A Multi-Layered Artificial Intelligence Framework for Detecting Coordinated Financial Crime Networks

Abstract

Coordinated financial crime involves distributed transactions, accounts, intermediaries, and payment vehicles, which form complex adaptive networks that cannot be inferred using transaction-level analytics. In this work, we designed a multi-layer artificial intelligence system and demonstrate its elements, including transaction attributes, contextual features, graph structure, temporal properties, and explainable risk scores in the task of detecting coordinated financial crime. We evaluated the system on two publicly available synthetic datasets. In Dataset I, Cash, Cheque, and Cross-border transactions were labeled as suspicious, whereas ACH Transfers, Credit Cards, Debit Cards, E-Wallet, and Mobile Payments were considered normal. This study explored the performance of a hierarchical detection model that employs a set of diverse evidence in a multi-layer system and the challenges of the model performance assessment in terms of leakage, network effects, temporal factors, and human-related aspects. The study highlighted the opportunities and limitations of the multi-layer approach.