Premium
Popular MRI de‐facing software does not sufficiently protect participants from re‐identification via face recognition
Author(s) -
Schwarz Christopher G.,
Kremers Walter K.,
Therneau Terry M.,
Sharp Richard R.,
Gunter Jeffrey L.,
Vemuri Prashanthi,
Arani Arvin,
Spychalla Anthony J.,
Kantarci Kejal,
Knopman David S.,
Petersen Ronald C.,
Jack Clifford R.
Publication year - 2020
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.045157
Subject(s) - artificial intelligence , facial recognition system , voxel , computer science , chin , population , software , face (sociological concept) , computer vision , pattern recognition (psychology) , medicine , anatomy , programming language , social science , environmental health , sociology
Background Data shared for research use must be de‐identified to protect participant confidentiality. A recent study has shown that face imagery from brain imaging studies could be used to re‐identify anonymous participants via automatic face recognition, linking them with all study‐collected protected health information. Here we measure how well existing software for removing MRI face imagery protects against this privacy threat. Method We recruited 153 participants with existing 3D FLAIR MRI and photographed their face from 5 varying angles. We used surf_ice to reconstruct each participant’s face from their MRI and tested whether free, publicly‐available face recognition software (Microsoft Azure) could match their photos to their correct MRI. We tested two existing programs for removing face voxels in MRI: mri_deface (FreeSurfer) and pydeface (python package), and one in‐house implementation. We observed that existing methods often left participants’ eyes intact while removing most of their nose/mouth/chin. This prevented Azure from detecting the reconstructions as faces, therefore preventing its attempt of face recognition. We hypothesized that the retained eye imagery alone could be sufficient for face recognition, if a face were detected. Therefore, we also reconstructed each de‐faced MRI after applying novel pre‐processing that replaced any removed/missing face voxels with those from our population‐average template, e.g. replacing missing noses/mouths/chins with population‐average noses/mouths/chins, before testing face recognition. Results Without any MRI de‐facing, Microsoft Azure correctly matched 78% of participants’ photos to their MRI. After mri_deface and pydeface, 8% and 6% were still identified, respectively. After preprocessing the de‐faced MRI to replace missing areas with those from a population‐average template prior to face reconstruction, the face imagery remaining in the de‐faced images was enough to allow re‐identification of 23% and 20% of participants, respectively. With an in‐house face removal implementation that removes the whole face from every image, 0% of faces were re‐identified from the standard reconstructions, and 5% were identified from the “missing‐parts‐replacement” reconstructions. Conclusions Even after applying popular MRI de‐facing software, Microsoft Azure face recognition automatically re‐identified up to 23% of participants. Better de‐facing software is needed to remove more of the face more reliably, without affecting brain measurements from de‐faced images.