AI Model Detects Crypto Money Laundering with High Accuracy
A new study published in Discover Artificial Intelligence introduces an advanced AI model designed to detect money laundering in cryptocurrency transactions. The research, led by Yong Shang of Henan Judicial Police Vocational College, presents GT-SSL, a graph-structured Transformer trained through self-supervised learning. The model aims to solve the difficult problem of identifying illicit transactions in blockchain data, where only a small fraction of transactions are labeled as illicit or licit.
The core challenge lies in the vast majority of blockchain transactions being unlabeled, making traditional supervised machine learning ineffective. GT-SSL converts raw blockchain records into a directed attributed graph, where nodes represent transactions and edges represent fund transfers. The model uses a biased restart random walk to generate fixed-length sequences of transactions, which are then processed by a Transformer. This approach helps preserve the local laundering path while maintaining neighborhood diversity.
The model's architecture includes a soft multi-hop structural bias, which ensures that attention weights are constrained by the topology of the underlying fund-flow graph. This design choice is crucial for detecting multi-level account transfers and fund splitting, common in laundering schemes. The study reports impressive results on two benchmark datasets, with GT-SSL achieving an F1-score of 95.80 percent and an AUC of 97.62 percent on the Elliptic dataset, and an F1-score of 93.78 percent and an AUC of 95.91 percent on the AML-Bitcoin dataset.
The model also demonstrated resilience in scenarios with very few labels, maintaining high accuracy and recall even when only 5 percent of training labels were visible. It outperformed several baseline models and reduced false positives and false negatives significantly. However, the study acknowledges that GT-SSL is a static graph model and may not fully capture the dynamic nature of blockchain transactions over time.