Geração de dados sintéticos para anonimização de dados de saúde por meio de redes adversárias generativas e uma função de perda customizada

Abstract

Health data may present vulnerabilities by containing private and sensitive information, which must be considered in contexts that require the manipulation of such data. One solution to the problem of exposing sensitive information is the generation of synthetic samples that accurately represent the dataset to be considered, allowing it to be replaced in the works proposed for a specific task. Among the various methods for generating synthetic data, the use of generative adversarial networks (GANs) stands out in the field of image generation. For tabular data, although studies are still limited, the possibilities are vast and demonstrate the flexibility of these networks in generating samples of lower dimensionality. The proposed method is based on a GAN architecture, supplemented by a training method that employs a custom loss function and different approaches for its application. The goal is to obtain a distribution of the synthetic samples as faithful as possible to the real ones. The main hypothesis is that GAN, combined with the proposed training method, would be capable of generating data whose distribution closely approximates that of the real data. The results indicate that the use of a loss function, based on the approximation of two distributions, promotes the generation of more realistic data, which can be used in the applications required by various machine learning tasks.