Ученые из Дьюкского университета создали алгоритм для выявления преступных операций со стейблкоинами

Researchers at Duke University built one of the first large-scale datasets of stablecoin transfers and used it to train machine-learning models that flag suspicious cryptocurrency wallets. Published in the Blockchain Journal, the study detects illicit activity with high accuracy and, importantly, tells apart different types of criminal behavior, helping compliance teams act on real threats while sparing innocent users from frozen assets. Almost every crime leaves a financial trail. Whether the offense is fraud, trafficking, or a cyberattack, criminals eventually need to move, hide, and cash out their proceeds. Increasingly they do it with stablecoins, digital currencies pegged to the U.S. dollar that move billions of dollars a day. In 2024, an estimated $51 billion was laundered through cryptocurrencies. A new Duke University study addresses this problem. As privacy tools make parts of the crypto ecosystem harder to observe, centralized stablecoins such as Tether's USDT and Circle's USD Coin (USDC) stay visible, because their issuers must keep an auditable record to remain convertible into regular money. The team used that visibility to build a large-scale dataset of Ethereum wallet transfers and to set a baseline for spotting money laundering from on-chain behavior alone. "Stablecoins have become the connective tissue of illicit finance, but that same visibility is what makes them auditable. We wanted to show you can catch laundering by studying behavior, not identities," says Luciano Juvinski, lead author. The team compared several families of artificial intelligence, from simple linear methods to deep neural networks and graph-based models. The clearest result: carefully engineered, behavior-based models known as tree ensembles beat the more complex graph approaches, which lost accuracy as the transaction network fragmented and grew harder to trace. The models do more than raise a single alarm. They separate distinct types of illicit behavior. Wallets tied to cybercrime move funds fast and scatter them across many addresses, while sanctioned or frozen wallets leave a constrained, static footprint. That difference matters in practice. Under new rules such as the European Union's MiCA and the U.S. GENIUS Act, compliance teams must decide which wallets to act on, and a wrong call can mean an innocent person's money is locked. By blocking laundering with mathematics, the researchers argue, it becomes possible to disrupt the flow of illicit funds and raise the cost of crime itself. The work offers a foundation that future systems can build on to move this kind of detection from research into everyday compliance. The study was conducted by Luciano Juvinski, Haochen Li, and Alessio Brini at the Duke University Pratt School of Engineering, and grew out of the Duke Master in FinTech program. The dataset and code are publicly available to support further research. The paper, "StableAML: machine learning for behavioral wallet detection in stablecoin anti-money laundering on Ethereum," was published in the Blockchain Journal. Juvinski L, Li H, Brini A. StableAML: machine learning for behavioral wallet detection in stablecoin anti-money laundering on Ethereum. Blockchain. 2026, https://doi.org/10.55092/blockchain20260007