As a test of their resulting AI tool, the researchers checked its results against a cryptocurrency exchange – which is not named in the study – and identified 52 suspicious transaction chains, all of which ultimately flowed into that exchange. It found that the exchange had already flagged 14 of the accounts that received those funds for suspected illegal activity, including 8 that it had identified as being linked to money laundering or fraud, based in part on the “Know Your Customer” information from account holders. Although they did not have access to this know-your-customer data or information about the origins of the funds, the researchers’ AI model was consistent with the conclusions of the exchange’s own investigators.
Correctly identifying 14 of 52 of these customer accounts as suspicious may not sound like a high success rate, but the researchers point out that overall, only 0.1 percent of the exchange’s accounts are flagged as potential money laundering. Their automated tool, they argue, has essentially reduced the search for suspicious accounts to more than one in four. “Going from ‘One in a thousand things we look at will be illegal’ to 14 out of 52 is a crazy shift,” says Mark Weber, one of the paper’s co-authors and a fellow at MIT’s Media Lab. “And now investigators are actually going to examine the rest of it to see, wait, did we miss something?”
Elliptic says it has already used the AI model privately in its own work. As further evidence that the AI model produces useful results, the researchers write that analyzing the sources of funds for some suspicious transaction chains identified by the model helped them discover Bitcoin addresses used by a Russian dark -Web market, a cryptocurrency “mixer” designed to obscure the trace of Bitcoins on the blockchain, and a Panama-based Ponzi scheme. (Elliptic declined to name any of these suspected criminals or services and told WIRED that it is not naming the targets of ongoing investigations.)
Perhaps more important than the practical use of the researchers’ own AI model, however, is the potential of Elliptic’s training data, which the researchers published on Google’s own machine learning and data science community site Kaggle. “Elliptic could have kept this to itself,” says MIT’s Weber. “Instead, there was a clear open source ethos here of contributing to the community that enables everyone, including their competitors, to become better at combating money laundering.” Elliptic notes that the data it publishes is anonymized and does not contain identifiers for the owners of Bitcoin addresses or even the addresses themselves, but only the structural data of the “subgraphs” of transactions that it has marked with its assessments of suspected money laundering.
This vast trove of data will undoubtedly inspire and enable much more AI-focused research on Bitcoin money laundering, says Stefan Savage, a computer science professor at the University of California, San Diego, who served as an advisor to the lead author of a groundbreaking Bitcoin traceability paper published in 2013. However, he argues that the current tool is unlikely to revolutionize anti-money laundering efforts in the crypto space in its current form, but rather will serve as a proof of concept. “I think an analyst will have a hard time using a tool like this Art “Sometimes that’s right,” Savage says. “I look at this as progress that says, ‘Hey, there’s something here.’ More people should work on this.’”