Research On Centralized And Local Differential Privacy Based Privacy Protection For Image Datas

With the rapid development of Internet and the growing popularity of personal devices with camera,digital images can be widely spread and occupy an important position in people's daily life.Because the images are inherently sensitive,the problems of privacy leakage are increasingly prominent during the image sharing and using.Therefore it is necessary to protect the privacy of images.Traditional image privacy protection schemes,such as access control,k-anonymity,Mosaic,due to the constant impact of state of the art techniques,has become no longer reliable.Differential privacy,which can provide strict and provable privacy protection,has been widely concerned and applied since it was proposed.While in the existing image privacy protection schemes that based on differential privacy,noise is often added directly to the whole image,the visual availability of the image is greatly reduced,which destorys the availability of image data.Thus,according to the two different research directions of differential privacy,two image privacy protection schemes based on differential privacy are proposed.In order to minimize the interference caused by adding noise to implement of differential privacy mechanism and take the published data availability and security into consideration,an image publishing scheme based on centralized differential privacy is proposed.Firstly,taking face and pedestrian as the typical sensitive content in the image,two different methods are provided to the users,which can locate and segment the sensitive region containing relevant content in the image and divide the image into two parts with image processing technology: manual selection and automatic recognition.Secondly,for the segmented image that contain the sensitive content,a privacy protection algorithm that satisfies the differential privacy is designed and processed.Finally,the two parts of the image are recombined and published.The effects of different privacy budget parameters and different identification methods on usability is analyzed through experiments.Compared with previous schemes,it is found that the scheme not only can reduce practial privacy risk,prevent privacy leakage in images,but also improve the visual availability of the published images.Meanwhile,in order to overcome the drawback that the traditional centralized differential privacy need to rely on the trusted data collectors which is assumed,an image privacy protection scheme based on local differential privacy is proposed.Before sending an image data to the untrusted data collectors,it was perturbed by random response on the client at first,then through statistics and correction by the data collectors to recover and collect the statistical data.The random response perturbation process for the image data satisfies ?-differential privacy is analyzed and proved,Experiments show that the visual effect of the processed images by the scheme is greatly reduced,making it impossible for neither the attackers nor the data collectors to obtain the privacy information directly,and retain enough statistical datas for the data collectors to publish.Some data availability is retained while the security is guaranteed.