Research On Key Technologies For Data Sharing Security Based On Blockchain And Differential Privacy

With the arrival of big data era,data sharing has become a major means of breaking down data silos.The typical implementation of data sharing is facilitated by cloud servers with powerful storage and computing capacity.However,as the central server for data sharing,cloud servers are inherently subject to the centralized credit model and face the risk of data sharing failure when a single point of failure occurs or when malicious attacks are encountered.Therefore,how to ensure the secure sharing of data has become an urgent issue to be addressed.Blockchain,as a distributed ledger with computing power,is decentralized,trustless,tamper-proof and anonymous,which can better fit the security and privacy protection needs in the field of data sharing.Based on the key tool of blockchain technology,this thesis combines with differential privacy and federated learning technology to studies key technologies for data sharing security based on blockchain and differential privacy,and main contributions are as follows:1.Blockchain-based data sharing method can effectively enhance the credibility of data.However,the limitation of storage capacity on blockchains and the lack of privacy protection in data have restricted the development of blockchains.In this thesis,we introduce a multi-level blockchain-based model for data sharing.Take the medical scenario as an example.By utilizing both on-chain and off-chain hybrid storage technology and multilevel blockchains,our model improves the efficiency of data storage obviously.Through encrypting personal medical data and their index information by AES encryption and attribute-based encryption algorithms,and uploading the encrypted data to the off-chain distributed storage and on-chain storage respectively,the proposed multi-level blockchain-based model guarantees the security of medical data efficiently.Moreover,our model enhances the protection of data privacy by employing differential privacy technique in transferring parameters of federated learning module.The experimental result demonstrates that the storage costs on the institution side are reduced by about 60%,and the storage costs on the user side are reduced by about 30% on average in comparison with Ethereum.2.The privacy protection of regression analysis as a typical algorithm of machine learning,has attracted much attention.Based on functional mechanism,we propose regression analysis with differential privacy preserving based on Gaussian noise.Differential privacy techniques provide an effective way to protect privacy in machine learning by adding noise to provide a quantifiable level of privacy to the model.However,the existing literatures on differentially private regression models are limited and most of them use Laplace noise mechanism,which affects the usability of the models.The paper proposes Zero-Concentrated Differential Privacy Functional Mechanism(z CDP-FM)which adding Gaussian noise in Functional Mechanism used for privacy-preserving regression analysis instead of Laplace noise.The thesis further proposes Gaussian Differential Privacy Regression(GSDPR)and zero Concentrated Differential Privacy Regression(z CDPR)models effectively balance privacy budgets for sensitive and non-sensitive attributes by adding different scales of Gaussian noise to the polynomial coefficients of the objective function of the regression model.Theoretical analysis and empirical evaluations demonstrate our approach not only prevents the leakage of data privacy effectively but also retains the utility of the model.3.For the privacy protection requirements of model parameters in federated learning scenarios,we propose a logistic regression privacy based on blockchain and federated learning,and it based on a multi-level blockchain-based model for data sharing and regression analysis with differential privacy preserving based on Gaussian noise.Federated learning,as a new distributed machine learning technique,achieves data sharing by using a central server to coordinate the participants and collaboratively modelling by sharing model parameters without data leaving the local area.Federated learning avoids the direct sharing of data between participants and can largely avoid the leakage of private data.But it relies on the central server for coordination,which is a typical centralized structure,the single point of failure on cloud server and the leakage of model parameters have restricted the development of federated learning.We take logistic regression,which is the classic machine learning for classification algorithms,as an example to propose a privacy-preserving logistic regression based on blockchain and federated learning.To ensure the trustability of the federated learning,the local model parameters are used to upload to the blockchain with decentralized features,and carry out model aggregation through smart contracts.To ensure the privacy of logistic regression and the usability of the federation model,For off-chain thraning we propose the differential privacy logistic regression(DPLR)algorithm based on the functional mechanism of privacypreserving logistic regression models,For on-chain aggregation,we designs and implements a commission model validation contract that effectively protects against model poisoning attacks in federal learning.Theoretical analysis and experimental results show that the scheme proposed in this chapter can effectively guarantee the privacy of federal learning while improving the utility of the model.