Research On Efficient And Privacy-preserving Single-layer Perceptron Learning Algorithm

With the popularity of social networks and the diversification of mobile devices,unprecedented amount of data is produced.Traditional data storage and processing technologies have been unable to meet existing requirements.Cloud computing emerged as a service model for on-demand access to shared configurable computing resources.At the same time,massive data contains a lot of valuable information.It has become a trend to extract useful information from big data warehouses and establish intelligent decision support models by using machine learning technologies.More and more clients with resource-constrained devices outsource tasks of data storage and complicated computation to the cloud server in a pay-per-use manner.However,the cloud server is not fully trusted.Once the data is uploaded,the ownership and management right will be separated.Therefore,the data stored on the cloud server faces a great security threat.At present,the issue of protecting privacy of training data and prediction model in the cloud environment has become one of the research hotspots in academia and industry.Under this background,this paper focuses on the widely used single-layer perceptron model.The existing privacy-preserving single-layer perceptron learning schemes either leak sensitive training data and prediction model,or have expensive computational overhead and communication cost,or lack of scalability when feature dimension changes.In order to solve the above problems,we propose an efficient and privacy-preserving single-layer perceptron learning scheme in cloud computing,which has been accepted by Journal of High Speed Networks and applied for a patent.The main contributions of this thesis are summarized as follows:1.We analyze the security of PSLP(PSLP:Privacy-Preserving Single-Layer Perceptron Learning for e-Healthcare)scheme proposed by Wang et al.,and point out the security weaknesses of PSLP scheme: In the training process,the honest-but-curious cloud server can not only get intermediate results and optimal weight vector,but also obtain the plaintext of medical training data by solving equations.Furthermore,the more times the scheme iterates,the more medical cases will be leaked.2.Based on the symmetric homomorphic encryption algorithm,we design a symbol classi- fication algorithm,and propose an efficient and privacy-preserving single-layer perceptron learning scheme.The scheme has the following advantages: In the honest-but-curious cloud server model,our scheme can protect the privacy of training data,intermediate results and optimal predictive model.In addition,based on the homomorphic addition,homomorphic multiplication,and homomorphic scalar multiplication of the symmetric homomorphic encryption algorithm,the calculations of each step are lightweight in our scheme.Therefore,compared with the state-of-the-art works,our scheme enjoys better training efficiency.Lastly,our scheme performs well in scalability.On the one hand,when the feature dimension changes,our scheme does not need to re-encrypt and upload the previous training data.On the other hand,each training sample is encrypted separately,which can be flexibly applied to the scenarios of training single-layer perceptron model using stochastic gradient descent and batch gradient descent.3.We deploy the programs of cloud server and client on the Linux platform and Windows platform,respectively.Besides,we evaluate the efficiency and accuracy of the proposed scheme by conducting experiments on two real datasets: Wisconsin Breast Cancer and Default of Credit Card Clients,and provide detailed comparisons with the PSLP and PPDP(PPDP: An Efficient and Privacy-Preserving Disease Prediction Scheme in Cloud-Based eHealthcare System)schemes.Experimental results show that our scheme offers significant time savings on the setup stage and learning phase,and also can achieve a comparable accuracy rate.