AdvGAN-RSSI: Untargeted Regression Attacks on Indoor Localization Models
IEEE Sensors Journal · 01.06.2026 · 26 (11) · 17902-17917
DOI: 10.1109/jsen.2026.3686200 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Sensors Journal, erschienen am 01.06.2026. Verfasst von Ayesha Ayub und 2 weiteren Autor:innen. Die Arbeit nennt 41 Literaturangaben. Gefördert von: Ministry of Higher Education (MoHE) Malaysia through the Fundamental Research.
Abstract (Originalsprache)
[] Indoor localization systems based on received signal strength indicator (RSSI) fingerprints are increasingly deployed in real-world environments; yet, their susceptibility to adversarial perturbations poses significant security concerns. This article investigates the vulnerability of regression-based indoor localization models to untargeted adversarial attacks under both glass-box and closed-box settings. We extend the AdvGAN framework to generate small but highly effective perturbations in two variants: 1) AdvGAN for static attacks and 2) LatentAdvGAN for dynamic, noise-conditioned attacks. Evaluations on two benchmark datasets—Bluetooth low energy (BLE) and SODIndoorLoc (Wi-Fi)—demonstrate severe localization degradation. On the BLE dataset, root-mean-square error (RMSE) increased from 0.427 to 14.148 m and success rate (SR) dropped from 98.11% to 2.71% under closed-box attack. On SODIndoorLoc (CETC331), RMSE rose from 2.251 to 10.544 m, and the floor classification accuracy decreased by over 91.88%. Despite their impact, all perturbations satisfied the $\ell _{{2}}$ norm constraint of $||\delta ||\lt {1}.{0}$ . A comparative study with gradient-based baselines (FGSM, PGD, MIM, and CW-L2) under identical perturbation budgets further confirms that GAN-based attacks achieve stronger, more transferable, and semantically consistent perturbations than conventional methods across both BLE and Wi-Fi environments. These findings highlight the real-world implications of adversarial vulnerability in indoor localization, with transferable attacks exposing significant weaknesses in current systems. Our results call for stronger defenses and robust training paradigms to safeguard localization accuracy against adversarial threats.
Förderung
- Ministry of Higher Education (MoHE) Malaysia through the Fundamental Research
Zitieren
APA
Ayub, A., Abidin, Z. Z., & Alhammadi, A. (2026). AdvGAN-RSSI: Untargeted Regression Attacks on Indoor Localization Models. IEEE Sensors Journal, 26(11), 17902-17917. https://doi.org/10.1109/jsen.2026.3686200
Vancouver
Ayub A, Abidin ZZ, Alhammadi A. AdvGAN-RSSI: Untargeted Regression Attacks on Indoor Localization Models. IEEE Sensors Journal. 2026;26(11):17902-17917. doi:10.1109/jsen.2026.3686200
BibTeX
@article{Ayub2026,
author = {Ayub, Ayesha and Abidin, Zuhairiah Zainal and Alhammadi, Abdulraqeb},
title = {AdvGAN-RSSI: Untargeted Regression Attacks on Indoor Localization Models},
journal = {IEEE Sensors Journal},
year = {2026},
volume = {26},
number = {11},
pages = {17902--17917},
doi = {10.1109/jsen.2026.3686200}
}
EUVIMED