iRogue: Identifying Rogue Behavior from App Reviews

An app user can access information of other users or third parties. We define rogue mobile apps as those that enable a user (abuser) to access information of another user or third party (victim), in a way that violates the victim's privacy expectations. Such apps are dual-use and their identification is nontrivial. We propose iRogue, an approach for identifying rogue apps based on their reviews, posted by victims, abusers, and others. iRogue involves training on deep learning features extracted from their 1,884 manually labeled reviews. iRogue first identifies how alarming a review is with respect to rogue behavior and, second, generates a rogue score for an app. iRogue predicts 100 rogue apps from a seed dataset curated following a previous study. Also, iRogue examines apps in other datasets of scraped reviews, and predicts an additional 139 rogue apps. On labeled ground truth, iRogue achieves the highest recall, and outperforms baseline approaches that leverage app descriptions and reviews. A qualitative analysis of alarming reviews reveals rogue functionalities. App users, platforms, and developers should be aware of such apps and their functionalities and take measures to curb privacy risk.