Secure Multi-Organization Healthcare Data Analysis Using Federated AI Architectures

Abstract

The difficulties of working in sectors with strict privacy constraints, such as healthcare, result from the fact that data is separated and distributed across multiple sources that are unable to exchange it. The solution to this dilemma is to allow multiple data owners to collaborate on a joint task by using shared/global models while keeping the data decentralized. Three paradigms of federated machine learning are proposed: horizontal federated learning, vertical federated learning, and federated transfer learning. Federated learning processes and architectures with privacy-preserving properties that also address data provenance, access control, and interoperability.

Integrating heterogeneous distributed data sources while maintaining privacy is a major challenge for modern data and AI analysis. Sensitive data in particular, such as healthcare data, privacy is guaranteed by law. Therefore, many healthcare data sources are available based on these data sharing and data monetization are a difficult integration task. Multiple organizations are willing to collaborate by sharing insights from their private datasets, but without sharing the original datasets because healthcare data is highly sensitive. However, existing guidelines permit the use of healthcare data in Europe and America only for either research or health benefits, and not even for commercial use. Nevertheless, several companies are attempting to break this barrier by using fake-generated data for AI modeling purposes. As a result, three main criteria must be considered during the establishment of collaboration for multi-organization AI data modeling: privacy, security, and monitoring.