Privacy-Preserving Face Recognition Using Noised Eigenvectors

Privacy-Preserving Face Recognition Using Noised Eigenvectors

Paper in proceedings outside WoS and Scopus

Widespread face recognition systems raise significant privacy concerns due to potential data exposure, especially with centralized data storage. We propose a privacy-preserving framework integrating k-same pixelation, Principal Component Analysis (PCA), and Differential Privacy (DP). Our pipeline applies k-same smoothing for initial feature averaging, uses PCA for dimensionality reduction while preserving essential facial features, and adds Laplace noise to the resulting projection vectors to achieve DP. This method masks biometric information, operating efficiently in the lower-dimensional PCA space, aiming to balance privacy protection with the utility needed for identity verification. Evaluations on the LFW dataset quantitatively analyze this trade-off using MSE and SSIM metrics. Results confirm integrating DP enhances privacy. Crucially, experiments show adding noise to lower-dimensional projection vectors preserves utility better than noising higher-dimensional eigenfaces. We identified parameters (k=10, PCA ratio=0.19, $\epsilon$n=0.24) yielding a practical balance (Avg. MSE 1499, Avg. SSIM 0.38), enabling effective machine recognition on the anonymized data and demonstrating the framework's viability.

Widespread face recognition systems raise significant privacy concerns due to potential data exposure, especially with centralized data storage. We propose a privacy-preserving framework integrating k-same pixelation, Principal Component Analysis (PCA), and Differential Privacy (DP). Our pipeline applies k-same smoothing for initial feature averaging, uses PCA for dimensionality reduction while preserving essential facial features, and adds Laplace noise to the resulting projection vectors to achieve DP. This method masks biometric information, operating efficiently in the lower-dimensional PCA space, aiming to balance privacy protection with the utility needed for identity verification. Evaluations on the LFW dataset quantitatively analyze this trade-off using MSE and SSIM metrics. Results confirm integrating DP enhances privacy. Crucially, experiments show adding noise to lower-dimensional projection vectors preserves utility better than noising higher-dimensional eigenfaces. We identified parameters (k=10, PCA ratio=0.19, $\epsilon$n=0.24) yielding a practical balance (Avg. MSE 1499, Avg. SSIM 0.38), enabling effective machine recognition on the anonymized data and demonstrating the framework's viability.

K-Same Pixel, Eigenface, Laplace Noise Addition, Differential Privacy, Facial Recognition, Biometric Authentication

K-Same Pixel, Eigenface, Laplace Noise Addition, Differential Privacy, Facial Recognition, Biometric Authentication

Bruce L'Horset, Charles Mailley, Elodie Chen, Sara Ricci

2026

29.04.2025

Brno University of Technology, Faculty of Electrical Engineering and Communication

Brno

978-80-214-6321-9

Proceedings I of the 31st Student EEICT 2025 (General Papers)

165

168

4

https://www.eeict.cz/eeict_download/archiv/sborniky/EEICT_2025_sbornik_1.pdf