Rapid technological advancements have significantly increased the significance of artificial intelligence (AI) in anti-money laundering (AML) systems/programs, which are crucial to preventing financial crimes, such as drug trafficking, terrorism financing, illegal arms sales, and human trafficking. AI can transform the existing AML efforts by more accurately and efficiently preventing and detecting suspicious activities. This article discusses the importance, applications, and challenges of AI in AML.
Image credit: EtiAmmos/Shutterstock
Importance of AI in AML
Traditional AML transaction monitoring involves monitoring high-risk entities, hidden relationships, transaction patterns, anomalies in behavior, and bank, credit, and debit cards. However, traditional AML efforts have several disadvantages, including higher operational costs and manual identification of new scenarios due to a lack of AML system capability, which is strenuous and inefficient.
AI algorithms can effectively identify anomalies and patterns indicative of money laundering activities, analyze substantial amounts of transactional data in real-time, and detect suspicious activities or transactions for more investigation, which can reduce false positives.
Financial institutions can employ AI to automate several manual tasks in AML, such as customer due diligence and high-risk customer and transaction monitoring, to improve the speed and accuracy of their AML efforts, reduce costs on compliance investigations and reviews, and allocate more resources for other critical tasks.
Customer due diligence involves customer identity verification and screening against the sanction list. Automating AML tasks also reduces the time needed for compliance reviews, allowing financial institutions to respond more quickly to potential threats.
Additionally, AI can continuously adapt to and learn new trends and patterns to identify previously unknown/emerging risks and threats, which improve the overall accuracy and effectiveness of AML programs. Moreover, AI can predict potential money laundering activities using predictive analytics based on historical behavioral and data patterns.
AI Techniques in AML
AI methods, specifically machine learning (ML) techniques/algorithms, including decision trees (DT), artificial neural networks (ANN), rule-based methods, support vector machines (SVM), outlier detection methods, random forest (RF), Naïve Bayes (NB), social network analysis (SNA), deep learning (DL), k-nearest neighbor (KNN), graph mining, one class SVM, and k-means clustering, have been utilized to identify money laundering patterns and groups and detect unusual behavior.
Among these techniques, RF, DT, SVM, and neural networks are used extensively in AML systems. For instance, the scalable graph convolutional neural network (SGCNN) can be used for forensic analysis of dynamic, dense, and large financial data.
The analysis outcome in visual form can serve as effective decision support for AML analysts who are involved in reviewing a large number of alerts generated by rule-based AML. Similarly, a multi-channel CNN-based sentiment classifier and a CNN-based sentiment classifier can be employed for financial news processing and social media data, respectively, to augment AML investigation and monitoring.
An unsupervised DL model using an autoencoder classifier has been used to investigate the possibility of committing export fraud by detecting anomalies in the regular data transaction patterns. Illicit transactions in the Bitcoin transaction graph can be predicted effectively using graph CNN (GCNN) and multi-layer perceptron (MLP). MLP can also accurately detect suspicious behavior and categorize money laundering-related crime.
Layering, integration, and placement are the three major phases in money laundering schemes. ML methods such as RF and SVMs can be utilized to classify fraud transactions using large annotated datasets. These data-driven approaches are often used for the layering and placement phases as the bank monitors the transaction data.
However, advanced AI methods such as entity relationship extraction from large news and social media data must be applied to AML in the final phase of integration due to the difficulty in detecting the final phase, as funds have already passed the fraud-detection mechanisms.
SVM/NB/DT, extreme gradient boosting (XGBoost), and an ensemble method using a bagging classifier, RF, and extra trees can be used for watch-list filtering automation to prevent false positives, detect illicit accounts involved in money laundering over Ethereum blockchain, and identification of illicit bitcoin transactions, respectively.
Similarly, the XGBoost-NB classifier, core DT-clustering algorithm, and clustering CLOPE algorithm can be employed to detect suspicious money laundering transactions, while potentially illicit behavior can be detected using logistic regression (LR) and XGBoost.
DT can be utilized to identify suspicious money laundering activities, with the most crucial classifiers for DT used in money laundering investigation and suspicious money laundering transactions. ANN, SVM, RF, DT, and Bayes LR, RF with minmaxscaler method, and an ensemble of isolation forests, one class SVM, GMM, and EM can be used to detect suspicious money laundering transactions, while SVM-RF and temporal-directed Louvain algorithm can be utilized to identify money laundering groups.
Suspicious money laundering accounts can be identified using a probabilistic relational model using the audit sequential pattern (PRM-ASP) mining, while suspicious money laundering transactions can be detected using density-based spatial clustering of applications with noise (DBSCAN)-link analysis and k-means clustering and back propagation neural network.
Dynamic Bayesian networks and neural network-genetic algorithm-clustering can be utilized to detect suspicious financial transactions and money laundering cases with investment activities, respectively. Isolation forest and one class SVM can be employed to identify anomalies in a set of transactions of a non-banking correspondent and detect hidden networks of money launderers. Outlier detection methods can be used to identify fraudulent financial transactions.
Navigating Challenges
AI algorithms primarily depend on high-quality data to make predictions with high accuracy. Thus, data quality issues, such as inaccurate or incomplete data, can lead to false negatives/false positives and reduce the AML program's effectiveness.
Financial institutions have to comply with ever-changing and complex AML regulations. Thus, implementing AI in AML programs by making significant changes to the existing systems and processes while maintaining compliance with regulations can be difficult for these institutions.
Although AI can automate several AML processes, human expertise is necessary to make AI-generated insight-based decisions. However, hiring qualified personnel who can leverage AI to improve AML efforts effectively can become a major challenge for financial institutions.
Bias is one of the key challenges of using AI, as AI algorithms can display bias when they are trained using biased data or not designed properly to address bias, leading to inaccurate predictions and discrimination. Moreover, AI algorithms can be difficult to comprehend, increasing the challenges for financial institutions to explain the making of decisions to auditors/regulators.
For instance, black-box models such as neural networks and gradient boosting models can predict highly accurate results that cannot be interpreted easily. However, local interpretable model-agnostic explanations (LIME), SHapley Additive exPlanations (SHAP), and partial dependence plots (PDPs) can be used to produce explanations on black box models using any data type.
Recent Developments
Financial data is often a target for cyber threat actors and the subject of regulatory controls. Thus, maintaining the confidentiality of substantial amounts of financial data needed to train the AI models is a significant challenge for implementing AI in AML.
In a paper recently published at the 2022 IEEE International Conference on Big Data, the authors proposed a scalable and secure architecture for AI implementation that utilizes confidential computing technology to ensure complete end-to-end protection of the intellectual property of AML algorithm developers and sensitive financial data.
GANs were demonstrated using Intel® Software Guard Extensions (Intel® SGX)-secured cloud infrastructure. The proposed solution architecture can be adapted to support federated ML between mutually distrusting institutions at scale with independent control of data security in transit, at rest, and in use by individual data owners.
Recently, Google Cloud has launched Anti Money Laundering AI (AML AI), an AI-powered product designed to assist financial institutions around the world to detect money laundering more efficiently and effectively.
The AML AI provides a consolidated ML-generated customer risk score based on the bank’s data, including know-your-customer (KYC) data, network behavior, and transactional patterns, as an alternative to rules-based transaction alerting.
The risk score can identify instances and groups of high-risk commercial and retail customers. Additionally, the product can adapt to changes in underlying data to provide more accurate results, improving operational efficiency and overall program effectiveness.
References and Further Readings
Han, J., Huang, Y., Liu, S., Towey, K. (2020). Artificial intelligence for anti-money laundering: a review and extension. Digital Finance, 2, 211–239. https://doi.org/10.1007/s42521-020-00023-1
Kute, D. V., & Pradhan, B., Shukla, N. (2021). Deep Learning and Explainable Artificial Intelligence Techniques Applied for Detecting Money Laundering–A Critical Review. IEEE Access. PP. 1-1. https://doi.org/10.1109/ACCESS.2021.3086230.
The Effects Of Artificial Intelligence In The Anti-Money Laundering [Online] Available at https://sanctionscanner.com/blog/artificial-intelligence-and-anti-money-laundering-17 (Accessed on 24 September 2023)
Google Cloud Launches AI-Powered Anti Money Laundering Product for Financial Institutions [Online] (Accessed on 24 September 2023)
Searle, R., Gururaj, P., Gupta, A., and Kannur, K. (2022). Secure Implementation of Artificial Intelligence Applications for Anti-Money Laundering using Confidential Computing. 2022 IEEE International Conference on Big Data, 3092-3098. https://doi.org/10.1109/BigData55660.2022.10021108.