The tool is implemented in a Python package for public use, leveraging easy parameter adjustments to reflect evolving payment trends.

Results

Simulations produce synthetic datasets with features such as transaction times, amounts, and fraud rates, calibrated to real-world data. Results show distinct differences between legitimate and fraudulent transactions in terms of payment amounts, distances, and time distributions, demonstrating the simulator's ability to model probabilistic fraud associations. Performance is computationally efficient, handling up to 3 million records in short runtime, but scale increases with parameters like the number of payers or days.

Limitations

CardSim simplifies fraud by not incorporating deterministic fraud typologies, limiting tests of rule-based approaches. It focuses on unauthorized fraud with basic feature sets and payer characteristics, potentially missing advanced modeling or demographic indicators. Simulators always involve trade-offs, and parameters can be fine-tuned for specific scenarios, but without real-world fraud labels, unsupervised detection approaches may need separate adaptations.

Conclusion

CardSim advances fraud detection research by enabling flexible, privacy-preserving simulations for testing machine learning models and interpretability tools. It highlights the potential for Bayesian methods in modeling payment systems. Future work should integrate fraud typologies and develop simulators for unsupervised learning to broaden applications in diverse payment systems.