WAPITI - A Weighted Bayesian Method for Private Information Inference on Social Ego Networks

Online social networking sites enable users to create profiles that often contain a broad range of potentially sensitive information, including relationship status, sexual orientation, religion, location, occupation, and interests. Despite the privacy controls offered by most social media platforms, these measures are often insufficient to safeguard users against profile attribute inference attacks, where adversaries infer private information by exploiting publicly accessible data. In this paper, we introduce a novel Bayesian model designed to infer the attributes of users on online social networks by leveraging publicly available profile data and topological features from the target user’s ego graph. Traditional Bayesian inference models rely on the assumption of attribute independence, meaning they treat all profile attributes as equally important and independent of one another, given a class label. However, we contend that this assumption is impractical in the context of social networks, as it directly contradicts the principle of homophily, which posits that individuals tend to form connections with those who share similar characteristics. To address this limitation, we propose WAPITI (Weighted Bayesian Model for Private Information Inference), a weighted Bayesian model tailored for attribute inference in social ego networks. Our approach incorporates attribute weighting to account for the dependencies between attributes, thus providing a more accurate representation of real-world social networks. Through extensive experimental evaluations on real-world datasets of user profiles and ego graphs, we demonstrate that the integration of attribute weighting substantially improves the overall performance of our inference model, yielding higher success rates compared to traditional baseline models.