On 5 March 2025, Marcel Grimmer successfully defended his thesis titled

Deep Generative Networks in Face Recognition with Application to Face Image Quality Assessment

for the Doctoral Degree in Information Security and Communication Technology.

Abstract

The widespread adoption of face recognition systems for authentication spans from convenience-based applications (e.g., smartphone unlocking) to security-critical applications (e.g., border control). This trend stems from the transition from traditional machine learning to deep learning-based recognition models, which have significantly improved recognition accuracy and robustness against various sources of variation. However, the effectiveness of deep neural networks largely depends on the availability of sufficient training and evaluation data. One possible solution to meet the high demand for data is to exploit recent advances in deep generative models, which can generate highly realistic facial images that are indistinguishable from images upon human inspection. Therefore, this thesis examines how large-scale data collection can be circumvented by using deep generative models for evaluating face recognition systems and deriving quality assessment measures.
 The specific contributions of this thesis can be summarised as follows:

• An investigation into how deep generative models can be used to create synthetic mated samples by editing facial attributes while preserving identity.
• A survey on deep learning-based face age editing techniques, relevant datasets, evaluation strategies, and open challenges.
• A survey on deep learning-based face age editing techniques, relevant datasets, evaluation strategies, and open challenges.
• Development of a diffusion-based face age editing technique that outperforms current state-of-the-art methods in identity preservation, photorealism, and age pattern diversity.
• Generation of synthetic evaluation datasets to estimate task-specific metrics, such as equal error rates or average similarity decreases.
• Development of quality component measures according to ISO/IEC 29794-5, for predicting a sample’s face recognition utility with respect to individual quality elements, facilitated by deep generative models.

Assessment Committee

The following committee was appointed to evaluate the thesis, trial lecture, and defence:

• 1. opponent:                Professor Alice O'Toole, The University of Texas at Dallas, USA

• 2. opponent:                Professor Adam Herout, Brno University of Technology, Czech Republic

• Internal member:         Professor Patrick Bours, NTNU, Norway 

Professor Patrick Bours, Department of Information Security and Communication Technology, NTNU, was appointed as the administrator of the assessment committee.

Supervisors

The doctoral work has been carried out at the Department of Information Security and Communication Technology.

The main supervisor was Professor Christoph Busch, and the co-supervisor was Professor Raghavendra Ramachandra.

Congratulations!