MMOF-AGM: Multimodel Adversarial Attack for Security Assessment of Neural Network-Based Radar Signal Recognition
IEEE Transactions on Aerospace and Electronic Systems · 2026 · 62 · 7794-7806
DOI: 10.1109/taes.2026.3674867 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Aerospace and Electronic Systems, 2026. Verfasst von Zhida Bao und 4 weiteren Autor:innen. Die Arbeit nennt 45 Literaturangaben. Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
Neural-network-based radar signal recognition has achieved significant progress in automatic feature extraction and classification accuracy, yet these systems face serious security threats from adversarial attacks that exploit gradient information to generate imperceptible perturbations. However, existing adversarial attacks, typically driven by single-model gradients, suffer from poor transferability due to architectural biases and suboptimal optimization trajectories. Facing this challenge, this article proposes the multimodel output fusion with average gradient mapping (MMOF-AGM), an adversarial attack method that enhances transferability and attack effectiveness through dual-level optimization. The core insight is that architectural bias from single surrogate models and magnitude loss from sign-function quantization compound each other, and addressing either alone yields limited gains. MMOF-AGM jointly resolves both through stochastic multimodel ensemble sampling that reduces directional bias, a continuous magnitude-preserving mapping that replaces the sign function, and neighborhood gradient smoothing that stabilizes both components, producing synergistic effects exceeding the sum of individual contributions. Extensive experiments on simulated radar signal datasets demonstrate that MMOF-AGM consistently achieves superior attack success rates and transferability compared to popular methods across multiple classification models in both glass-box and closed-box scenarios, while maintaining lower perturbation magnitudes and better concealment properties. The approach provides insights for developing more secure automatic radar signal recognition systems and enhancing communication system security in critical electromagnetic spectrum applications.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Bao, Z., Zhang, S., Yang, Y., Yang, S., & Lin, Y. (2026). MMOF-AGM: Multimodel Adversarial Attack for Security Assessment of Neural Network-Based Radar Signal Recognition. IEEE Transactions on Aerospace and Electronic Systems, 62, 7794-7806. https://doi.org/10.1109/taes.2026.3674867
Vancouver
Bao Z, Zhang S, Yang Y, Yang S, Lin Y. MMOF-AGM: Multimodel Adversarial Attack for Security Assessment of Neural Network-Based Radar Signal Recognition. IEEE Transactions on Aerospace and Electronic Systems. 2026;62:7794-7806. doi:10.1109/taes.2026.3674867
BibTeX
@article{Bao2026,
author = {Bao, Zhida and Zhang, Sicheng and Yang, Yandie and Yang, Songlin and Lin, Yun},
title = {MMOF-AGM: Multimodel Adversarial Attack for Security Assessment of Neural Network-Based Radar Signal Recognition},
journal = {IEEE Transactions on Aerospace and Electronic Systems},
year = {2026},
volume = {62},
pages = {7794--7806},
doi = {10.1109/taes.2026.3674867}
}
EUVIMED