Lens Privacy Sealing: A New Benchmark and Method for Physical Privacy-Preserving Action Recognition.

Area of Science:

• Computer Vision
• Privacy-Preserving Technologies
• Surveillance Systems

Background:

• RGB camera systems are crucial for public safety and healthcare but pose significant privacy risks.
• Current privacy solutions are often post-capture, failing to protect data during acquisition.
• There is a need for effective, low-cost, pre-sensor privacy protection methods.

Purpose of the Study:

• To introduce Lens Privacy Sealing (LPS), a hardware-based solution for pre-sensor privacy protection in RGB cameras.
• To develop a robust action recognition framework capable of handling LPS-induced video degradation.
• To create a comprehensive dataset for evaluating privacy-preserving action recognition.

Main Methods:

• Lens Privacy Sealing (LPS) utilizes adjustable laminating film for stochastic multi-layer scattering, providing irreversible privacy.
• Introduction of the P³AR dataset (P³AR-NTU and P³AR-PKU) with privacy annotations for action recognition research.
• Development of MSPNet, a single-stage framework with Inter-Frame Noise Suppressor (IFNS) and Cross-Frame Semantic Aggregator (CFSA), enhanced by contrastive language-image pre-training.

Main Results:

• MSPNet with IFNS and CFSA demonstrated nearly double the action recognition accuracy compared to baseline methods.
• Identity recognition was suppressed to low levels, indicating effective privacy preservation.
• LPS achieved a superior privacy-utility trade-off compared to state-of-the-art hardware methods and resisted reconstruction attacks.

Conclusions:

• Lens Privacy Sealing (LPS) offers a cost-effective and robust hardware solution for pre-sensor privacy protection.
• MSPNet effectively addresses video degradation caused by LPS, enabling high-accuracy action recognition.
• The P³AR dataset and proposed methods advance the field of privacy-preserving computer vision.