Co-designed pre-capture privacy optics for computer vision.

Area of Science:

• Optics
• Computer Vision
• Cybersecurity

Background:

• Privacy concerns in digital imaging are increasing.
• Existing privacy methods often compromise data utility.
• Computer vision (CV) tasks require high-quality image data.

Purpose of the Study:

• To develop a novel pre-capture privacy optical system.
• To jointly optimize this system with a CV task.
• To achieve a balance between privacy preservation and CV performance.

Main Methods:

• A metasurface-based optical system was designed for pre-capture image degradation.
• The system was co-designed with a binary classification CV model.
• Performance was evaluated against both privacy attackers and the CV model using obscured and unobscured images.

Main Results:

• Privacy attacker accuracy dropped from 75.4% to 17.5% with obscured images.
• CV task performance decreased from 95.7% average precision (AP) to 67.0% AP.
• The joint optimization successfully favored CV performance while enhancing privacy.

Conclusions:

• The proposed co-design scheme effectively enhances pre-capture privacy.
• Metasurface optics offer a viable solution for privacy-preserving CV.
• This approach demonstrates a practical trade-off between privacy and data utility.