Reducing bias and increasing utility by federated generative modeling of medical images using a centralized adversary