Private data publishing method, system and device based on generative adversarial network

Abstract

The invention provides a privacy data publishing method, system and device based on a generative adversarial network, and the method comprises the steps: building a generator cluster and a discriminator based on adversarial learning by taking an original data set to be published and a small amount of public data sets which have the same attributes as the original data set and are subjected to privacy processing as training data; a random vector is used as the input of a generator cluster to fit published data which conforms to the attribute characteristics of an original data set and meets the privacy requirement, in the training process of a generator cluster module, generator performance is evaluated through a fitness evaluation function, and generators are reserved and eliminated based on evaluation scores, so that the privacy requirement of the generator cluster is met. And generating a new generator through crossover and variation, and adding the new generator into a cluster for subsequent iterative training. In the joint training process of the generator cluster module and the discriminator module, adaptive gradient cutting and noise introduction are carried out based on a differential privacy technology to protect sensitive data information. According to the method, the privacy protection intensity can be ensured, and meanwhile, the high effectiveness of the published data can be maintained.