3FG-Defense: Multi-Domain Adversarial Detection for Face Recognition via Cross-Domain Consistency Modeling
IEEE Access · 2026 · 14 · 92765-92785
DOI: 10.1109/access.2026.3704402 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Asma A. Alhashmi und 7 weiteren Autor:innen. Die Arbeit nennt 60 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Northern Border University, Saudi Arabia, Prince Sattam bin Abdulaziz University, University of Bisha through the Fast-Track Research Support Program.
Abstract (Originalsprache)
Adversarial attacks represent a critical security vulnerability for face recognition systems deployed in high-stakes applications, including border control, financial verification, and access management. In addition, such systems are increasingly relevant to smart mining environments, where biometric authentication is used to protect access to restricted zones, control rooms, mineral-processing facilities, and digitally connected operational assets. Current defense mechanisms face a critical trade-off: diffusion-based purification methods provide robust protection but suffer from prohibitive computational costs, while single-domain detectors operate efficiently but remain vulnerable to adaptive and unseen attack vectors. In this work, we propose 3FG-Defense, a lightweight multi-domain adversarial detection framework that uses cross-domain consistency across multiple complementary facial representations: 3D Morphable Model (3DMM) for geometric coefficients, Fast Fourier Transform (FFT) for spectral features, and Graph Convolutional Network (GCN) for facial landmark topologies. The proposed approach relies on a basic concept, that a pixel level adversarial perturbation creates measurable inconsistencies in these disparate domains with different geometric, spectral and anthropometric constraints. The pairwise cosine similarities measure and quantify the consistency of the data across domains and then are added with the extracted features to be used by a lightweight Multi-Layer Perceptron (MLP) classifier for binary classification of adversarial data. We run extensive experiments on four face benchmarks CelebA-HQ’s, CASIA-WebFace, LFW, and IJB-C’s to show that 3FG-Defense performs significantly better than the existing single-domain and multi-domain baselines under multiple attack families, like Projected Gradient Descent (PGD), Carlini & Wagner (C&W), and AutoAttack than equally well with respect to demographic subgroup and consistently in both controlled and unconstrained real-world setups. It demonstrates good performance to unseen attacks and adaptation to different recognition backbones, with a two order of magnitude speed up in inference time compared to diffusion-based purification techniques. The proposed framework is therefore also applicable to smart mining and mineral-processing security, where lightweight and robust biometric defense is required under operational and resource constraints.
Förderung
- Northern Border University, Saudi Arabia
- Prince Sattam bin Abdulaziz University
- University of Bisha through the Fast-Track Research Support Program
Zitieren
APA
Alhashmi, A. A., Alharbi, K. N. R., Alshammari, A., Abdullah, M., Alghawli, A. S. A., Sallami, C., Ebad, S. A., & Darem, A. A. (2026). 3FG-Defense: Multi-Domain Adversarial Detection for Face Recognition via Cross-Domain Consistency Modeling. IEEE Access, 14, 92765-92785. https://doi.org/10.1109/access.2026.3704402
Vancouver
Alhashmi AA, Alharbi KNR, Alshammari A, Abdullah M, Alghawli ASA, Sallami C, et al. 3FG-Defense: Multi-Domain Adversarial Detection for Face Recognition via Cross-Domain Consistency Modeling. IEEE Access. 2026;14:92765-92785. doi:10.1109/access.2026.3704402
BibTeX
@article{Alhashmi2026,
author = {Alhashmi, Asma A. and Alharbi, Khalid N. R. and Alshammari, Ahmed and Abdullah, Monir and Alghawli, Abed Saif Ahmed and Sallami, Chams and Ebad, Shouki A. and Darem, Abdulbasit A.},
title = {3FG-Defense: Multi-Domain Adversarial Detection for Face Recognition via Cross-Domain Consistency Modeling},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {92765--92785},
doi = {10.1109/access.2026.3704402}
}
EUVIMED