Introduction

In this article, I am summarizing the content of How We Built It: Stripe Radar, which Stripe company has published on their official blog.

What is Stripe Radar

Stripe Radar is a fraud prevention solution provided by the online payment platform Stripe. This tool operates within a short span of time from the moment a user clicks the "purchase" button until the transaction is confirmed. Stripe Radar evaluates over 1,000 features and determines whether a transaction is fraudulent or not within 100 milliseconds. Specifically, this judgment is made within 100 milliseconds.

Online payment fraud poses a significant challenge for both businesses and consumers. Effective fraud prevention tools need to be accurate, fast, and operate at low cost per transaction. Furthermore, they must strike a balance between blocking bad transactions while minimizing false positives (blocking legitimate payments). Stripe Radar is designed to meet these requirements.

How We Built It: Stripe Radar introduces the following lessons for building ML systems like Stripe Radar.

Lesson 1: Avoid Stagnation in Machine Learning Architecture

During the initial stages of Stripe Radar's development, relatively simple machine learning (ML) models (e.g., logistic regression) were used. However, as Stripe's network grew and ML technology evolved, the transition was made to more advanced models (e.g., deep neural networks).

The latest architecture change took place in mid-2022. With this change, the transition was made from the Wide & Deep model consisting of an XGBoost model and deep neural network (DNN) to a pure DNN model. As a result, model training speed improved, and scalability increased.

It's important not to get too comfortable with the current ML architecture and to continually ask oneself, "If starting over today, what kind of model would I build?"

Lesson 2: Continuously Explore New Machine Learning Features

Enhancing the performance of fraud prevention models requires feature engineering. The Stripe Radar team extensively investigates past fraud attacks to understand fraudster thought patterns.

Several processes are necessary to effectively identify and implement features. The Stripe Radar team conducts weekly research on new fraud trends and devises features based on the results. Once features are determined, prototypes are created to understand their impact.

Lesson 3: Explanations Are as Important as Detection

It's crucial not only to detect fraud but also to explain to users how the determination was made. Particularly in cases of false positives or instances where fraudulent transactions are mistakenly allowed, users want to know the reasons behind these decisions. To address this challenge, Stripe Radar introduced the "Risk Insight" feature.

The Risk Insight feature displays contributing features to users, explaining why a transaction was denied. Examples include matching the cardholder's name with the provided email address or the number of cards associated with the same IP address. This enhances users' understanding of specific denial contexts.

Conclusion

The development process of Stripe Radar offers significant lessons in line with the evolution of fraud prevention technology. The first lesson is to avoid stagnation in ML architecture. Just as technology evolves, models must evolve as well. The second lesson is to continuously explore new features. Feature engineering is crucial for improving fraud prevention model performance. Finally, transparency is equally important as detection. Transparency to users is essential for building trust and understanding.

References

https://stripe.com/blog/how-we-built-it-stripe-radar
https://arxiv.org/abs/1606.07792