A Privacy Preservation Pipeline for Personally Identifiable Data in Images Using Convolutional and Transformer Architectures
Karla Brkić, Tomislav Hrkać, Zoran Kalafatić · 2022 45th Jubilee International Convention on Information, Communication and Electronic Technology (MIPRO) · 2022
Image and video data of people, shared voluntarily and involuntarily, is ubiquitous. There is an increased need for techniques that enable privacy protection via removal of personally identifiable information in such data, spurred by regulatory interest and increased social awareness of privacy implications. In this paper, we introduce a privacy preservation pipeline that enables de-identifying personal data in images and videos via replacement image synthesis while retaining data utility. We utilize the recently proposed convolutional VQGANs with autoregressive transformers to synthesize realistic and fully de-identified images of people that are then blended with the original scene. Experimental results show that the method provides strong de-identification while retaining the realism of the scene.