Machine learning in fraud prevention

Machine learning in fraud prevention

The sun is shining, you’re drinking a cool beverage and your earphones are playing the perfect tune. It’s a song you’ve never heard before, but you instantly like it. How did Spotify know? You’re listening to a playlist that was made especially for you, thanks to the well-trained machine learning algorithms of Spotify.

So yes, machine learning is everywhere. Or should we say artificial intelligence? Does it really matter? We’ll dive deeper into what machine learning is, what it can do in fraud prevention now and towards the future and find out how merchants are currently using it within their own fraud prevention strategies.

Machine learning versus artificial intelligence

First, let’s look at some basics: what are we talking about when we discuss machine learning and artificial intelligence and what is the difference between the two? Artificial intelligence (AI) can be described as the science and engineering of making computers mimic ‘cognitive’ functions like learning and problem solving. Machine learning is an application of AI based around the idea that a machine is able to learn and improve through experience by inputting data. In this article we will talk about how AI is applied to contribute to fraud prevention, meaning we’ll talk mostly about machine learning. However, it’s good to know that in practice, both terms are often used for the same thing.

Card not present fraud

As there are many types of fraud, there are also many ways to use machine learning to prevent fraud from happening. We will focus on online card not present (CNP) fraud, meaning fraudulent transactions where the online payment does not require the debit or credit card to be presented at the moment of purchase (more information can be found here). The most common types of fraud seen in this category are friendly fraud, triangulation, clean fraud, card testing and account takeover. Merchants combat all these variations by using fraud prevention systems that help them detect fraudulent behavior, block transactions and therefore protect their revenue.

Machine learning in fraud prevention

In general, machine learning is applied in fraud prevention by using it to detect patterns in the incoming and processed transactions. By learning from this input, the algorithm is able to tell a fraudulent transaction from a genuine. Okke Formsma, Lead Machine Learning at Perseuss explains. “By analyzing the combined data from our merchants, we optimize the machine learning system to find the right balance between catching fraud and false positives. Because we leverage our data across merchants, you don’t have to become a victim of new fraudulent patterns before you can protect yourself against it.”

This addresses a concern fraud analysts express in the Perseuss Global Survey; how does machine learning cope with new fraud patterns? One of the respondents says: “Fraudsters change patterns so fast, even algorithms need time to learn.” “This is true”, Okke confirms, “when the system has never encountered this new pattern, it is unable to recognize it as fraud. It needs a bit of input and then it will adjust and start recognizing the new fraud.”

Drayton Williams, Investigations Manager at Momentum Travel, shares: “We use a small number of rules to supplement our machine learning model. This allows us to force manual review or to reject a transaction if a new fraud trend appears that our model hasn't learned from yet.” Another respondent comments: “When new fraud types or attack routes are encountered, manual review is always needed”. The value of the ‘human expert eye’ is emphasized by almost all respondents. That makes sense when you ask fraud experts to share their opinion on the relevance of their daily work. Okke again agrees with the respondents: “In most cases it’s the fraud analysts that fill the temporary gap in detecting new patterns, but they can only do so when supplied with enough relevant data to research and analyze a transaction.

That is why it’s so important to show more than just a recommendation or a score. When the machine learning system is a complete black box and there is no relevant outcome, due to the lack of data, there is absolutely no added value of it for the fraud analyst. Perseuss shows detailed information on every transaction and allows the fraud analyst to do broader searches both in the platform and on external websites.” This is underlined in a recent article published on It’s the fraud analyst that in the end is responsible for the outcome of transactions and therefore both needs to have enough trust and visibility on the main triggers causing the machine learning algorithms to flag a transaction as risky.

The strength of machine learning

The absolute strength of machine learning in fraud prevention, is its ability to analyze big bulks of data quickly and detect patterns that no human eye is able to distinguish. In the bigger scope of a merchants’ fraud prevention strategy, all available prevention methods can complement each other: machine learning takes care of the bulk, the manual review team detects new patterns and implements rules to bridge the time the system needs to adopt and recognize the new patterns. For most merchants machine learning forms their first line of defense. The machine learning analyzes all incoming data and advises the manual review team which transactions need an extra review. This highly increases the efficiency of the team. A respondent comments: “Humans do not perform well in repeated tasks. When the number of transactions is very high and most of them are handled automatically, the actual manual review will be more efficient for the few cases that need the extra attention.”

How does it work?

All machine learning tools are different. That’s not a surprise. The data sets available to train and maintain the algorithm are different. To better understand how the algorithms can be trained, it’s important to know what the most commonly used types of learning for a machine are:

  • Supervised learning: the machine is given a data set with examples of all possible inputs and desired outputs and generates patterns and rules of behavior based on that.
  • Semi-supervised learning: the machine is given a small data set with inputs and desired outputs and a bigger data set with unlabeled, uncategorized data.
  • Unsupervised learning: the machine is given a data set with unlabeled data, so the algorithms act on the data without any prior training.

The challenge for data scientist is to find the best fit for their tool. As you can imagine, algorithms trained by a fully labeled data set can very accurately detect known fraud patterns, where the unsupervised algorithms are able to discover abnormalities and previously undiscovered underlying connections. As customers change their behavior over time, the algorithms need a constant feed of fresh data and need to be flexible in adapting to those new behaviors quickly. Same goes for detecting new fraud patterns as mentioned before. There lies an important job for data scientist to adjust their models and algorithms to find a perfect fit and detect as much fraud as possible with little false positives.

Future of machine learning

Machine learning has shown remarkable progress in the past few years. It has become a reliable and indispensable part of fraud prevention and its contribution will only grow towards the future. To keep the development going and to improve the accuracy of the algorithms, machine learning specialists and fraud analyst are always looking for more data to feed their systems. New techniques like behavioral analytics are interesting options, because they make it possible to identify customer behavior patterns. It can also be expected that more biometric data will be added in the future. Using this information will allow machine learning algorithms to increase its accuracy in distinguishing a real customer from a fraudster.

Fraud prevention is an ever changing business which requires merchants to stay agile and alert. Luckily there are many ways to fight fraud and many tools available to help. Do you want to learn more about how Perseuss can help you? Contact us and we’re happy to tell you more on how we apply machine learning in calculating the risk level of your transactions.

Share this post on

Back to news overview