TOWARDS ROBUST AND PRIVACY-PRESERVING ANTI-MONEY LAUNDERING SYSTEMS: A SYSTEMATIC REVIEW OF FEDERATED LEARNING AND GRAPH NEURAL NETWORKS FOR FINANCIAL CRIME DETECTION

Abstract

Financial systems are undergoing rapid digital transformation, resulting in an unprecedented increase in the scale, complexity, and sophistication of money laundering and financial crime activities. Traditional rule-based Anti-Money Laundering systems are increasingly ineffective against evolving criminal typologies, high-frequency transactions, and cross-border illicit financial networks. Although Artificial Intelligence has emerged as a promising solution for enhancing AML capabilities, existing AI-driven approaches continue to face significant challenges, including limited adaptability to dynamic transaction behaviors, restricted inter-institutional collaboration due to strict data privacy regulations, lack of explainability, and susceptibility to adversarial attacks. This paper presents a systematic review of advanced AI-based methodologies for financial crime detection, with particular emphasis on Graph Neural Networks and Federated Learning. GNN-based models enable the representation of financial ecosystems as interconnected transaction graphs, facilitating the identification of complex relational dependencies and evolving illicit patterns. Simultaneously, FL supports collaborative model training across distributed financial institutions without requiring direct sharing of sensitive customer data, thereby preserving privacy and regulatory compliance. The study critically analyzes recent state-of-the-art approaches by comparing their detection accuracy, scalability, robustness, interpretability, and real-time operational capabilities. Furthermore, major research challenges including high false-positive rates, handling of unstructured and heterogeneous financial data, computational overhead, and vulnerability to adversarial manipulation are comprehensively examined. Based on the findings, the paper outlines future research directions involving explainable AI, zero-trust cybersecurity architectures, adversarial robustness, and multimodal financial intelligence systems. The review provides valuable insights for researchers, cybersecurity professionals, financial institutions, and policymakers seeking to develop scalable, secure, and privacy-preserving next-generation AML frameworks