A Conceptual AI/ML Framework for Fraud Detection and Anti-Money Laundering in Fintech: Architecture, Algorithms, and Regulatory Alignment

Abstract

The amount of yearly transactions on financial technology platforms is counted in billions, and as much as the rapidity and accessibility make fintech attractive to legit users, it also makes it attractive to fraudsters and money launderers. The rule-based transaction monitoring tools, which still comprise the core of most companies' compliance programs, generate huge numbers of false positives and are unable to adjust to new fraud types, while black box machine learning models which detect fraud more accurately fail to provide case-to-case explanations which are necessary for Suspicious Activity Reports according to FATF and other regulations. The paper provides a conceptual 5-layer framework which comprises the layers of data ingestion and KYC processing, feature engineering, multi-family modeling layer containing supervised, unsupervised and graph-based algorithms, explainability and risk scoring layer involving SHAP and LIME and compliance/regulatory reporting layer with the concept drift feedback loop between the last two layers allowing for analyst decision routing to the retraining of models. The paper takes advantage of the knowledge about deep learning models, graph neural networks, mitigation of class imbalance problem and federated learning for anti-fraud and AML purposes in order to substantiate each of the framework layers. The taxonomies of candidate algorithms, mappings of framework layers to the corresponding regulatory requirements and a 5-phase implementation roadmap are provided as practical deliverables which an actual fintech company can utilize in its deployment. Consistent with its stated scope, this paper is conceptual: it does not report empirical detection performance on any dataset, and Section 11 explicitly identifies empirical validation as the required next step before any component of the framework could be considered production-ready.