Privacy-preserving face verification with fully homomorphic encryption

Details

Supervisors

Standaert, François-Xavier ; Peters, Thomas

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

enThis thesis investigates the use of Fully Homomorphic Encryption (FHE) for privacy-preserving one-to-one face verification. Motivated by the sensitivity of biometric data and privacy regulations such as the General Data Protection Regulation (GDPR), we focus on enabling secure computation directly on encrypted embeddings using the Fast Fully Homomorphic Encryption over the Torus (TFHE) scheme and Zama’s Concrete framework. To address FHE’s computational overhead, we propose an optimized preprocessing pipeline comprising Principal Component Analysis (PCA) for dimensionality reduction, min–max normalization, and integer quantization. We compare global and per-feature (per-dimension) strategies and find that global normalization with 4-bit global quantization offers the best trade-off between efficiency and performance metrics. Our system performs encrypted comparisons in approximately 0.142 seconds, with ciphertext sizes as small as 1.27 KB. We further demonstrate practical viability by integrating this pipeline into a proof-of-concept client-server communication scenario using Elliptic Curve Diffie–Hellman (ECDH) key exchange, Schnorr zero-knowledge identification, and Hash-based Message Authentication Codes (HMAC). Although slower than some previous approaches, our system achieves better performance on several metrics while significantly reducing memory usage, demonstrating the practicality of FHE for biometric authentication in low-throughput, high-security environments.