隐私保护深度学习研究综述OA北大核心CSTPCD

In deep learning as a service(DLaaS),private data constantly flow among various parties,which inevitably leads to privacy risks.On one hand,data owners worry about the possible exposure of their private data to service providers when they directly upload the data in plaintext.On the other hand,model owners are concerned about that adversaries may steal their costly trained models during extensive data interactions.As a result,combining privacy protection with deep learning has become a hot research topic these days.This paper reviews the research results in privacy-preserving deep learning since 2016,and categorizes the techniques into linear and nonlinear computations,which are the two basic building blocks in deep learning models.Specifically,the pros and cons of diverse tech-niques used in different layers are presented according to time and quantity statistics.In addition,the evolutionary directions of every technique are clarified by tracing their optimization routes.Following a comprehensive overview of each representative research scheme,the hurdles of privacy-preserving deep learning are listed and the resolution as well as promising directions for further research are proposed.