A Study On User Data Privacy-Preserving Mechanism With Differential Privacy

Recent years,the rapid spread of personal smart devices bring people great con-venience in daily life.Users will generate various data proactively or passively when they use these smart devices.The data may include selfies,voice,chat history,location information etc,which greatly boost the development of various data mining or machine learning applications.These applications motivates companies or institutions to collect massive data from their users as much as possible.However,from the privacy perspective,these data usually contains highly sensitive information about users,and directly pulish these data or analyze these data will raise users' serious concern about their sensitive information may be disclosed.Therefore,both companies and users need a technique that can provide strong privacy protection.In this paper,we consider to use the state-of-the-art privacy preserving technique,differential privacy,and focus on two challenging scenario,real-time spatio-temporal data publishing and collaborative deep learning system.For the first problem,we present a framwork for privacy preserving real-time spatio-temporal data publishing,E-RescueDP.E-RescueDP consists of six mechanisms,statis-tics prediction,sampling,adaptive budget allocation,dynamic grouping and filtering.Our scheme utilizes data dynamics and solve the challenge of data sparsity.Theoret-ical analysis proves that our scheme achieves w-event e-differential privacy.Extensive experiments shows E-RescueDP achives the best performance among all existing works.For the second problem,we propose a privacy preserving collaborative deep learning system,SecProbe.We for the first time consider the existence of irregular participants and desgin several mechansims to solve this problem.Extensive experiments shows SecProbe achieves better performance than exsiting works.