Secure Two-party Neural Network Computing And Learning

Privacy preserving data mining is a hot research topic in recent years. The ap-proaches of solving the privacy issue of data mining are two: 1) Data Fuzzy; 2) SecureMultiparty Computation. After R. Agrawal, R. Srikant and Y. Lindell, B. Pinkas pro-posed the notion of privacy preserving data mining in 2000 respectively, many cryp-tography researchers were interested in solving the privacy issue of data mining bysecure multiparty computation, and security constraints were added to several miningtechniques, such as decision trees, artificial neural networks, support vector machines,bayesian classifiers, clustering, Statistic Analysis, and so on.The issues of secure two-party neural network computing and learning are stud-ied in this paper, and the main results are as follows:1. Oblivious infinite derivable evaluation protocol and secure multiplicationevaluation protocol are constructed. They are the underlying protocols for study-ing secure two-party neural network computing and learning in this paper. Amongthem, the former one, which is applied to oblivious Sigmoid function evaluation inneural network computing, is based on the oblivious polynomial evaluation protocolfor ?oating-point numbers and the property that any infinite derivable functions canbe Taylor expanded, while the latter protocol, based on the private scalar product pro-tocol, is applied to oblivious supervised learning.2. Oblivious threshold function evaluation protocol and oblivious Sigmoid func-tion evaluation protocol are constructed, which are for solving the problem of securetwo-party neural network computing, i.e. the secure computation problem of activa-tion function(threshold function, Sigmoid function, ect.).3. The secure computation of neural network learning algorithm is studied, in-cluding supervised learning, Competitive Learning, organization learning, and so on.4. Oblivious BP neural network classifier protocol is constructed. Neural net-work can be used to predict and classify, and we realize the secure classifier of BPneural network based on the existing results of this paper―oblivious Sigmoid func-tion evaluation protocol and secure multiplication evaluation protocol.