How did Stripe using artificial intelligence increased cheating detection rates from 2 % to 5 %?

The use of artificial intelligence in Stripe to enhance the accuracy of fraud detection and improve payments is a valuable example of the practical applications of artificial intelligence in the payment industry. This experience shows how artificial intelligence can play a vital role in preventing financial problems, where it causes the most harm to people.

STRIPE, one of the largest online payment service providers, which holds customers such as Openai, Amazon, Google and Apple, at the conference Stripe sessions That was held last month from a new artificial intelligence model called Payments Foundation Model Unveiled. The model was introduced by STRIPE as the world’s first fundamental model for payments and aims to improve fraud and enhance security in payment operations.

Identify credit card fraud with the new model

This model was one of the complex challenges in the payment industry, namely Credit Card Test Manage better than previous methods. Credit Card Test refers to a situation where cyber attackers are trying to test the stolen cards to confirm their credit for illegal purchases.

But how does this model work?

Transformer model in the service of the payment industry

Gatam KodiaIn a post in LinkedIn, the applied machine learning department manager at LinkedIn explained that the previous standard machine learning models at STRPE had been able to reduce the fraud. However, these models needed dedicated training for various tasks such as payment approval, fraud detection, dispute resolution, and the like.

“Given the power of transformer -based public architectures, we thought about whether a similar approach could be used here,” Kodia explained. At first glance, it was not clear that this method would work, because payments are only similar in some respects. “

Stripe developed a fundamental model for payments as a result of this idea; A well -trained model and process the data of each transaction as dense and multipurpose vectors similar to language models. This model based on tens of billions of trained transactions and key signals of each payment in the form of one The unit vector Summary.

“This method leads to the creation of a multidimensional vector space, which represents the position of each transaction in this space indicating rich data and relationships between different elements,” Codia explained.

He added: Transactions with common characteristics are naturally grouped in this vector space. For example, transactions issued by a particular bank are close to each other, and payments that use a similar email address are almost indivisible.

These groupings allowed Stripe to identify and identify sophisticated patterns in suspicious transactions and More precise classification models Build on the features of each payment and its relationship with other payments.

Dramatically increase the accuracy of fraud detection

In the past two years, Stripe has been able to reduce credit card tests for its users by up to 5 %. But more sophisticated attacks, hiding new patterns of attacks in the high volume of large corporate transactions, remains a serious challenge.

“We have developed a classification model that receives a set of vectors from the fundamental model and predicts whether the transaction volume is under attack,” Codia explained. This model A moment It operates to stop the attacks before the businesses are harmed.

The result of this approach is to increase the rate of detection rates of credit card tests from 2 % to 2 % Was only in one night.

Codia went on to point out that the success of STRipe shows payment transactions may have meaning Semantic Be. Just like words in one sentence, transactions have complex dependencies and hidden interactions that are not identifiable by manual characteristics.

Recovery of $ 2 billion in rejected wrong transactions

Another notable achievement of STRPE in Year 2 was the use of artificial intelligence to recover transactions that were mistakenly rejected by card exporters. This effort made more than $ 5 billion Recover for Stripe users.

Product Adaptive AcceptanceWhich uses artificial intelligence to identify rejected wrong transactions, has played a key role in this success. Stripe has announced that the system could identify specific patterns in transaction data that indicate the mistake of a valid payment.

Earlier, the Stripe used the Gradeian Reinforced Tree Model, but then migrated to a deep neural network based on the Tabransformer, called Tabransformer+. This system performs better in modeling complex interactions among hundreds of factors affecting the success of transactions.

With these changes, the accuracy of the new model in identifying valid transactions has been rejected to 5 % increase Found. These advances have helped Stripe recover more for their users, while at the same time the number of re -efforts for transactions Reduce 2 %.

Radar’s 2 % decrease in cheating

In addition, the Stripe cheat prevention tool, called RadarIt has been updated with new capabilities, including automatic authentication. This tool can now activate the two -step authentication flow for transactions and prevent cheating attacks. Early users of this utility tool reduced 5 % Cheating has been eligible for transactions.

Artificial Intelligence, Future Payments Processing

The success of STRPE in the use of artificial intelligence shows the increasing importance of this technology in the payment industry. Other companies, such as Razorpay, also use artificial intelligence to solve problems such as delaying customer payment, simplifying payment settings, and reducing problems related to return on goods.

Overall, using advanced models of artificial intelligence has not only increased payments’ security, but it has been able to significantly improve its user experience. This approach can be a model for other companies active in this field.