There is a lasting buzz about AI becoming more and more infiltrating almost every part of a company's operations. However, some of the most valuable concrete results can be achieved if AI helps prevent scenarios where it hurts people in the most painful places.
Stripe's use of AI to enhance fraud detection and increase security measures is an important case study on how AI is used in payment processing, especially given the size it operates, along with customers such as Openai, Amazon, Google, and Apple.
At the keynote speech at Stripe Sessions Product last month, the company unveiled the Payments Foundation model as part of its efforts to boost fraud detection and enhance security. Stripe also promoted it as “the world's first basic model built for payments.”
This foundation model highlighted the company's success in identifying card testing. There, bad actors will determine whether the stolen card information is valid and make it possible to purchase and purchase.
Stripe has built a transformer-based “Payments Foundation Model”
Gautam Kedia, who leads applied machine learning at Stripe, explained the company's methodology in a LinkedIn post.
Kedia has revealed that standard machine learning models can help reduce Stripe fraud, but each requires task-specific training for activities such as approval, fraud detection, dispute resolution, and more.
“Given the learning capabilities of generalized transformer architectures, I wondered whether an LLM-style approach would work here. It wasn't clear that it would do that.
Therefore, the company decided to build a basic payment model, a self-monitoring model that learns “a dense, general purpose vector of all transactions,” which is very similar to a language model that embeds words. The model is said to be trained in hundreds of billions of transactions, distilling the key signal for each charge into a single embedding.
“You can think of the results as a huge distribution of payments in a high-dimensional vector space. The location of each embedding captures a wealth of data, including how different elements relate to each other.”
Kedia further explained that payments with shared characteristics are naturally grouped. For example, transactions from the same card issuer are clustered, transactions from the same bank even closer, and payments using the same email address appear to be indistinguishable.
He further showed that these “rich embeddings” make it easier to spot subtle adversarial transaction patterns, making it easier to build a more accurate classifier based on both the capabilities of individual payments and the relationship between other payments in order.
Over the past two years, the company has reduced card testing for Stripe users by 80%. But more refined attacks in which card testers hide new attack patterns in large corporate volumes make them difficult to find in traditional ways, Kedia suggested.
“We have built a classifier that ingested an embedded sequence from the underlying model and predicted whether the traffic slice was in attack,” he said. Kedia added that this works in real time, allowing Stripe to block attacks before hitting business.
“This approach has improved the detection rate of card test attacks against large users from 59% overnight to 97%.”
He also said that Stripe's success may suggest that payment activities contain semantic meaning. “Like sentence words, transactions have complex continuous dependencies and interactions of potential features that cannot be captured simply by manual functional engineering,” he said.
Kedia also shared a table comparing the performance of Stripe's own basic model and “in-service” machine learning models.
Stripe recovered $600 million in a legal transaction that was rejected
In particular, the company also reported using AI to prevent redundant transactions and identify which transactions actually have to be retried.
“This has recorded legitimate transactions with over $6 billion in 2024 legal decline transactions recovered by users.
The company said it resolved the issue through adaptive acceptance, a product that uses AI to automatically identify transactions that were mistakenly rejected.
Stripe said it can recognize different patterns of transaction data. This indicates that a legitimate payment was mistakenly rejected by the issuer as suspected of fraud.
Previously, Stripe used the gradient boost tree model xgboost, but later moved to tabfolker-based deep neural networks. “The system is excellent at modeling complex interactions between hundreds of factors that affect transaction success,” the company says.
He also said the new architecture also features a high-dimensional embedding, allowing payment patterns to be mapped and the model can capture and analyze signals that affect payment outcomes. This allows the model to make “more subtle decisions” that reject transactions to retry, and “more subtle decisions” about how to adjust them for a higher approval chance.
“Based on these improvements, Adaptive Acceptance's new AI model achieved 70% accuracy in identifying legal transactions that have been incorrectly reduced. This increased accuracy allowed us to reduce retry by 35% while recovering more revenue than last year,” the company said.
What's more, Stripe's fraud prevention tool radar has been updated with an automatic authentication feature. Two-factor authentication flow can be enabled so that it can trigger 3DS, the security layer of Stripe's additional layer. Stripe is also backed by a new multi-head model and a decision-making layer, with early users seeing a 30% reduction in eligible transition fraud.
All things considered, AI is widely used in the payment processing industry. Several major global giants, including Stripe's competitor Razorpay, are using AI to tackle delays in customer payments, simplify payment gateway setups, and reduce the issue of return to origin.
