The growing popularity of cryptocurrencies has given rise to new forms of cybercrime. The partial anonymity afforded by cryptocurrencies has become an advantage for perpetrators of financial crimes such as money laundering. Law enforcement agencies are often at a disadvantage when identifying cases of money laundering via cryptocurrencies, as it is extremely difficult to trace suspects through large amounts of data on the blockchain. .
However, AI and machine learning are particularly capable of analyzing large amounts of data. Thus, notable developments have been made in the field of AI to tackle the issue of financial fraud through cryptocurrencies. Notably, Elliptic, a cryptocurrency intelligence company focused on protecting the cryptocurrency ecosystem from criminal activity, published a paper with the MIT-IBM Watson AI Lab. This paper investigates how machine learning models can identify transactions that may be instances of money laundering. This dataset has the potential to identify Bitcoin flows that may be associated with money laundering activities by detecting instances of cryptocurrency chains that are converted into legitimate currency. This dataset, called Elliptic2, has been made publicly available to facilitate further research into financial crime detection tools.
Some background on money laundering
There are multiple ways to launder money on the blockchain. Bad actors may use cryptocurrencies to convert laundered funds into cash and legitimize transactions. You can also transfer funds overseas or to other accounts without being tracked. Scammers often create wallets to make multiple small transactions from large amounts of funds. This makes it exponentially more difficult for law enforcement to track down a single criminal. According to Chainalysis, an additional $22.2 billion worth of cryptocurrencies were transferred from illegal services in 2023, indicating money laundering activity.
Regulations are in place to address this. That's why Bitcoin is not anonymous, but rather pseudo-anonymous, according to guidance issued by the Financial Crimes Enforcement Network (FinCEN) on how the Bank Secrecy Act of 1970 (BSA) applies to cryptocurrencies. , because networking is a must in the United States. “Knowing your customers well enough to determine what level of risk they represent to your financial institution.”
In accordance with FinCEN, networks are required to have an anti-money laundering (AML) program. This requires conducting a proper risk assessment of your network and determining both the identity and profile of your customers. This may require you to implement know-your-customer (KYC) services and classify potentially illegal accounts. However, this process presents significant challenges. Especially when it comes to contacting individuals. Similarly, identify fraudulent transactions from large and growing datasets. This often leads to high false positive rates and can be a burden on time and resources. By utilizing AI, the effort required for this process can be reduced.
Classifying fraudulent transactions using AI
In 2019, Elliptic published a paper detailing how it uses machine learning to classify accounts as illegal and legal based on their transaction history. Elliptic created the dataset using a large amount of publicly available raw Bitcoin transaction data. The data set consisted of 200,000 Bitcoin transactions over the specified time period. The dataset was a “graph network” of Bitcoin transactions classified as either “legal” or “illegal.” Transactions linked to entities such as exchanges, wallet providers, miners, and other trusted sources were classified as legitimate. Similarly, transactions related to fraud, malware, terrorist organizations, ransomware, Ponzi schemes, etc. were also tagged as illegal.
The set created these binary classifications through a “heuristic-based reasoning process.” For example, accounts that reuse the same address and have a large number of entries are typically associated with legitimate activity. This is because these transactions reduce the anonymity of the entity signing the transaction. Similarly, accounts that consolidate funds managed at multiple addresses into a single transaction, thereby reducing anonymity safeguards for a large number of users, are more likely to be legitimate exchanges.
Conversely, accounts that tend to prefer transactions with fewer inputs will reduce the impact of de-anonymization. Therefore, they are more likely to be illegal. Elliptic used Logistic Regression (LR), Random Forest (RF), and Multilayer Perceptron (MLP) as classification techniques, and employed Graph Convolutional Networks (GCN) for scalability. This study found that random forests significantly outperformed logistic regression and GCN.
At the time, this dataset was the largest labeled dataset of Bitcoin transactions. It was then released to the public to facilitate further research. This allows us to track accounts related to illegal activities.
Identifying money laundering transactions
On May 1, 2024, they published another paper on enhancing money laundering detection using AI. New dataset combats financial crime by identifying whether a particular Bitcoin flow may be associated with money laundering activity, rather than identifying transactions by illegal actors . This is done by observing whether transactions follow strange patterns and whether cryptocurrencies are converted to cash at multiple transaction points.
This model uses “subgraph representation,” a technique for analyzing local structure (or shape) within complex networks. This technique can be used to identify anomalous sequences of chain transactions or “shapes” that clearly resemble money laundering patterns.
The paper concludes that “paths on the blockchain that connect illegal clusters to legitimate clusters without changing ownership of funds are likely to represent money laundering activities by criminals or organizations. ” is based on the theory. Therefore, the chain of transactions from legitimate accounts to illegitimate accounts, referred to as a “multi-hop” laundering process, creates “shapes” that can be identified as subgraphs that are known to be associated with illicit activities such as money laundering. may be created.
Elliptic2 is a dataset of 200 million classified and labeled Bitcoin transactions. This dataset defines fraudulent transactions by time frame, maximum number of hops between accounts, and conditions under which a change of ownership is likely to occur. For this dataset, the time frame was chosen to be one year of blockchain data, and the maximum number of hops was 6. They were also asked to define each step of the transaction as “legal” or “illegal.”
Three methods were used to train the model: GNN-Seg, Sub2Vec, and GLASS. All methods were able to converge with less than 8 hours of inference time and several days of training. However, GLASS was used for further experiments.
In our experiments, we selected 52 subgraphs that were considered suspicious and were redeemed on the exchange. Exchanges were then asked to conduct an assessment of the legitimacy of these accounts based on their due diligence practices. According to the exchange, 14 of the 52 accounts may have been involved in illegal activities. Similarly, another experiment tested the model at scale by looking to determine the source of funds flowing into subgraphs that appeared suspicious. Of these subgraphs, 182 were identified to be related to financial fraud.
Additionally, the model was also able to identify known money laundering techniques, namely chain stripping and nesting services. “Peeling Chain” means that a User reduces or “peeles” a smaller amount from a larger amount to another address and sends the remainder to the User's address. This chain continues, decreasing in quantity and becoming harder to trace. Eventually, the scammers convert these amounts into cash on an exchange. A “nested service”, on the other hand, is a business that maintains an account on a large cryptocurrency exchange and allows for liquidity in the account without having to interact directly with the exchange. These can be exploited by money launderers to cash out cryptocurrencies without trading directly with an exchange.
The Elliptic2 dataset and model are now widely available to facilitate further research.
why is it important
Money laundering through cryptocurrencies has become a major concern for law enforcement agencies, increasing the demand for cutting-edge technology. For example, Chennai Police recently announced a tender seeking tools to analyze cryptocurrency transactions to combat financial fraud.
Furthermore, identifying illegal activity may require precautions such as KYC, which defeats the purpose of anonymity promised by cryptocurrencies. Therefore, the development of AI in this field is important.
Also read:
Stay up to date with tech news: Get a daily newsletter from MediaNama with the day's top stories delivered to your inbox before 9am. Click here to register now!
