Feature/SK-1261 | Split Learning in FEDn

[FrankJonasmoelle]

This is a first implementation of Split Learning (for vertical federated learning) in FEDn.

The implementation includes:

• A new session "splitlearning_session", which can be executed from the API client via "start_splitlearning_session"
• A new helper (splitlearninghelper) responsible for saving & loading embeddings and gradients
• A new aggregator (splitlearningagg) that concatenates client embeddings, continues the forward pass and calculates gradients based on the labels it receives
• The logic is based on a forward and backward pass
• In the forward pass, clients generate embeddings and send them to a combiner, where they are concatenated and the gradients are calculated
• In the backward pass, the calculated gradients are sent back to the clients, who update their model parameters by continuing the backward pass
• A vertical federated learning example, based on the titanic data set, for demonstration.

Limitations:

• The current implementation works for only one combiner
• It is assumed that the data & labels are sorted by index
• Training data is processed in one batch