Privacy-Preserving Visual Learning Using Doubly Permuted Homomorphic Encryption (Supplementary Material)

We are interested in leveraging ‘private’ images, which are not shared publicly but just saved on a personal storage privately, for visual learning. In this section, we would like to provide some statistics that motivate our work. Based on the recent report from Kleiner Perkins Caufield & Byers [14], the number of photos shared publicly on several social networking services (Snapchat, Instagram, WhatsApp, Facebook Messenger, and Facebook) per day has reached almost 3.5 billion in 2015. It also shows that the number of smartphone users in the world was about 2.5 billion in the same year. From these statistics, if everyone takes three photos per day on average, about four billion photos in total would be stored privately everyday. Some prior work [9, 10] has shown that such private photos still contained meaningful information including people, faces, and written texts, as well as some sensitive information like a computer screen and a bedroom. Our privacy-preserving framework is designed to learn visual classifiers by leveraging this vast amount of private images while preserving the privacy of the owners.