Directed-Hypergraph Analytics and Fragility Testing for Anti-Money Laundering

Abstract

Illicit activity in financial networks is difficult to detect. We exploit the idea that the chance of a transaction being illicit depends partly on the concurrent positions and recent actions of counterparties, as well as the higher-order organization of fund flows. We combine hypergraph features and machine learning to demonstrate practical ef fectiveness. Specifically, we use the Elliptic++ Bitcoin dataset to show that graph and directed-hypergraph-based features improve illicit-address classification. The suc cess of the predictive model then allows us to examine potential money-laundering chains are structurally fragile. Several illicit-to-licit transaction motifs with unknown intermediary addresses specified, and the fragility of these motifs is evaluated through progressive removal of the highest-risk intermediaries. Surprisingly, these motifs are fragile, suggesting that money laundering could be disrupted via the identification and exclusion of owners of a limited number of crypto wallets.