Threshold Proximity and Spoofing Vulnerability of Open-Source Facial Recognition Systems to Low-Cost Printed Fabric Masks: A Preliminary Empirical Study

Version 1.0 (Preliminary Empirical Study) Baseline cross-system evaluation prior to the full 10-mask, 240-trial benchmark Presentation Attack Detection (PAD) in facial recognition largely relies on the assumption that effective three-dimensional (3D) facial spoofing requires expensive, custom-molded silicone or resin prosthetics. In this preliminary empirical study, we investigate whether ultra-low-cost, off-the-shelf printed fabric masks featuring celebrity faces can erode biometric verification margins and approach false positive decision thresholds across diverse deep learning architectures. Unlike flat prints or digital screen replays, wearable stretchable textiles conform naturally to the attacker's 3D facial geometry and head movements without specialized fabrication. We evaluate two consecutive physical captures of a commercial printed fabric celebrity mask against a reference enrollment image across eight widely used open-source facial recognition models (ArcFace, FaceNet, FaceNet-512, VGG-Face, SFace, GhostFaceNet, OpenFace, and DeepID). Experimental measurements show that while strict default thresholds prevented outright false positives in single-reference baseline trials, improving capture alignment reduced cosine distances to within critical proximity of acceptance thresholds—reaching 0.6564 against a 0.6500 threshold on GhostFaceNet (a margin of 0.0064) and 0.0267 against 0.0150 on DeepID. These initial findings challenge the traditional cost-barrier assumption and establish the empirical foundation for our ongoing 10-mask, 240-trial cross-system benchmark.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24244927
Primary Topic
Biometric Identification and Security
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Threshold Proximity and Spoofing Vulnerability of Open-Source Facial Recognition Systems to Low-Cost Printed Fabric Masks: A Preliminary Empirical Study

Velasco Gómez Miguel
Zenodo (CERN European Organization for Nuclear Research)
Biometric Identification and Security
preprint

Threshold Proximity and Spoofing Vulnerability of Open-Source Facial Recognition Systems to Low-Cost Printed Fabric Masks: A Preliminary Empirical Study

Velasco Gómez Miguel
preprint en

Abstract

Version 1.0 (Preliminary Empirical Study) Baseline cross-system evaluation prior to the full 10-mask, 240-trial benchmark Presentation Attack Detection (PAD) in facial recognition largely relies on the assumption that effective three-dimensional (3D) facial spoofing requires expensive, custom-molded silicone or resin prosthetics. In this preliminary empirical study, we investigate whether ultra-low-cost, off-the-shelf printed fabric masks featuring celebrity faces can erode biometric verification margins and approach false positive decision thresholds across diverse deep learning architectures. Unlike flat prints or digital screen replays, wearable stretchable textiles conform naturally to the attacker's 3D facial geometry and head movements without specialized fabrication. We evaluate two consecutive physical captures of a commercial printed fabric celebrity mask against a reference enrollment image across eight widely used open-source facial recognition models (ArcFace, FaceNet, FaceNet-512, VGG-Face, SFace, GhostFaceNet, OpenFace, and DeepID). Experimental measurements show that while strict default thresholds prevented outright false positives in single-reference baseline trials, improving capture alignment reduced cosine distances to within critical proximity of acceptance thresholds—reaching 0.6564 against a 0.6500 threshold on GhostFaceNet (a margin of 0.0064) and 0.0267 against 0.0150 on DeepID. These initial findings challenge the traditional cost-barrier assumption and establish the empirical foundation for our ongoing 10-mask, 240-trial cross-system benchmark.

Zenodo (CERN European Organization for Nuclear Research)
Biometric Identification and Security
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon