A Non-centralized Federated Learning Architecture to Obtain Accurate Privacy Preserving Results
Jesús Bobadilla, Abraham Gutiérrez, Santiago Alonso · 2024
Federated learning was introduced in 2016 by Google researchers to train machine learning models to preserve user privacy. Basically, applying the federated learning concept, the companies’ servers receive partially trained models instead of data from the users. These pre-trained models have been run on user devices and then sent to the central server that is in charge of aggregating the local models of the users and sending the resulting global model back to them. Some distributed versions of federated learning have been proposed to exploit data locality and increase fault tolerance, at the expense of reducing accuracy. Our proposed federated learning architecture is in between centralized and distributed approaches, trying to catch the better features of them. Basically, our proposed approach, in each federated training loop, randomly selects a user to play the role of the centralized server. Experiments run in this paper show that this non-centralized fault tolerant strategy returns results as accurate as the centralized federated learning approach; additionally, it is required less federated loops to train the model. The source code of our proposed architecture is provided in GitHub. Some promising feature works are open to test the impact of adjusting the random choice of users to the capability and availability of their device.