| Deep learning provides huge development space for all fields,but also brings potential privacy threats.Face images used for face detection and training of deep learning models are facing the threat of identity disclosure,and the protection of identity privacy of face images has become an important research topic.At present,the image generated by face identity de-recognition algorithm has complete facial structure and natural face,but there are still some problems such as expression,skin color and other attributes destruction,and insufficient security of protection model.In this regard,the following research work is carried out in this paper:(1)A strong facial feature analysis algorithm is proposed.According to the characteristics of face recognition technology,Euclidean Distance was used to calculate the matching degree between each pixelated facial feature image and the face image,and the features whose matching degree was lower than the set threshold value were considered as strong facial features.(2)Design a face privacy protection model IF-CIAGAN based on strong facial feature analysis.The CIAGAN model image input form was improved,and a strong facial feature analysis algorithm was embedded into the preprocessing layer of CIAGAN model to reduce the loss of facial attributes caused by the overall facial anonymity.(3)The differential privacy algorithm based on deep neural network is studied.Differential privacy adjusts the degree of privacy protection through privacy budget.Differential privacy stochastic gradient Descent(DP-SGD)adds differential privacy noise to the training gradient in the form of grouping,so that the deep neural network can satisfy the differential privacy and control the privacy and availability of generated images.(4)The face image privacy protection model IFDP-CIAGAN based on differential privacy is proposed.In the process of training the model,the differential privacy noise was added to the gradient,and the dynamic adjustment was used to control the size of the added noise,and the IF-CIAGAN model was improved to make the generated image meet the differential privacy.Experiments show that the image generated by IF-CIAGAN retains more facial texture features of the original image.In order to enhance the privacy protection capability of IF-CIAGAN,the IFDP-CIAGAN model based on differential privacy is proposed.The generated image retains more features of the original image under the condition of realizing face deidentification,and the IFDP-CIAGAN model is protected from the threat of inversion attack. |
