Hybrid-FGPG: a novel fusion of gradient and geometric transformations for adversarial vulnerability evaluation

Dimitrios-Christos Asimopoulos; Panagiotis Radoglou-Grammatikis; Vasileios Argyriou; Thomas Lagkas; Pantelis Angelidis; Vasileios Vitsas; Panagiotis Fouliras; Ioannis Ktenidis; Panagiotis Sarigiannidis

doi:10.1109/GCWkshps68340.2025.11590917

Hybrid-FGPG: a novel fusion of gradient and geometric transformations for adversarial vulnerability evaluation

Dimitrios-Christos Asimopoulos
, Panagiotis Radoglou-Grammatikis
, Vasileios Argyriou
, Thomas Lagkas
, Pantelis Angelidis
, Vasileios Vitsas
, Panagiotis Fouliras
, Ioannis Ktenidis
, Panagiotis Sarigiannidis

International Hellenic University
University of Western Macedonia
K3Y Limited
Democritus University of Thrace
University of Macedonia

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

Deep neural networks (DNNs) are known to be vulnerable to adversarial examples—small, carefully crafted perturbations that can mislead model predictions. In this work, we introduce Hybrid-FGPG, a novel white-box adversarial attack that combines traditional gradient-based perturbations with geometric transformations such as pixel-wise rotation. Unlike classical methods such as FGSM and PGD, our approach explores additional perturbation subspaces, leading to more effective and less perceptible attacks. We evaluate Hybrid-FGPG across three diverse image datasets—CIFAR-10, GTSRB, and EuroSAT—using a ResNet-18 backbone. Comparative results against FGSM and PGD show that Hybrid-FGPG achieves higher misclassification rates while preserving visual fidelity. We also present qualitative visualizations and perturbation heatmaps that reveal the hybrid attack’s stealthy nature. Our findings suggest that integrating geometric components into gradient-based methods significantly enhances adversarial efficacy and transferability.

Original language	English
Title of host publication	2025 IEEE Globecom Workshops (GC Wkshps)
Publisher	IEEE
Pages	2000-2005
Number of pages	6
ISBN (Electronic)	9798331567415
ISBN (Print)	9798331567422
DOIs	https://doi.org/10.1109/GCWkshps68340.2025.11590917
Publication status	Published - 7 Jul 2026
Event	IEEE Global Communications Conference: Sustainable Communications for Ubiquitous Intelligence - Taipei International Convention Center, Taipei, Taiwan, Province of China Duration: 8 Dec 2025 → 12 Dec 2025 https://globecom2025.ieee-globecom.org/workshops

Publication series

Name	IEEE Globecom Workshops
Publisher	IEEE
ISSN (Print)	2166-0069
ISSN (Electronic)	2166-0077

Conference

Conference	IEEE Global Communications Conference
Country/Territory	Taiwan, Province of China
City	Taipei
Period	8/12/25 → 12/12/25
Internet address	https://globecom2025.ieee-globecom.org/workshops

Keywords

Adversarial attacks
Black-box
evasion
surrogate-model
transferability
white-box

Access to Document

10.1109/GCWkshps68340.2025.11590917

Cite this

Asimopoulos, D.-C., Radoglou-Grammatikis, P., Argyriou, V., Lagkas, T., Angelidis, P., Vitsas, V., Fouliras, P., Ktenidis, I., & Sarigiannidis, P. (2026). Hybrid-FGPG: a novel fusion of gradient and geometric transformations for adversarial vulnerability evaluation. In 2025 IEEE Globecom Workshops (GC Wkshps) (pp. 2000-2005). (IEEE Globecom Workshops). IEEE. https://doi.org/10.1109/GCWkshps68340.2025.11590917

@inproceedings{c3b0717d49674a38a0a9f0645d0035bb,

title = "Hybrid-FGPG: a novel fusion of gradient and geometric transformations for adversarial vulnerability evaluation",

abstract = "Deep neural networks (DNNs) are known to be vulnerable to adversarial examples{\textemdash}small, carefully crafted perturbations that can mislead model predictions. In this work, we introduce Hybrid-FGPG, a novel white-box adversarial attack that combines traditional gradient-based perturbations with geometric transformations such as pixel-wise rotation. Unlike classical methods such as FGSM and PGD, our approach explores additional perturbation subspaces, leading to more effective and less perceptible attacks. We evaluate Hybrid-FGPG across three diverse image datasets{\textemdash}CIFAR-10, GTSRB, and EuroSAT{\textemdash}using a ResNet-18 backbone. Comparative results against FGSM and PGD show that Hybrid-FGPG achieves higher misclassification rates while preserving visual fidelity. We also present qualitative visualizations and perturbation heatmaps that reveal the hybrid attack{\textquoteright}s stealthy nature. Our findings suggest that integrating geometric components into gradient-based methods significantly enhances adversarial efficacy and transferability.",

keywords = "Adversarial attacks, Black-box, evasion, surrogate-model, transferability, white-box",

author = "Dimitrios-Christos Asimopoulos and Panagiotis Radoglou-Grammatikis and Vasileios Argyriou and Thomas Lagkas and Pantelis Angelidis and Vasileios Vitsas and Panagiotis Fouliras and Ioannis Ktenidis and Panagiotis Sarigiannidis",

year = "2026",

month = jul,

day = "7",

doi = "10.1109/GCWkshps68340.2025.11590917",

language = "English",

isbn = "9798331567422",

series = "IEEE Globecom Workshops",

publisher = "IEEE",

pages = "2000--2005",

booktitle = "2025 IEEE Globecom Workshops (GC Wkshps)",

note = "IEEE Global Communications Conference : Sustainable Communications for Ubiquitous Intelligence ; Conference date: 08-12-2025 Through 12-12-2025",

url = "https://globecom2025.ieee-globecom.org/workshops",

}

Asimopoulos, D-C, Radoglou-Grammatikis, P, Argyriou, V, Lagkas, T, Angelidis, P, Vitsas, V, Fouliras, P, Ktenidis, I & Sarigiannidis, P 2026, Hybrid-FGPG: a novel fusion of gradient and geometric transformations for adversarial vulnerability evaluation. in 2025 IEEE Globecom Workshops (GC Wkshps). IEEE Globecom Workshops, IEEE, pp. 2000-2005, IEEE Global Communications Conference, Taipei, Taiwan, Province of China, 8/12/25. https://doi.org/10.1109/GCWkshps68340.2025.11590917

Hybrid-FGPG: a novel fusion of gradient and geometric transformations for adversarial vulnerability evaluation. / Asimopoulos, Dimitrios-Christos; Radoglou-Grammatikis, Panagiotis; Argyriou, Vasileios et al.
2025 IEEE Globecom Workshops (GC Wkshps). IEEE, 2026. p. 2000-2005 (IEEE Globecom Workshops).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - Hybrid-FGPG

T2 - IEEE Global Communications Conference

AU - Asimopoulos, Dimitrios-Christos

AU - Radoglou-Grammatikis, Panagiotis

AU - Argyriou, Vasileios

AU - Lagkas, Thomas

AU - Angelidis, Pantelis

AU - Vitsas, Vasileios

AU - Fouliras, Panagiotis

AU - Ktenidis, Ioannis

AU - Sarigiannidis, Panagiotis

PY - 2026/7/7

Y1 - 2026/7/7

N2 - Deep neural networks (DNNs) are known to be vulnerable to adversarial examples—small, carefully crafted perturbations that can mislead model predictions. In this work, we introduce Hybrid-FGPG, a novel white-box adversarial attack that combines traditional gradient-based perturbations with geometric transformations such as pixel-wise rotation. Unlike classical methods such as FGSM and PGD, our approach explores additional perturbation subspaces, leading to more effective and less perceptible attacks. We evaluate Hybrid-FGPG across three diverse image datasets—CIFAR-10, GTSRB, and EuroSAT—using a ResNet-18 backbone. Comparative results against FGSM and PGD show that Hybrid-FGPG achieves higher misclassification rates while preserving visual fidelity. We also present qualitative visualizations and perturbation heatmaps that reveal the hybrid attack’s stealthy nature. Our findings suggest that integrating geometric components into gradient-based methods significantly enhances adversarial efficacy and transferability.

AB - Deep neural networks (DNNs) are known to be vulnerable to adversarial examples—small, carefully crafted perturbations that can mislead model predictions. In this work, we introduce Hybrid-FGPG, a novel white-box adversarial attack that combines traditional gradient-based perturbations with geometric transformations such as pixel-wise rotation. Unlike classical methods such as FGSM and PGD, our approach explores additional perturbation subspaces, leading to more effective and less perceptible attacks. We evaluate Hybrid-FGPG across three diverse image datasets—CIFAR-10, GTSRB, and EuroSAT—using a ResNet-18 backbone. Comparative results against FGSM and PGD show that Hybrid-FGPG achieves higher misclassification rates while preserving visual fidelity. We also present qualitative visualizations and perturbation heatmaps that reveal the hybrid attack’s stealthy nature. Our findings suggest that integrating geometric components into gradient-based methods significantly enhances adversarial efficacy and transferability.

KW - Adversarial attacks

KW - Black-box

KW - evasion

KW - surrogate-model

KW - transferability

KW - white-box

U2 - 10.1109/GCWkshps68340.2025.11590917

DO - 10.1109/GCWkshps68340.2025.11590917

M3 - Conference contribution

SN - 9798331567422

T3 - IEEE Globecom Workshops

SP - 2000

EP - 2005

BT - 2025 IEEE Globecom Workshops (GC Wkshps)

PB - IEEE

Y2 - 8 December 2025 through 12 December 2025

ER -

Hybrid-FGPG: a novel fusion of gradient and geometric transformations for adversarial vulnerability evaluation

Abstract

Publication series

Conference

Keywords

Access to Document

Fingerprint

Cite this