Facial identity protection using deep learning technologies: an application in affective computing

Abstract

Automatic prediction of human attributions of valence and arousal using facial recognition technologies can improve human–computer and human–robot interaction. However, data protection has become an issue of great concern in affect recognition using facial images, as the facial identities of people (i.e. recognising who a person is) could be exposed in the process. For instance, malicious individuals could exploit facial images of users to assume their identities and infiltrate biometric authentication systems. Possible solutions to protect the facial identity of users are to: (1) extract anonymised facial features in users’ local machines, namely action units (AU) of facial images, discard their facial images and send the AUs to the developer for processing, and (2) employ a federated learning approach i.e. process users’ facial images in their local machines and only send their locally trained models back to the developer’s machine for augmenting the final model. In this paper, we implement and compare the performance of these privacy-preserving strategies for affect recognition. Results on the popular RECOLA affective datasets show promising affect recognition performance in adopting a federated learning approach to protect users’ identities, with Concordance Correlation Coefficient of 0.426 for valence and 0.390 for arousal.