Federated learning without communication

Hi All,

I’m looking into the feasibility of using flower for use in the heathcare setting.
The title is a bit misleading, but what I wanted to say is that we want to handle the transfer of the weights on our own (read manually) through our own data sharing portals across various sites that are involved in the federated learning process.

  1. Before, I jump into the code, I wanted to know if the communication part of Flower is separate from the aggregation algorithms, put differently, can I transfer weights manually and still use Flower only for the aggregation part? Or are the communication and aggregation modules too intertwined for that to be possible?

  2. This is more of a general question that you may not have an answer to: In your view, is manual transfer of weights feasible? How many weight transfers do you generally need to in your use cases?

I know the second question is very broad, but just wanted to get your input anyway!

Thank You,
catnip

1 Like

Hi @catnip , welcome to the Flower Community!

Great question! Flower’s communication layer is completely independent of the core framework. There’s even an API that can be used to run the Flower SuperLink and the Flower SuperNode on top of other infrastructure.

The API is generally referred to as “gRPC Adapter”. It’s a bit hidden because it’s not aimed at regular users. But it’s used by some notable projects in the Flower ecosystem.