Research On Privacy-preserving Machine Learning Models In The Application Of Internet Of Things

With the rapid development of the Internet of Things(IoT)and artificial intelligence(AI),intelligent applications of the IoT are becoming more and more popular.Applications such as speech and image recognition can be deployed on IoT devices to analyze and process data to provide people with intelligent services.However,in a complex IoT environment,these applications are vulnerable to malicious attacks.And attacks may occur at any stage,including data release,model training,and model prediction stages,leading to the leakage of user data and application models.Therefore,in the application field of the IoT,it is of great practical significance to study the privacy-preserving machine learning models to meet people's privacy needs.This thesis focuses on secret sharing technique and studies the privacy-preserving machine learning model in IoT applications.The main research contents are as follows:(1)A privacy-preserving speech recognition model based on BiLSTM(PSRBL)is proposed.Based on the idea of polynomial fitting function,PSRBL uses piecewise function to approximate Sigmoid and Tanh activation functions of BiLSTM,avoiding intensive and complex calculations.At the same time,PSRBL constructs a secure BiLSTM neural network for model training through an additive secret sharing protocol to avoid leakage of model parameters and voice data.Finally,experiments show that the training effect of PSRBL is similar to that of the existing privacy-preserving outsourcing speech recognition model,and the model training time is reduced by 30%.(2)A privacy-preserving image recognition model for based on federated learning(PIRBL)is proposed.Based on the secret sharing technique,PIRFL first splits the gradient parameters of the client's local model into multiple copies,and then obfuscates the gradient parameters among the clients.Finally,the server executes the secure aggregation calculation of the obfuscated gradient parameters to avoid the leakage of local model gradient parameters and image data.At the same time,At the same time,PIRFL designs a secure communication mode for the gradient parameter confusion process to avoid model training errors caused by incomplete aggregation of server's gradient parameters.Experimental results show that PIRFL is correct and secure,can effectively defend against inference or reconstruction attacks against federated learning and protect the privacy of image data.