Payment fraud detection is the check that scores each payment before the money moves, then approves it, asks for more proof, sends it to review or declines it. It runs at the merchant’s checkout, at the payment gateway and at the card issuer or bank. Rules, machine learning and outside signals, such as stolen cards offered for sale, feed that score. Prevention is the other half: controls that make fraud harder to start.
The losses keep growing. Consumers told the US Federal Trade Commission about $15.9 billion in fraud losses in 2025, from 3 million reports. Payment fraud in the European Economic Area (EEA) reached €4.2 billion in 2024, up from €3.5 billion in 2023, according to the joint EBA and ECB payment fraud report. In the UK, UK Finance counts £1.28 billion stolen through payment fraud in 2025, while banks stopped another £1.68 billion of unauthorized fraud.
What Is Payment Fraud Detection?
Payment fraud is any attempt to take money or goods through a payment channel with stolen data or deception. Detection finds those attempts while they are running. For each payment request it answers one narrow question: is this the real customer, paying a real payee, of their own free will?
Every check ends in one of four outcomes:
- Approve: the payment goes through with no extra step.
- Step up: the customer is asked for more proof, such as a one-time passcode or a biometric check.
- Review: the payment is held for an analyst.
- Decline: the payment is refused, and the device or card may be blocked.
Detection and prevention are different jobs. Prevention makes fraud harder to start, with strong authentication, tokenized card numbers and warnings before a first payment to a new payee. Detection catches what gets past those controls. Fraud and risk management in digital payments is the program around both: the risk a firm accepts, its controls, its metrics and its response when losses rise.
Which Types of Payment Fraud Do You Need to Detect?
Each fraud type leaves a different trace, so one score cannot catch them all. The table lists the main types and the first signal a team can watch for each.
| Fraud type | How it works | Signal to watch |
|---|---|---|
| Card-not-present fraud | Stolen card details are used to buy online, by phone or by mail order. | New device, billing and shipping data that do not match, card country far from the IP address |
| Card testing (enumeration) | Bots try many card numbers or security codes with small payments to find cards that work. | Bursts of small authorizations, many cards from one device, high decline rates |
| Account takeover | An attacker signs in to a real account with stolen or reused credentials, then moves money or changes payout details. | Login from a new device, then a password, email, phone or payee change |
| Passcode theft and wallet enrollment | A scammer tricks the customer into sharing a one-time passcode, then adds the card to a digital wallet. | Wallet enrollment soon after a passcode was sent, an old card on a new device |
| Authorized push payment (APP) scams | The victim is manipulated into sending a real transfer to the criminal. | First payment to a new payee, unusual amount, payee name that does not match the account |
| Invoice redirection | A fake supplier email changes the bank details on a real invoice. | Bank details changed just before a payment run, sender domain one letter off |
| Synthetic identity fraud | A mix of real and made-up personal details builds a new identity that opens accounts. | Thin credit file, phone or address shared across applicants |
| Chargeback (friendly) fraud | A buyer disputes a genuine purchase as unauthorized. | Repeat disputes from one customer, delivery proof that contradicts the claim |
| E-skimming at checkout | Malicious script on a payment page copies card data as the customer types it. | New or changed scripts on payment pages, calls to unknown domains |
Card-not-present fraud is where most card losses sit. In the EEA, remote card payments were only about 28% of card payment value in 2024, but around 83% of card fraud. Their fraud rate by value was 0.091%, against 0.007% for card payments made in person. UK Finance counts nearly 3.2 million remote purchase fraud cases in 2025, up 13%, with losses of £423.5 million.
The average loss per card case is falling. UK Finance puts it at £132 for remote purchase fraud in 2025, down from £300 in 2016, and credits strong authentication and one-time passcodes for the drop. Criminals have answered with social engineering: they trick customers into sharing those passcodes, then use them to register digital wallets or make one-off purchases.
Bank transfers show the same shift. In the EEA, manipulation of the payer caused 74% of the value of fraudulent credit transfers in 2024, up from 65% in 2023. In the UK, APP scam losses rose 19% to £576.4 million in 2025, and 66% of APP fraud began online. That is why business email compromise and fake shops matter to a payments team, not only to the email team.
Phishing feeds several rows of the table. CISA describes phishing as email or websites that pose as a trustworthy organization to get personal information. It warns that a fake site’s address may differ by one spelling or by its domain ending, such as .com against .net. Credentials taken this way, or by malware, are the usual start of account takeover fraud.
How Does Fraud Detection Work in a Payment Gateway?
A payment gateway carries the transaction from the merchant’s checkout to the acquirer, then over the card network to the issuer. Fraud screening at the gateway runs before the request leaves for authorization. It is one of several checkpoints, and each one sees different data. The table follows one online card payment from start to finish.
| Checkpoint | Who runs it | What it checks | Possible outcome |
|---|---|---|---|
| 1. Checkout page | Merchant | Repeated card entries from one session, scripts on the payment page, security code and billing address | Limit attempts, add a CAPTCHA, block the session |
| 2. Gateway fraud screen | Merchant’s payment provider or gateway | Risk score from device, IP address, card issuing country, order history and velocity | Approve, send to 3-D Secure, hold for review, decline |
| 3. EMV 3-D Secure | Card issuer, with data sent by the merchant | Transaction, payment method and device data | Pass without friction, or a challenge such as a one-time passcode |
| 4. Authorization | Card issuer, over the card network | The issuer’s own fraud model, the account’s limits and history | Approve or decline |
| 5. After authorization | Merchant, acquirer and issuer | Chargebacks, fraud reports and refund patterns | Labels that retrain rules and models, card reissue |
EMV 3-D Secure is the step that brings the issuer into an online purchase. EMVCo says it passes data about the transaction, the payment method and the device to the issuer. For many purchases the customer simply clicks to buy. For higher-risk ones, the issuer can ask for a challenge such as a one-time passcode, a knowledge-based question or a biometric check.
Card testing shows why one checkpoint is not enough. Visa’s fall 2025 Biannual Threats Report says enumeration attacks now spread their attempts across merchants worldwide at low speed. Each merchant sees too few attempts to trip its velocity rules, so traditional velocity controls often miss the coordinated attack. Merchants still need their own limits, and confirmed attacks should go to their acquirer.
Bank transfers have their own checkpoints. The sending bank screens the payment and checks the payee’s name: the UK runs Confirmation of Payee, and the euro area requires verification of the payee under Regulation (EU) 2024/886. The same regulation says an instant euro transfer must reach the payee’s account within 10 seconds, or the payer’s bank restores the money. There is no overnight window for a second look.
The receiving bank has a job too, because mule accounts collect the money. Under the UK’s APP reimbursement rules, the sending and receiving firms split the cost of a refund 50:50, according to the Payment Systems Regulator. That gives the receiving bank a direct reason to score incoming payments and new accounts.
How Do Rules and Machine Learning Score a Payment?
A score is only as good as the signals behind it. Most systems draw on the same families of data:
- Payment: amount, currency, merchant category, and the card’s issuing country against the shipping and IP country.
- Device and network: device fingerprint, IP reputation, proxy or emulator use, and how many cards one device has tried.
- Account history: account age, usual amounts and payees, and recent password, email or phone changes.
- Behavior: typing and navigation patterns, and the time spent on the page before paying.
- Outside intelligence: cards, logins and domains tied to your customers that show up on criminal markets or phishing sites.
Two approaches turn these signals into a decision, and most teams run them together.
| Question | Rules | Machine learning |
|---|---|---|
| How does it decide? | Fixed conditions written by analysts | Patterns learned from past transactions |
| What does it catch? | Known tricks, such as many orders from one card in a minute | Unusual behavior nobody wrote a rule for |
| How fast can it change? | A new rule can go live the same day | Needs fresh labels, retraining and testing |
| Main weakness | Fraudsters learn the rules and stay under them | Needs good labeled data and regular retraining |
Rules are easy to explain and quick to change after a new scam appears. Models find patterns nobody wrote down, such as an unusual mix of device, amount and payee. A hybrid setup lets rules block known attacks while the model scores everything else. Both need feedback, and that feedback is slow: a chargeback can arrive weeks after the payment.
The rules below show how a team might write three common checks. They are pseudocode, not a product’s syntax. The thresholds (N, M, R) and the time windows are examples: set them from your own data, and test them on past traffic before you switch them on.
# Card testing: many cards from one device in a short window
IF attempts(device_id, last 10 min) > N
AND distinct_cards(device_id, last 10 min) > M
AND decline_rate(device_id, last 10 min) > R
THEN block device_id; require CAPTCHA at card entry
# Passcode theft: wallet enrollment right after a passcode was sent
IF event = wallet_provisioning
AND otp_sent(card, last 15 min)
AND device_id NOT IN known_devices(customer)
THEN hold enrollment; confirm in the banking app
# APP scam: large first payment to a new payee
IF payee_first_seen(customer) < 24 h
AND amount > p95_amount(customer)
THEN show scam warning; delay release or call back
Speed matters on both sides. A stolen card can be used within minutes, so a check that runs before authorization can stop the loss, while a review the next day can only report it.
Stolen data can also wait. Visa studied payment data recovered from the BidenCash and B1ack’s Stash markets. It found that 85% to 93% of the exposed accounts were linked to enumeration attacks at least 12 months earlier.
How Do You Prevent Payment Fraud Before It Starts?
Prevention controls cut the number of attempts that ever reach the score. These are the main ones:
- Strong authentication. In the EU, Delegated Regulation (EU) 2018/389 sets the rules for strong customer authentication (SCA) under PSD2. The EBA and ECB found that card payments verified with SCA were less likely to be fraudulent. Card fraud was 17 times higher when the payee was outside the EEA, where SCA is not legally required.
- Tokenization. An EMV payment token replaces the card number and should not be usable beyond a specific merchant, device or payment scenario. A stolen token is worth far less than a stolen card number.
- Checks at card entry. Ask for the security code and billing address, limit attempts per session, and add friction when one device tries many cards.
- Payment page control. PCI DSS requirements 6.4.3 and 11.6.1 ask merchants to authorize every script on a payment page, check its integrity and get alerts when the page or its security headers change.
- Payee checks and warnings. Check the payee’s name before a first transfer, and warn the customer at the moment an APP scam depends on.
- A second channel for wallet enrollment. Confirm a new digital wallet in the bank’s own app, not only with a passcode sent by text.
Authentication helps, but it is not a wall. The EBA and ECB note that new fraud targets payments that use an SCA exemption, or manipulates real users into approving fraudulent payments themselves. Behavior checks, payee warnings and outside intelligence still matter after SCA is in place.
What Does Threat Intelligence Add to Payment Fraud Detection?
Transaction data shows what happens at checkout. Threat intelligence shows what happens before it: where a card or login was stolen, who is selling it, and which fake pages are collecting more. That earlier view lets a bank or merchant act before the first fraudulent payment, not after the chargeback.
The gap between theft and use is often long. UK Finance says card data stolen in a breach “can be used for months or even years after the incident”. Visa’s 12-month finding points the same way. A card that sits on a market for a year can be found, matched to your customers and reissued before anyone uses it.
Five outside signals matter most for payments:
- Stolen cards offered in carding forums and card shops, matched against your card ranges so the cards can be reissued early.
- Customer logins in stealer logs and combolists, which feed account takeover and credential stuffing. Leaked credential monitoring finds them.
- Phishing kits and panels built to capture card data and one-time passcodes for your brand. See what a phishing kit contains.
- Lookalike domains and fake payment pages that copy your checkout or send fake invoices. See phishing and lookalike domain monitoring.
- Cash-out recruitment: posts that hire people to lend bank accounts or withdraw cash, the mule side of APP and account takeover fraud.
What Brandefense Blackhole Records Show
Brandefense Blackhole, our cybercrime database, holds the offers behind these signals. The fraud chapter of the Cybercrime Report 2026 compared selected posts from 1 January to 27 October 2026. They are advertisements, not verified thefts, but each one points a fraud team at the next payment to watch:
- A 20 October card ad sold refundable stock from $1.50 and non-refundable stock from $0.05, at half price for cards near expiry. A 9 August ad promised a replacement or refund if a card could not be added to a digital wallet. Watch wallet enrollment on exposed cards, and read what stolen card ads promise.
- A 2 February listing offered bank login and email access for $95, next to a claimed balance of $7,871. Email access can let an attacker reset passwords and read alerts, so watch email and phone changes on bank accounts.
- A phishing ad promised to bypass two-factor authentication, yet described waiting for the user’s approval or an SMS. The customer is still part of the attack, so a warning at the moment of approval can break it. Our analysis of phishing offers for sale compares more of these claims.
- A cash-withdrawal recruitment pitch appeared in 831 of 2,448 retained financial-account posts (33.9%). That counts posts, not people, but it is a reason to score the accounts that receive money, not only the ones that send it.
Brandefense covers this ground with its Digital Risk Protection and Cyber Threat Intelligence products. Its platform uses 35+ typosquat algorithms, checks 2.4M+ domains a day and contacts registrars and hosts to take fake sites down, in about 24 hours on average. Findings can reach fraud systems through the API and MCP. Banks and payment firms can read more about threat intelligence for financial services.
Which Metrics Show That Fraud Controls Work?
Blocking fraud is only half the job. A control that also blocks good customers costs sales, and that cost can be larger than the fraud it stops. Track these four together, by channel and by country:
- Fraud rate: fraud losses or fraudulent transactions as a share of payment value or count. It lags, because confirmed fraud arrives late.
- False-positive rate: the share of legitimate payments that were declined or challenged by mistake.
- Chargeback rate: disputed transactions as a share of all transactions, which card networks and acquirers watch closely.
- Approval rate: the share of payment attempts that go through.
Read them as a set. If fraud and approvals rise together, the rules are probably too loose. If approvals fall while fraud stays flat, they are probably too strict.
A simple sum shows the stakes. If 0.1% of payments are fraud and a rule wrongly declines 1% of good ones, it turns away about ten good customers for every fraudster.
Two more measures help the team itself. Review rate and time to decision show whether analysts can keep up. Time from exposure to block shows whether outside signals arrive in time: a card reissued after a market listing never shows up in the fraud rate at all.
Which Standards and Rules Apply to Payment Fraud Controls?
The PCI DSS standard sets baseline technical and operational requirements to protect payment account data. The PCI Security Standards Council retired version 4.0 on 31 December 2024, and version 4.0.1 is a limited revision with no added or deleted requirements. Its future-dated requirements took effect on 31 March 2025, including 6.4.3 and 11.6.1 against e-skimming. The Council has also published separate guidance on those two requirements.
PCI DSS protects card data. It does not by itself stop account takeover or payment scams, so the payment rules below matter as well:
- EU, PSD2: strong customer authentication under Delegated Regulation (EU) 2018/389, the EBA’s technical standards.
- EU, instant payments: a free check of the payee before the payer authorizes a credit transfer, under Regulation (EU) 2024/886. Payment providers in euro-area countries had to offer it by 9 October 2025, and those in other EU countries must do so by 9 July 2027.
- UK, APP reimbursement: for Faster Payments and CHAPS, the sending and receiving firms share refunds 50:50, and most victims should be repaid within five business days.
What Should a Fraud Team Do When an Attack Starts?
Most payment fraud incidents first show up as a pattern: a spike in small declines, a cluster of new payees, or a fake page with your logo. These steps, in this order, fit most of them:
- Confirm the pattern. Group declines and disputes by device, IP address, card range, merchant ID and payee. One source trying many cards points to card testing. Many customers paying one new payee points to a scam or a mule account.
- Contain it. Tighten the matching rule, add friction to the targeted page or endpoint, and rotate API keys if the attack hits your payment API directly.
- Tell your partners. Report card testing to your acquirer or payment provider, and report mule accounts to the bank that holds them.
- Protect exposed customers. Reissue tested or listed cards; our guide on how to respond to stolen payment cards sets out the first 24 hours. For accounts found in stealer logs, force a password reset and end all sessions. Warn customers who paid the scam payee.
- Remove the source. Take down the phishing page or lookalike domain that collected the data, and look for its copies.
- Feed the result back. Label confirmed cases so rules and models learn from them, and record the time from first signal to block.
Start with what is cheapest to check: which lookalike domains and leaked logins already point at your brand. Atlas runs a free external scan of your domain in about 30 seconds, and the domain scanner lists typosquat domains that could host a fake payment page. For continuous monitoring and takedown, book a Brandefense demo. For fraud beyond payments, read what digital fraud is and how it works.
Frequently Asked Questions
What is payment fraud detection?
Payment fraud detection is the check that scores each payment before the money moves. It then approves the payment, asks for more proof, holds it for review or declines it. It uses rules, machine learning, device data and outside signals such as stolen cards offered for sale.
What is digital payment fraud?
It is any attempt to take money or goods through a payment channel with stolen data or deception. It includes unauthorized card use, account takeover, scams that trick people into sending money, and the data theft that feeds them.
What are the types of fraud in the digital payment industry?
Card fraud includes card-not-present fraud, card testing, passcode theft used to add cards to digital wallets, and e-skimming at checkout. Account takeover, authorized push payment scams, invoice redirection, synthetic identity fraud and chargeback fraud complete the list. Each leaves a different signal, so each needs its own check.
How does machine learning detect payment fraud?
It learns patterns from past transactions and scores new ones against them. It can flag a mix of device, amount and payee that is unusual for the customer, including tactics no analyst wrote a rule for. Models need good labeled data, and labels arrive late because a chargeback can come weeks after the payment.
What is payment fraud detection in a payment gateway?
It is the fraud screen that runs at the gateway before a card payment goes to the issuer for authorization. It scores the request with rules, models, device data and velocity, then approves it, sends it to EMV 3-D Secure, holds it for review or declines it. The issuer still runs its own check at authorization.
What is card testing?
Card testing, also called enumeration, is the use of bots to try many card numbers or security codes with small payments to find cards that work. Visa reports that these attacks now spread their attempts across many merchants at low speed, so one merchant's velocity rules may not see them.
What are the most effective strategies to prevent payment fraud across digital payment channels?
Use strong customer authentication such as EMV 3-D Secure, tokenize card numbers, check device and behavior signals, and limit attempts per device. For transfers, check the payee's name and warn customers before a first payment to a new payee. Add threat intelligence to find stolen cards, logins and fake payment pages before they are used.
How does threat intelligence help detect payment fraud?
It shows what happens before the fraudulent payment: stolen cards and logins on criminal markets, infostealer logs, phishing kits, lookalike domains and mule recruitment. A bank or merchant can then reissue cards, reset passwords and take fake sites down before the fraud is used.
What is the difference between fraud detection and fraud prevention?
Prevention makes fraud harder to start, with controls such as strong authentication, tokenization and payee checks. Detection finds the attempts that get past those controls while the payment is being made. A fraud program needs both, plus metrics that show when either one slips.
What is PCI DSS?
PCI DSS is a payment card security standard. The PCI Security Standards Council describes it as a baseline of technical and operational requirements to protect payment account data. Version 4.0.1 replaced version 4.0 at the end of 2024, and its future-dated requirements, including 6.4.3 and 11.6.1 on payment page scripts, took effect on 31 March 2025.
Sources
- European Central Bank and EBA: 2025 report on payment fraud
- UK Finance: Annual Fraud Report 2026
- Visa: Biannual Threats Report, Fall 2025
- EMVCo: EMV 3-D Secure
- PCI Security Standards Council blog: Just Published, PCI DSS v4.0.1
- PCI Security Standards Council blog: Payment Page Security and Preventing E-Skimming
- EUR-Lex: Regulation (EU) 2024/886 on instant credit transfers in euro
- European Banking Authority: RTS on strong customer authentication under PSD2
