Research And Implementation Of Privacy Preserving Algorithms Based On Deep Learning

Big data and artificial intelligence have become the main focus and research direction of social technology development.In the era of information sharing,both Internet companies and scientific research institutions are committed to researching new technologies that can help social development and ensure social stability.For example,machine translation,automatic driving,security monitoring,medical analysis,etc.In these applications,image and video play an important role.Computer vision are more and more widely used in social development.With the development and popularization of artificial intelligence,some data related to sensitive information has not been well controlled.Leakage and abuse infringe the privacy rights of individuals or companies.How to protect the visual data effectively while using it rationally has become the goal of researchers.Common privacy protection technologies include encryption,video steganography and so on.Therefore,this thesis focuses on the problems of poor practical application and high complexity of visual privacy protection algorithm.(1)An individual protection algorithm based on instance segmentation is proposed.The algorithm is divided into two stages:individual location and protection.In the first stage,the instance segmentation technology is used innovatively.The network structure includes backbone network,feature pyramid network,prototype mask generation network and output network.Backbone network can effectively select features by adding attention mechanism.Feature pyramid and prototype mask generation network add the underlying features for multi-scale fusion.The output network uses three main loss functions to help network training and reverse transmission.In the localization stage,instance segmentation technology is used to reduce the inference speed of the model,and it can help to identify the gaps in the classification.In the second stage,foreground and background analysis are used to hide the image,so as to achieve the final purpose of individual protection.The algorithm can effectively solve the problems of high complexity and unclear boundary description of visual privacy protection algorithm,increase the practicability and improve the operation speed.(2)A face protection algorithm based on differential privacy is proposed.The purpose of the algorithm is to protect the face data of the training set and prevent the information from being tampered.Firstly,the algorithm improves the network structure of face recognition,and uses linear transformation instead of convolution operation to reduce the amount of network parameters.While ensuring the recognition rate of the algorithm,it can significantly improve the speed of the inference stage,so as to alleviate the high computational complexity of the privacy protection training mechanism.Secondly,the sensitive data and non-sensitive data in the training set are obtained.Using the "teacher-student" fusion architecture,the final training data is obtained by training multiple face recognition teacher network on sensitive data and using voting mechanism to label non-sensitive data.Finally,the data is trained as the input of the student model to get the final model.The algorithm can protect the sensitive data of the training set by protecting the parameters of the neural network,effectively improve the security of the algorithm and prevent the model from being attacked and tampered.(3)Build privacy protection system,embed face protection and individual protection algorithm.The system uses client-side and server-side connection to meet the needs of users.The individual protection part is mainly divided into privacy positioning module,image processing module and training module.According to the user's requirements,the system hides information in the form of adding individuals.In the face protection part,sensitive data is isolated by online training and offline training,and privacy protection training mechanism is used to protect datasets and prevent from being tampered.