Research On Privacy Protection Methods For Facial Image Based On Machine Learning

In recent years,deep neural networks have been widely applied in many fields,and their accuracy in face recognition has been comparable to that of human vision.While the application of artificial intelligence brings convenience to the public,it also faces the risk of leaking personal information.Personal sensitive information such as facial images,fingerprints,and voice can be collected.Once information leakage occurs,it will have a great impact on the rights and interests of citizens.In fact,the leakage of facial image is particularly serious,it is necessary to protect personal sensitive data through effective privacy protection methods.At present,privacy protection methods for facial image have received widespread attention.Generally,their research goals are face anonymization and data availability protection.Face de-identification,which uses image processing methods on sensitive areas of the face,such as covering,blurring,distortion,is an intuitive method.Although these methods have a simple implementation,they are usually not sufficiently anonymizing when considering image distortion,thus it is difficult to meet the requirements of privacy protection in deep neural networks.Moreover,it does not explicitly perform diversity preservation of attributes such as emotions,expressions,and races,thus it is impossible to perform data analysis tasks on non-identity attributes.In order to improve the anonymization of facial images in deep neural networks,this thesis proposes a generator for facial image with privacy based on adversarial samples.First,the image distortion and identity distance are comprehensively considered in the formula.Further,the targets are transformed into the problem of generating adversarial samples based on the mini-network for face recognition,and an adaptive optimization method is applied to generate an image that is anonymous to the deep neural network.Then,comparison experiments are designed and tested under various models,such as MobileNet,VGG16,FaceNet,etc.The results show that the algorithm is low-distortion,anonymous and transferable to many different unknown black-box models,thus has a sufficient performance of privacy protection in deep neural networks.Then,in order to perform diversity preservation,this thesis proposes a generator for facial image with diversity,which is suitable for expression classification.First,the adaptive algorithm of generating image based on the previous research is applied to the targets of privacy protection.Further,additional factors are introduced into the formula to control the loss of expression classification based on the mini-network for face expression.Then the evaluation experiments shows that the algorithm has a sufficient performance of diversity preservation and privacy protection in deep neural networks.Therefore,facial images with diversity and privacy are still suitable for analysis tasks such as emotions and expressions.