GAN Latent Space Manipulation and Aggregation for Federated Learning in Medical Imaging

Abstract

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 synthetic images by interpolating the cluster centroids, thus reducing the possibility of collision with latent vectors corresponding to real samples and a consequent leak of sensitive information. The proposed approach is tested over two X-ray datasets for Tuberculosis classification to simulate a realistic scenario with two different nodes and non-i.i.d. data. Experimental results show that our approach yields performance comparable to, or even outperforming, training on the full joint dataset. We also show quantitatively and qualitatively that images synthesized with our approach are significantly different from original images, thus limiting the possibility to recover original data through attacks.