Chao Kongin Beijing
Researchers from China’s national police academy have developed an AI framework that can detect illicit cryptocurrency transactions with nearly 90 per cent overall accuracy, according to a recent study.
Cryptocurrency transactions are growing rapidly, but their pseudonymous and cross-border nature can provide channels for money laundering and other illegal activities.
In an article published in the peer-reviewed Chinese publication Journal of Intelligence, corresponding author Dr Sun Jingchao wrote that the study “provides a precise, generalisable and interpretable solution for detecting illicit cryptocurrency transactions”.
It also “provides an innovative technological pathway for regulatory authorities to combat illicit cryptocurrency transactions and economic crimes”, Sun, a researcher specialising in criminal investigation and cybersecurity, wrote.
The May study, conducted by researchers from the People’s Public Security University of China, which is affiliated with the Ministry of Public Security, comes as authorities continue to focus on cryptocurrency-linked financial crime.
n March, the Supreme People’s Procuratorate reported that prosecutors indicted 3,259 people in 2025 for money laundering involving virtual currencies and underground banks.
On July 25, Chinese digital news publication The Paper reported that a court in Inner Mongolia concluded a case involving nearly 3 billion yuan (US$444 million) in funds allegedly used to settle online gambling debts, potentially linked to betting during the Fifa World Cup.
However, such cases may represent only a fraction of illicit activity, as tracing cryptocurrency flows can be difficult even when transactions are publicly recorded
Bitcoin transactions are recorded on a public blockchain, but wallet addresses are not directly linked to real-world identities, while illicit funds can move through complex networks of addresses.
Traditional rule-based systems can struggle to detect such evolving patterns.
The Chinese team’s system combines a dynamic graph neural network with a memory mechanism and a large language model (LLM) to analyse cryptocurrency transactions.
The dynamic graph neural network treats the cryptocurrency ecosystem as a constantly changing network, analysing transaction structures, money flows and their evolution over time.
The system then uses a memory module to retrieve historical illicit transaction patterns that resemble a new case, allowing it to compare current activity with previously identified behaviour.
The researchers also introduced an LLM, which combines extracted transaction features, historical patterns and structural information from the network.
The LLM then turns these complex signals into a natural-language risk assessment, producing not only a legal-or-illicit classification but also a risk score, key evidence and a reasoning chain.
“Our method has significantly enhanced the model’s generalisation ability to identify new and evolving illegal trading practices,” Sun wrote.
The researchers tested the system on the public Elliptic bitcoin transaction dataset, which contains 203,769 transaction nodes, 234,355 edges and 166 features for each node.
The model achieved an overall accuracy of 89.4 per cent, while its precision for illicit transactions reached 89.1 per cent and its recall was 64.5 per cent.
The 89.1 per cent precision meant that nearly nine in 10 transactions flagged as illicit by artificial intelligence (AI) were indeed unlawful in the actual data set, while the 64.5 per cent recall indicated that the model identified roughly two-thirds of labelled illegal transactions.
“Experimental results on a public bitcoin transaction dataset demonstrate that the proposed framework outperforms mainstream baseline models,” Sun wrote.
The researchers argued that the memory component could help the system recognise new or evolving laundering techniques by matching current money flows, transaction timing and network structures with similar historical patterns.
“Through the semantic reasoning capabilities of large language models, the system provides intuitive and traceable decision-making evidence for high-risk alerts, enhancing its operability and interpretability in real-world regulatory scenarios,” they wrote.
For More News Updates Follow Us On www.tconews.in
