Privacy-Preserving Face Recognition Method Based on Randomization and Local Feature Learning.

Area of Science:

• Computer Science
• Biometrics
• Cybersecurity

Background:

• Personal privacy protection is crucial, especially for face recognition applications.
• Traditional methods like encryption or perturbation require image restoration for recognition.
• Existing methods face challenges in balancing privacy and recognition accuracy.

Purpose of the Study:

• To develop a novel privacy-preserving face recognition method.
• To achieve high recognition accuracy while effectively eliminating sensitive facial information.
• To ensure both revocability and irreversibility of privacy protection.

Main Methods:

• A new privacy-preserving face recognition approach combining random convolution and self-learning batch normalization.
• Generation of privacy-preserved scrambled facial images with a high degree of fuzziness.
• Direct recognition of scrambled images on the server without image restoration.

Main Results:

• The proposed method achieves recognition accuracy comparable to normal facial images, exceeding 99% on multiple datasets.
• Visual information elimination is near-complete, approximating an encryption effect.
• Outperforms existing privacy-preserving face recognition techniques in both accuracy and information elimination.

Conclusions:

• The novel method effectively protects face privacy while maintaining high recognition performance.
• It offers a robust solution for privacy-preserving face recognition systems.
• The technique ensures that privacy protection is both revocable and irreversible.