Research On Core Technology Of Data Security In New Network Environments

The continuous development of new network environments has driven the rise of new architectures,such as cloud computing and crowdsensing.However,threats of data security and privacy in the new network environments,especially under cloud computing and crowdsensing,are increasingly diversified,complicated,and large-scale,which poses serious challenges for the secure collection,storage,and use of data.Specifically,in the phase of secure data collection,the smooth execution of existing works stems from the frequent interaction between users and servers,which cannot adapt to the task of data collection under abnormal networks.In the phase of secure data storage,most of the existing works are devoted to the design of text-based data retrieval,which results in relatively few solutions for the secure query of spatial data and DNA data.Moreover,some fundamental problems such as low efficiency and weak access control in the query are also not well solved in the above two fields.In the phase of secure use of data,especially for deep neural network training based on large-scale data sets,the current distributed privacy-based training techniques do not consider two fundamental issues: one is the verifiability of the results returned by the server,and the other is how to deal with the inconsistency of the quality of data held between users.In view of the above research status,this dissertation delves into the security problems of data collection,storage,and use in the new network environment(this article focuses on cloud computing and crowdsensing).The contributions of this dissertation can be summarized as follows:Research on secure truth discovery in the phase of data collection:(1)We propose an Efficient and Privacy-preserving Truth Discovery(EPTD)approach in mobile crowdsensing systems.EPTD can efficiently support users dropping out with proper modification.We design a double-masking with a one-time pads protocol to achieve the high aggregated accuracy along with privacy protection on both users' data and reliability information under the working process.(2)We propose V-PATD,the first verifiable and privacy-aware truth discovery protocol in crowdsensing systems.In V-PATD,we propose a publicly verifiable approach in the privacy-preserving truth discovery process.Our solution is the first approach to meeting the requirements of publicly verifiable,low cost,support for non-fixed outsourced functions,and multiple data contributors.Research on searchable encryption for DNA data and spatial data in the phase of data storage:(1)We propose a secure,efficient,and accurate range query scheme(EGRQ)over cloud data,where the polynomial fitting technique is exploited to create trapdoors,which can significantly reduce the local storage overhead in the process of index and trapdoor generation;We propose a novel data access control strategy to refine user's rights in our EGRQ.(2)We design EFSS,the first efficient and non-interactive DNA similarity search framework with data access control over encrypted cloud data.In EFSS,we introduce a private approximation algorithm to convert the edit distance computation problem to the symmetric set difference size approximation problem,which can significantly reduce the number of elements that need to be matched under ciphertext.We design a novel Boolean search method to achieve the complicated logic query such as mixed “AND” and “NO” operations on genes;EFSS is able to support data access control during the query process,where each DNA sequence is accessible to users who are authorized.Research on secure deep learning in the phase of data use:(1)We propose Verify Net,the first privacy-preserving approach supporting verification in the process of training neural networks.Verify Net allows users to verify the correctness of results returned from the server with acceptable overhead;We propose a double-masking protocol to guarantee the confidentiality of users' local gradients during the federated learning.It can endure a certain amount of users exiting for some reasons during the training process,and the privacy of these exiting users are still protected.(2)We propose PPFDL,an efficient and Privacy-Preserving Federated Deep Learning framework while maintaining a high data utility.The framework protects the privacy of all user-related information.We design a novel strategy to reduce the negative impact of irregular users on the accuracy of training results.Moreover,for practicality,our PPFDL is also robust to users dropping out during the entire training process for various unpredictable reasons.For all the above schemes,we have conducted a full security analysis and proved the security of the above schemes under the given threat models.In addition,extensive experimental analysis and comparison with existing solutions also demonstrate the superiority of the proposed solution in terms of performance.