Privacy-Preserving Machine Learning Prediction Algorithm Based On Secure Multi-Party Computation

With the rapid development of big data and artificial intelligence,intelligent technology has penetrated into all aspects of our lives.Intelligent technology has brought great convenience to our lives,and it has also caused many problems such as privacy leakage.The risk of privacy leakage affects people’ trust and use of intelligent technology,and also limits the long-term development of intelligent technology.Solving the risk of privacy leakage in machine learning has become an important issue.Cryptography is a powerful means to ensure data security.The machine learning technology for privacy protection based on cryptography has received extensive attention.From general solutions to specific solutions,a series of excellent results have been achieved in protocol design and protocol optimization.However,these protocols often require a lot of communication and computing costs to ensure security.Designing efficient practical solutions has always been a bottleneck in the field of privacy-preserving machine learning.Different cryptographic primitives have different overheads for communication and computation.Suitable machine learning algorithms and scenarios are also different to find.In research,it is important to better utilize the advantages of different cryptographic primitives in the protocol according to different scenarios.Compared with the model training stage,the machine learning prediction stage requires interaction between the service provider and the customer.This thesis mainly studies the privacy leakage problem in this stage.Considering that the ensemble learning algorithm has significant advantages compared with a single machine learning algorithm,and the privacy protection problem is more prominent.We choose the Adaboost ensemble learning algorithm as the research support.In order to improve the efficiency of the privacy-preserving machine learning prediction stage,the two works are as follows.Firstly,with the development of the network and computing environment.The cost of network communication resources has gradually become the focus of optimization.The mode of using local computing resources to exchange network communication resources has become a popular direction for the practical application of cryptographic protocols.The multi-party computing privacy-preserving fast text classification protocol is improved.For the feature vector extraction protocol with communication complexity of 0(n2),this thesis presents two improvement schemes.Service providers and customers use cuckoo hashing and naive hashing technologies to save their data in buckets,and only need to call the equivalent test protocol bucket by bucket to complete the feature vector extraction task.Using the bucketing idea of hash technology,the scope of the equivalence test is reduced,and the complexity is reduced to O(nlogn).The service provider calculates the dazed polynomial locally,encrypts the polynomial parameters homomorphically and sends it to the customer,and the customer substitutes their own data into the polynomial calculation.The feature vector extraction task only needs to be performed once per round to perform the equivalent testing protocol.Using the idea of dazed polynomial hybrid homomorphic encryption,the complexity is reduced to O(n).The two improved schemes presented in this thesis both significantly reduce the communication complexity.Secondly,this thesis introduces the secure three-party computing model in the field of secure multi-party computing,and improves the traditional service provider-client model.Based on the secret sharing protocol of secure three-party computing,the idea of outsourcing computing is introduced,and the computing in the prediction stage is outsourced to three servers that do not collude with each other.The protocols such as vector dot product and security comparison in the secure tripartite computing model are applied,and component protocols such as feature extraction and secure transformation under the tripartite computing architecture are designed.So as to design the privacy-preserving classification prediction protocol under the secure tripartite computing architecture.The service provider and the customer upload the model parameters and data to the server,and the server jointly completes the calculation of the classification prediction task and returns the result to the customer.During the whole calculation process,the relevant information of model parameters and customer data will not be leaked,and the security and privacy are guaranteed.The schemes and conclusions proposed in this scheme have been experimentally verified,which significantly improves the efficiency.At the same time,during the forecasting process,service providers and customers do not have to maintain online communication,which is more practical and efficient than the traditional service provider-customer model.