On September 24, 2025, Maximilian Russo successfully defended his BSc thesis entitled „Multi-Instanz-Ansatz zur Dimensionalitätsreduktion DNN-extrahierter Fingerabdrucksmerkmale“.
Abstract:
Ongoing digitalization demands authentication methods that are both secure and user-friendly. Conventional passwords and PINs are either insecure when chosen simply or impair usability when made complex. Biometric systems offer a promising alternative because they rely on unique, body-bound characteristics and can provide both strong security and ease of use. Fingerprints are among the most widely deployed biometric traits. However, storing and processing biometric data imposes strict requirements on privacy and security, since these traits, unlike passwords, cannot be readily replaced. The General Data Protection Regulation (GDPR) therefore explicitly requires that biometric data be protected with state-of-the-art technical measures, in particular strong encryption. Homomorphic encryption is especially suitable in this context because it enables comparisons directly on encrypted data, but its computational cost currently limits practical deployment. This work investigates how to reduce computational and storage overhead in processing fingerprint features without substantially degrading recognition performance. Feature vectors extracted with deep neural networks (DNNs) were subjected to various forms of dimensionality reduction and subsequently evaluated for classification performance. Considered were both single-instance approaches, which process a single fingerprint, and multi-instance approaches, which combine multiple fingerprints. The results show that dimensionality can be reduced without significant losses in recognition accuracy. In particular, the single-instance approach with binary quantization and interleaving enables drastic reductions, both the number of bits per entry and the number of features can be lowered with only minor impact on classification quality. Within the multi-instance setting, the AddTwo approach, which utilizes element-wise addition of the feature vectors of two successive fingerprints, proved especially promising, as it preserves information from two fingers at the same memory and computational cost as processing a single finger. It produced stable results even at reduced feature dimensionalities. However, multi-instance approaches do not provide a clear advantage over single-instance methods and can even yield slightly worse results in some cases.