Money laundering is legally defined as transferring illegally obtained money through legitimate people or accounts so that its original source cannot be traced.
The International Monetary Fund (IMF) estimated in December, 2018 the criminal proceeds laundered annually between 2 and 5 percent of global GDP, or $1.6 to $4 trillion a year. The profits of money laundering are often used to finance crimes, including terrorism, human trafficking, drug trafficking, and illegal arms sales.
To combat this, financial institutions such as banks and other institutions are required to implement anti-money laundering (AML) systems. Falling short of the required AML standards is a form of corporate wrongdoing and poses reputational risk to these financial institutions. Despite current efforts, several multinational financial institutions have been heavily fined by AML regulators for ineffective AML practices in recent years.
Applying Artificial Intelligence to AML Procedures
The introduction of AI for AML purposes enhances and facilitates the overall decision-making process, whilst remaining compliant with policies such as GDPR. AI can help to minimize the number of transactions falsely labeled as suspicious, achieve a demonstrable quality of compliance with regulatory expectations and improve the productivity of the operational resources.
Money laundering schemes can be broken down into three phases: placement, layering and integration.
- Proceeds from criminal activities enter the placement phase, where they are converted into monetary instruments or otherwise deposited in a financial institution (or both).
- The layering phase refers to the transfer of funds to other financial institutions or individuals via wire transfers, checks, money orders, or other methods.
- In the integration phase, funds are used to purchase legitimate assets or to continue financing criminalized enterprises. Once here, illegally obtained money officially becomes part of the legitimate economy.
AI approaches may be applied to identify money laundering activities in each of these phases. Common machine learning methods such as support vector machines (SVMs) and random forest (RF) can be used to classify fraud transactions, using large, annotated bank datasets. These data-driven approaches are normally used for the placement and layering phases because the transaction data is monitored by the bank. The final phase of integration is difficult to detect because funds have already passed fraud-detection mechanisms.
At present, the typical AML frameworks can be decomposed into four layers.
- The data layer, in which the collection, management, and storage of relevant data occurs. This includes both the internal data from the financial institution and external data from sources such as regulatory agencies, authorities, and watch-lists.
- Next, the screening and monitoring layer screens transactions and clients for suspicious activity. This layer has been mostly automated by financial institutions into a multistage procedure often based on rules or risk analysis.
- If suspicious activity is found, it is passed on to the alert and event layer for further inspection.
- Finally, in the operational layer, a human analyst decides whether or not to block or approve a transaction.
The Many AI Approaches You Can Take to AML
Natural language understanding (NLU) and ontology engineering—both subfields of AI—can help relieve the work burden by providing human experts with a score and link relationship visualization based on news data (e.g., the bank’s news database and traditional or social media news sources) concerning the potential fraud entity.
Linked Knowledge Graph Over Entities
From our own experience with these kind of projects, we have realized one of the most efficient approaches to identifying money laundering is to define a linked knowledge graph over entities. Entity recognition is a set of algorithms capable of recognizing relevant entities (e.g., persons, positions, and companies) mentioned in an input text string. Relation extraction detects the relationship between two named entity nouns (e1, e2) in a given sentence—typically expressed as a triplet [e1, r, e2], where “r” is a relationship between e1 and e2.
Entity resolution determines whether references to mentioned entities in various records and documents refer to the same or different entities. For example, the same person can be mentioned in different ways, and an organization could have different addresses.
Expert.ai’s hybrid approach has overcome the main challenges in graph learning for AML (e.g., graph learning/parsing speed and graph size) by using Fast-Graph Convolutional Networks (fast GCN), which dramatically increase training speeds compared with conventional GCNs.
Sentiment analysis can be useful for AML by shortening the investigation period of a compliance officer. It can be applied at different levels of representation, including backlog management, client onboarding and client profile-monitoring stages.
In this context, the goal of a sentiment analysis system is to monitor sentiment trends associated with a client to identify important patterns. When AML investigators identify a company that has potentially been involved in a suspicious transaction, they generally consult the Internet for evidence.
Analyzing sentiment levels from news articles concerning a specific organization can reveal meaningful insight in support of AML procedures. AI-based sentiment analysis can screen thousands of articles in seconds, significantly improving the efficiency and accuracy of the investigation process.
Sentiment analysis can also be used in client profile-monitoring and onboarding processes to research and identify specific pain points of a client and their associations with negative articles. In terms of AI, numerous different techniques have been used for sentiment analysis, including SVM, conditional random fields, and deep neural networks such as CNNs and RNNs.
Explainable AI Methods Are Essential to AML
At expert.ai, we believe the effectiveness of an AI system is contingent upon its ability to explain a specific decision that it made or predicted. The nature of explanation varies according to differences in data and algorithms and, unfortunately, there remains no common or standard framework of explanation.
Communication with analysts is of utmost importance when designing any AML system, because users make the final decision. Explainable AI solutions, such as the expert.ai Platform, operate by providing users with clear information on why a prediction was made (e.g., why the system believes a transaction is suspicious) to assist users in making a decision and furthering their understanding of the technology.
It is important that any system can explain its decisions in a user-friendly way. European policy and GDPR regulations emphasize the need for financial institutions to provide explainable and human-authorized decisions. Thus, it is critical that any future AML method incorporates a human analyst and ensures that the analyst clearly understands the information presented to him or her. A black-box system that labels a transaction as “fraudulent” with no further insight is unacceptable.
AML systems should be cyclical, rather than linear, in which automated methods communicate with and learn from the analysts. In this future communal working environment, a final decision made by the human (who may or may not support the system prediction) is then back-propagated to the system to improve its decision-making capabilities. Only through this joint effort of humans and artificial intelligence can AML procedures achieve true success.