On the Use of "deep" Features for Online Image Sharing

Abstract

Online image sharing in social networking sites such as Facebook, Flickr, and Instagram can lead to unwanted disclosure and privacy violations, when privacy settings are used inappropriately. Despite that social networking sites allow users to set their privacy preferences, this can be cumbersome for the vast majority of users. In this paper, we explore privacy prediction models for social media that can automatically identify private (or sensitive) content from images, before they are shared online, in order to help protect users' privacy in social media. More precisely, we study "deep" visual features that are extracted from various layers of a pre-trained deep Convolutional Neural Network (CNN) as well as "deep" image tags generated from the CNN. Experimental results on a Flickr dataset of thousands of images show that the deep visual features and deep image tags can successfully identify images' private content and substantially outperform previous models for this task.