da/sec scientific talk on Biometrics

Topic: Benchmarking Quantization Schemes for Deep Features: Quantization-Aware Training vs. Post-Hoc Compression

by Florian Bayer
D19/2.03a (also online via the corresponding BBB room), November 27, 2025 (Thursday), 12.00 noon

Keywords — Neural Network Quantization, Embedding Compression, Face Recognition

Abstract

„Biometric authentication has become ubiquitous in modern security systems, with deep neural network (DNN)-extracted features consistently outperforming traditional computer vision approaches. State-of-the-art feature extractors produce highly discriminative fixed-length real-valued embeddings for identity verification. However, these biometric features present a dual challenge: they constitute sensitive personal data under GDPR regulations, and recent research has demonstrated their reversibility—raw biometric images can be reconstructed from feature vectors. Therefore, they must be protected.

Traditional cryptographic methods are impractical for biometric information protection as they typically require discrete inputs while biometric features are inherently fuzzy and continuous. Addressing these challenges requires compact, stable and discrete representations to enable cryptographic protection while maintaining efficient processing.

This work presents a benchmark of quantization strategies for deep biometric features, comparing three paradigms: (1) post-hoc quantization applied after training, (2) quantization-aware training (QAT) with learnable parameters, and (3) Index-of-Maximum (IoM) hashing. We evaluate these methods across multiple bit-widths (1-bit to 8-bit) on two popular face recognition datasets, demonstrating that in-network quantization achieves 4-32× compression while mostly preserving accuracy, enabling privacy-preserving biometric systems.“