How do BigData and AI help Anti-Money Laundering (AML)?

Money laundering is one of the main threats for the reliable operation of banking and financial institutions. AML is used to describe all the processes aiming to convert large amounts of money and assets from illegal activities (e.g., smuggling, drug trafficking, illegal gambling, black salaries, illegal weapon deals, terrorism funding) into legitimate wealth.

According to United Nations Office on Drugs and Crime (UNODC), money laundering accounts for the largest part of wealth gained through criminal activities and is estimated reaching USD 2 trillion (close to 5% of global GDP) per year. The consequences are multifaceted. The banks and other financial institutions lose their credibility against their customers and receive hefty fines for failing to run effective AML policies and uncover these illegal transactions. In parallel, citizens are being deprived of better service from their governments, due to the non-taxation of the illegally obtained money, while terrorist groups can support their criminal activities.

Despite strict laws, regulations and counteractive investments, money laundering hasn’t seen any serious reduction over the years. This is mainly because of last decade’s technological developments. Digital banking and novel decentralized financial transaction methods (e.g., cryptocurrencies and other blockchain applications) have opened new possibilities for illegal activities.  This has resulted in cybersecurity being raised to top priority and attracting ludicrous funding. According to a recent research by Precedence Research, the global AML software market is estimated to increase from USD 2.8 billion in 2021 to USD 7.4 billion by 2030.

In the newest AML software, there is extensive use of artificial intelligence (AI) and machine learning (ML) methods. This is normal, as AI/ML technologies can empower financial organizations to combine information from recent entries and data legacy systems, enhance customer identity management by profiling behaviours and transactional patterns, monitor compliance with AML regulations, track the wealth paths, and identify unusual money transfers and illegal assets’ acquirement. AI-enhanced AML software can be trained to identify risks, using a pool of historical data as a baseline. Moreover, risk identification can be combined with new information towards establishing a feedback-learning loop that refines and improves the risk-identification capabilities of the software.

One of the most prominent functionalities of the AI/ML-assisted AML software is the detection of anomalies and outliers. This is based on the application of methods that facilitate the monitoring and identification of data that fall out of admissible datasets and specific rules. In an AML context, anomaly detection functions flag as outliers any data that:

• Do not follow the prescribed money transfer path.
• Deviate from the predictable customers’ behaviour.
• Violate rules and legitimate behaviours described in the official AML guidelines.

Nevertheless, AML solution integrators must be cautious on what they label as “anomalous”: The more anomalies that are identified, the bigger will be the process load and alarm-signals generation. Moreover, dealing with a humongous volume of potential violation incidents can be overwhelming, not only for humans, but for machines as well. Also, there is a need for a constant feeding of AI/ML systems with accumulated knowledge from successful operation and the periodical revision of the “anomaly” labelling. Overall, the input and output data should be of very high quality, to boost the anomaly identification process and the general effectiveness of the prediction models.

Anomaly detection techniques can be differentiated based on the technology that they employ, for instance whether they employ AI/ML (supervised or semi-supervised learning). Another important factor for categorizing the anomaly detection systems for AML, relates to the dimensions of anomaly identification (univariate or multivariate) and to the complexity of the information structures that they consider. For instance, the following information can be considered:

• Point: A single-dimensioned incident value that differs significantly from the acceptable dataset.
• Contextual: An incident that differs in context from the rest of the data, hence considered as noise.
• Collective: Multi-dimensional outlier incidents that are usually consisted of subsets of outliers.

Furthermore, anomaly detections systems for AML can be classified as deterministic and stochastic. Systems employing deterministic modelling produce the same outcome for specific input datasets, regardless of the number of times the modelling process is run. Hence, any unknown component is immediately considered an anomaly. On the other hand, stochastic techniques allow unknown components into the model. Therefore, for the same inputs with a deterministic model, they can provide a range of possible outcomes. AML is a heavily regulated space, which makes deterministic models highly relevant. Indeed, deterministic models, help solution integrators to build effective and clear-cut software solutions. Nevertheless, AI/ML-based stochastic models are more efficient when there is a need to identify novel, interesting, and potentially hidden patterns of money laundering behaviour.

H2020 INFINITECH is a flagship EC co-funded project on BigData and AI solutions for digital finance. The INFINITECH partners have already delivered a range of solutions for AML, leveraging the project’s cutting edge BigData and AI technologies. Here are some prominent examples of INFINITECH’s developments in the AML space:

• Nexi S.p.A has created a rule-based anomaly engine that forms and analyses deterministic, rule-based anomalous patterns, corelated with more complex ones, to retrieve suspicious events and subjects related to money laundering crime. It further uses graph database technology to model links between subjects involved in payments transactions.

• Bogasizi University has developed a scalable transaction graph analysis component, that can be used to construct the transaction graph from blockchain transactions, analyse this graph using graph algorithms and trace blockchain-based, fraudulent activities. This component enables financial organizations to analyse blockchain transactions like Bitcoin transactions for fraudulent activities.
• Jožef Stefan Institute has collaborated with the Central Bank of Slovenia in the development of a screening tool framework, which uses anomaly/pattern detection and social cognitive theory to provide transaction screening functionalities. The solution is provided as a docker-file that imports transaction data and metadata, fuses the data together, prepares a suitable data model and imports it for further use.

To learn more about INFINITECH’s solutions: