Our new research: Enhancing blockchain analytics through AI

• Elliptic researchers have advanced the use of AI to detect money laundering in Bitcoin. A new paper describing this work was co-authored with researchers from the MIT-IBM Watson AI Lab.
• A deep learning model is used to successfully identify proceeds of crime deposited in a crypto exchange, new patterns of money laundering transactions, and previously unknown illegal wallets. These outputs are already being used to improve Elliptic’s products.
• Elliptic has also made the underlying data publicly available. Containing over 200 million transactions, it will allow the wider community to develop new AI techniques to detect illicit cryptocurrency activity.

At Elliptic, we have always pushed the boundaries of blockchain analytics, to enable our clients to more accurately and efficiently assess risk in cryptoassets. Part of this innovation is exploring how artificial intelligence can be used to improve the detection of money laundering and other financial crimes on blockchains.

Blockchains provide fertile ground for machine learning techniques, thanks to the availability of transaction data and information about the types of entities that transact, which we and others collect. This is in contrast to traditional finance where transaction data is typically isolated, making these techniques difficult to apply.

Machine learning on the blockchain

We first published research on this topic in 2019, co-authored with researchers from the MIT-IBM Watson AI Lab. A machine learning model is trained to identify Bitcoin transactions made by illegal actors, such as ransomware groups or darknet markets. The training data, compiled by Elliptic and containing over 200,000 bitcoin transactions, was made publicly available to encourage further experimentation and collaboration in this emerging field. That paper has now been cited nearly 400 times by researchers around the world, demonstrating the impact it has had and continues to have in the fields of machine learning and anti-money laundering.

We have now published further research, applying the new techniques to a much larger dataset, containing nearly 200 million transactions. This paper is again co-authored by researchers from the MIT-IBM Watson AI Lab. Instead of identifying transactions made by illegal actors, a machine learning model is trained to identify “subgraphs,” chains of transactions that represent Bitcoin laundering. By identifying these subgraphs rather than illegal wallets, this approach allows us to focus on the “multi-hop” laundering process in general, rather than the behavior of specific illegal actors in the chain.

The training data contains “subgraphs”: chains of transactions, some of which are known to constitute money laundering.

Testing our results

We worked with a cryptocurrency exchange to test whether this technique could be used to identify money laundering attempts through that business. Of the 52 subgraphs of “money laundering” predicted by the model and which ended up with deposits on this exchange, the stock exchange confirmed that 14 were received by users who had already been marked as linked to money laundering. On average, less than one in 10,000 of these accounts are flagged as such, suggesting that the model is working very well^*. Importantly, the exchange’s insights are based on off-chain information, suggesting that the model can identify money laundering that could not be identified using traditional blockchain analytical techniques alone.

We also investigated the types of money laundering patterns identified by the trained model. This revealed known money laundering patterns such as “skinning chains”, which can already be automatically detected in Elliptic’s transaction and wallet verification tools. However, it also identified new patterns such as the use of intermediary “nested services” in specific ways. Knowledge of these money laundering behaviors is of value to AML practitioners and can add to the set of behaviors that can be detected using Elliptic’s tools.

A machine learning model can also be used to identify previously unknown illegal wallets. When the model predicts that a given subgraph is an example of money laundering, it implies that the funds potentially originated from some type of illegal activity. Targeted research can then be performed on these wallets to try to identify them. This approach has already allowed us to identify a number of previously unknown wallets used by illegal actors, including ponzi schemes and darknet markets.

We share our data with the community

In addition to publishing our research, we also made the underlying data publicly available. The largest public data set of its kind, “Elliptic2” will enable the development of new techniques for detecting illicit cryptocurrency transactions by the wider community. It will also help develop fundamental graph neural network methods, which are used in applications including drug discovery, physics and computer vision.

This new work shows that artificial intelligence methods can be applied to blockchain data to identify illicit wallets and money laundering patterns, which were previously hidden from view. This is made possible by the inherent transparency of blockchain and shows that cryptoassets, far from being havens for criminals, are far more susceptible to detection of AI-based financial crime than traditional financial assets. We’ve barely scratched the surface of what’s possible in this domain, but this work has already led to benefits for Elliptic users. Further cooperation and data sharing will be key to further improving these techniques and fighting financial crime in crypto-assets.

You can read the full research paper here and the Elliptic2 dataset is now available to access. To discuss the research and learn more about how we are applying these new techniques to improve our products, please contact us.