AI Potluck
Contribute
Back to Gap Map Model components / Federated learning

NEBULA

CyberDataLab
open source / Overall score: 2.0

NEBULA is a platform for decentralized federated learning, formerly called Fedstellar, developed at the University of Murcia with armasuisse and the University of Zurich. A web frontend and a controller set up a federation on star, ring, mesh or custom topologies and deploy it as local processes, Docker containers or physical devices, with a core running on each participant. It adds differential privacy through Opacus, simulated attacks with reputation-based defenses, and trustworthiness metrics.

The slug carries a `-dfl` suffix, as the PyPI package does, because the bare name is a common word.

Openness

5 high confidence
5.0
license
AGPL-3.0(OSI
source
public(the frontend, controller and participant core)
core features withheld
no — the proprietary Enterprise Edition is named as a separate commercial offering whose features are "To be determined", and the tree has no enterprise path

NEBULA's frontend, controller and participant core are AGPL-3.0 in one public repository. Its developers offer a proprietary Enterprise Edition beside it, but their commercial FAQ lists its features as not yet determined, and nothing in the repository is held back for it.

Adoption

1 medium confidence
1.0

PyPI downloads of `nebula-dfl`, the platform's own package.

Capability

3 low confidence
3.0

NEBULA can run a federation across real devices as well as emulate one, and it trains with differential privacy. Its controller generates every participant's certificate itself, and the thirteen setup steps in its user guide include none for an organization to admit or authenticate its own node. On that documented evidence it sits a step below Flower; whether the participant runtime enforces certificate admission is not documented.

Verified 2026-09-28