Mastercard Innovates with an AI Model to Combat Digital Fraud

Mastercard Innovates with an AI Model to Combat Digital Fraud

A Fundamental Model for Payment Security

Mastercard has recently introduced an advanced tabular model designed to analyze transaction data at scale to enhance security and authenticity in the realm of digital payments. Unlike language models that focus on text or images, this model is specifically trained on transactional data. Currently, it is based on the analysis of billions of card transactions, with the ambition to cover hundreds of billions in the future.

The data used to train this model includes various payment-related events, such as:

• Merchant locations
• Authorization processes
• Fraud incidents
• Chargebacks
• Activity related to loyalty programs

Mastercard assures that all personal information is removed before training the model, allowing for a focus on behavioral patterns rather than individual identities. This approach aims to reduce privacy risks, a concern often associated with other forms of artificial intelligence in the financial sector. While anonymization may eliminate some potentially useful signals for risk assessment, Mastercard argues that the breadth of behavioral data more than compensates for this loss.

Understanding the Large Tabular Model (LTM)

Mastercard's Large Tabular Model (LTM) differs from large language models (LLMs) in its approach. While LLMs rely on unstructured data to predict sequences of words, the LTM examines relationships between different fields in multidimensional data tables. This method is closer to pure machine learning than traditional artificial intelligence.

By analyzing raw data, the LTM learns to identify predictable relationships, enabling it to detect abnormal patterns that would not be captured by predefined rules. Mastercard describes this model as an "insights engine" capable of enriching existing products and improving current workflows. The technical infrastructure of the LTM relies on Nvidia technologies for the computing platform and Databricks for data engineering and model development.

Deployment and Practical Applications

Cybersecurity is the primary area where Mastercard has actively deployed this technology. Like many other institutions, Mastercard uses several fraud detection systems that analyze transactional data. These systems require initial human intervention and ongoing monitoring to define what constitutes suspicious behavior.

Among the monitored behaviors are:

• Sudden increases in the number of transactions
• Purchases made by the same user in different regions of the world within a short time

Early results show that this model improves performance compared to traditional techniques in certain cases. For example, it appears to better distinguish legitimate transactions from anomalies in the case of high-value but infrequent purchases. Mastercard plans to integrate this model into hybrid systems, combining established methods with this new technology to comply with strict regulatory standards.

The model could also be used to analyze loyalty program activity, portfolio management, and other internal analyses, where structured data is abundant. Currently, companies often use multiple models tailored to each task, which can lead to high training and validation costs. A single foundational model, adjustable for different tasks, could simplify these processes and reduce costs.

Challenges and Future Perspectives

The multifunctional approach of the LTM carries risks: a failure in a widely deployed model could have systemic consequences. That is why Mastercard is cautiously applying this technology to existing detection systems for now.

Mastercard envisions scaling the data used and enhancing the model's sophistication. It also plans to provide APIs and SDKs to enable internal teams to develop new applications. Privacy compliance, transparency, explainability, and auditability of the models are key responsibilities highlighted by Mastercard.

Any system influencing credit decisions or fraud outcomes will be subject to regulatory scrutiny. Highly structured data, rather than text or images, is at the core of the LTM. Tabular models could mark the beginning of a new era for AI systems in the banking and payments sector.

Current evidence relies primarily on vendor reports, so performance claims should be taken with caution. Robustness against adverse conditions, long-term costs post-training, and regulatory acceptance are challenges that these models will need to overcome. These factors will determine the pace and extent of their adoption, but Mastercard seems confident in the potential of these innovations.