A Data Augmentation Method for Vertical Federated Learning

Abstract

Federated learning (FL) enables various organizations to jointly train one single model without revealing their private data to each other. The FL can be classified as horizontal federated learning (HFL) and vertical federated learning (VFL) according to the distribution of overlap samples and overlap features in the dataset. VFL allows various organizations to share machine learning based on the overlap samples, each one of which has the same identity. However, VFL suffers from insufficient number of overlap samples among all participants. Hence, the shortage of overlap data results in a worse performance of the global model. In this article, we propose a data augmentation method, FedDA, which is based on the generative adversarial network (GAN) to increase the number of training data. We generate more overlap data by learning the features of finite overlap data and many locally existing nonoverlap data, which expand the availability for training the overlap dataset. A series of experiments were executed on both MNIST and CIFAR-10. The results show that FedDA can efficiently utilize nonoverlap samples to enhance the effect of the data augmentation. It can generate high-quality overlap samples and expand the set of overlap samples. Thus, when the VFL is short of overlap samples, FedDA can provide abundant training data to improve the performance of the VFL model.