Partially-federated learning: A new approach to achieving privacy and effectiveness

In Machine Learning, the data for training the model are stored centrally. However, when the data come from different sources and contain sensitive information, we can use federated learning to implement a privacy-preserving distributed machine learning framework. In this case, multiple client devices participate in global model training by sharing only the model updates with the server while keeping the original data local. In this paper, we propose a new approach, called partially-federated learning, that combines machine learning with federated learning. This hybrid architecture can train a unified model across multiple clients, where the individual client can decide whether a sample must remain private or can be shared with the server. This decision is made by a privacy module that can enforce various techniques to protect the privacy of client data. The proposed approach improves the performance compared to classical federated learning.