Binarized Neural Network Inference Protocol Based On Secure Multi-party Computation

Nowadays,neural networks have been applied in image classification,speech recognition,financial network detection,mobile payment,etc.,and are already an important technology to support our society.Therefore,the application of security issues in neural network inference is currently based on cryptography.The research on the cryptography framework of neural network model inference includes SecureML,ABY3,etc.But often because the communication volume is too large,the latency in the inference of the data is too high.This paper proposes a two-party binarized neural network inference protocol and a three-party binarized neural network inference protocol.Parties enter the parameters of the network and the data to be predicted,and after the protocol is executed,each participant can obtain the secret share of the prediction result of the binarized neural network.This paper proposes a new parameter encoding method for binarized neural network,optimizes the MSB(Most Significant Bit)calculation process and applies it to the two-party protocol and the three-party protocol.In the case of two parties,the offline computation cost is further reduced compared to MiniONN;in the case of three parties,the participants no longer need to perform offline sports.In addition,this article proposes a new three-party oblivious transfer protocol which is secure against semihonest adversary.The protocol is invoked in the process of computing MSB,which further optimizing the computation cost and communication cost of the protocol in the three-party setting.This paper experiments on the two proposed schemes using multiple neural structures on different data sets,and calculates the running time,communication cost and accuracy rate,and compares the experiments with previous work in academia,and analysis theoretically.Compared with the SecureML and ABY3 schemes,the running time and communication cost of the two schemes proposed in this paper are greatly improved.