On the Recognition Performance of BioHashing on state-of-the-art Face Recognition models

Face recognition has become a popular authentication tool in recent years. Modern state-of-the-art (SOTA) face recognition methods rely on deep neural networks, which extract discriminative features from face images. Although these methods have high recognition performance, the extracted features contain privacy-sensitive information. Hence, the users' privacy would be jeopardized if the features stored in the face recognition system were compromised. Accordingly, protecting the extracted face features (templates) is an essential task in face recognition systems. In this paper, we use BioHashing for face template protection and aim to establish the minimum BioHash length that would be required in order to maintain the recognition accuracy achieved by the corresponding unprotected system. We consider two hypotheses and experimentally show that the performance depends on the value of the BioHash length (as opposed to the ratio of the BioHash length to the dimension of the original features). To eliminate bias in our experiments, we use several SOTA face recognition models with different network structures, loss functions, and training datasets, and we evaluate these models on two different datasets (LFW and MOBIO). We provide an open-source implementation of all the experiments presented in this paper so that other researchers can verify our findings and build upon our work.