How Stripe Detects Fraudulent Transactions Within 100 ms

2025-12-15 · Source: ByteByteGo Newsletter · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, FinTech & Digital Financial Services · Depth: Advanced, long

Summary

Stripe's Radar fraud detection system processes billions of online payments, achieving 99.9% accuracy in identifying legitimate transactions within 100 milliseconds. The system's architecture has evolved significantly, moving from simpler logistic regression models to a complex Wide & Deep ensemble, and then to a custom deep neural network architecture called Shield NeXt, inspired by ResNeXt. This transition involved removing an XGBoost component that, while accurate, created operational bottlenecks, slowing retraining and limiting advanced ML techniques like transfer learning and embeddings. Shield NeXt reduced training time by over 85%, enabling faster experimentation and the exploration of multi-task learning. Radar's competitive advantage also stems from its vast network data, which provides rich training labels and enables powerful features like network-wide aggregates and embeddings for geographic fraud knowledge transfer. The system balances false positives and false negatives through configurable risk thresholds and manual review queues, and offers explainability features like risk insights and Elasticsearch integration to help merchants understand and collaborate with the model.

Key takeaway

For AI Engineers building high-stakes, real-time ML systems, consider that architectural decisions impacting operational efficiency and future extensibility can be more critical than marginal, immediate accuracy gains. Your team should prioritize scalable training, data advantage through network effects, and robust explainability features to ensure long-term adaptability and merchant trust, even if it means refactoring components that currently contribute to performance.

Key insights

Evolving ML architectures and leveraging vast network data are crucial for high-performance fraud detection.

Principles

• Operational efficiency can outweigh marginal accuracy gains.
• Network effects enhance feature informativeness.
• Explainability fosters user collaboration and data quality.

Method

Stripe's Radar uses a multi-branch deep neural network (Shield NeXt) for fraud detection, leveraging network-wide aggregate features and embeddings. It balances precision and recall via configurable thresholds and manual review, and provides explainability layers.

In practice

• Prioritize architectural flexibility for future ML advancements.
• Implement embeddings for categorical data and knowledge transfer.
• Offer configurable risk thresholds for diverse business needs.

Topics

• Stripe Radar
• Fraud Detection Systems
• Machine Learning Architecture
• Deep Neural Networks
• Feature Engineering

Best for: AI Engineer, AI Product Manager, Machine Learning Engineer, MLOps Engineer, AI Architect

Related on AIssential

• Getting a transit-fraud detector into production by fixing the features, not the model — Effective fraud detection prioritizes feature engineering over model complexity, especially for evolving adversarial patterns.
• We Committed Fraud with OpenAI's New Image Model (and Called Mum) - EP99.38 — AI image generation has reached a level of realism enabling highly convincing forgeries, posing significant societal and economic challenges.
• Why Powerful ML Is Deceptively Easy — Part 2 — Spatial ML models require rigorous evaluation frameworks to ensure true generalization beyond observed data.
• Shapley Value-Guided Adaptive Ensemble Learning for Explainable Financial Fraud Detection with U.S. Regulatory Compliance Validation — Explainable AI models can achieve high fraud detection accuracy while meeting stringent U.S. financial regulations.
• What AI builders can learn from fraud models that run in 300 milliseconds — AI models can detect fraud in milliseconds by analyzing individual transaction patterns and user behavior.
• Enhancing AI Interpretability and Safety through Localised Architectures — Localized hardware ML architectures can enhance AI interpretability and efficiency by leveraging higher per-node expressivity, particularly for smaller datasets.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by ByteByteGo Newsletter.