The Challenge
Hospitals, factories and energy operators generate the kind of data that is too sensitive to centralize: patient records, machine telemetry, plant capacity signals. Organizations require the ability to train AI models collaboratively without transferring sensitive data outside their infrastructure.
The added constraint is that many of the devices intended to run these models—such as wearable monitors, industrial sensors, and drones—operate with limited computational resources, strict latency budgets, and unreliable connectivity. In these environments, reducing inference latency and increasing throughput are essential to meet real-time performance requirements. A standard Federated Learning setup does not fit within these constraints, and without model compression, efficient edge deployment is not realistic.
Our Solution
Multiverse Computing is developing FMLDATA, a Federated Learning platform that compresses AI models with tensor networks and quantum-inspired techniques, so training can run across distributed nodes and inference can run on constrained edge hardware.
- Federated training built on Multiverse’s FL core with connectors and a management module for data space integration.
- Tensor network and quantum matrix decomposition techniques to compress models for deployment on low-capability edge devices without significant accuracy loss.
- Layered privacy: differential privacy, homomorphic encryption, Trusted Execution Environments and zero-knowledge proofs on top of secure aggregation (FedAvg, FedAdam).
Target Outcomes
FMLDATA has been developed under España Digital 2026 grant program. The acceptance criteria defined for the project are the next ones:
- Federated training of AI models in secure environments.
- A compressed, optimized model demonstrated running on an edge device such as a mobile phone.
- Platform integrated with a data space connector which is part of the Centro de Referencia de Espacios de Datos (CRED).
Strategic outcome
A federated AI training service that preserves data privacy by never requiring participants to share their raw data, combined with model compression techniques that enable significant gains in latency and throughput.