Collaboration has been identified by the AI Commission as a key enabler for the successful adoption of AI in Sweden. However, collaborating on tasks involving sensitive data, such as data subject to GDPR, presents significant challenges. A critical issue is the difficulty in assessing the risk of revealing sensitive information when sharing trained machine learning models or synthetic data.

To address this, we are developing LeakPro, an open-source tool designed to evaluate the risk of sensitive data leakage when sharing models or synthetic data. LeakPro supports various healthcare-relevant scenarios, including API-based model sharing, federated learning, and synthetic data generation. It handles multiple data modalities, such as images, text, tabular data, and graphs.

In this talk, we will introduce LeakPro and demonstrate its capabilities through examples from healthcare applications.