Valentina Fohr successfully finished her Bachelor Thesis on Evaluation of Fusion Methods for Multi-Biometric Cryptosystems

Biometric Cryptosystems have become increasingly popular due to their ability to preserve privacy for biometric data, e.g., iris or fingerprints. However, the entropy of a single biometric characteristic is limited regarding recognition performance and security. Consequently, to improve the recognition performance and security of Biometric Cryptosystems, a fusion of multiple biometric characteristics is required. While various fusion methods exist, this work focuses on the concatenation, interleaving and randomly shuffled methods. This work aims to provide insights into which of these fusion methods is the most effective regarding recognition performance and security of Biometric Cryptosystems, specifically within the framework of the fuzzy commitment scheme. In order to accomplish this aim, the following steps are performed. First, monomodal biometric databases from distinct biometric characteristics are created. The creation of databases from different modalities with different extractors poses a challenge as such extracted modalities result in non-uniform representation vectors. To address this challenge, datasets generated by Convolutional Neural Networks are used. Next, fused biometric databases are created by fusing the embeddings of the monomodal biometric databases using the concatenation, interleaving and randomly shuffled fusion methods. Afterwards, the fused databases are evaluated with respect to their recognition performance and security utilizing bit-level and block-level error correction codes. The findings show that overall, the most effective fusion method regarding recognition performance is the random shuffling method, closely followed by the interleaved method. Whereas the concatenation method performs poorly in comparison to the other two methods. The findings also reveal that security depends on the block size and number of blocks in the bit-level error correction, and on the number of correctable blocks in the block-level error correction, but not specifically on the fusion method.