Payment fraud detection is software that gives every card or online payment a score for how likely it is to be fraud. It is also called fraud scoring. The score is calculated in a few milliseconds, before your bank approves or declines the payment.
Say your debit card is normally used in Pune, for small amounts. Suddenly, at 2 a.m., someone tries to pay ₹80,000 with it on a foreign website. The fraud detection system compares this payment with your past payments. The amount, the time, the place and the type of shop are all unusual. So it gives the payment a high score, for example 95 out of 100. Your bank sees the high score. It then declines the payment, or it asks for an OTP first.
Fraud detection runs at several points. The card network scores every card payment that passes through it. Your bank runs its own system as well. For online payments, the payment gateway runs one more check before it sends the payment on. Each system uses a machine learning model. The model has been trained on millions of past payments, and some of those payments are known cases of fraud.
The models use more than the payment itself. Device fingerprinting tells them whether the phone or laptop has been seen before. A graph database stores the links between cards, phones and addresses. So it can find a group of accounts that one fraudster controls.
A fraud system can make two kinds of mistake. If it misses a fraud, someone loses money. If it blocks a genuine payment, the customer is annoyed and the shop loses a sale. Engineers adjust the models to keep both kinds of mistake low.
A newer question is how to score a payment when an AI agent buys something on your behalf. In that case, no person is at the keyboard, so many of the usual signals are missing.