Siren Song: Acoustic Attacks on Pose Estimation in XR Headsets

Zijian Huang, Yicheng Zhang, Sophie Chen, Nael B. Abu-Ghazaleh, Jiasi Chen

TL;DR

Researchers have demonstrated "Siren Song," a novel out-of-band attack that uses resonant acoustic signals to interfere with the MEMS sensors in XR headsets like the HoloLens 2. By injecting sound at specific frequencies, an attacker can trick the headset's pose estimation system (SLAM) into resetting, drifting, or misplacing virtual objects, leading to high-level UI exploits like clickjacking and "zone invasions."

What Problem Does This Paper Address?

Extended Reality (XR) headsets rely on a process called Simultaneous Localization and Mapping (SLAM) to understand where the user is in physical space. To do this accurately and smoothly, they fuse data from cameras with high-speed readings from an Inertial Measurement Unit (IMU)—specifically accelerometers and gyroscopes.

While software-based attacks are well-studied, the physical sensors themselves represent a "hardware-level" attack surface. Most IMUs use Micro-Electro-Mechanical Systems (MEMS), which contain tiny physical springs and masses. This paper explores whether an adversary can use sound waves to physically vibrate these components at their resonant frequencies, injecting "fake" motion data into the XR processing pipeline without ever touching the device's software.

Methods

The authors took a multi-stage approach to validate this threat:

  1. Acoustic Profiling: They performed frequency sweeps (2-30 kHz) on commercial-off-the-shelf (COTS) hardware, specifically the Microsoft HoloLens 2 and Meta Quest 3, to identify the exact resonant frequencies that cause sensor "spikes."
  2. Modeling & Simulation: Using the data-driven insights from the hardware profiling, they created a Gaussian Mixture Model (GMM) to simulate realistic sensor perturbations. They then applied these perturbations to open-source SLAM frameworks (ORB-SLAM3 and ILLIXR/OpenVINS) to see how different algorithms handle corrupted data.
  3. Physical Proof-of-Concept: They mounted a HoloLens 2 on a remote-controlled car (to simulate user movement) and used a portable speaker to launch real-world attacks. They focused on the "Snapback" effect—where the SLAM system detects a tracking failure and resets the user's coordinates to the origin (0,0,0).

Datasets

  • EuRoC MAV Dataset: A standard industry dataset for visual-inertial odometry used to evaluate the attack's impact on ILLIXR.
  • Custom Traces: The authors recorded their own walking and movement traces using a RealSense D435i camera and the HoloLens 2 (via the hl2ss library) to test specific attack scenarios.

Key Insights and Ideas

  • The "Snapback" Vulnerability: One of the most interesting findings is that the attack exploits a security/stability feature. When ORB-SLAM3 (and the HoloLens 2) detects that IMU data and camera data are wildly inconsistent, it assumes it is "lost" and resets the pose to the origin. An attacker can trigger this reset intentionally to move virtual objects relative to the user.
  • Algorithm-Dependent Failures: Not all SLAM is created equal. While ORB-SLAM3 "snaps back," OpenVINS (used in ILLIXR) tends to "drift away" to infinity. This suggests that the impact of an acoustic attack depends heavily on the specific software implementation of the XR runtime.
  • Motion as a Catalyst: The attack is significantly more effective when the headset is moving. When stationary, the headset can easily filter out the acoustic noise as an outlier. However, when the user is moving, the "fake" acoustic signals mix with legitimate inertial signals, making it much harder for the system to distinguish between real movement and the attack.
  • Audible vs. Ultrasonic: While some resonant frequencies are in the ultrasonic range (inaudible to humans), the HoloLens 2 accelerometer was found to be vulnerable at ~2.6 kHz—well within the range of human hearing, though it can be masked by background music.

Key Results

  • Resonance Identified: The HoloLens 2 was found vulnerable at 2.65 kHz (accelerometer) and 17.7 kHz (gyroscope).
  • High Success Rate: In real-world tests with the HoloLens 2, the researchers achieved a >90% success rate in triggering the "snapback" effect using an 85dB sound source.
  • Four Novel Attack Scenarios:
    • Manipulating User Input: Forcing a virtual car in a racing game to reset its position, effectively "steering" the user's progress.
  • Clickjacking: Causing a virtual keyboard to drift or oscillate, tricking the user into clicking the wrong keys.
  • Denial of Interaction: Using the snapback effect to keep an "invisible wall" (a virtual blocking object) permanently in front of the user's view.
  • Secure Zone Invasion: Moving a "virtual bomb" or unauthorized object from a public space into a user's private "secure zone" by manipulating the coordinate system.

Limitations and Future Work

  • Control Granularity: While the authors can trigger resets and drifts, they haven't achieved "fine-grained" steering (e.g., making a user walk exactly 2 meters to the left). This would require even more precise manipulation of the IMU's phase and amplitude.
  • Hardware Mitigations: The paper suggests that better physical shielding or Low-Pass Filters (LPFs) in the sensor's amplifier stage could mitigate these attacks, but these are difficult to implement in existing COTS devices.
  • Cross-Device Testing: While the Quest 3 showed less obvious vulnerability in initial sweeps (likely due to better physical dampening), the authors suggest further study is needed to see if higher-volume or different frequency ranges could still penetrate its casing.

Summary updated

Abstract

Extended Reality (XR) experiences involve interactions between users, the real world, and virtual content. A key step to enable these experiences is the XR headset sensing and estimating the user's pose in order to accurately place and render virtual content in the real world. XR headsets use multiple sensors (e.g., cameras, inertial measurement unit) to perform pose estimation and improve its robustness, but this provides an attack surface for adversaries to interfere with the pose estimation process. In this paper, we create and study the effects of acoustic attacks that create false signals in the inertial measurement unit (IMU) on XR headsets, leading to adverse downstream effects on XR applications. We generate resonant acoustic signals on a HoloLens 2 and measure the resulting perturbations in the IMU readings, and also demonstrate both finegrained and coarse attacks on the ORB-SLAM3 and an open-source XR system (ILLIXR). With the knowledge gleaned from attacking these open-source frameworks, we demonstrate four end-to-end proof-of-concept attacks on a HoloLens 2: manipulating user input, clickjacking, zone invasion, and denial of user interaction. Our experiments show that current commercial XR headsets are susceptible to acoustic attacks, raising concerns for their security.
BibTeX
@inproceedings{huang2025siren,
  author = {Zijian Huang and Yicheng Zhang and Sophie Chen and Nael B. Abu-Ghazaleh and Jiasi Chen},
  title = {{Siren Song: Acoustic Attacks on Pose Estimation in XR Headsets}},
  booktitle = {ISMAR},
  year = {2025},
  url = {https://openreview.net/forum?id=J4cehjURtr},
}

Loading related content...

Figure