Mastercard Deploys Large Tabular Model to Detect Payment Fraud

Mastercard has announced the development and deployment of a large tabular model (LTM), trained on extensive transaction datasets, as part of its ongoing efforts to improve security and authenticity in digital payments.

Unlike large language models (LLMs), which are designed to process unstructured data such as text or images, Mastercard's LTM is tailored for structured data found in transaction tables. The model learns to identify relationships and patterns within highly organised datasets, allowing it to better detect anomalies that might indicate fraudulent activity.

Training for the LTM draws on billions of card transactions, encompassing payment events and related data points such as merchant locations, authorisation flows, fraud reports, chargebacks, and loyalty programme activity. Critically, Mastercard states that all personal identifiers are removed prior to model training. The focus is on parsing behavioural trends while minimising privacy risks—an important distinction for AI applications in the financial sector.

The company claims that the volume and diversity of the data compensate for the lack of individual-level information sometimes used in conventional risk models. While anonymisation can reduce certain detection signals, Mastercard asserts that larger and more representative datasets enable the LTM to draw commercially valuable inferences.

LTMs differ from more widely discussed LLMs in both architecture and function. Instead of generating predictions based on token sequences, LTMs analyse relationships across multiple fields in structured data tables. This approach is rooted in traditional machine learning, rather than natural language processing. The model identifies predictable relationships and highlights transaction patterns that deviate from normal behaviour.

Mastercard describes its LTM as an 'insights engine' that can be integrated into existing tools and workflows. The technical infrastructure supporting the LTM includes platforms from Nvidia and Databricks, powering both computing and data engineering.

The LTM has initially been put to use in Mastercard's cybersecurity systems, which oversee multiple fraud detection mechanisms. Traditional systems rely on rules and manual input to flag suspicious behaviours, such as rapid increases in transaction frequency or geographically dispersed transactions in a short time frame. Early results from Mastercard suggest the LTM outperforms conventional approaches in discriminating between legitimate and fraudulent high-value, low-frequency transactions.

To mitigate risk, Mastercard will use the LTM alongside its established systems rather than relying on it exclusively. The company intends to expand the model's reach by increasing the volume of training data to encompass hundreds of billions of transactions, while also exploring API and SDK offerings to support new internal applications.

Areas beyond fraud detection, including portfolio management, loyalty programme analytics, and internal reporting, may also benefit from the LTM's ability to process large volumes of structured data. The company highlights the potential for a single foundation model to reduce development and monitoring costs across diverse business applications.

Nevertheless, Mastercard recognises the inherent risks of deploying a multi-functional model at scale, including the potential for widespread effects should a failure occur. Regulatory scrutiny is expected, particularly given the influence of such systems on credit decisions and fraud outcomes. Model explainability, transparency, and robust data governance remain central concerns as adoption progresses.

While LTMs represent a novel approach to AI in payments and banking infrastructure, independent validation of effectiveness remains limited. Issues such as model robustness against adversarial actions, ongoing operational costs, and eventual regulatory acceptance will shape broader deployment in the financial sector.

artificialintelligence-news.com