An Empirical Evaluation of Grad-CAM and Optimal-Transport Self-Explaining Gates
CENTRAL ASIAN JOURNAL OF MATHEMATICAL THEORY AND COMPUTER SCIENCES · 2026
DOI: 10.51699/n1x0jf96 ↗Kurz gefasst
Fachartikel in der Zeitschrift CENTRAL ASIAN JOURNAL OF MATHEMATICAL THEORY AND COMPUTER SCIENCES, 2026. Verfasst von Maryam Suleiman und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Automated facial recognition (AFR) systems built on deep convolutional and transformer backbones achieve near-human verification accuracy, yet they remain vulnerable to Physical Adversarial Attacks (PAAs) — adversarial eyeglasses, printed patches, cosmetic occlusion, and high-fidelity presentation media that manipulate a camera's optical input before any digital pixel is captured. Reactive Presentation Attack Detection (PAD) systems treat this problem as static artefact classification and fail against unmodeled, "zero-day" attack vectors. This paper reframes post-hoc explainability as an active authentication primitive. We design and empirically evaluate a Self-Explaining Gate (SEG): a plug-and-play verification layer that cross-examines a recognition backbone's internal attention against a zero-background anatomical prior mask covering the eyes, nose, and facial region, expressed through a bounded Anatomical Consistency Score (ACS). Two attribution mechanisms are compared — Gradient-weighted Class Activation Mapping (Grad-CAM) and a Vision-Transformer attention field aligned via Sinkhorn Optimal Transport (ViT-OT) — across the CASIA Face Anti-Spoofing Dataset (CASIA-FASD) and the Spoof in the Wild (SiW) dataset. Results reveal a pronounced three-way trade-off between security, usability, and generalization. The Grad-CAM gate achieves a strong intra-dataset defensive posture (False Acceptance Rate as low as 1.19% under zero-background masking) but suffers severe false rejection of genuine users (up to 99.83%) and its accuracy collapses to 22.87% on the unseen SiW dataset. In particular, the ViT-OT gate leads to a significant cut in false rejections (11.16% on SiW) but its reliance solely on geometric alignment opens up to spoof media (media where facial structure is preserved) that can trick the system into accepting a false match. So, False Acceptance Rates skyrocketing up to 91%. It is a fact as well as the results obtained that a non-zero background weight in the validation mask creates a kind of silent "mathematical floor" which inflates the consistency scores, and removing these weights leads to a considerable improvement in the security-utility trade-off. For instance, the ablation study shows that the Half-Error (Half-Total Error) Rate dropped from 7.56% to 2.30%. Finally, our analysis reveals that aligning attention with anatomical features is a required but not entirely sufficient condition for robustness against physical adversarial examples, and to have a reliable deployment semantic sanitization of doubtful regions should be used with the spatial consistency scores.
Zitieren
APA
Suleiman, M., Ogedebe, P., & Idris, U. A. (2026). An Empirical Evaluation of Grad-CAM and Optimal-Transport Self-Explaining Gates. CENTRAL ASIAN JOURNAL OF MATHEMATICAL THEORY AND COMPUTER SCIENCES. https://doi.org/10.51699/n1x0jf96
Vancouver
Suleiman M, Ogedebe P, Idris UA. An Empirical Evaluation of Grad-CAM and Optimal-Transport Self-Explaining Gates. CENTRAL ASIAN JOURNAL OF MATHEMATICAL THEORY AND COMPUTER SCIENCES. 2026. doi:10.51699/n1x0jf96
BibTeX
@article{Suleiman2026,
author = {Suleiman, Maryam and Ogedebe, Peter and Idris, Usman Abubakar},
title = {An Empirical Evaluation of Grad-CAM and Optimal-Transport Self-Explaining Gates},
journal = {CENTRAL ASIAN JOURNAL OF MATHEMATICAL THEORY AND COMPUTER SCIENCES},
year = {2026},
doi = {10.51699/n1x0jf96}
}
EUVIMED