Frequency-Constrained Iterative Adversarial Attacks for Automatic Modulation Classification

Abstract

Although adversarial attacks present a significant threat to intelligent models based on deep learning (DL) for automatic modulation classification (AMC). However, the existing works for electromagnetic signal adversarial attacks introduce high frequency components in the frequency domain, which causes spectral mismatch and glitch problems, degrading the attack success rate after transmission through a band-limited channel. In this letter, we propose a frequency-constrained iterative adversarial attacks (FCIAA) algorithm which can suppress the high frequency components and optimize adversarial perturbations during the iterative process to alleviate such problems. The experiments using qualitative and quantitative indicators demonstrate that the proposed algorithm can effectively constrain out-of-band perturbation energy, which improves both the time and frequency domain concealment quality of the adversarial signals and enhances adversarial attacks effect.