Abstract:
The use of online social networks is growing rapidly. With this rapid increase, preserving privacy of users is becoming harder and harder. Typically, social networks address the privacy problem by asking users to define their privacy constraints up front. However, many times deciding on whom to show a post is dependent on the post itself and its context. Hence, users are forced to configure each post specifically, which is both cumbersome and prone to error. Accordingly, this study first proposes an approach that suggests privacy configurations for each post. The suggestions are based on learning from users’ previous posts and configurations. However, when the user does not have many previous posts, recommendations need to take other information into account. We propose a multiagent system architecture where agents of the users consult other users’ agents about possible privacy rules they can take into account. In addition to privacy-based decision making, we also propose another approach that considers users’ utility of sharing each post since users also regard its benefits when they decide to share. The system aims to estimate benefits such as the number of likes that a post gets, the number of comments that it receives and how many times it is shared again. These estimations are built on learning from users’ previous posts and their benefits. These privacy-based and utility-based approaches are combined in order to assist users in their sharing decisions. We evaluate single-agent and multi agent privacy-based approaches on a benchmark dataset that we created based on content from Flickr and Reuters while we evaluate the latter one on a dataset that is collected with our Facebook application.
