Introduction to federated learning, split learning and splitfed learning

Federated learning (FL) and split learning (SL) are two prevalent methods of distributed machine learning (ML). In both approaches, users train and test ML models without sharing raw data. However, SL offers enhanced model privacy compared to FL due to its separation of the ML model architecture between clients and the server. Furthermore, because of the split model, SL is a superior choice for resource-constrained applications. However, because of the relaybased training across numerous clients, SL performs slower than FL. In this regard, this chapter introduces splitfed learning, a new technique that incorporates the two approaches while eliminating the resulting drawbacks, as well as a refined structural configuration including distinct security and PixelDP to improve data privacy and model its durability. Splitfed is useful in resourceconstrained contexts when complete model training and installation are impossible and quick model training time is necessary to continually update the global model on an evolving dataset. These contexts define different domains.