How much of a face is a face: exploring reidentification potential with generative AI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Clinical photographs are crucial in medicine, with deidentification historically using simple masking like black bars over eyes.
• This traditional deidentification method is increasingly questioned due to evolving technologies.

Purpose of the Study:

• To evaluate the risk of generative artificial intelligence (AI) reconstructing faces from partially obscured clinical photographs.
• To assess the effectiveness of current deidentification practices against advanced AI.

Main Methods:

• Generative AI was used to reconstruct 10,000 facial images from the Synthetic Faces High Quality dataset.
• 14 different regional masking strategies were applied, including single-feature and multi-feature occlusions.
• Reconstructed images were analyzed for face mesh distortion and structural similarity, and verified against original images using the Dlib face verification model.

Main Results:

• Masking only the eyes or any single facial feature allowed for highly identifiable reconstructions, with low distortion and high similarity to original faces.
• A face verification model successfully matched 97.98% to 99.93% of reconstructed images to their originals when only one feature was masked.
• Removing all major facial features (eyes, nose, mouth, eyebrows) reduced face verification rates significantly, from 98.87% to 33.93%.

Conclusions:

• Generative AI poses a significant reidentification risk for clinical photographs deidentified with traditional single-feature masking.
• Quantitative metrics demonstrate the high success rate of AI in reconstructing faces even with partial obscuration.
• Current deidentification guidelines may require revision to address the capabilities of modern AI technologies.