Artificial Intelligence–Enabled Facial Privacy Protection for Ocular Diagnosis: Development and Validation Study
Abstract BackgroundFacial biometric data, while valuable for clinical applications, poses substantial privacy and security risks. ObjectiveThis paper aims to address the privacy and security concerns related to facial biometric data and support auxiliary diagnoses,...
Saved in:
| Main Authors: | , , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
JMIR Publications
2025-07-01
|
| Series: | Journal of Medical Internet Research |
| Online Access: | https://www.jmir.org/2025/1/e66873 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract
BackgroundFacial biometric data, while valuable for clinical applications, poses substantial privacy and security risks.
ObjectiveThis paper aims to address the privacy and security concerns related to facial biometric data and support auxiliary diagnoses, in pursuit of which we developed Digital FaceDefender, an artificial intelligence–driven privacy safeguard solution.
MethodsWe constructed a diverse set of digitally synthesized Asian face avatars representing both sexes, spanning ages 5 to 85 years in 10-year increments, using 70,000 facial images and 13,061 Asian face images. Landmark data were separately extracted from both patient and avatar images. Affine transformations ensured spatial alignment, followed by color correction and Gaussian blur to enhance fusion quality. For auxiliary diagnosis, we established 95% CIs for pixel distances within the eye region on a cohort of 1163 individuals, serving as diagnostic benchmarks. Reidentification risk was assessed using the ArcFace model, applied to 2500 images reconstructed via Detailed Expression Capture and Animation (DECA). Finally, Cohen Kappa analyses (n=114) was applied to assess agreement between diagnostic benchmarks and ophthalmologists’ evaluations.
ResultsCompared to the DM method, Digital FaceDefender significantly reduced facial similarity scores (FDface vs raw images: 0.31; FLAME_FDface vs raw images: 0.09) and achieved zero Rank-1 accuracy in Pose #2-#3 and Pose #2-#5, with near-zero accuracy in Pose #4 (0.02) and Pose #5 (0.04), confirming lower reidentification risk. Cohen Kappa analysis demonstrated moderate agreement between our benchmarks and ophthalmologists’ assessments for the left eye (κ=0.37) and right eye (κ=0.45; both P
ConclusionsIn summary, Digital FaceDefender effectively balances privacy protection and diagnostic use. |
|---|---|
| ISSN: | 1438-8871 |