Multi-Rail, Multi-Bank Approaches to Payment Fraud Detection
By Florian Loecker (Chief Product & Technology Officer)
Real-time payment fraud detection systems in a multi-rail, multi-bank setting are typically discussed in two flavors: Federated Learning and Graph Machine Learning. In this blog post, we explain why FNA has opted for the second approach.
First, let us start by discussing the purpose of these systems – their purpose is to alert the sending financial institution about potentially fraudulent payments through the use of risk models and rules, historically at the Financial Institution (FI) level. In so doing, financial loss can be prevented before it occurs and, therefore, does not incur costs associated with post-fraud investigation and recovery of funds. It helps FIs maintain positive customer relationships, is a required activity in many jurisdictions, averts the liability of FIs, and, most importantly, keeps participants in the financial system safe from significant financial harm.
In the context of Fraud Portals and data sharing initiatives, there are two notable approaches under discussion:
Federated learning
Graph Machine Learning
Both approaches share a common objective, which is to make machine learning and rules-based systems deployed at banks better at detecting fraudulent transactions. This means fewer false positives and false negatives.
However, they are quite different from a deployment and methodology perspective. Let’s start with federated learning – this enables multiple financial institutions to collaboratively train a machine learning model on sensitive customer or transaction data without ever sharing the raw data itself. Each bank trains the model locally on its own data — such as transaction histories, fraud labels, or behavioral features — and sends only the updated model parameters (not the data) to a central aggregator. The aggregator combines these updates into a new global model, which is then sent back to all participants. This process repeats iteratively until the model converges. Importantly, the actual execution of the model for transaction scoring happens locally at the bank.
The key benefit is that the banks retain full control over their data, minimizing regulatory, privacy, and competitive risks while still improving the accuracy of the shared model.
This contrasts with approaches based on graph machine learning. In this case, transaction data and fraud labels are shared with a central utility (after processing through privacy-enhancing technologies, more on that below). The central utility deploys graph machine learning methods, such as graph neural networks, on the pooled data. This produces a fundamentally different vantage point based on multi-rail, multi-bank networks, which cannot be reproduced using models deployed at the bank level (even those enhanced by federated learning). In this way, the model can detect not only behavioral features of accounts in isolation, but can also take into account any coordination of potentially illicit behavior with its transaction partners (and their transaction partners, and so forth). In practice, this approach can significantly increase model performance, as demonstrated by the BIS Innovation Hub in Project Aurora on synthetic data and by FNA on real data.
Banks would call this utility as part of their own counter-fraud workflows through an API, e.g. at the confirmation of payee stage. This API calls the Fraud Portal and retrieves features and scores. These can be merged with existing bank-level transaction scoring outputs and used as part of the bank’s existing workflows.
On the downside, this model poses additional constraints on privacy. Privacy-enhancing technologies such as homomorphic encryption and secure multiparty computation provide theoretical solutions (see our previous blog post), but they fail to perform at a scale of tens of millions of payments per day and response time SLAs below 10ms. Instead, secure cryptographic hashing can meet these requirements and can effectively obscure sensitive personal data such as account numbers.
Due to the significant increase in predictive performance, an increase that cannot be achieved solely by improving bank-level models, FNA has built its approach on graph machine learning. Get in touch if you would like to learn more about this next generation of transaction monitoring systems.
