Quantum-Enhanced AML: Hybrid Quantum-Classical Graph Learning With Entity Resolution and Evidence Subgraph Discovery for Transaction Screening

Abstract

Anti-money laundering (AML) monitoring must detect hidden illicit fund flows from large, noisy, and highly connected financial transactions. In practice, laundering behavior is rarely visible in a single transaction; it appears as multi-hop paths across accounts, devices, merchants, and services. However, real AML graphs are unreliable because identities are fragmented (duplicate customers, shared devices, reused phone numbers), and many systems still generate alerts without strong evidence trails. This reduces both detection quality and investigation speed, especially when the network structure changes over time. To address this, we propose Quantum-Enhanced AML, a hybrid quantum-classical graph learning pipeline for transaction screening. First, we build a transaction-entity graph and apply graph-based entity resolution to merge duplicate identities and recover a cleaner actor-level structure. Next, we learn temporal graph embeddings using a classical encoder enhanced with a lightweight quantum embedding layer, so the model captures compact relational patterns even under noisy links. We further add a contrastive self-supervised objective to stabilize representations across time windows. Finally, for every flagged case, the system extracts an evidence subgraph that highlights key paths and neighbors responsible for the risk decision, which supports audit and analyst review. Experiments on a time-aware split show that Quantum-Enhanced AML achieves strong screening performance, with F1 up to 0.969 and AUC up to 0.995, while consistently producing case-ready evidence graphs. Ablation results confirm that entity resolution improves actor consistency and that temporal self-supervision improves generalization. Overall, the proposed pipeline provides accurate detection with practical explainability for compliance-focused deployment.