Churn prediction and fraud detection are standard machine learning use-cases in the banking and insurance industries. As with all machine learning applications, the model performance strongly correlates with the available data. In particular smaller banks often suffer from too few datapoints to build meaningful models. Together with potentially lacking data science expertise, this drives a desire for collaboration between such organizations. However, banking and financial data are highly sensitive and cannot simply be shared between organizations.
Our platform allows multiple organizations to collaboratively train machine learning models on sensitive datasets without any of the organizations having to trust any other with their data. The massively increased amount of training data results in improved fraud detection and churn prediction models which in turn result in reduced costs for the organizations.
The basic development of the model architecture (requiring interactive iterations) is done by one of organizations on a local dataset. After going through the process of remote attestation, the individual organizations provide cryptographic keys to our platform which enables the platform to decrypt their separately uploaded datasets. Additionally, the training code (based on the model architecture) are uploaded into our platform. The platform then runs the training and outputs the resulting model to all organizations.