Оценка эффективности технологии FedAVG при федеративном обучении генеративных нейронных сетей

Training neural networks often requires very large amounts of image data. However, sharing images owned by different institutions can be problematic. One of the known solutions could be to train neural networks jointly by way of iterative sharing neural models, but not the restricted sets of training images. Such an approach is known as Federated Learning. In this paper, we present results of an experimental study of the efficiency of Federated Learning of generative neural networks of the DCGAN type. Specifically, the FedAVG approach has been investigated based on large collection of medical images, including chest x-ray images, axial slices of 3D computed tomography images, and Hematoxylin-Eosin stained histology images. The results of the FedAVG approach were found to be highly dependent on the homogeneity of the image datasets. Among the images being employed, the best potential for federated training was demonstrated by chest x-ray images, while the routine histology images were found to be unsuitable for FedAVG training. The 2D computed tomography image slices were situated somewhere in-between of these two image types and showed characteristically unstable behavior. The period of aggregation of training results on the federated server should be reasonably short and repeated after every 1-3 epochs performed on the local image datasets of federated clients.