Intra-Class Universal Adversarial Attacks on Deep Learning-Based Modulation Classifiers

Most existing adversarial attack methods generally rely on ideal assumptions, which is unreasonable for practical applications. In this letter, a practical threat model which utilizes adversarial attacks for anti-eavesdropping is proposed and a physical intra-class universal adversarial perturbation (IC-UAP) crafting method against DL-based wireless signal classifiers is then presented. First, an IC-UAP algorithm is proposed based on the threat model to craft a stronger UAP attack against the samples in a given class from a batch of samples in the class. Then, we develop a physical attack algorithm based on the IC-UAP method, in which perturbations are optimized under random shifting to enhance the robustness of IC-UAPs against the unsynchronization between adversarial attacks and attacked signals. Finally, the numerical results corroborate the effectiveness of the proposed approach based on the benchmark dataset.

Paper

Full text

PDF

Intra-Class Universal Adversarial Attacks on Deep Learning-Based Modulation Classifiers

Semantic Scholar · Computer Science · 2023

Abstract

Most existing adversarial attack methods generally rely on ideal assumptions, which is unreasonable for practical applications. In this letter, a practical threat model which utilizes adversarial attacks for anti-eavesdropping is proposed and a physical intra-class universal adversarial perturbation (IC-UAP) crafting method against DL-based wireless signal classifiers is then presented. First, an IC-UAP algorithm is proposed based on the threat model to craft a stronger UAP attack against the samples in a given class from a batch of samples in the class. Then, we develop a physical attack algorithm based on the IC-UAP method, in which perturbations are optimized under random shifting to enhance the robustness of IC-UAPs against the unsynchronization between adversarial attacks and attacked signals. Finally, the numerical results corroborate the effectiveness of the proposed approach based on the benchmark dataset.

Similar papers

© 2026 NYSGPT2525 LLC