Outpacing fraudsters with AI

Online payments have been growing steadily for years. As a result, the risk of fraud is increasing, but proving it is challenging. Artificial intelligence (AI) can provide a solution. In the AI Lab for Sustainable Finance, researchers collaborate with the banking sector to use intelligent algorithms to better detect criminal activities. However, AI is not a panacea, warns university lecturer Ioana Karnstedt-Hulpus. Five questions for the coordinator of the AI Lab for Sustainable Finance.

Why is it so difficult to detect fraud in digital transactions?

Ioana Karnstedt-Hulpus: “The number of online payments is growing. Especially in the last couple of years, we have witnessed a significant surge in digital transactions, thereby escalating the risk of fraud. When something unusual occurs in these transactions, such as a sudden change in behavior, the fraud detection system raises an alert. Analysts then investigate these transactions to determine if something is genuinely wrong.

With the increasing number of online payments, this alert is raised more frequently, creating a huge workload for analysts. Moreover, these alerts do not always prove accurate. For some models, 99.9 percent of the alerts turn out to be false positives. Consequently, analysts spend an immense amount of time examining legitimate transactions of legitimate users. This allows genuinely problematic transactions to slip through, due to lack of time.”

How can artificial intelligence help in this regard?

“Artificial intelligence can assist in alleviating the workload of analysts. AI can improve detection or give faster results with less effort from the human analyst. A computer model can be trained to identify the characteristics of fraudulent transactions. For instance, how often money is transferred from an account, historical knowledge of an account, whether there is a relationship between the accounts involved in the transaction, and the characteristics of the account holder. This is known as supervised machine learning, where we provide the algorithm with examples of past fraud and instruct it to learn patterns that it can later use to detect fraudulent cases in real-time.

Another approach is what we call unsupervised machine learning. In this case, we do not show the system any examples. We just instruct it to compute certain properties of transactions, and then let it detect statistically anomalous transactions. The hypothesis here is that a fraud case will always stand out because there is something about the transaction that does not add up.”

Millions of online payments are legitimate transactions, and most account holders are legitimate individuals.

In what way does the AI & Sustainable Finance Lab contribute to fraud detection?

“In the AI & Sustainable Finance Lab, we contribute to fraud detection by bringing together experts from the financial sector with our scientists to collaboratively seek solutions. We aim to be a meeting place where financial institutions, such as pension funds and banks, can address the problems and challenges they face. Our expertise is not necessarily related to financial crime. For example, our researchers can also work with financial institutions to find solutions on how to build a future proof investment portfolio, applying the results of academic research directly in practice.

At the time, we have two doctoral candidates collaborating with ING Bank in the fraud detection project, creating graphical representations of the network of digital transactions. In these graphical representations, nodes represent individuals or companies, and the lines between them depict financial transactions. In this network of digital transactions, we use artificial intelligence to discover patterns. For instance, a sudden spike in the quantity of transactions between otherwise unrelated communities could indicate fraud.”

So, in the future, will there be no scams thanks to artificial intelligence?

“If AI manages to eradicate humans from the banking sector, then that might indeed happen. But I do not think that is a clever idea at all. Utopia aside, the road towards significantly reducing fraud is quite challenging.

One problem is that the algorithm needs to be trained with data. Training is challenging due to a lack of data. Thanks to our collaboration with ING, our researchers have access to the actual financial data of a bank. As academics, we are fortunate to be in this position. Most researchers do not have access to real data as the bank – rightfully so – must adhere to privacy laws and regulations. This makes the field advance slower than others.

To overcome this, we also work with synthetic datasets. By collaborating with fraud experts at the bank, our researchers gain a good understanding of the domain, allowing them to build a dataset with properties that mimic the real dataset. They can then experiment and search for patterns. This is an interesting area of research that will receive much more attention in the coming years.

Additionally, the lack of fraudulent data is a significant issue. Millions of online payments are legitimate transactions, and most account holders are legitimate individuals. Therefore, it is challenging for machine learning to identify patterns indicative of fraud because there are so few examples.

Furthermore, criminals are very apt at changing their modus operandi and finding new ways to exploit the systems and the victims. As long as AI models learn patterns from the past, we are a step behind the fraudsters. That is why at the AI Lab for Sustainable Finance we put a particular focus on the unsupervised models that do not get stuck in past behavior.”

Criminals are very apt at changing their modus operandi

What developments do you expect in the near future?

“First, we need to see that the financial industry is heavily regulated. As such, the adoption of AI solutions is still in its infancy in many departments. This brings a lot of opportunities, and I expect to see a lot of progress in optimizing decision making processes.

But there are also many challenges. Current trends in AI are very data hungry. As it is of crucial importance to preserve the privacy of financial matters, I anticipate that AI models in the financial industry will be powered by algorithmic solutions rather than data-intensive approaches. When I mention ‘data-intensive’, I am primarily referring to the process of data collection.

Also, the research community is generally very eager to try new methods. But the most recent AI models have this reputation of being ‘black boxes’, because it is at the moment impossible to decipher how they reach their results. This is certainly not acceptable when we talk about decisions such as accusing someone of fraud. So, I also expect the financial AI community to play a significant role in advancing explainable AI. We certainly aim for that in our Lab.”