To enable broader use of face recognition in scenarios with severe computational constraints, this thesis presents a series of efficient face recognition models called MixFaceNets. With a focus on automated network architecture design, this work is the first to apply Neural Architecture Search to develop a set of compact network architectures, referred to as PocketNets, for face recognition.

Furthermore, the thesis introduces a new training paradigm based on knowledge distillation, called multi-step KD, which improves the verification performance of compact models. To further enhance face recognition accuracy, the thesis also presents a novel margin-penalty softmax loss function, ElasticFace, which removes the limitation of a fixed penalty margin.

The occlusion of parts of the face by face masks during the recent COVID-19 pandemic posed a new challenge for face recognition systems. This thesis presents an approach that mitigates the effects of masks on recognition performance and thereby improves accuracy. The proposed solution builds on existing face recognition models and thus avoids additional computational overhead from retraining or implementing a separate solution specifically for masked faces.

With the goal of introducing biometric recognition into new embedded domains, this thesis is the first to propose the use of head-mounted displays for identity verification of users of virtual and augmented reality applications. For this purpose, a compact solution for eye segmentation as part of the recognition pipeline is also proposed. In addition, an identity-preserving approach for generating synthetic eye images is developed to address potential privacy concerns related to access to real biometric data and to facilitate further advancement of biometric recognition in new application areas.