Riemannian Adversarial Attacks on Subspace Detectors via the Grassmann Manifold
IEEE Signal Processing Letters · 2026 · 33 · 3117-3121
DOI: 10.1109/lsp.2026.3716159 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von George Atia. Die Arbeit nennt 27 Literaturangaben. Gefördert von: National Science Foundation.
Abstract (Originalsprache)
We study adversarial vulnerabilities of the Matched Subspace Detector (MSD), a classical GLRT-based detector whose test statistic depends on an estimated signal subspace. Because the space of $k$ -dimensional subspaces of $\mathbb {C}^{n}$ is the Grassmann manifold $\text{Gr}(k,n)$ , geometrically meaningful perturbations must respect its intrinsic Riemannian structure. We propose Grassmannian PGD (Gr-PGD), which minimizes the test statistic over a geodesic ball via Riemannian gradient descent with a closed-form projection. For the isotropic signal model we prove that the equal-angle allocation uniquely minimizes the aggregate retained signal power across all $k$ principal angles and derive a closed-form missed detection probability $P_{MD}(\epsilon)$ for this attack, together with a characterization of vulnerability regimes in terms of SNR and subspace dimension. Experiments confirm the analytical results and demonstrate that snapshot-aware Gr-PGD greatly exceeds the average-case analytical benchmark.
Förderung
- National Science Foundation
Zitieren
APA
Atia, G. (2026). Riemannian Adversarial Attacks on Subspace Detectors via the Grassmann Manifold. IEEE Signal Processing Letters, 33, 3117-3121. https://doi.org/10.1109/lsp.2026.3716159
Vancouver
Atia G. Riemannian Adversarial Attacks on Subspace Detectors via the Grassmann Manifold. IEEE Signal Processing Letters. 2026;33:3117-3121. doi:10.1109/lsp.2026.3716159
BibTeX
@article{Atia2026,
author = {Atia, George},
title = {Riemannian Adversarial Attacks on Subspace Detectors via the Grassmann Manifold},
journal = {IEEE Signal Processing Letters},
year = {2026},
volume = {33},
pages = {3117--3121},
doi = {10.1109/lsp.2026.3716159}
}
EUVIMED