On January 22, 2026, Valentina Fohr successfully defended his MSc thesis entitled „Security and Performance Evaluation of Biometric Fusion in Dual-User Authentication“.
Abstract:
Biometric Cryptosystems are widely used for secure authentication but traditionally assume a strict one-to-one relationship between a user and their biometric template. This assumption limits their applicability in scenarios requiring account sharing or delegated access, such as shared household devices, collaborative work environments or power of attorney. To address this gap, this work investigates how biometric fusion techniques can enable Dual-user authentication, while maintaining acceptable recognition performance and practical security within a Biometric Cryptosystem. While various fusion methods exist, this work focuses on Sensor-level fusion, implemented through facial Image morphing, and Feature-level fusion, performed by averaging facial embeddings. Feature extraction is performed using the state-of-the-art Deep learning model MagFace. This work aims to provide insights into the effectiveness of these fusion methods regarding recognition performance and security within the Fuzzy Vault Scheme. To accomplish this aim, the following steps are performed. First, Image preprocessing and facial feature extraction are carried out. Next, multiple Dual-user datasets are constructed, including Image-fused and Feature-fused (average, worst-case, random) datasets, and a baseline Single-user dataset. These datasets are then analysed using cosine similarity and Hamming distance metrics, followed by cross-metric comparisons. Moreover, the Dual-user dataset is evaluated regarding loss in recognition performance and similarity balance of the contributing subjects. Finally, the fusion methods are evaluated within the Fuzzy Vault scheme. The findings reveal that, overall, the most effective fusion method for Dual-user authentication is Feature-level fusion through averaging. Further, the findings indicate a clear trade-off between recognition performance and security for the Image-level fusion method. Image-level morphing can achieve acceptable recognition performance only at the expense of practical security, whereas configurations that ensure practical security fail to meet acceptable recognition performance.