The Dark Side of Metaverse: Unmasking the Threats from eXtended Reality (XR) Systems

Overview

Extended Reality (XR) devices are best understood not as conventional mobile terminals but as full-fledged cyber-physical systems. Following the NSF's characterization of CPS as engineered systems built from and depending upon the seamless integration of computation, networking, and physical components, an XR platform tightly couples a physical sensing layer that continuously measures and illuminates the user's body, a networking layer that sustains immersive rendering through behavior-dependent wireless transmission, and a computation layer that fuses these streams into real-time interactions in virtual scenes. Precisely because this integration is seamless, every layer becomes an attack surface. Physical emanations escape into the ambient environment, network traffic modulates the surrounding radio medium, and inbuilt sensors capture far more of the user than applications require, allowing adversaries to breach privacy non-intrusively without malware installation, physical constraints, or even line-of-sight access.

In this talk, I will unveil the emerging privacy threats in the Metaverse by systematically characterizing the leakage across these cyber-physical layers, and illustrate them through signal modeling, analysis, and reconstruction. Specifically, I will introduce a non-intrusive and unconstrained keystroke inference attack that exploits infrared emanations from the constellation tracking system to recover virtual keyboard inputs (NDSS'25), and a long-range, through-wall eavesdropping attack that leverages the wireless power side channel to uncover multi-level user behaviors, ranging from websites and apps to unseen in-game activities and avatar-based keystrokes, at distances of 5 to 8 meters (CCS'26). Then, I will present our recent study that pushes privacy inference beyond observable behaviors to unobservable brain-level perceptions, reconstructing EEG-correlated representations from unrestricted motion sensors to reveal what a user is perceiving inside the virtual scenes (S&P'26). Finally, I will discuss effective countermeasures and outline future directions toward trustworthy XR platforms.