Privacy-Aware Graph Embeddings for Anti-Money Laundering Pipelines

Abstract

This article introduces a novel approach to anti-money laundering (AML) that combines graph neural networks (GNNs) with homomorphic encryption (HE) to detect suspicious financial patterns while preserving personally identifiable information (PII). Current AML systems face significant challenges in cross-border financial networks due to privacy regulations and data protection concerns. The proposed architecture enables financial institutions to analyze encrypted transaction graphs using privacy-preserving GNN inference, generating intermediate embeddings that retain predictive value without exposing raw identities. By performing computations directly on encrypted data, the system prevents the disclosure of sensitive customer information while maintaining detection capabilities. Experimental results demonstrate complete elimination of PII exposure incidents while substantially improving detection precision compared to baseline methods. Additionally, the system achieves notable reductions in false positive alerts, decreasing the manual review burden for financial institutions. This work addresses a critical gap in existing AML pipelines by supporting encrypted, privacy-safe graph analytics at scale and presents a three-phase implementation roadmap for integration with international banking systems.