On Privacy of Socially Contagious Attributes

A common approach to protect user's privacy in data collection is to perform random perturbations on user's sensitive data before collection in a way that aggregated statistics can still be inferred without endangering individual secrets. In this paper, we take a closer look at the validity of Differential Privacy guarantees, when sensitive attributes are subject to social contagion. We first show that in the absence of any knowledge about the contagion network, an adversary that tries to predict the real values from perturbed ones, cannot train a classifier that achieves an area under the ROC curve (AUC) above 1-(1-δ)/(1+e
^ε
), if the dataset is perturbed using an (ε,δ)-differentially private mechanism. Then, we show that with the knowledge of the contagion network and model, one can do substantially better. We demonstrate that our method passes the performance limit imposed by differential privacy. Our experiments also reveal that nodes with high influence on others are at more risk of revealing their secrets than others. Our method's superior performance is demonstrated through extensive experiments on synthetic and real-world networks.