Gender Inference for Facebook Picture Owners.

Social media such as Facebook provides a new way to connect, interact and learn. Facebook allows users to share photos and express their feelings by using comments. However, its users are vulnerable to attribute inference attacks where an attacker intends to guess private attributes (e.g., gender, age, political view) of target users through their online profiles and/or their vicinity (e.g., what their friends reveal). Given user-generated pictures on Facebook, we explore in this paper how to launch gender inference attacks on their owners from pictures meta-data composed of: (i) alt-texts generated by Facebook to describe the content of pictures, and (ii) comments posted by friends, friends of friends or regular users. We assume these two meta-data are the only available information to the attacker. Evaluation results demonstrate that our attack technique can infer the gender with an accuracy of 84% by leveraging only alt-texts, 96% by using only comments, and 98% by combining alt-texts and comments. We compute a set of sensitive words that enable attackers to perform effective gender inference attacks. We show the adversary prediction accuracy is decreased by hiding these sensitive words. To the best of our knowledge, this is the first inference attack on Facebook that exploits comments and alt-texts solely.