Fraunhofer ISST and Fujitsu Research Introduce Federated Unlearning to Safely Remove Data from Collaborative AI Models

© Fraunhofer ISST
Using federated unlearning, decentralized AI models revert to the state they were in before a data provider joined when that provider leaves. From that point on, they are retrained.

(IN BRIEF) Fraunhofer ISST and Fujitsu Research have developed a federated unlearning method that allows companies to remove their data from decentralized AI models when leaving collaborative projects. By rewinding the training process to a point before the data was introduced and retraining the model without it, the system ensures complete data removal while preserving performance. The approach reduces the need for full retraining, minimizes disruption and supports compliance with regulations such as GDPR. With applications in areas like manufacturing, the technology aims to increase trust and adoption of collaborative AI by ensuring data sovereignty and flexibility for participating organizations.

(PRESS RELEASE) MUNICH, 1-Apr-2026 — /EuropaWire/ — Fraunhofer ISST and Fujitsu Research have developed a new approach to managing data in collaborative artificial intelligence systems, enabling precise removal of a participant’s data from decentralized AI models without compromising performance.

© Fujitsu Research
Unlearning at your fingertips: screenshot of the interactive demo on federated unlearning, which experts of Fraunhofer ISST and project partner Fujitsu Research will present at Hannover Messe 2026.

Collaborative AI development often benefits from multiple organizations contributing data, resulting in more robust and accurate models. To maintain data privacy, many companies adopt federated learning, where data remains locally stored and only model parameters are shared. However, a long-standing challenge has been what happens when a participant leaves such a collaboration, as their data remains embedded within the trained model and cannot easily be removed.

To address this, researchers have introduced a method known as federated unlearning. This technique effectively rewinds the training process of a decentralized AI system to the point before a specific partner contributed data, and then resumes training without that data. The result is a model that no longer contains any trace of the departing partner’s information, while retaining most of its learned capabilities.

Florian Zimmer explained that the approach avoids rebuilding the model from scratch, significantly reducing computational effort while maintaining performance. Although some reduction in model accuracy may occur after removing data, the system can quickly recover through continued training.

The concept has clear industrial applications, particularly in manufacturing environments where multiple companies collaborate using shared AI models. For instance, data from different partners can help predict machine failures or optimize performance. If one partner exits, federated unlearning allows the model to adapt without requiring a full retraining cycle.

Janosch Haber noted that traditional methods would require rebuilding models entirely, leading to temporary performance degradation. The new method minimizes disruption and enables faster recovery of high-quality results.

Beyond operational efficiency, federated unlearning also strengthens trust in collaborative AI by ensuring companies retain control over their proprietary data. This is particularly important for organizations operating under strict data protection frameworks such as the General Data Protection Regulation.

Researchers believe the approach could significantly increase adoption of AI in cross-company partnerships by removing concerns around data ownership and compliance. The technology will be demonstrated at Hannover Messe 2026, where experts will showcase its practical applications for decentralized AI systems.

Media Contact:

Britta Klocke
Corporate communications
Fraunhofer Institute for Software and Systems Engineering
Speicherstraße 6
D-44147 Dortmund
Phone +49 231 97677-160
Fax +49 231 97677-199
britta.klocke@isst.fraunhofer.de

SOURCE: Fraunhofer-Gesellschaft