Research On Secure Outsourcing Of Machine Learning Algorithm

Recently,machine learning has become a research hotspot in the fields of artificial intelligence and big data analysis,which demonstrates impressive performance,especially in fault detection,face recognition and spam detection.Due to the characteristics of the structure of machine learning algorithms,the training process of machine learning models often consumes a lot of computing resources,which makes some users unable to afford such heavy computing tasks.The emergence of outsourcing computing provides a solution to the above problems.Users can outsource heavy machine learning training tasks to cloud servers with sufficient computing resources to reduce their own computing burden.Although outsourcing computing has many advantages,its unique service outsourcing and other characteristics also bring unprecedented security challenges to users.First of all,cloud servers cannot be completely trusted.In the process of outsourcing computing,there may be a risk of disclosure of users’ private information.Second,software bugs and malicious attacks may make the outsourced computing results received by users erroneous.Therefore,more and more researchers have begun to focus on how to design a secure and practical outsourcing solution to machine learning algorithms.The existing machine learning outsourcing schemes are designed for specific algorithms.This is because the structure of each machine learning algorithm is different,so it is difficult to design a general security outsourcing scheme for machine learning algorithms.This thesis designs the privacy protection outsourcing scheme based on a single malicious cloud server for the broad learning and the extreme learning machine in the machine learning algorithm,including:(1)The first secure outsourcing algorithm for broad learning is proposed.This algorithm allows users to outsource the training process of the broad learning algorithm to a cloud server,so that the broad learning algorithm can be widely used in some resource-constrained devices.In this algorithm,all inputs and outputs obtained by the cloud server are obscured,so the cloud server cannot obtain any useful information from it.In addition,client can effectively verify whether the calculation results returned by the cloud server are correct.This algorithm can greatly reduce the training time of broad learning and save local computing resources.Theoretical analysis and experiments proved the correctness,confidentiality,reliability and effectiveness of the algorithm.(2)The first privacy protection scheme for extreme learning machine algorithms is proposed.In order to protect the private data in the algorithm,the scheme designs a new matrix transformation method that greatly saves the computational cost required to protect the user’s input privacy and output privacy.In the solution,resource-constrained IoT devices can also complete the training process of the extreme learning machine.Experiments show that the computational cost of the proposed scheme on the user side is far less than the computational cost of the original extreme learning machine algorithm.Detailed theoretical analysis and experimental results prove the security,correctness,and verifiability of the algorithm.