Research On Multi-Media Preference Protecting Scheme Based On Adversarial Example

With the popularity of mobile terminals and the upgrading of Internet bandwidth,people share more and more multi-media data on various social networks.The multi-media social networks such as Facebook and Tiktok have also become popular applications.However,the privacy problem in multi-media data also arises.Because of the diversity of multi-media data,such as image or video data,it was difficult to mine the user's privacy information directly from these data in the past.But with the development of artificial intelligence technology,especially after the great progress of deep neural network in the field of image processing,the feasibility of mining user's privacy information directly from image or video has increased a lot.Unfortunately,there is still a lack of research on how to protect the privacy information contained in the multi-media data of users against AI technology.Users on social media are lack of protection from others mining their privacy in image and video data by artificial intelligence.Aiming at this particular problem,we propose a solution based on adversarial example technology,and design corresponding user preference protecting scheme for multi-media data.The specific work of this paper is as follows:(1)We propose a multi-media preference protecting scheme based on adversarial example technology.The scheme protect user's preference information mainly by using adversarial example to disrupt the mining behavior of artificial intelligence.We test the scheme in experimental environment and practical production environment on consideration of three indicators below,i.e.generating time,success rate and transferability.Our experiments show that by adding adversarial perturbations on the multi-media data uploaded by users in advance,the user preference protecting scheme proposed in this paper can successfully reduce the leakage of user preference information in the multi-media data;(2)In consideration of that current adversarial example generating method for videos can hardly meet actual demand,we propose a new video adversarial example generating algorithm which combines targeted attack and non-targeted attack,that is,double step sparse advanced performances for videos.Our experiments show that the success rate of the proposed DS-SAPV algorithm is only slightly lower than the success rate of sparse adversarial attacks for videos algorithm while its generating speed is greatly improved.Therefore,before users uploading their videos,using our scheme to add perturbations on their videos can quickly and effectively prevent the user's preference information from being mined.