Deep Index-of-Maximum Hashing for Face Template Protection
Jiandong Cui, Andrew Beng Jin Teoh
Abstract
Face authentication is one of the most common biometrics available nowadays. Ensuring security of facial templates is vital to circumvent impersonation and privacy invasion. One of the notable remedies for facial template protection is cancelable biometrics whereby the compromised template can be revoked and replaced. In this work, we propose a cancelable facial template technique based on the Index-of- Maximum (IoM) hashing by means of deep neural networks, termed as Deep IoM (DIoM) hashing. Unlike data-agnostic IoM hashing, the DIoM hashing is data-driven and trained by supervision to render a discriminative cancelable facial template. The DIoM hashing relies upon a permutable pretrained deep feature learning network and a hashing network responsible for optimizing the DIoM hash codes. The hashing network consolidates maxout and softmax function, namely softmaxout to approximate the discrete DIoM hash code. A dedicated loss function is designed in order to achieve similaritypreserving learning, code balancing and quantization. The proposed network is assessed on unconstraint Labeled Faces in the Wild dataset and shown outperformed vanilla IoM hashing significantly.