Research On Privacy Protection Mechanism Based On Support Vector Machine Classification In Cloud Environment

In recent years,with the continuous growth of data resources and computational demands,the field of machine learning has obtained rapid development.As a kind of mass storage and powerful computing ability of cloud platform has become the first selection of machine learning applications.Particularly,Support Vector Machine(SVM)algorithm are widely adopted,and the implementation of multidimensional data classification functions in the cloud has become user requirements.However,the exchange of massive amounts of data also brings more privacy risks.For instance,in cloud computing environments,classification tasks often involve processing and transmission of sensitive information.Consequently,users are concerned about personal information leakage while expecting high-quality services.Additionally,prediction models obtained through machine learning training belong to their owners’ assets and require protection against potential leaks.Aiming at the above problems,this thesis studies the privacy protection mechanism based on support vector machine classification in the cloud environment.The privacy of data and models is protected based on homomorphic encryption.On the one hand,it protects the data privacy of users,on the other hand,it ensures the classification performance of support vector machine while ensuring the privacy of service providers’ models.The research content mainly includes the following aspects:(1)In order to ensure the privacy of data and the model parameters in cloud environments,a multi-user dual-cloud privacy protection support vector machine classification service scheme is proposed.The scheme adopts the distributed double trapdoor public key cryptosystem(DT-PKC).Using the homomorphism of DT-PKC system,the cloud platform of SVM service request,which can support multi-users in different encryption domains,is designed.Meanwhile,the trusted third party key distribution center is introduced to reduce the communication cost of each participant.The dimension of support vector is expanded,so that the cloud platform encrypts the expanded support vector by using the user’s public key,it prevents the computer service provider to guess the original support vector.Privacy analysis shows that DT-SVM scheme has efficient semantic privacy preserving,learning process privacy preserving,and classification result privacy preserving.At the same time, compared with the existing schemes,this scheme has better noise insensitivity,and its prediction accuracy in the ciphertext domain will not be damage.(2)In order to further ensure the privacy of user data and model parameters,a privacy-preserving scheme based on federated learning for support vector machines(SVM)has been proposed.This scheme utilizes the Okamoto-Uchiyama(OU)homomorphic cryptosystem to encrypt users’ data and employs SVM algorithm to obtain classification results.The scheme effectively protects user privacy while enabling them to swiftly obtain classification results.A novel lightweight approach is designed to acquire users’ classification results,which conceals sensitive information of the federated learning model,thereby preserving the privacy of its parameters.Through security analysis,functional comparisons,and experimental evaluations,this scheme achieves high-precision classification.Compared with traditional methods,it exhibits robust security capabilities against vector insertion attacks and vector reconstruction attacks.