Objectives
Validate, within a sandbox environment, a distributed AI architecture integrating Edge AI, Federated Learning and homomorphic encryption, reaching TRL 4. More specifically:
- define the requirements, application scenario and hardware specifications of the edge device;
- develop the Federated Learning platform for both edge and cloud environments;
- integrate a homomorphic encryption scheme to encrypt model weights;
- integrate and test the solution on real SECO hardware;
- assess performance metrics, including model convergence, training and aggregation times, and cryptographic overhead;
- document the solution with a view to future industrialisation in regulated sectors.
Initial challenge
The project addresses the growing need, across regulated sectors such as healthcare, fintech, industry and public administration, for AI solutions that are effective, inherently secure, privacy-preserving and fully compliant with the European regulatory framework, including the GDPR and the AI Act. Traditional AI architectures based on data centralisation are not suitable for highly sensitive data. Federated Learning alone is also insufficient, as the model weights exchanged with the cloud may expose information about the training data. What is missing is a replicable, validated, end-to-end privacy-by-design solution.
Solution
The solution designed and developed as part of the project consists of a secure federated architecture for distributed AI processing across edge nodes. Sensitive data remains confined to the edge device through Edge AI; models are trained collaboratively through Federated Learning without sharing raw data; and local model weights are encrypted using homomorphic encryption before being sent to the cloud, where aggregation is performed directly on the encrypted data.
A final controlled decryption process enables the updated model to be verified and redistributed to the nodes. The approach is privacy-by-design, compliant with the GDPR and the AI Act, and based on open-source technologies.
Benefits
The expected benefits of the project span several areas, from data security to sustainability:
- End-to-end privacy for sensitive data, covering both raw data and information that could be inferred from model weights
- Full compliance with the GDPR and the AI Act
- More than 90% reduction in data traffic to the cloud compared with centralised solutions
- Greater energy efficiency through execution on low-power edge devices
- Scalability and replicability across regulated sectors, including healthcare, Industry 4.0, fintech and public administration
- Enabling new privacy-preserving AI business models
- Strengthening data sovereignty and public trust in AI
IFAB’s role
IFAB played a dual role in the project as both funding body and operational coordinator. As a funder, IFAB selected the project as part of its investment activities in research, innovation and technology transfer, recognising its scientific and practical value and enabling the technical partners to carry out the planned activities.
From an operational perspective, IFAB managed financial reporting and budget monitoring, supervised relations with the partners, organised and led project meetings, and oversaw progress monitoring and the collection of results. This comprehensive involvement confirms IFAB’s mission as an accelerator of high-impact initiatives for the region and its productive ecosystem.
For further information, please contact: projects@ifabfoundation.org
