AudioUnlock: Device-to-Device Authentication via Acoustic Signatures and One-Class Classifiers

Area of Science:

• Cybersecurity
• Signal Processing
• Machine Learning

Background:

• Device-to-device authentication leverages manufacturing variations in microphones and speakers.
• Previous methods focused on multi-class recognition, unsuitable for large, unknown device pools.
• Real-world authentication requires recognizing a single device from numerous untracked devices.

Purpose of the Study:

• To develop a robust device-to-device authentication system using acoustic fingerprints.
• To implement one-class classification for recognizing a single legitimate device.
• To explore cloud-based deployment for efficient computation and data storage.

Main Methods:

• Utilized one-class classification algorithms: one-class Support Vector Machine and Local Outlier Factor.
• Trained models on acoustic fingerprints derived from manufacturing variations.
• Deployed and tested the system on smartphones and an automotive headunit.

Main Results:

• Achieved recognition rates between 50% and 100% across various devices.
• Evaluated performance under diverse environmental conditions: distance, altitude, and component aging.
• Demonstrated the feasibility of cloud-based processing for authentication tasks.

Conclusions:

• One-class classification effectively enables single-device recognition from large pools using acoustic fingerprints.
• The proposed system shows promise for secure authentication in various environments, including in-vehicle systems.
• Environmental factors and component aging present challenges, but solutions are proposed within the threat model.