MLEDGE — Federated Learning-as-a-Service platform

A federated-learning project doesn’t fail at the algorithm — it fails when several organisations have to agree on how their machines talk to each other, who can read what, which cloud bills it lands on, and what happens when someone tries to poison the model. MLEDGE’s cloud-infrastructure workpackage was about building the platform that makes those decisions tractable. Acuratio built it.

The result is a five-layer FLaaS stack. The execution layer runs locally on each participant’s hardware or in their cloud, packaged in Docker for portability, with Python interfaces for data scientists sitting on top of a Rust core for the performance-sensitive components — and JupyterLab for hands-on exploration. It supports horizontal and vertical FL: Federated Averaging, XGBoost for tabular data, Split Neural Networks (from Acuratio’s research with the MIT Media Lab), Federated Multi-Task Learning, plus federated analytics like Private Set Intersection and Private Join-and-Compute over Secure Multi-Party Computation.

The communications layer is a WireGuard mesh that punches through corporate NATs and firewalls without exposing inbound ports; private keys never leave the node that owns them. The coordination layer never sees raw data or gradients — it only tells nodes when to start a round and who else is participating. The governance layer is a full RBAC system with per-organisation isolation, so the same platform can host many simultaneous projects without leakage. The cost layer wraps the AWS Pricing, GCP Cloud Billing Catalog, and Azure Retail Prices APIs into a single dashboard that compares the same workload across providers, regions, and instance families — turning multi-cloud arbitrage into a default rather than an after-the-fact exercise.

On top of this, Acuratio integrated two security primitives developed with IMDEA Networks. FedQV applies quadratic voting to the aggregation step: nodes spend a vote budget weighted by how anomalous their cosine similarity to the global model looks, which reduces the success rate of poisoning attacks by up to a factor of four versus FedAvg while preserving its accuracy in benign conditions. FedRM-RR combines repeated-median regression with a subjective-logic reputation model that decays over time; it converges 1.6× to 2.4× faster than baseline robust aggregation while remaining resilient to adversarial clients.

The platform is what powers MLEDGE’s other two use cases — the steam-boiler federation with Inmarepro and the epidemic-risk analytics with Orange España. MLEDGE is funded by the Spanish Ministry of Economic Affairs and Digital Transformation through EU NextGeneration-EU; the work was reviewed by IMDEA Networks.