ANONY-MED 2 is working to make health data securely usable for AI-based diagnoses. The project, funded by the Federal Ministry of Research, Technology and Space BMFTR, implements two key approaches: first, the anonymization of datasets through privacy-preserving data synthesis; and second, the anonymization of AI model parameters through privacy-preserving model training. The goal is to develop an open toolkit and a practical handbook to help hospitals get started with modern anonymization technologies. Using the specific case of stroke therapy as a model, demonstrators are being developed using modern security concepts and architectures. Under the consortium leadership of the Fraunhofer Institute for Applied and Integrated Security AISEC, partners include Charité – Universitätsmedizin Berlin and the software company Smart Reporting GmbH.

AI-based diagnoses using health data are expected to improve personalized therapies for patients in the future. However, this requires that sensitive data be protected (in particular, in compliance with the GDPR). ANONY-MED 2 addresses this challenge: The research project is developing solutions to anonymize raw data through secure synthesis and to secure AI models in a manner compliant with data protection regulations without restricting their usability. It demonstrates that data utility and data protection can be ensured simultaneously through innovative technology.  

This will be achieved through two complementary approaches:

1.    Generation of synthetic data: Anonymization of raw data through privacy-preserving data synthesis – genuine clinical patterns are preserved, but the original data is not disclosed.

2.    Anonymization of downstream AI model parameters – the internal, modifiable variables of an AI model – ensures that models are protected while results remain usable. 

How ANONY-MED 2 Works 

The project combines security architectures such as confidential computing (computing in a protected environment), Trusted Execution Environments (TEEs) (secure areas within processors), and homomorphic encryption (HE) (computations on encrypted data) with privacy-preserving machine learning technologies such as Differential Privacy (random noise to protect individual data points) and Federated Learning (decentralized learning without a central database).

These technologies are being implemented in demonstrators for stroke therapy: from the generation of synthetic data to the training of anonymized AI models for decision support, all the way to an AI-based reporting tool. Refined metrics and evaluation frameworks ensure that remaining data protection risks and clinical utility become traceable and quantifiable. Another work package investigates the extent to which the developed methods enable the sharing of data and models in a manner that is compliant with data protection laws.

Open Toolkit and Practical Handbook 

The project’s components are to be bundled – wherever possible as open-source software – into a modular toolkit. This includes, among other things, software modules for generating synthetic image and routine data, reference architectures for secure AI training, evaluation tools for data protection risks and usability, as well as practical example workflows for clinical applications. This will make it easier for hospitals and healthcare facilities to transparently apply state-of-the-art anonymization technologies in real-world scenarios. In the long term, the ANONY-MED 2 approaches are intended to be transferable to many other medical specialties, such as cancer detection or the treatment of rare diseases.

"ANONY-MED 2 aims to establish a robust, scalable approach to integrating medical data into AI-supported medicine in compliance with data protection regulations – not only in Germany but throughout Europe. The goal is to comply with data protection requirements, harness innovation potential, and improve the quality of care through better AI-supported decisions", says Prof. Dr. Gerhard Wunder, head of the Cognitive Security Technologies department at Fraunhofer AISEC.

ANONY-MED 2 is part of the German federal government’s IT security research framework program "Digital. Secure. Sovereign." [in German] of the Federal Ministry of Research, Technology and Space (BMFTR) within the Anonymization Research Network [in German] and builds on the work of its predecessor project, ANONY-MED [in German].