FPGAN: Face de-identification method with generative adversarial networks for social robots.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Privacy-Preserving Technologies

Background:

• Visual facial privacy is a growing concern, especially with the increasing use of social robots and AI systems.
• Existing face de-identification methods often struggle with maintaining image quality and robust privacy protection.

Purpose of the Study:

• To propose and validate a novel, end-to-end face de-identification method called FPGAN.
• To enhance the quality and privacy protection capabilities of generative adversarial networks for facial data.

Main Methods:

• Developed FPGAN, an end-to-end generative adversarial network (GAN) for face de-identification.
• Utilized an improved U-Net generator and two seven-layer discriminators for enhanced feature extraction.
• Introduced pixel loss, content loss, and adversarial loss functions with a specific optimization strategy.

Main Results:

• FPGAN achieved superior performance compared to four baseline methods on CelebA, MORPH, RaFD, and FBDe datasets.
• The proposed method was successfully applied to face de-identification in social robots.
• A new face de-identification evaluation protocol was proposed and validated.

Conclusions:

• FPGAN offers an effective solution for visual facial privacy protection.
• The developed method and evaluation protocol advance the field of privacy-preserving AI.
• FPGAN shows significant potential for real-world applications requiring secure facial data handling.