Research On Privacy Protection Technology Of Training Data In Deep Learning

The accuracy of deep learning model depends on a large number of high-quality data sets,but the existing deep learning does not pay special attention to the privacy protection of user data used in training model.In medical related fields,with the help of deep learning,doctors can diagnose medical images and other related information more conveniently.But under the premise of protecting personal privacy,how to obtain a large number of high-quality training data sets is still a challenge.This paper studies the privacy protection of medical image data sets,and proposes an aggregation model based on deep learningThere are three main reasons for the difficulty in obtaining high-quality medical data sets:first,the data in the medical field relies on highly specialized doctors for processing,but their time is limited,and they can not mark enough samples for the use of the model;second,the medical information needs to be kept secret,and the medical institutions can not simply share the data with their cooperators;finally,the deep learning models have the ability to remember the training data.If the network is attacked,the privacy data of users in the training data set may be leaked.In view of the above problems,this paper starts from the deep learning algorithm,studies the integrated learning and knowledge transfer,and proposes an improved model to aggregate the knowledge of multiple participants not only to protect the privacy of data in the model.This model uses the teacher-student mode.First,the teachers use the privacy data to train independently.The knowledge obtained is aggregated by voting method.Second,the teachers' information is added with noise.Finally,the information protected by noise is transferred to the students and the students are trained,so as to achieve the effect of protecting the data privacy of the training set.In this paper,the teacher network transmits convolution results to the student network,and obtains the relevant characteristics of the teacher network by probability selection according to the accuracy of the teacher,and trains the student network Because of the aggregation of data from many teachers,students recognize the features better than teachers.Using the data set HAM10000 of dermatoscope image,a deep convolution neural network is designed to analyze and process the data,and the model structure suitable for this data set is obtained.Then the optimization of the model is completed by adjusting the parameters and model structure,and the results are applied to the improved teacher student network.The experimental results show that the improved teacher student network can maintain privacy without reducing the accuracy of the model prediction results.