Privacy Protection System Base On Generative Adversarial Network

With the development of Internet technology,big data technology has brought great convenience to all aspects of people’s life.But at the same time,the problem of data privacy leakage has also received a lot of attention,and more and more research on data privacy protection has been conducted.However,since the risk of leakage is gradually increasing and the attackers’ methods are becoming more and more diverse,there is a higher demand for the capability of privacy protection algorithms.However,since traditional privacy protection approaches require strict limitations on attackers’ capabilities,and cryptography-based approaches are criticized for their time and space consumption,attempts have been made to incorporate deep learning into privacy protection methods.Among deep learning,GAN,as the most important model proposed in recent years,has shown its powerful ability in the generation of images,text,speech,video,etc.,by using the concept of adversarial.Therefore,our approach of combining the application of generative adversarial networks with privacy protection in order to reduce the risk of data privacy leakage is of great relevance.We plan to use generators in GAN to transform the original data and use the privacy and utility requirements of the data as discriminators for adversarial purposes,and finally obtain a privacy-preserving generator that can carry out specific processing of the original data,and the processed data can satisfy the established privacy and utility requirements.Based on this idea paper proposed a PPGAN framework by combining a VAE-like generator with three different discriminators,and two sets of loss functions.It is experimentally demonstrated that the framework in this paper can reduce the classification accuracy of privacy attributes to 10% while the utility accuracy is almost constant,and the performance remains superior compared with other similar models.In this paper,we have done the following work: 1.Firstly,we have studied the knowledge about deep learning and GAN,and we have understood the traditional privacy preserving methods.2.We have analyzed the great challenges faced by the current privacy-preserving methods and the ways to deal with them.Based on this,we have proposed a privacy-preserving model,PPGAN,and then we have experimentally demonstrated the effectiveness of this one in privacy preservation.The model is able to satisfy the need to remove knowledge useful for inferring privacy attributes while retaining application-specific knowledge.3.Using the PPGAN model,a tool system is designed and implemented that can be used by users have different privacy and utility requirements.