OPOM: Customized Invisible Cloak Towards Face Privacy Protection

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Face recognition technology offers convenience but poses significant privacy risks on social media.
• Current systems lack robust defenses against surreptitious identity analysis from images and videos.
• Protecting user privacy against advanced face recognition is a critical technological challenge.

Purpose of the Study:

• To develop a technological solution for face privacy protection against malicious face recognition.
• To propose a novel method for generating person-specific universal masks to prevent identity disclosure.
• To evaluate the effectiveness and practical applicability of the proposed privacy-preserving method.

Main Methods:

• Introduced the 'one person one mask' (OPOM) method for generating class-wise universal masks.
• Optimized training samples to move away from source identity feature subspaces.
• Investigated modeling techniques like affine hulls, class centers, and convex hulls for feature subspace description.

Main Results:

• Demonstrated the effectiveness of OPOM against black-box face recognition models on diverse datasets.
• Validated the method's performance across different loss functions and network architectures.
• Showcased practical utility through an application study on the Sherlock video dataset.

Conclusions:

• The proposed OPOM method offers a viable approach to enhance face privacy in digital media.
• Person-specific universal masks effectively prevent identity recognition by malicious systems.
• The study highlights the potential for practical application in video privacy protection.