0
FaceObfuscator: A New Solution for Facial Privacy Protection from Zhejiang University andAlibaba
This article reports on a new facial privacy protection solution called FaceObfuscator, developed by Zhejiang University and Alibaba Security. It addresses the growing concern of facial data security and the vulnerability of facial recognition systems to reconstruction attacks.
KeyPoints:
- The Problem: Existing facial recognition systems store facial features in databases, which are susceptible to leakage. Malicious actors can use these features toreconstruct original faces through powerful AI techniques, compromising privacy.
- FaceObfuscator’s Solution: The solution involves two key steps:
- Redundant Information Removal: It removes redundant visual information from facial images,preserving only the essential features for recognition.
- Facial Feature Obfuscation: It introduces randomness to the facial features, making it impossible for attackers to reconstruct the original face.
- Advantages:
- Strong PrivacyProtection: Effectively prevents reconstruction attacks.
- High Recognition Accuracy: Maintains high recognition accuracy comparable to mainstream models.
- High Efficiency: Requires minimal storage space and computational resources.
- Applications: FaceObfuscator can be widely applied in various scenarios, including surveillance, facial payment,and access control, benefiting industries like security, finance, and education.
Technical Details:
- Frequency Domain Transformation: The solution utilizes Discrete Cosine Transform (DCT) to convert images into frequency domain features, allowing for the identification and removal of redundant information.
- Randomness Introduction: Randomnessis introduced to the facial features through both direction and scale transformations, hindering the reconstruction process.
- Server-Side Reversal: The server-side reverses the scale transformation to maintain recognition accuracy.
Experimental Results:
- Robustness: FaceObfuscator effectively prevents reconstruction attacks, significantly reducing thecosine similarity between reconstructed and original images and minimizing the success rate of replay attacks.
- Performance: The solution maintains high recognition accuracy, low storage overhead, and efficient computational speed.
Conclusion:
FaceObfuscator offers a promising solution for facial privacy protection, effectively addressing the vulnerabilities of existing facial recognition systems. Its combination of strong privacy protection, high recognition accuracy, and efficient performance makes it a valuable tool for ensuring secure and reliable facial recognition applications.
【source】https://www.ithome.com/0/791/494.htm
Views: 0
0