By Dr. Jochen Papenbrock, Head of Financial Technology, EMEA, NVIDIA, and Prabhu Ramamoorthy, Partner Developer Relationship Manager, Financial Services, NVIDIA
Financial institutions globally, including banks, are subject to the Financial Action Task Force (FATF) for combating financial crime, terrorist financing, and preventing money laundering. Artificial intelligence (AI) is increasingly being used to fight financial crime and offers new opportunities to help detect and prevent fraud. AI and associated machine learning (ML) or deep learning (DL) models provide bank case managers, regulators and compliance officers with powerful new capabilities.
AI is the capability of a computer program, referred to as a machine, to think and learn and take actions without being explicitly encoded with commands. AI can be thought of as the development of computer systems that can perform tasks autonomously, ingesting and analyzing enormous volumes of data, then recognizing patterns in that data. ML, a subset of AI, is the practice of using algorithms to parse data, learn from it, and then determine or predict next steps. Improving on the traditional rules-based approach or the hand-coding of software routines with a specific set of instructions to accomplish a particular task, AI algorithms are trained using large amounts of data, enabling them to learn a task. DL is a subset of machine learning, with the difference that DL algorithms can automatically learn representations from data such as images, video, or text, without introducing human domain knowledge. The word “deep” in deep learning represents the many layers of algorithms, or neural networks, that are used to recognize patterns in data. DL’s highly flexible architectures can learn directly from raw data, similar to the way the human brain operates, and can increase their predictive accuracy when provided with more data
The easiest way to think of their relationship is to visualize them as subsets within overlapping AI — the idea that came first — the largest, then machine learning — which blossomed later, and finally deep learning — which is driving today’s AI explosion — fitting inside both.
Why is AI an Effective Tool?
AI models and algorithms can consume and synthesize massive volumes of data. Furthermore, AI can ingest the data and act on it in near-real time, enabling authorities to stay in step with the movements of bad actors rather than remaining days or weeks behind.
AI models are designed to detect anomalies in the patterns of data they are ingesting by scoring those behaviors relative to expected benchmarks, so that banking compliance officers are alerted when potentially nefarious interactions occur. The investigations tied to these alerts are often led by compliance personnel within banks, and are therefore time-consuming and costly.
Rules- vs. Model-based Approaches to Combating Money Laundering
Money laundering is a process that criminals use to hide the illegal source of their funds. By passing money through multiple, sometimes complex, transfers and transactions, the money is “cleaned” of its illegitimate origin and made to appear as legitimate business profits.
Technological advances in digital banking, online account opening, open banking and cryptocurrency have made tracking the source of funds and uncovering suspect patterns and behaviors far more resource-intensive for financial institutions and their regulators. Traditional methods of automation are unable to keep up with the increasingly sophisticated ways the financial system is abused, so the FATF has encouraged the digital transformation of Anti-Money Laundering (AML) and Counter-Terrorist Financing (CTF) efforts.
Legacy rules-based AML systems have a false positive rate > 90%, meaning investigators’ valuable time is wasted on the wrong transactions. These approaches rely on databases of human-engineered rules that are used to spot patterns indicative of fraud. Figure 2 illustrates a rules-based approach to identifying suspicious financial transactions. Here, a large set of rules is defined and then applied to all financial transactions. If a transaction matches any of the rules, an alert is triggered.If the alert is incorrectly triggered (false positive), it incurs a cost. If no alert was triggered, but one should have been (false negative), a new rule is designed. But fraudsters are becoming savvy at avoiding patterns that rules-based systems can easily recognize.
Many financial institutions are therefore transitioning to model-based systems that reduce false positives by more than 60% and increase anomaly detection by 200%. Using historical actual or suspected financial crime data, the machine learns what is considered normal and suspicious behavior and predicts the risk of money laundering at a more accurate rate, reducing detection from hours to milliseconds and stopping crime in its tracks. Rather than looking for patterns that exactly match pre-defined rules, model-based systems can learn to generalize and identify new fraud schemes that might be new interpretations of old ones. This makes it harder for criminals to avoid detection. They are no longer able to make small adjustments to get around a relatively static set of rules.
In addition to this, graph neural networks (GNNs) can be used by investigators to evaluate relationships between any number of parties to flag potential money laundering behavior. The concept is to construct a heterogeneous graph from tabular data and train a model to detect suspicious transactions and complex laundering activities, as criminals work collaboratively in groups to hide their abnormal features but leave some traces of relationships.
Generative AI and Large Language Models in Fraud Detection and Prevention
ChatGPT was the iPhone moment for AI, and brought Generative AI to the forefront of public discourse. In the financial services industry, Generative AI will play a key role in fraud detection.
Large Language Models (LLMs) can be useful in fraud detection because they can retrieve and analyze information from large unstructured data sources in a short timeframe. Unstructured data can be news, social media or internal organization-generated data in the form of contracts, onboarding documents, audio calls, satellite imagery, trade documents, and invoices/payments. The models can identify and connect different legal entities (companies, individuals, etc.) with associated actions, understand that context, and incorporate this complex information into an information retrieval algorithm for identifying suspicious activities.
Another fraud detection application of LLMs is to monitor conversations, chats and trade activities. Deep Neural Network (DNN) models can analyze the data, draw conclusions and trigger alerts to compliance officers.
Software and Hardware Needs for AI
The implementation of Generative AI requires dedicated hardware and software. On the hardware side, an AI acceleration platform – GPU accelerated compute, accelerated software libraries, data, and storage – is needed for unstructured and semi-structured data models. On the software side, AI frameworks and data scientist resources to code and deploy solutions at scale are needed.
Accelerated platforms make it possible to process several tasks simultaneously in a significantly shorter time. Before accelerated computing, unstructured and semi-structured language models took weeks or months to train, with results also taking time to be returned. Today, Generative AI LLMs and GNNs can be trained in hours or days, and their results returned in milliseconds. As the amount of data, including transactions and related contracts to those transactions, grows exponentially, more and more advanced models are trained. It is not only the models but also the data processing needed for the models that need to be accelerated to process large volumes of unstructured, tabular data to be useful for real time fraud applications. This is why, in turn, state-of-the-art computational accelerators such as NVIDIA are needed with integrated hardware and software solutions for end-end analytics and ML/AI pipelines.
AI is Here to Stay
Financial regulators, bank executives, and risk and compliance officers are prioritizing an investment in AI for financial compliance and fraud detection. AI technologies can help by analyzing large amounts of data and using advanced algorithms to identify suspicious activity. AI can adapt to changes without human intervention catching fraudsters in the act of committing financial crimes. For a successful implementation, financial institutions typically rely on a combination of models for use cases in different stages – for example, Generative AI and LLMs for unstructured data analysis, and ML and Graph Neural Networks for analyzing and visualizing transactions.
AI is here to stay and will continue to unlock opportunities for the financial services industry. Beyond its use in preventing financial crime, firms are leveraging AI to create more relevant customer experiences, improve risk management, drive operational efficiency, and deliver more value to customers.
