Abstract

Hospitals possess the necessary data for training accurate diagnostic models, but privacy regulations prevent this from happening. Combining the approach of federated learning (FL) with generative models allows for expanding and balancing the training data without the need to change the location of patient data. However, the given paper describes five recent studies in the sphere of combining the above-mentioned studies: an approach of differentially private federated GAN, a review of the latest approaches in federated generative models, an empirical study of the use of federated models in creating tabular data in the field of leukaemia, a model of latent diffusion instead of using the FL approach in MRI of the brain and a one-shot approach of FL based on the use a flow model with rectified flow. In this case, the studies have been compared according to the type of federation, type of generative model, as well as the practice of privacy and evaluation. The main conclusion of the comparison is that while the way of using the FL changes from iterative one to one-shot one, the types of evidence used in privacy claims are different (e.g. formal differential privacy or any empirical or information.

Keywords

How to Cite This Article

Rights