Mastercard Introduces Generative AI System Built On Transaction Data To Power Security, Insights, And Personalization

In Brief

Mastercard is developing a generative AI foundation model trained on anonymized transaction data to improve insights, fraud detection, and payment services while preserving user privacy.

Technology company and global payments network Mastercard introduced a generative AI system designed as a large-scale foundation model intended to support a wide range of applications. The model is being trained on proprietary datasets derived from billions of payment transactions, with personal identifiers removed to preserve user privacy. By analyzing anonymized patterns within this data, the system is designed to generate insights and anticipate future transaction behavior.

The approach draws a parallel to modern conversational AI systems, which predict subsequent words in a sequence, though in this case the model is not intended for dialogue generation. Instead, it is being developed as an analytical engine to enhance existing services, including cybersecurity measures, customer loyalty programs, and tools for small businesses.

The system is being developed with the support of major computing and data infrastructure providers, including Nvidia and Databricks, enabling large-scale processing and accelerated model training. The company has indicated that results from this work are expected to be presented at an upcoming industry conference.

Foundation AI Model Built On Structured Transaction Data To Enhance Payments And Security

The underlying architecture differs from commonly used large language models, which are trained on unstructured data such as text, images, and video. Instead, this model belongs to a category known as large tabular models, which are trained on structured datasets organized in tables. The training process incorporates transaction data at scale, with plans to expand into broader datasets such as merchant location information, fraud indicators, authorization records, chargeback data, and loyalty program activity.

The increased scope of data is intended to improve the model’s ability to identify patterns and generate more accurate predictions. One of the primary areas of focus for application is cybersecurity, where existing systems are already used to detect and prevent fraud. The integration of this new model is expected to strengthen these capabilities through improved pattern recognition and reduced false positives.

Current cybersecurity models typically rely on engineered features created by data scientists to highlight specific signals within transaction data, such as sudden changes in spending behavior. In contrast, the new system is designed to learn these patterns with minimal manual feature engineering, allowing it to identify relationships in the data that may not be immediately apparent through traditional methods.

Initial testing suggests improved performance compared with conventional machine learning approaches, particularly in reducing false positives in scenarios involving legitimate but uncommon transactions. The system has demonstrated the ability to better distinguish between unusual yet valid activity and potentially fraudulent behavior.

Additional potential applications include enhancements to personalization systems, optimization of rewards programs, improvements in portfolio analysis, and more advanced data analytics capabilities. The model is also expected to reduce the need to maintain large numbers of specialized models across different regions and use cases.

Plans for future development include expanding the model’s capabilities, refining its architecture, and introducing application programming interfaces and developer tools to enable broader use across the organization. Continued collaboration with technology partners is expected to support ongoing advancements.

The initiative is being developed in line with established data governance principles, emphasizing privacy protection, responsible data use, and transparency. As development progresses, the model is expected to contribute to increased efficiency, improved security, and enhanced intelligence within payments and commerce systems.