AI and fraud prevention. 6 chapters and a final quiz.

Chapter 1. The cost of fraud and the limits of rules.

Payment fraud covers transactions the account holder did not authorize, as well as those they authorized because they were manipulated. Two reports give a sense of scale, one global and one European. According to the Nilson Report (2024), global card fraud losses reached $33.83 billion in 2023, and the firm projects cumulative losses of more than $400 billion over the next decade. In Europe, the joint ECB/EBA report on payment fraud (August 2024) puts total fraud at €4.3 billion in 2022 across all payment instruments. Broken down by acceptance channel, fraud is clearly concentrated. Card fraud is overwhelmingly card-not-present (remote payments), which accounts for nearly 79% of card fraud by value in the EEA.

Strong authentication moved fraud elsewhere; it didn't eliminate it

In Europe, PSD2 and strong customer authentication (SCA), rolled out through 3-D Secure, sharply reduced fraud on authenticated transactions. Other markets rely on different authentication mechanisms. Fraud moved to flows outside the scope of SCA. These include cross-border transactions with an acquirer outside the EEA, merchant-initiated transactions (MITs), and transactions exempted under transaction risk analysis (TRA). Two other routes bypass authentication without having to defeat it. Social engineering leads cardholders to authenticate a payment they believe is legitimate, and account takeover puts the account holder's authentication factors in the fraudster's hands. SCA did not solve the problem. A single checkpoint at the start of the customer journey no longer covers these cases. Detection must therefore score every transaction continuously: before, during, and after authorization.

• Rule: fixed, explicit logic and a binary decision, easy to get around once the threshold is known
• Statistical/ML model: learns correlations across thousands of transactions and produces a continuous risk score
• In practice, you combine both: rules handle known patterns and regulatory requirements (sanctions lists, limits), while the model handles nuanced and novel cases

Chapter 2. Features: the raw material of detection.

A feature is a numerical variable that describes one aspect of a transaction and serves as input to the model. All the features for a transaction make up its feature vector: the transaction expressed as tens or hundreds of numerical variables. Feature quality determines 80% of performance, far more than the choice of algorithm. Features fall into a few broad families. Many only become meaningful over time and across the network.

Device fingerprinting and behavioral fingerprints

A device fingerprint identifies a device without a cookie by combining browser configuration, screen resolution, time zone, fonts, and sensors. The same device fingerprint showing up on different cards is a classic red flag. Behavioral biometrics models how the user interacts: typing speed, how they hold the phone, and how they navigate. A fraudster who has the username and password rarely reproduces the account holder's gestures. Both signals are passive. They add no friction for legitimate customers.

Chapter 3. Models: rules, gradient boosting, and graphs.

Fraud detection in production runs several models at once, organized as a stack of complementary layers. Each layer answers a different question and covers the others' blind spots. None of them covers everything on its own. That's why vendors talk about a “decision engine” rather than an “algorithm.”

Why gradient boosting (GBM) dominates

Gradient-boosted trees (GBM), implemented as XGBoost and LightGBM, are the most widely used approach in fraud detection, ahead of deep neural networks and single decision trees. The method builds hundreds of decision trees in sequence, each correcting the errors of the one before. Three properties explain its dominance. It excels on tabular, heterogeneous data (numerical + categorical) and handles missing values natively. It tolerates extreme class imbalance, since fraud is often less than 1% of transactions. And it lends itself to explainability through SHAP values, which show which features pushed the score up.

Graphs: seeing organized fraud

Graph analysis represents the entities in the payment system as a network and looks for fraud in the links between them. A GBM scores a transaction almost in isolation. It doesn't use the fact that a card, a disposable email address, and a device all belong to the same money mule ring. The model is built from nodes and edges: nodes are cards, accounts, devices, and IP addresses; edges are shared transactions. You then look for suspicious substructures, such as densely connected clusters, accounts linked to an entity already known to be fraudulent, or risk spreading from one neighbor to the next. Graph neural networks (GNNs) automate this analysis. This layer is what distinguishes opportunistic fraud, such as a stolen card, from industrial-scale fraud, such as a money laundering operation.

{
  "transaction_id": "tx_8842",
  "fraud_score": 0.87,
  "decision": "STEP_UP",
  "top_features": [
    { "feature": "distinct_cards_device_24h", "contribution": 0.34 },
    { "feature": "amount_vs_avg_order_value", "contribution": 0.21 },
    { "feature": "distance_to_flagged_entity", "contribution": 0.18 }
  ],
  "latency_ms": 42
}

Chapter 4. Real-time scoring and the false-positive trade-off.

Real-time scoring means computing a risk score while the authorization itself is in progress. The score must be produced before the authorization request goes out to the network, which leaves a latency budget of a few tens of milliseconds. The entire pipeline has to fit in that window: fetching rolling aggregates, computing features, running model inference, and applying rules. This is a real-time engineering problem as much as a data science one, requiring a low-latency feature store, compiled models, and cached velocity counters.

Precision, recall, and the decision threshold

• Recall: the share of actual fraud that gets caught, to be maximized without letting false positives explode
• Precision: the share of alerts that are actual fraud; low precision overwhelms manual review teams
• The threshold is not a technical setting: it is a business decision that reflects what a euro of fraud costs compared with a euro of lost sales
• Manual review (3DS challenge, step-up, customer call) is the third option between approving and declining: it turns doubt into an informed decision

Production engines split the score range into several bands instead of using a single threshold. Above a high score, the transaction is declined automatically; below a low score, it is approved without friction; in the gray zone, it gets an authentication step-up or a review. This segmentation reduces friction for the vast majority of legitimate transactions and focuses verification effort on doubtful cases. Because fraud patterns change, the band boundaries are constantly recalibrated. That recalibration relies on confirmed fraud and declines assessed after the fact, which only the feedback loop provides.

Chapter 5. The feedback loop: chargebacks and labels.

Labels are the “fraud” or “legitimate” tags attached to transactions after the fact, and they are what a supervised model learns from. This ground truth arrives late and incomplete. Its main source is the chargeback (a disputed transaction reversed through the card network). Network reports supplement it by passing on fraud declared by issuers: TC40 at Visa and SAFE (System to Avoid Fraud Effectively) at Mastercard. Both sources have one thing in common. They describe the transaction after the fact, and only once a third party has reported it: the cardholder who disputes it or the issuer who declares it.

The virtuous cycle and where it leaks

Two phenomena limit how well this loop works. Biased feedback arises because the true outcome is only known for approved transactions. A declined transaction produces neither a chargeback nor a confirmation, so the model gets no information about unjustified declines. Drift comes from changes in both the transaction mix and fraud tactics, which erode the performance of a model left unchanged. Both call for active labeling campaigns and ongoing monitoring of the alert rate, realized precision, and feature distributions.

Chapter 6. Fraudsters' generative AI, tools, and governance.

Generative AI refers to models that can produce text, images, voice, and video. Fraudsters use AI too, and generative AI has industrialized their arsenal. It lets them write flawless, personalized phishing messages at scale, create synthetic identities, get past KYC checks with deepfakes (face and voice), and automate entire campaigns. Malicious models sold on the dark web, such as FraudGPT and WormGPT (2023), lower the technical barrier. A scammer with no coding skills can now build a fraud kit with a few prompts.

The tooling landscape

The market includes the card networks (Visa, Mastercard), which score at the network level, and e-commerce specialists (Forter, Signifyd, Riskified, Sift, Ravelin, SEON). Then come AML/banking fraud platforms (Feedzai, Featurespace, NICE Actimize, SAS) and components built into PSPs (Radar at Stripe). Consolidation is accelerating. Visa's acquisition of Featurespace (announced in 2024, completed in 2025) shows how scoring is being vertically integrated into the networks themselves.

Governing an automated decision model