GARG-AML against Smurfing: A Scalable and Interpretable Graph-Based Framework for Anti-Money Laundering

Researchers Bruno Deprez, Bart Baesens, Tim Verdonck, and Wouter Verbeke introduced GARG-AML, a scalable graph-based framework for anti-money laundering that specifically targets 'smurfing'—a tactic where criminals break large sums into numerous small transactions to evade detection. Unlike complex deep learning models that act as black boxes, GARG-AML maps an account's immediate and secondary connections (its second-order neighborhood) into an adjacency matrix and measures the density of specific blocks within that matrix to flag suspicious patterns. The method boosts performance further using decision trees and gradient-boosting classifiers, matching or beating state-of-the-art results across all tested datasets, including synthetic and open-source data.

The framework's key innovation is translating human expert knowledge of smurfing directly into a simple network representation, avoiding the interpretability trade-offs of deep learning. By leveraging only basic network features like adjacency matrix density, GARG-AML easily processes the massive transaction graphs typical of large financial institutions. This design prioritizes both speed and clarity—critical for investigators who need to understand why a risk score was assigned. The authors highlight that this approach can be seamlessly integrated into existing AML solutions, offering a transparent alternative that balances detection power with real-world business demands for scalability and interpretability.