GPSniffer: Group-based Location Inference Attack in Augmented Reality using Network Traffic Analysis
Abstract
Augmented Reality (AR) applications increasingly rely on cloud-based processing, where camera frames are frequently uploaded for visual understanding and content generation. While such designs enable rich functionality, they also introduce new privacy risks through unintended side channels. In this paper, we present GP-Sniffer, a novel group-based location inference attack that exploits encrypted network traffic patterns generated by AR applications. The key insight is that co-located users capture similar visual scenes, resulting in correlated network traffic patterns that can be used to infer their physical proximity. GPSniffer leverages network metadata and zero-permission IMU sensors to identify co-located users. By propagating location information from a small subset of users who reveal their GPS data (anchor users), GPSniffer can infer the locations of a larger population of users (target users). We implement GPSniffer and evaluate it using a dataset collected across multiple real-world locations. Experimental results show that GPSniffer achieves high inference accuracy (0.949), outperforming prior AR-specific attacks, while eliminating their reliance on predeployed artifacts. Our findings highlight a previously overlooked privacy vulnerability in immersive systems and call for new defenses against traffic-based inference attacks.