Privacy-Preserving Federated Artificial Intelligence for Cross-Institutional Anti-Money Laundering Detection in the United States: A Review

Money laundering moves an estimated 2%-5% of global gross domestic product through the financial system each year.Its central tactic, layering funds across multiple institutions, limits detection because each bank observes only part of the transaction chain.Machine learning has improved single-institution transaction monitoring, and graph-based methods exploit the network structure of illicit flows, but models trained within one institution remain constrained when laundering activity spans several institutions.Centralizing standardized and appropriately linked transaction data could improve cross-institutional visibility, but legal, privacy, operational, and competitive constraints make such pooling difficult in the United States.Federated learning offers an alternative by allowing institutions to train a shared model on local data while exchanging model updates rather than raw records.This narrative review synthesizes 45 scholarly sources published between 2013 and 2025, together with US statutes and supervisory guidance retrieved separately from official sources.It examines the evolution of anti-money laundering detection from rules to graph neural networks, compares horizontal, vertical, and graph-federated architectures, evaluates privacy risks and protections against information leakage, considers separate defenses against poisoning and backdoor attacks, and maps these techniques onto the US regulatory environment.The reviewed evidence shows that the core federated-learning, cryptographic, and statistical methods are well established, but AML-specific validation remains limited.Existing studies rely mainly on synthetic data, cryptocurrency networks, challenge datasets, or single-institution records.Standard federated learning can support shared model training but does not reconstruct cross-institutional transaction paths, which additionally require privacy-preserving entity resolution and secure cross-boundary graph protocols.None of the studies included in this review validated privacy-preserving federated anti-money laundering detection using real multi-institutional US banking data.