Split learning

Split learning is a different and more novel approach in the distributed machine learning domain.

This approach offers a unique way to handle privacy concerns. Unlike FL, where model training happens locally on each device, and only model parameters are shared, split learning splits the neural network. The network is divided between the client and the server: the client processes data up to a specific layer (the cut layer), and only the outputs at this layer are sent to the server, which completes the remaining computation.

This method ensures that raw data never leaves the client’s side, enhancing privacy.

Additionally, split learning is beneficial in scenarios where clients have limited computational resources, as the server performs the bulk of the computation.

However, it requires a more consistent connection between client and server than FL.

Split learning is less adopted compared to FL but emerges as an effective alternative, especially when...