I. Introduction
The efficacy of deep learning-powered face recognition algorithms has received much acknowledgment due to their remarkable accuracy when applied to various near-frontal face datasets such as Labeled Faces In The Wild (LFW) [1] and VGGFace [2]. The datasets mentioned above, accessible to the general public, exhibit common traits, including near-frontal positions, less than 45 degrees yaw-angle, persons of Caucasian or African heritage, and the lack of head cover attributes. It is widely acknowledged that numerous deep face recognition systems exhibit suboptimal performance when confronted with datasets without those properties. Regrettably, many faces obtained in unconstrained environments lack this attribute, leading to the compromised accuracy of face recognition algorithms when deployed in real-world scenarios.