View article

Federated learning aims at improving data privacy by training local models on distributed nodes and at integrating information on a central node, without data sharing. However, this calls for effective integration methods that are currently missing as existing strategies, e.g., averaging model gradients, are unable to deal with data multimodality due to different distributions at multiple nodes. In this work, we tackle this problem by having multiple nodes that share a synthetic version of their own data, built in a way to hide patient-specific visual cues, with a central node that is responsible for training a deep model for medical image classification. Synthetic data are generated through an aggregation strategy consisting in: 1) learning the distribution of original data via a Generative Adversarial Network (GAN); 2) projecting private data samples in the GAN latent space; 3) clustering the projected samples and generating …