How to Save Your Face: a Facial Recognition Method Robust Against Image Reconstruction

Nowadays systems based on biometric authorization are widely applied in the industry, financial and government resorts or even as continuous identification of security or CCTV live feed. In the case of security breach in a database, the image of users' “biometric token” (e.g. face, palm, iris) may be recovered from the feature vector stored in the database. Such image reconstruction presents an additional level of privacy threat that is enhanced further by the rapid progress of the popularity of neural networks, e.g. generative adversarial networks. We propose a method of biometric identification based on facial features that aims at privacy-by-design, i.e. is robust against some classes of face reconstruction attacks. Our method identifies displacements of facial landmarks in a short input video and utilizes the displacements' vectors instead of landmark positions. The displacements' vectors of two samples are compared by the DTW algorithm and a final decision of similarity of the two samples is undertaken by the SVM trained using similarity vectors. Tests were made on 34 volunteers, with each taking part in 5 to 7 video samples. Depending on the number of enrolled samples, we achieved the identification rate equal to 70.79% or 69.09%, whereas a random guessing method gives 2.94% rate. We also got accuracy equal to 68.54% utilizing only small subsets of 44 features. The method could be used as an auxiliary biometric system, guaranteeing the presence of the identified person and improving their privacy protection. Our system could be also considered as cancelable biometrics, where a given sentence is a key used for transformation/encryption.