Research On User Data Abuse Prevention In Online Deep Learning Inferring Services

In recent years,deep learning technology has achieved rapid development,and online deep learning inferring service(DLIS)is one of the most important application scenarios of deep learning.In this scenario,a service provider trains an inferring model for a specific task and deploys the model on a cloud server.To use this inferring service,users need to upload their data to the cloud server.In this service architecture,user data is extremely vulnerable to be abused by service providers,that is,service providers can easily use collected user data in various forms without obtaining user notification to get the highest business benefit,which will greatly infringe the rights of users and the security of user data.In this paper,we study the user data abuse problem in DLIS scenario,and carry out following work.First of all,we investigate the relevant literature about user data security in DLIS scenario,and make a categorization of their technologies.We analyze the basic principles and application forms of each technology,and summary their advantages and disadvantages,along with the mechanism that causes these features.After that,we propose the first practical data abuse prevention scheme: DAPter.This is a user-side DLIS input converter,which can transform the data in advance before it being uloaded to the server.The transformed data can still be used for deep learning and inferring,but it will be difficult for service providers to abuse it.The core of DAPter is a lightweight generation model,which is generated through an innovative training process,and can minimize the information that can be used for data abuse in user data.At the same time,DAPter does not require any changes to the existing DLIS inferring architecture and inferring model.Finally,we design a comprehensive evaluation system for the anti user data abuse scheme in DLIS scenario.We implement DAPter,and comprehensively evaluate its security and practicability.The experimental results show that DAPter can greatly reduce the risk of user data abuse with almost no impact on the accuracy of inferring service by adding a small amount of calculation to the user side.DAPter achieves much higher overall performance than existing technologies.