Abstract
Anti-money laundering (AML) systems play a critical role in safeguarding the global economy. As money laundering is considered one of the top group crimes, there is a crucial need to discover the money laundering sub-network behind a particular transaction for a robust AML system. However, existing rule-based methods for sub-network discovery are heavily based on domain knowledge and may lag behind the evolving modus operandi of launderers. In this work, we address the money laundering sub-network discovery problem with a neural network-based approach and propose an AML framework, AMAP, equipped with an adaptive sub-network proposer. Specifically, we design an adaptive sub-network proposer guided by a supervised contrastive loss to discriminate money laundering transactions from massive benign transactions. Extensive experiments on real-world datasets from AliPay demonstrate the effectiveness of AMAP in both transaction detection and sub-network discovery. The framework yields a more comprehensive risk coverage and deeper insight into money laundering strategies.
Introduction
Money laundering (ML) is a massive global issue, accounting for an estimated 2-5% of global GDP. Modern launderers do not simply move money once; they utilize complex, multi-hop transaction chains to disguise the illicit source of funds. Existing AML systems face two major hurdles: multi-hop complexity, where transactions are spread across many layers, and camouflage, where illicit transactions are mixed with a massive volume of benign ones to blend in.
Traditional rule-based systems are often too rigid to keep up with these evolving tactics. While recent Graph Neural Network (GNN) approaches have improved node classification for flagging single transactions, they often fail to discover the surrounding sub-network—the actual organizational structure of the crime. This leaves financial institutions with an incomplete picture of risk.
Methodology: The AMAP Framework
The authors propose AMAP (AML framework with an Adaptive sub-network Proposer), which consists of two main stages designed to move beyond individual detection toward full network discovery.
1. Adaptive Sub-network Proposer
Instead of analyzing a fixed neighborhood, AMAP starts at a suspicious node and iteratively "grows" a sub-network. It uses a GNN to create representations of the current sub-network and its neighbors. To handle camouflage, the model is trained using Supervised Contrastive Learning, which pulls ML transactions closer together in the embedding space while pushing benign transactions away.
A key innovation is the Virtual Threshold Node (VTN). Rather than using a blunt global similarity threshold to decide which neighbors to include, AMAP learns an adaptive threshold. A neighbor is only added to the sub-network if it is more similar to the current sub-network than it is to this learned VTN.
2. Dual-View Fusion Classifier
Once a potential ML sub-network is proposed, AMAP employs a dual-view approach to validate it. It considers both the "cleaned" potential ML sub-network and the original, noisy ego-network. An attention mechanism weighs these two views to produce a final fused embedding, which is used to classify the original transaction.
Experimental Evaluation
The researchers evaluated AMAP using two massive, real-world datasets from AliPay (Ant Group), denoted as M6 and M12. These datasets contain transactions from February 2022, featuring 400 attributes per transaction, such as sender/receiver profiles and transaction amounts. The data is highly imbalanced, reflecting the rarity of ML transactions in the real world. The "ground truth" for sub-networks was defined as the 3-hop ego-graph of known ML transactions.
Results and Insights
AMAP was tested against state-of-the-art GNNs (e.g., GAT, GIN) and Graph Anomaly Detection (GAD) methods.
- ML Transaction Detection: AMAP achieved an average improvement of 3.02% in AUC-ROC and 3.74% in AUC-PR over the best baselines. On the M12 dataset, it reached a high ROC of 0.9678.
- Sub-network Discovery: This is where AMAP demonstrated its primary strength. Compared to GNN "explainer" methods, AMAP improved the F1-score by 56.68% and Accuracy by 18.52%.
Ablation studies confirmed that removing the VTN (the adaptive threshold) caused a significant drop in discovery performance, proving that personalized thresholds are vital for navigating complex, noisy transaction networks.
Conclusion
AMAP represents a significant advancement in AML technology by explicitly modeling sub-network discovery as a learning task. By moving from blocking single transactions to uncovering entire illicit chains, it provides a tool for more comprehensive risk coverage. While challenges remain regarding label scarcity and the computational overhead of iterative processing, AMAP's ability to denoise complex networks through contrastive learning and adaptive thresholding sets a new benchmark for financial security systems.
Summary updated