EUVIMEDEUVIMED entdecken
Crossref

CMo-IABA: Constrained Multi-Objective Invisible and Adaptive Backdoor Attack for Deep Neural Networks-Based SAR Image Classification

Guo-Qiang Zeng, Hai-Nan Wei, Kang-Di Lu, Guang-Gang Geng

IEEE/CAA Journal of Automatica Sinica · 2026-07 · 13 (7) · 1731-1746

DOI: 10.1109/jas.2025.125888 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE/CAA Journal of Automatica Sinica, 2026-07. Verfasst von Guo-Qiang Zeng und 3 weiteren Autor:innen. Die Arbeit nennt 66 Literaturangaben. Gefördert von: National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Deep neural networks (DNNs) have been widely applied in the field of synthetic aperture radar (SAR) image while they are facing more and more serious threats from a variety of malicious attacks. As one of the malicious attacks with strong destructiveness and stealth, backdoor attacks have severely affected DNNs, but there are no related research studies concerning the backdoor attacks against the DNNs-based SAR image classification models. In this work, we make the first attempt to automatically design a constrained multi-objective invisible and adaptive backdoor attack termed as CMo-IABA for DNNs-based SAR image classification. In the CMo-IABA, we firstly generate an initial trigger-based backdoor attack randomly by a random combination of pixels with random noise conforming to the Gaussian distribution. Then, we design multi-objective functions by considering the trade-off between maximizing the attack success rate and minimizing $L_{2}$ distance-based invisibility. The classification error between the backdoor DNN and the clean model is considered as the constraint to maintain the original performance of the model. To solve the optimization problem, a discrete non-dominated sorting genetic algorithm-II is introduced as the search engine with the developed crossover operation and mutation operation. The superiority of the proposed CMo-IABA to five state-of-the-art backdoor attacks on six different types of DNNs-based SAR image classification models has been demonstrated by the experimental results on Fudan University SAR (FUSAR)-ship and moving and stationary target acquisition and recognition (MSTAR) datasets in terms of attack success rate and $L_{2}$ distance-based invisibility.

Förderung

  • National Natural Science Foundation of China

Zitieren

APA

Zeng, G., Wei, H., Lu, K., & Geng, G. (2026). CMo-IABA: Constrained Multi-Objective Invisible and Adaptive Backdoor Attack for Deep Neural Networks-Based SAR Image Classification. IEEE/CAA Journal of Automatica Sinica, 13(7), 1731-1746. https://doi.org/10.1109/jas.2025.125888

Vancouver

Zeng G, Wei H, Lu K, Geng G. CMo-IABA: Constrained Multi-Objective Invisible and Adaptive Backdoor Attack for Deep Neural Networks-Based SAR Image Classification. IEEE/CAA Journal of Automatica Sinica. 2026;13(7):1731-1746. doi:10.1109/jas.2025.125888

BibTeX
@article{Zeng2026,
  author = {Zeng, Guo-Qiang and Wei, Hai-Nan and Lu, Kang-Di and Geng, Guang-Gang},
  title = {CMo-IABA: Constrained Multi-Objective Invisible and Adaptive Backdoor Attack for Deep Neural Networks-Based SAR Image Classification},
  journal = {IEEE/CAA Journal of Automatica Sinica},
  year = {2026},
  volume = {13},
  number = {7},
  pages = {1731--1746},
  doi = {10.1109/jas.2025.125888}
}
Kostenlos starten →