RISK IDENT CEO Roberto Valerio sheds light into the way Artificial Intelligence and Machine Learning work to prevent fraud, going way beyond their buzzwordy meaning
It seems everyone is talking about artificial intelligence and machine learning, especially within the fraud prevention sphere. But despite all the buzz, it’s not always clear how these intelligent elements actually help curb fraud rates.
First things first: though they are often used interchangeably, artificial intelligence (AI) and machine learning (ML) are not the same thing. AI refers to machines that are able to carry out tasks in a human way, while machine learning is a component of AI that involves giving a machine access to large amounts of data and allowing it to learn for itself and solve problems based on that data and patterns the machine recognizes. The concepts of Artificial Intelligence and machine learning have been around since the 1950s. However, only recently have they become a reality for businesses due to advanced developments in the field and newfound affordability.
Fraud is advancing at an alarming rate, and one-size-fits-all programs that solely use traditional rules-based systems such as Boolean (true or false) and weighted rules (a percentage of riskiness) can’t always keep up. This is partly due to the fact that today’s fraudsters are not only well-financed, but also come to the table with matching data abundance and technology. And the only way to fight fire is with, well…fire.
Rules-based versus machine learning fraud tools
Fraud prevention tools powered by machine learning begin working the moment a customer initiates a transaction on a website. The software gathers hundreds of pieces of information about the said transaction and analyzes hundreds of risk attributes in search of red flags that may indicate fraud. The machine filters through the data in real-time and delivers results instantly. This, coupled with its ability to recognize brand new fraud patterns on its own, arguably makes it superior to legacy fraud prevention programs.
Some key pieces of transactional information that traditional rules-based programs often examine include:
- the purchase price;
- whether or not the transaction matches a user’s previous purchase history;
- if the customer is using a different device than usual;
- if the delivery address differs from the billing address;
- if there’s a proxy or VPN to mask the user’s IP address;
- if the user’s account has experienced a recent password or address change;
- and if the account recently experienced failed login attempts.
Machine learning systems also examine the transaction features above, but ML programs are able to dig deeper into the information and transaction patterns — both past and present — to better decipher if a purchase is legitimate or fraudulent. Take this scenario, for example:
A trustworthy customer named Simon realizes he forgot his mother’s birthday and needs to buy her a gift. He’s at work and uses his company computer rather than his home computer to purchase the present, which he needs to rush deliver to her holiday home in Morocco in order for it to arrive on time. Simon doesn’t have his list of passwords with him, so he uses three guesses before finally putting in his correct account information.
Boolean and weighted-rules based fraud prevention programs would likely flag Simon’s purchase as suspicious because several features of the transaction make it appear fraudulent: the purchase comes from a different device the customer usually uses, he chose rush delivery to a foreign address and used the wrong account password several times.
Machine learning programs might also flag the complicated transaction as suspicious, but the chances of that happening would be much, much lower. This is because machine learning fraud prevention solutions instantly delve into a wealth of historical data and past patterns of both the customer as well as those from his or her peer group. For example, an ML program might recognize that Simon has bought similar gifts from this particular shop in the past or that other customers within his area and his age group often buy the same item he chose to buy. Using this information — statistical correlations, as it’s called — gives ML programs an advantage over rules-based programs in determining if purchases are fraudulent or not.
The human touch
While machine learning provides exemplary results in predicting fraud cases, a human being is still needed to make manual decisions and thereby train the model. For example, fraud prevention experts are in charge of deciding which machine learning algorithms to use for their products and which data is worth feeding into the system. As we’ve explained before, data quality and relevance are much more important than the amount of data when it comes to refining machine learning systems. Therefore, a machine or any type of artificial intelligence can never replace a human being with years of experience. But when humans and machines work together, criminals have no choice but to admit defeat.