1) A purchase occurs - the payer uses their own card (friendly fraud) or stolen data (criminal fraud); in card testing these are mass small transactions. 2) The goods/service are delivered or the resource (credits/tokens, compute) is consumed. 3) The payer or the legitimate cardholder files a dispute with the issuing bank, contesting the transaction. 4) The bank reverses the funds to the payer; the merchant loses the transaction amount, the delivered resource and a chargeback fee. A high chargeback ratio risks penalties from payment processors.
From the merchant's perspective this is a threat, not a solution: it causes financial losses, fees and loss of goods/services. It is modeled in order to design defenses (fraud detection, 3-D Secure, KYC).
The payer's own card (friendly fraud) or stolen card data (criminal fraud).
Official
Purchase and delivery of goods/service or consumption of credits/tokens (compute).
Official
Contesting the transaction with the issuing bank to reverse the payment.
Funds reversed to the payer plus a handling fee charged to the merchant.
Overly aggressive blocking rejects honest customers and reduces revenue.
With first-party fraud it is hard for the merchant to prove the customer actually received the goods/service.
Automated small transactions validate stolen cards and generate costs and penalty risk.
In AI services, consumed compute is a cost that cannot be recovered after a chargeback.
The U.S. Fair Credit Billing Act establishes the cardholder's right to dispute charges - the basis of the chargeback mechanism.
The growth of online (card-not-present) commerce makes friendly fraud a dominant share of disputes at many merchants.
Selling AI-model credits/tokens becomes a target of card fraud that feeds access-reselling (relay) markets.
Time complexity: Nie dotyczy (zjawisko ekonomiczne, nie algorytm). Space complexity: Nie dotyczy.
The key defense constraint is distinguishing fraud (especially friendly fraud) from legitimate disputes without blocking honest customers; card testing also generates a high volume of events to analyze in real time.
Criminal (stolen card), friendly (first-party) or card testing.
Most often card-not-present (online); affects available protections (e.g. 3-D Secure).
Single disputes vs mass, automated card testing.
This describes the defense layer (fraud detection); the fraud itself is not a compute paradigm.
Anti-fraud systems conditionally route transactions to approval, manual review or rejection based on risk signals.
Attacks (especially card testing) are highly automated and parallel; detection also runs in parallel in real time.
The phenomenon itself is hardware-independent; the defense (ML fraud detection) may use accelerators but does not require them.
