SCAFDS: Graph Attention Network for Interbank Fraud Detection with SAR Generation
A recent scholarly article presents SCAFDS (Systemic Contagion-Aware Fraud Detection System), a comprehensive seven-stage framework designed for detecting fraud between banks. This innovative system tackles five key shortcomings found in earlier graph neural network (GNN) models, which inadequately represent credit contagion through misaligned credit distress signals for fraud investigations. SCAFDS captures fraud-specific interbank structures by utilizing fraud co-occurrence frequency metrics from FinCEN SAR registry data and incorporates edge-feature-informed graph attention. It produces Suspicious Activity Report (SAR) narratives that link forensic traceability to specific numerical detection results, addressing regulatory auditability issues in reports submitted to FinCEN. The U.S. financial sector handles around 1.3 million interbank transactions each day, yet no current system effectively models fraud spread using fraud co-occurrence edge features.
Key facts
- SCAFDS is a seven-stage integrated surveillance pipeline for interbank fraud detection.
- It addresses five structural limitations of prior GNN architectures.
- Fraud-specific interbank topology is encoded using fraud co-occurrence frequency metrics from FinCEN SAR registry records.
- Edge-feature-informed graph attention is used.
- SCAFDS generates SAR narratives with per-assertion forensic traceability.
- The U.S. financial system processes about 1.3 million interbank transactions daily.
- No existing system models fraud propagation using fraud co-occurrence edge features.
- The paper is available on arXiv with ID 2605.18913.
Entities
Institutions
- FinCEN
Locations
- United States