StripeStripe Radar: fraud detection as a developer primitive.

Radar is Stripe's real-time fraud detection system. Because Stripe handles a substantial percentage of internet commerce, its model has access to data from millions of global companies. When a card is flagged for fraud on one merchant, Radar instantly applies that knowledge to block the card across every other merchant on the network.

Stripe represents payment requests as complex entities in a global transaction graph. The graph maps connections between cards, cardholder names, email domains, bank routing numbers, physical locations, and IP addresses. This network-level dataset allows Radar to identify coordinated card-testing rings that standalone fraud providers would miss.

Fraud patterns change rapidly. To catch card testing attacks (where bots test thousands of stolen cards on a single merchant), Stripe uses an advanced streaming feature pipeline that aggregates transaction velocity (number of purchases in the last 60 seconds) and updates scores in real time.

Using Apache Flink, Stripe's feature store computes rolling aggregations over sub-second event logs. When a new transaction request arrives, the fraud model queries the Flink feature store to read features like 'number of failed attempts from this IP in the last 10 minutes'. The model returns a fraud probability score before authorizing the transaction.

Stripe balances model stability and adaptability by utilizing a hybrid learning approach. A deep neural network is retrained in batch mode weekly on historic transaction logs, while a lightweight linear online model adjusts its weights continuously as merchants flag fraud events.

This hybrid approach ensures that Radar remains stable against seasonal shopping shifts while adapting to new, emerging fraud patterns in real time.

In fraud detection, Gradient Boosted Decision Trees (GBDTs) are often preferred for their speed. However, Stripe implemented a custom Deep Learning structure to model complex temporal patterns (such as keypress latency and browser scrolling behavior). This sequence modeling capability helps Radar distinguish between human shoppers and automated checkout scripts.

Stripe chose not to sell Radar as an expensive add-on. Instead, it bundled core fraud prevention into the base payment fee. This ensures that every transaction feeds the global model, making the network's predictive capabilities stronger with every transaction processed.

Stripe is integrating Generative AI features directly into its developer interfaces. Using GPT-4 models integrated into Stripe Sigma, developers can now write natural language questions (e.g. 'what is the average chargeback rate of customers using dynamic currency conversion in the UK?') and instantly receive ready-to-run SQL audit scripts.