posted on 2019-12-01, 00:00authored byGabriele Galfre
Automatically detecting the private nature of images posted in social networks such as Facebook, Flickr, and Instagram, is a long-standing goal considering the pervasiveness of these networks. Several prior works to image privacy prediction showed that object tags, either manually annotated or automatically extracted from images, are highly informative about images' privacy. However, we conjecture that other aspects of images captured by abstract concepts (e.g., religion, sikhism, spirituality) can improve the performance of models that use only the concrete objects from an image (e.g., temple and person). Several experimental setups have been defined to investigate how the usage of these type of information influence the performance of privacy classifiers. Results on a Flickr dataset show that the abstract concepts are better capturing the privacy nature of images, but with concrete object tags they complement each other, yielding the best performance when used in combination as features for image privacy prediction.