The Federated Learning allows training of machine learning (ML) models jointly among parties along the computing continuum. The tool leverages federated training algorithms to extract the information from local distributed data to improve ML models globally. The tool is designed to reduce as much as possible the exchange of sensible and private data among learners, by exploiting advanced algorithms to avoid directly sharing the weights or the gradients.
Open source / proprietary
The Federated Learning is an open-source tool, accessible through the Apache 2.0 licence.
The Federated Learning System is a separate component and it will be provided via software stubs in the standard deep learning frameworks targeted by the project i.e., PyTorch and TensorFlow. The following figure depicts the FL architecture of the first prototype. Instead of sharing the models directly, it shares generative models trained on local private data.