Detecting Payment Fraud
Payment fraud is the movement of money out of an account, or onto a payment instrument, by deception or without the holder's authority. It is the widest of the fraud use cases: account takeover, mule accounts and velocity patterns are all ways of committing it, and this page covers what they have in common.
What this fraud pattern looks like
Payment fraud falls into two families. In the first, the payment is unauthorised: someone other than the holder initiates it, using stolen card details, a compromised wallet or a hijacked online banking session. In the second, the payment is manipulated: the holder initiates it themselves, but only because someone deceived them, such as an invoice-redirection email, a fake investment, or a caller posing as the bank.
The channel changes what the fraud looks like. A card-not-present attack appears as purchases at online merchants the cardholder has never used. A bank-transfer attack appears as a payment to a new beneficiary. A wallet attack appears as a transfer followed quickly by a cash-out. The common thread is that value leaves the account in a way its owner did not intend.
Why it is hard to detect
The difficulty is that a fraudulent payment is built from the same parts as a genuine one. It has a valid instrument, a plausible amount, a real beneficiary and, usually, a correct set of credentials. Nothing in the payment message says it is fraudulent, so detection has to work from context rather than from the message.
Manipulated payments are harder still. The holder authenticated, used their own device, from their usual location, and approved the payment. Every control that asks whether the person is genuine answers yes, because they are. What is wrong is the purpose of the payment, which is not in the data.
Two further problems compound this. Fraudsters probe for limits and stay under them, so a fixed threshold teaches them where the line is. And genuine customers occasionally do unusual things, such as a large one-off purchase, so a check sensitive enough to catch the fraud also fires on legitimate behaviour, and the alerts that result must be worked by people.
The signals that expose it
No single attribute is conclusive, so detection relies on signals read together:
- Amount relative to the account. A payment far larger than anything the account has sent before, or one just under a known limit.
- Beneficiary age and novelty. A payee added minutes before the payment, or one the account has never paid.
- Pace. Several payments in a short window, or cumulative value that would be unremarkable over a month but is not over an hour.
- Device and session context. A device the account has not used, or a session whose behaviour differs from the holder's usual one.
- Channel and timing. Activity in a channel the account does not normally use, or at hours that do not match its history.
- What happens next. Funds moved onward or cashed out almost immediately after arriving.
Each is weak alone and strong in combination. A large payment to an old, familiar beneficiary is ordinary; the same payment to a payee added an hour earlier from a new device is not.
The rules that catch it
A rule turns one of these signals into a testable condition. The high-value transaction rule compares an amount with a limit, or with the account's own history, and is the most direct control on large unauthorised payments. Several neighbouring rules cover the other signals:
- The new beneficiary rule flags payments to payees the account has only just added.
- The transaction velocity rule aggregates activity inside a window, which exposes payments kept small to avoid a threshold.
- The new device transaction rule flags payments from a device the account has not used before.
- The unusual transaction amount rule compares an amount with the account's own baseline rather than a shared limit.
Composite rules combine these. Requiring a high amount and a new beneficiary together is far more specific than either alone, and that specificity is what keeps alert volume workable.
Investigating it
An analyst reviewing a payment alert asks first whether the holder made the payment. If the holder can be reached, that is often the fastest answer. If not, the evidence is in the surrounding activity: when the beneficiary was added, whether the device and session were familiar, whether the payment fits the account's history, and where the funds went afterwards.
The outcome sorts the case into one of the two families above, and that matters for what follows: an unauthorised payment points to compromised credentials or a compromised device, while a manipulated one points to a deception the holder may still be under. Accounts that received the funds are worth examining too, since they may belong to the wider mule account network. Confirmed outcomes should feed back into the rules, so a pattern seen once is easier to find the next time.
How Akoben handles it
Akoben receives payment events asynchronously, over REST or as batch files, and enriches each with the account's recent history and the context the sending system supplies. It compares the payment with a behavioural baseline built from that account's own history, and evaluates it against the active ruleset. A payment that matches a rule raises an alert, which Akoben routes into a case queue, where it can be assigned, escalated and tracked against the queue's SLA.
An analyst can change rules directly. Each change creates a new ruleset, applied as a whole and reverted to the previous version if the result is undesirable, without an engineering release. A new ruleset can be validated against historical transactions in batch, and run in alerting-only mode while a team checks what it would flag. Tags applied during an investigation carry forward, so later transactions on a tagged account are scored differently. Rule-effectiveness and false-positive-rate reports show which rules are earning their alerts.
See transaction monitoring for the full description, or case management for the investigation workflow.
Common questions
What is payment fraud?
Payment fraud is the use of a payment instrument or channel to take money by deception or without the account holder's authority. It covers unauthorised payments, such as card details used by someone who stole them, and manipulated payments, where the genuine holder is tricked into sending money. Both appear as transactions in the payment data, which is where monitoring looks for them.
The two kinds need different detection. In an unauthorised payment the question is whether the person initiating it is the holder. In a manipulated payment the holder really did initiate it, so the evidence lies in how the payment differs from what that account normally does.
What is the difference between unauthorised and authorised payment fraud?
In unauthorised payment fraud, the account holder did not make the payment: a criminal used stolen card details or a hijacked account. In authorised payment fraud, the holder made the payment themselves after being deceived, for example into paying a fake supplier. The first fails the question of who initiated it. The second passes it, which is why it is the harder pattern to detect from the transaction alone.
Authorised push payment fraud is the best-known form of the second kind. Because the holder authenticates honestly, credential checks give no signal, and only behavioural context around the payment can.
Read the glossary definition of authorised push payment fraud
Discuss your transaction monitoring requirements
Tell us which transaction flows you need to monitor, which systems Akoben would integrate with, and how your fraud team works today.
Request a demo