Research On Personalized Image Privacy Detection Method

With the rapid development of electronic information technology and Internet technology,more and more people choose to share the images by smart devices such as mobile phones and cameras to social network.Users often inadvertently expose private information from the images when sharing the images.To improve users' awareness of privacy protection and avoid users' unconscious disclosure of privacy,the privacy detection technologies for images are emerged.However,the privacy problem is very subjective,it is affected by people's growing environment and social role.A general privacy detection method cannot meet the diversified needs of privacy protection,which is the main difficulty in the technology of image privacy detection.This paper focuses on this difficulty,discusses the potential significance of user privacy and the character of personalization,and innovates on how to design personalized privacy detection scheme for images sharing on social network.The main achievements are as follows:1.Considering that it is difficult to obtain personalized privacy data marked by users,and the scale of high-quality data that can be obtained is relatively small,we propose a method “Personalized Domain Adaptive Tradaboost(PDAtradaboost)”.This method can make reasonable use of the effective information of the privacy annotation data of large-scale generic domain and help the training of privacy detection model on the small-scale privacy image data in the personalized domain.PDAtradaboost is improved based on the traditional Tradaboost method.Tradaboost simply takes largescale user image data as auxiliary data and personalized image data as target field data for weight update.It is easy to make the target field data drown in the huge amount of auxiliary field data,and fail to highlight the importance of the target field data.Aiming at the limitations of Tradaboost,in PDAtradaboost,we design a method “User Privacy View Cluster(UPVC)”.Firstly,UPVC clusters the data of the general domain according to the privacy view,and takes the data of the personalized domain as the leading part.Then it selects the most similar privacy view cluster in the general domain as the auxiliary data to join the iterative training,which can effectively solve the problem of the dilution of the data features in the target domain existed in the original Tradaboost method.2.Considering that users' concerns about privacy are personalized,and privacy is not only related to a single object,but also related to the relevance of multiple objects.PDAtradaboost provides a method of personalized image privacy detection,but it can only detect the privacy at the image level.In order to further explore the privacy detection at the image-object level,a personalized image privacy object detection algorithm based on association rules is proposed.The algorithm first uses Deep Lab to identify the object sets containing context semantics in the image.Then,through the connection between semantic object set and privacy label in image set,the privacy rules can be extracted efficiently from a small number of image sets where users have given image-level privacy tags.The privacy rules can intuitively explain users' personalized privacy attitude and help users judge the privacy of images from the object level of images.