Research On Decision Tree Evaluation Mechanism Of Outsourced Privacy Protection Based On Homomorphic Encryption

In the era of big data with abundant information sources and rapid developing machine learning technology,cloud computing has become an ideal platform for machine learning due to its characteristics of large storage capacity and strong computing power,while privacy issues are also increasing.Decision tree model,as a widely used machine learning classification model,can be trained and deployed in the cloud to provide classification services to users.However,on the one hand,direct delivery of users' query data to the cloud may reveal their privacy,at the same time,the evaluation results obtained by the user may involve their sensitive information.On the other hand,the training models obtained by machine learning belong to the model owner who hopes that these models are unique and will not be disclosed.As a result,there are more and more scholars studying the area of privacy protection decision tree evaluation.Based on security comparison protocol,two encrypted data can be compared without decryption.In order to ensure the practicability and accuracy of the model,this paper protects the data and models based on homomorphic encryption.The following three aspects are included in the main research contents.(1)Based on the dual cloud model,an efficient privacy protection decision tree evaluation scheme is proposed in view of the security and inefficiency of the privacy decision tree assessment.In this scheme,a new type of homomorphic encryption method,distributed two-trapdoor public-key cryptosystem is used to encrypt the data in the scheme.The master key is split into two parts,and each of the two cloud servers holds part of the key,so they can only partially decrypt,and neither of them can get the final plaintext,which will improve the security of the scheme.The evaluation process is completely carried out by cloud interaction to support users' online or offline evaluation,increasing convenience for users and has higher efficiency.By establishing a secure XOR protocol,the existing security comparison protocol is improved to make the protocol more secure and efficient.The scheme protects the privacy of users' queries,classification results and decision tree models.The decision tree model is converted into a linear function,so that the homomorphic operation only needs to be done through additive operation,which improves the efficiency of the scheme.(2)In view of the high cost of user computing in the evaluation process,a privacy protection decision tree evaluation scheme based on authorization is proposed.Introduced a third-party application,which authorized by the user,can share the evaluation results with the user and provide better analysis and suggestions to the users.And the application,as an intermediary,reduces the interaction between the user and the cloud server.Users simply send feature vectors and authorization information to the application,and send the authorization status of the application to the cloud server,and no longer participate in subsequent processes,which reduces user's computing costs,and makes the model closer to reality.In this scheme,the security comparison protocol is further improved and at the same time,random thresholds are added to the comparison protocol to hide the comparison times of decision tree nodes and protect the privacy of model parameters.Moreover,the scheme also protects users' query information and evaluation results.(3)In view of the high cost of key interaction process between the parties,a trusted third-party key distribution center is introduced,which reduces the cost of communication.After the key is distributed,the key distribution center remains offline.Through security analysis,performance analysis and experimental analysis,it is proved that the proposed scheme supports multi-key,user online/offline evaluation,application and other functions,and has higher security,lower computing cost and communication overhead.