CREDIT CARD FRAUD PREVENTION
Milliseconds to decide: Fraud or no fraud?
Interview with Roy Prayikulam, SVP Risk & Fraud at INFORM
CREDIT CARD PREVENTION
„Decide in milliseconds: Fraud or no fraud?“
Interview with Roy Prayikulam, SVP Risk & Fraud at INFORM
Pull out your card, enter your PIN, and get the goods - cashless payment, whether on the Internet or in a store, is quick and easy. The shopper is unaware of the many processes that take place in the background of each transaction. Ultimately, the goal is to protect customers, merchants, and banks from fraud. In this interview, Roy Prayikulam explains how AI-powered software makes this work with millions of daily transactions.
Worldwide, banks and credit institutions process over 3 billion cashless payments for their customers daily. How can these payments be effectively protected from fraud?
At peak times, large banks process up to 5,000 payments per second via Internet banking, online shopping, or mobile banking. Protecting against fraud requires correlating massive amounts of data from customer profiles, past financial transactions, and Internet activity. Within milliseconds, a decision must be made based on these correlations: Fraud or no fraud? This is the only way banks can stay one step ahead of fraudsters. Modern software systems combine AI processes, such as machine learning, with a system of rules based on human knowledge.
VIDEO: AI IN EVERYDAY LIFE
Credit card fraud
In this video, our colleague Tyrone Castelanelli, AI Catalyst at INFORM, explains how AI is working to protect customers, merchants, and banks from credit card fraud.
So, it’s all about generating critical knowledge from large data sets?
Exactly! On the one hand, you need to identify and block fraudulent and suspicious transactions. On the other hand, most transactions are legitimate and should be processed as quickly as possible. Fraud detection systems access a wide range of data to achieve this selectivity. This includes information such as address changes, credit limits, or requests for new passwords. They also look at the amounts and locations where a customer typically makes payments. Anything out of the ordinary could indicate fraud. This is complemented by up-to-date information about a transaction. For example, it may be known that a customer is about to make a purchase at a bookstore. In addition, it is known that payments at a bookstore are rarely large. Therefore, an unusually large payment at a bookstore also indicates possible fraud. These examples illustrate the principle but not the complexity. There are thousands of different decision variables to consider depending on the type of payment.
How does machine learning help?
It allows banks to infer the likelihood of fraud in a transaction from these large data sets. This is done by training an algorithm with historical transactions that have been flagged as fraudulent to recognize the fraud patterns contained in the data sets. In this way, fraud attempts can be detected automatically.
How transparent and accountable are the decisions made by such an algorithm?
Every payment transaction identified as suspicious is reviewed by a human expert at the banks. In order to understand why a transaction is rejected, the algorithm must be transparent and not hidden in a black box. After all, a fraud pattern is a combination of many individual factors. Which parameters in which areas best represent the pattern with a high degree of accuracy is a task for the algorithms. The sheer number of combinations is beyond the capacity of humans.
What role does human expertise still play?
It still plays an important role because a machine cannot replace the years of experience of an expert. We rely on gut instinct and expert knowledge, especially in new environments where there are no existing fraud patterns. Let’s take two examples: If two transactions are made with a credit card in two different locations within a short time, at least one is fraudulent. An AI can recognize this and react accordingly.
And where does AI reach its limits?
To stay with the example: If, for example, a new payment method for online purchases using cryptocurrency comes into play, an expert can store new rules in the system based on experience and evaluation of the transaction data. Fraud patterns that emerge from this need to be repeated several times before the system recognizes them as such. The financial damage would be enormous if you had to wait for this machine-learning model training time for every new fraud pattern.
Is there a trend toward combining machine learning and knowledge-based rules?
Criminals are constantly developing new fraud patterns. Only with such a combination is it possible to react quickly. That’s why the systems developed at INFORM feature so-called “hybrid AI,” which integrates the data-driven and knowledge-driven methods described above in a single piece of software. This makes it possible to search huge amounts of data on a daily basis for recurring correlations and patterns that indicate criminal behavior. This is the only way to keep the percentage of fraudulent transactions so low that fraud occurs no more than once in a million.