Abstract

Image privacy protection is an important topic in Human-Computer Interaction and usable security. Researchers have examined different aspects of image privacy by collecting samples by themselves. However, there does not exist a publicly-available dataset on image privacy, which prevents these efforts from sharing common technical foundations. We introduce DIPA, an open source dataset that provides content-level annotations that specifically focus on image privacy. We include 1,495 images from two existing datasets in DIPA, and augment them with 5,671 annotations. Each annotation includes reasons why the associated visual content can be privacy-threatening, a rating of how informative annotators thought the associated content is to threaten privacy, and another rating of how broadly the image could be shared. We also collected annotations from people living in Japan and UK to enable researchers and developers to perform analysis from the perspective of cultural differences. In this paper, we present the construction procedure of DIPA and report high-level statistics of the data we obtained. We hope that DIPA would accelerate various future research, including quantitative understandings of cultural differences on perceptions of image privacy and the development of robust recognition models for image privacy protection.