SkullID: Through-Skull Sound Conduction based Authentication for Smartglasses

Document Title

SkullID: Through-Skull Sound Conduction based Authentication for Smartglasses

Document Information

• Subject Area: Authentication Technology, Smart Device Security, Bone-Conduction Audio Signal Processing
• Keywords: Smartglass Authentication, Bone Conduction, Acoustic Response, Biometrics, Security, Usability, Spoofing Attacks, Multi-Microphone Fusion

Research Background and Problem

• Problem or Challenge:

  • Smartglasses, as an emerging technology, face the challenge of user authentication. However, traditional authentication methods (e.g., passwords or fingerprint recognition) are not suitable for their compact form factor.
  • Some biometric-based authentication methods (e.g., facial recognition, iris scanning) require specialized sensors and may be affected by environmental lighting or user behavior.
  • Acoustic signals for authentication are promising, but existing solutions lack robustness against noise and flexibility in sensor placement.

• Research Importance:

  • Authentication systems in smartglasses need to ensure user privacy while being efficient (fast authentication), adaptive (to environmental changes), and resistant to attacks.

• Research Motivation and Related Work:

  • Building on previous works like SkullConduct and similar studies, it has been observed that audio signals transmitted through the skull can carry biometric pattern information.
  • Existing studies primarily use single sensor positions or air microphones, lacking comprehensive evaluation of different microphone combinations and multi-scenario (noise, repeated use) performance.

Solution

• Proposed Method:

  • SkullID utilizes bone-conducted audio signals for user authentication, with signals transmitted through the skull and captured at multiple microphone positions.
  • The study employs a right mastoid bone sensor as the audio source and places microphones on the forehead and left mastoid positions.
  • A support vector machine (SVM)-based classifier is developed for user identity authentication.

• Innovations:

  • First exploration of authentication performance using multiple sensor position combinations.
  • Proposed an optimization method to reduce the amount of training data, significantly enhancing applicability.
  • Comprehensive analysis tested the system's robustness against environmental noise, vibration interference, and spoofing attacks.

• Implementation Steps and Techniques:

  • Validation was conducted through three prototypes (headband prototype, smartglass frame prototype, integrated smartglass prototype).
  • Experiments utilized various audio signals (e.g., linear chirps, wake-up tones, speech audio).
  • Data processing involved feature extraction using power spectral density (PSD) and Mel-frequency cepstral coefficients (MFCC), combining features from multiple sensors.

Research Results

• Specific Results:

  • In preliminary experiments (25 participants), the equal error rate (EER) dropped to 2.35% when using microphones on the forehead and left mastoid, achieving the best results with multi-sensor fusion.
  • In subsequent long-term tests (30 participants), the EER remained low (2.72%) even across multiple recall sessions.
  • In another test, the system maintained stable EER (2.94%) during multi-day usage (27 participants).

• Advantages:

  • The system significantly reduced sensitivity to background noise (compared to SkullConduct, which used air microphones and had an EER of up to 65.61% in noisy environments).
  • Through mixed learning (including attack samples), the system effectively resisted signal replay and spoofing attacks.

• Experimental and Evaluation Results:

  • Multiple experiments demonstrated that combining two microphones (forehead + mastoid) significantly improved reliability.
  • Users rated the system highly for usability: the System Usability Scale (SUS) score averaged over 76 after the experiments.
  • Performance and training time were feasible even on resource-constrained devices (e.g., Raspberry Pi), with classifier training taking only 30 seconds.

• Limitations and Future Directions:

  • Current tests primarily focused on Asian participants; the system's performance across other ethnic groups requires further validation.
  • Broader usage scenarios, such as complex environments like public transportation, need to be studied.
  • Additionally, future work could explore combining bone-conduction microphones with air microphones to further enhance system reliability.