On the Feasibility of Privacy-Secured Facial Authentication for low-power IoT Devices - Quantifying the Effects of Head Pose Variation on End-to-End Neural Face Recognition

Abstract

Recent low-power neural accelerator hardware provides a solution for end-to-end privacy and secure facial authentication, such as smart refueling machine locks in shared accommodation, smart speakers, or televisions that respond only to family members. This work explores the impact that head pose variation has on the performance of a state-of-the-art face recognition model. A synthetic technique is employed to introduce head pose variation into data samples. Experiments show that the synthetic pose variations have a similar effect on face recognition performance as the real samples with pose variations. The impact of large variations of head poses on the face recognizer was then explored by further amplifying the angle of the synthetic head pose. It is found that the accuracy of the face recognition model deteriorates as the pose increases. After fine-tuning the network, the face recognition model achieves close to the accuracy of frontal faces in all pose variations, indicating that the face recognition model can be tuned to compensate for the effect of large poses.