Learn under heterogeneity
Study algorithms for non-IID data, limited edge resources and variable connectivity. Compare accuracy, communication cost and computational requirements.
MIMIT · STEP · Industrial research & experimental development
FLINT
Federated Learning for INdustrial Twins
Connecting local learning, secure model aggregation and adaptive Digital Twins in edge-based environments
UNINA Principal Investigator · Francesco Piccialli
- Programme
- MIMIT · STEP / EU co-financing
- Planned duration
- 36 months
- Partnership
- TPS · UNINA · OPIWARE
- Validation target
- TRL 7
The challenge
Industrial twins that learn across data boundaries
FLINT addresses the combination of distributed industrial data, limited edge resources and the need for adaptive Digital Twins. Its platform is designed to train models locally and coordinate their updates, rather than transferring raw operational data to a central training repository.
Federated learning is connected to Digital Twins that use learned predictions and new observations to update their state. The project brings together TPS for platform requirements and edge infrastructure, UNINA / M.O.D.A.L. for federated algorithms and privacy mechanisms, and OPIWARE for Digital Twin integration and operational validation.
Planned platform architecture
Local data, coordinated intelligence
An explanatory view of the intended components and information flow, based on the project work plan.
Local site APhysical system + edge learner
Local observations and training data
Raw data stays at the siteLocal site BPhysical system + edge learner
Heterogeneous devices and data
Raw data stays at the siteLocal site CPhysical system + edge learner
Site-specific operating conditions
Raw data stays at the site Model updates · Protected aggregation · Model redistribution
UNINA / M.O.D.A.L. · OR2Federated algorithms & secure aggregation
Heterogeneity · Robustness · Privacy–utility trade-offs
Model services, observations and feedback
Adaptive Digital TwinsMonitoringPredictionDecision support
UNINA / M.O.D.A.L. · OR2
The scientific core: federated algorithms and secure aggregation
The group’s industrial-research objective spans the full project. Its work combines algorithm design, robust protocols, privacy mechanisms and experimental validation.
Build robust aggregation
Implement federated protocols and investigate aggregation strategies, including weighted and clustered approaches suited to heterogeneous data and models.
Evaluate privacy mechanisms
Investigate differential privacy and secure aggregation, together with threats such as membership inference and model inversion. Quantify protection rather than assuming it follows from decentralisation.
Validate reproducibly
Compare solutions using public and synthetic benchmarks, varying federation size, data heterogeneity and privacy settings. Produce experimental code and validation evidence.
Tasks and month ranges follow the project proposal, OR2. They describe the planned work rather than completed milestones. The scope is also reflected in the public UNINA research notice.
Consortium responsibilities
Five connected research and development objectives
OR1TPS · Industrial research
Requirements & logical architecture
Define technical-scientific requirements, platform components and functional specifications.
OR2UNINA · Industrial research
Federated algorithms & secure aggregation
Develop and validate the learning protocols and privacy mechanisms.
OR3OPIWARE · Industrial research
Adaptive Digital Twins
Integrate distributed AI with twin models, interfaces and adaptation mechanisms.
OR4TPS · Experimental development
Edge infrastructure & orchestration
Develop the testbed, node management and federated orchestration.
OR5OPIWARE · Experimental development
Operational validation
Evaluate the integrated platform in productive environments and demonstrators.
Evaluation and intended impact
From controlled experiments to operational evidence
FLINT targets applications in manufacturing, energy and healthcare. The planned validation examines predictive quality, efficiency, robustness and privacy at the algorithm level, then the behaviour of the integrated edge and Digital Twin platform.
Expected outputs include a federated-protocol library, privacy modules, reproducible benchmark experiments, adaptive twin prototypes and an integrated testbed. TRL 7 is the project’s validation target, to be assessed through operational evidence and demonstrators.
Related MODAL research
Explore FL at MODAL →Federated Learning in depth
Discover the methods and publications behind the group’s research direction