SALT-GNN Improves AML Detection in Dense Financial Networks

Automated anti-money laundering (AML) detection is crucial for financial stability, with Graph Neural Networks (GNNs) increasingly employed to identify relational patterns indicative of illicit activities. However, current AML GNNs often struggle with "dense recipient contexts," where high-activity accounts receive numerous transactions, making it difficult to discern suspicious signals from legitimate noise. This issue is often masked by aggregate performance metrics. This research introduces a new evaluation methodology that stratifies AML metrics by recipient-context density, revealing consistent performance drops in dense areas. The degradation is attributed to GNN limitations: multiset non-discriminability, cardinality blindness, and attention attenuation of weak multi-hop signals in dense neighborhoods. To counter this, the paper proposes SALT-GNN, a lightweight architecture that integrates degree-aware statistical aggregation with attention at each message-passing layer. SALT-GNN's design ensures that distributional and cardinality information influences the node states before subsequent attention steps, leading to more robust signal isolation. Benchmarking on three datasets shows SALT-GNN significantly improves F1 scores in dense contexts, using fewer parameters than transformer-based baselines. The gains are consistent across different attention operators, highlighting the effectiveness of its statistical-attentional fusion.