da/sec scientific talk on Biometrics
Topic: Feature Truncation for Multi-Biometric Workload Reduction
by Florian Bayer
D19/2.03a, May 14, 2025 (Wednesday), 12.00 noon
Keywords — Multi-Biometrics, Feature Truncation, Multi-Modal Fusion
Abstract
„Given the widespread use of biometric recognition, it is imperative to address concerns regarding privacy and security of extracted templates. However, Biometric Template Protection (BTP) schemes, particularly those employing Homomorphic Encryption (HE), pose significant challenges due to an increase in computational workload. The utilization of deep neural networks (DNNs) in the domain of face recognition has attracted considerable attention, with recent advancements resulting in the presentation of DNN-based feature extractors for the widely used modalities face, fingerprint and iris. Additionally, biometric sensors have become cost-effective and prevalent. Therefore multi-modal fusion can be used to combine features from different modalities in order to increase security. The objective of this study is to demonstrate preliminary findings regarding biometric performance of reduced template sizes. The experiments were conducted on an in-house virtual multi-biometric database, derived from DNN-extracted features from three modalities (face, fingerprint, and iris) of the FRGC, MCYT and CASIA databases. The approaches evaluated include methods that are (i) explainable and straightforward to implement, (ii) training-free, and (iii) capable of generalization. The dimensionality reduction of feature vectors results in a reduction of the number of operations in the HE domain, thereby facilitating more efficient encrypted processing while maintain biometric accuracy and, consequently, security at a level that is equivalent to or exceeding single-biometric recognition.“