Abstract

Cross border payment flows and the expansion of decentralised finance (DeFi) create new challenges for anti money laundering (AML) and fraud prevention. While artificial intelligence, including machine learning, deep learning and natural language processing, delivers strong detection performance, four critical unresolved tensions persist: susceptibility to adversarial evasion, limited explainability for regulatory auditing, tensions between large scale data aggregation and data protection legislation, and performance deterioration on sparse, highly imbalanced DeFi datasets.This doctoral research proposes a hybrid privacy preserving explainable AI (XAI) AML framework integrating federated learning, SHapley Additive exPlanations (SHAP) and on chain blockchain analytics. Adopting a mixed methods approach, the work combines a systematic literature review, empirical model benchmarking, adversarial robustness testing and regulatory compliance evaluation across traditional banking transaction data, Suspicious Activity Report (SAR) narratives and public Ethereum ledger data. Logistic regression, gradient boosted trees, LSTM networks and BERT serve as baseline reference models.It hypothesises the framework will achieve competitive detection (AUROC ≥ 0.94; F1 score ≥ 0.90) without centralising raw financial data, improve robustness under adversarial perturbations (ε = 0.1) relative to centralised deep learning baselines, and produce feature attributions compliant with FATF and EU AML Directive audit standards. Key predictive indicators are expected to include large value transactions, geographic anomalies, abnormal transaction volumes and atypical wallet behaviour. Limitations comprise computational overhead of federated aggregation and scarce public on chain labelled data. Future extensions may incorporate reinforcement learning and zero knowledge proofs. Outputs will deliver actionable technical and policy guidance for financial institutions, RegTech developers and regulators.