2025-06-16-国际清算银行-Hertha项目_识别实时零售支付系统中的金融犯罪模式（英）_30页_2mb

Motivation

Project Hertha addresses the challenge of detecting financial crimes, like money laundering and fraud, in real-time retail payment systems. With global losses estimated at $3 trillion annually, it explores how network-wide data from payment systems, combined with AI, can enhance detection beyond individual institutions using synthetic data to preserve privacy.

Methodology

The project utilized synthetic data representing a national retail payment ecosystem (1.8 million accounts, 308 million transactions). It tested AI models, including machine learning (e.g., XGBoost) and deep learning, to identify financial crime typologies, comparing results against a benchmark of isolated bank/PSP monitoring.

Results

Payment system analytics identified 12% more illicit accounts on average and showed a 26% improvement for previously unseen patterns. Collaboration between banks/PSPs and payment systems, with a feedback loop for model refinement, yielded the best outcomes. Supervised AI performed better for detection stability, but unsupervised methods faced high false positives.

Key Insights

• Successful use of minimal data points for explainable AI, potentially reducing privacy risks.
• Found robust feedback loops essential for model calibration, as unsupervised methods severely penalized lack of labeled data.
• Limited scope included legal and policy implications, but highlighted the value of explainable AI for regulatory and investigative purposes.

Areas for Further Research

The project outlined experiments for tracing financial crime networks, collaborative investigations across entities, and extending transaction analytics to other payment systems like cryptocurrencies or large-value transfers.