JMIR Publications

ABSTRACT

Background:

Facial biometric data, despite commercial value, poses significant privacy and security concerns.

Objective:

To address these concerns and support auxiliary diagnoses, we developed Digital FaceDefender, an AI-driven solution.

Methods:

To ensure privacy protection, we generated a diverse set of virtual Asian-face avatars representing both genders, spanning ages from 5 to 85 years in 10-year increments, utilizing 70,000 facial images and 13,061 Asian faces images. Landmark data were separately extracted from both the preprocessed patient images and the avatar images to accurately delineate the eye region. Affine transformations were applied to align the eye regions for image fusion, followed by color correction and Gaussian blur to enhance the quality of the finally fused images. For auxiliary diagnosis, we established 95% confidence intervals (CIs) for pixel distances within the eye region on a cohort of 1,163 individuals, serving as diagnostic benchmarks. Reidentification risks were assessed using ArcFace on 2,500 Detailed Expression Capture and Animation (DECA)-reconstructed images. Finally, Cohen’s Kappa analyses, conducted on 114 individuals, were used to evaluate the agreement between these diagnostic benchmarks and ophthalmologists' assessments.

Results:

Compared to traditional method, Digital FaceDefender significantly enhances privacy protection while maintaining essential ocular diagnostic features. Similarity score and Rank-1 accuracy analyses further confirm its efficacy in minimiznig reidentification risks across various facial poses. The Cohen's Kappa results indicate excellent agreement between the developed diagnostic benchmarks and ophthalmologists' assessments. The convenient Digital FaceDefender platform is established and accessible.

Conclusions:

In summary, Digital FaceDefender offers a robust solution for safeguarding privacy while supporting auxiliary diagnoses in ocular disease.