Research On Data Privacy Preservation Technologies Using Secure Multi-Party Computation

The rapid development of mobile computing,the internet of things and the big data technology has facilitated various applications.Upon collecting an amount of data,these applications can provide more user-friendly services and bring great convenience.However,these data may involve sensitive information,and privacy concerns arise during the collecting,exchanging,analyzing,and utilizing of the data.Data privacy leaks have occurred frequently,many countries and regions worldwide are proposing laws and regulations to hasten data privacy protection.Therefore,achieving efficient data privacy protection and giving full play to data value is a challenging problem.This thesis takes the idea of secure multi-party computation and conduct the research on the data confidentiality and privacy-preserving calculation for three data application scenarios,i.e.,data quality estimation,neural network prediction model and neural network training model.The main content and innovations of this thesis are summarized as follows:(1)Aiming at the privacy protection during the estimation of data reliability,this thesis proposes a series of efficient three-party calculation toolkits,and further constructs a privacy-preserving real-time incentive system for crowdsensing(abbr.EPRICE).The EPRICE scheme can estimate the reliability of sensing data in a privacypreserving manner.The theoretical analysis demonstrates that our proposed scheme meets the privacy-preserving properties in practical applications.The experimental findings indicate that our proposed EPRICE system significantly decreases the computation costs.(2)To improve robustness of secure computation,this thesis further design a decentralized and privacy-preserving scheme(abbr.Secure NLP)for LSTM-based sequenceto-sequence with attention model for natural language processing.Specifically,for non-linear functions such as Sigmoid and Tanh,this thesis designs two secure three-party protocols,which are used to construct the privacy-preserving long short-term memory network,and privacy-preserving sequence to sequence transformation.This thesis proves that Secure NLP can protect the data privacy against a semi-honest adversary who has corrupted n-1 parties.The findings from the experimental evaluation demonstrate the utility of the protocols in cross-domain NLP tasks.(3)To improve scalability of secure computation,this thesis designs communicationfriendly and computation-friendly multi-party calculation toolkits for hybrid network architecture.Then this thesis also proposes a privacy-preserving deep neural network training scheme(abbr.Pp NNT)for industrial Io T environment.The proposed Pp NNT can flexibly support machine learning in the multi-party setting with high security,efficiency,and scalability,as evidenced by the theoretical security proof and experimental results on the CIFAR10 dataset.