p-Blend: Privacy- and Utility-Preserving Blendshape Perturbation Against Re-Identification Attacks in Virtual...

Area of Science:

• Computer Science
• Virtual Reality Security
• Human-Computer Interaction

Background:

• Facial blendshape data in virtual reality (VR) is vulnerable to re-identification attacks.
• Streaming VR facial data to third-party applications poses significant privacy risks.
• Existing privacy methods may compromise animation utility and naturalness.

Purpose of the Study:

• To introduce p-Blend, a novel mechanism for blendshape perturbation in VR.
• To enhance privacy protection against intra- and cross-app re-identification attacks.
• To balance privacy preservation with the utility and naturalness of facial animations.

Main Methods:

• Developed p-Blend, a blendshape perturbation technique considering both privacy and utility.
• Implemented perturbation on raw and statistical blendshape data.
• Conducted extensive empirical evaluations and user studies with a large-scale dataset.

Main Results:

• p-Blend significantly reduced re-identification accuracy across various machine learning models.
• Effectively mitigated re-identification risks from statistical features, unlike pure-random perturbation.
• User studies confirmed superior smoothness and naturalness of animations generated by p-Blend compared to random methods.

Conclusions:

• p-Blend offers an effective defense against VR facial re-identification attacks.
• The mechanism successfully preserves animation quality and user experience.
• p-Blend provides a practical solution for secure blendshape data streaming in VR applications.