Research On The Security Outsourcing Of Two Machine Learning Classification Algorithm

As an important core technology in the field of artificial intelligence,machine learning has been rapidly developed and widely used in computer vision,medical diagnosis and other fields.The model training and classification stage of machine learning often involves a large amount of data and computation.It becomes a challenge to execute on the resource-constrained user’s equipment.A common solution is to combine machine learning with cloud/edge computing by outsourcing complex computing tasks or models to cloud/edge servers.Although cloud/edge computing can reduce the users’ computing burden,improve the users’ efficiency,and realize the sharing of machine learning models,it inevitably leads to some security problems.Because it is almost impossible to find a fully trusted cloud server in reality,this may threaten the privacy of the user’s input information and the privacy of machine learning models.Due to the widespread application of convolutional neural network classification algorithm and naive Bayesian classification algorithms,this thesis conducts research on these two types of classification algorithms.Under the semi-honest model,the dual server model is used to design the corresponding secure outsourcing scheme.The specific work includes:(1)A secure outsourcing scheme of convolutional neural network classification algorithm is proposed.In this scheme,the linear layer,which is computationally heavy,is outsourced to the edge servers for computing.Besides,the nonlinear layer,which is computationally efficient,is left locally for computing.This makes it possible for some resource-constrained devices to complete the classification process of convolutional neural network safely and efficiently.At the same time,this scheme adopts lightweight random number blinding technology to protect the privacy of the input data,the privacy of the convolutional neural network model and the privacy of the results computed by the edge server.Through formal theoretical analysis,it has been proven that the scheme is correct and secure.The efficiency and scalability of the scheme are demonstrated through experimental analysis on three real datasets.Compared with the scheme without outsourcing computing,the efficiency of the proposed secure outsourcing scheme can be improved by 11 times.(2)A secure outsourcing scheme of naive Bayesian classification algorithm is proposed.This scheme realizes the sharing of naive Bayesian models with cloud computing,so that the users without models can also use the shared models on the cloud to obtain classification results.This scheme only uses lightweight addition secret sharing technology and lightweight cryptographic primitives to protect the privacy of the model,the privacy of users’ input data,and the privacy of the classification results computed by the cloud server.When implementing this secure classification scheme on the cloud,the scheme does not require model providers and users to remain online to participate in computation,which is convenient for model providers and users.A formal security proof is provided to demonstrate that the scheme is secure.The experimental results show that this scheme can efficiently handle the classification requests in the test dataset in less than100 milliseconds.