Artificial Intelligence Supported Anti-Money Laundering Detection Through Multi-Scale Temporal Encoder and Graph Representation Learning

Abstract

A layered-screening mechanism via Trade-based money laundering (TBML) targets money laundering by improving anomaly detection in high transaction volumes and adapting to evolving criminal methods, addressing limitations of current deep learning techniques with sparse data. To overcome these issues a new framework has developed AI with Multi-Scale Temporal Encoder and Graph Representation Learning (AMTEGR). Real-time fraud detection in supply chain transactions requires precision. The dataset has 180,000 transactions, 3% of which are fraudulent. A multi-scale temporal encoder with TCNs uses a multi-branch convolutional architecture and kernel sizes of 4, 8, and 14 to capture dependencies at different time scales: intra-day patterns, local transaction patterns, and recurring transactions. Receptive field and dilation are employed to improve long-range temporal dependency capture. Attention-based feature refinement employs a multi-head attention mechanism to identify key feature associations in transactions and enhance representations through normalization, input fusion, and residual connections. Graph Representation Learning (GRL) nodes representing transactions, customers, orders, and products are interconnected through a graph, enabling a model to understand complex supply-chain and financial interactions. Convolutional Neural Network (CNN) for AML Detection for temporal and graph representations are applied in a convolutional network to identify suspicious and fraudulent transaction patterns. The performance analysis is training scalability, inference scalability, F1-Score Vs threshold value calculations.