Currently Flower seems to only offer a centralized Federated Learning pattern, which relies on a central entity (server) for model aggregation.
Are there any plans to implement a peer-to-peer mode where the nodes can auto-discover, elect a leader and reelect if the leader drops off, or have some other implementation which eliminates the central entity?
If not yet and if you believe this is a good idea, would there be openness from Project maintainers for a discussion on the topic and a comntribution?
Hi,
With the current architecture, it is not possible. And I’m not aware of any near-future plans to provide support for it.
Contributions are always welcome; just note that this issue would require both the knowledge of the current architecture and desired major changes; if you’re still up for it, I’d advise opening a GitHub issue with a very detailed plan for the changes.
I will access a few other options for the particular project I am doing as I am on a short timeline, but if I have to settled on using FloweAI, will get back to you with a proposed architecture and/or a PR in case I am able to find a way to neatly provide this capacility.