CPCAA: Cross-Domain Power Constrained Adversarial Attacks Against Automatic Modulation Classification

automatic modulation classification (AMC) plays a vital role in enhancing spectrum utilization efficiency in Cognitive Internet of Things (CIoT) scenarios. However, it introduces security vulnerabilities by enabling eavesdroppers to accurately identify modulation schemes of intercepted signals through deep neural network (DNN)-based classifiers. Considering that DNN performance is highly susceptible to adversarial examples, this work aims to design adversarial perturbations for signal transmitters to effectively prevent eavesdroppers from intercepting and identifying target user signals. Unlike image classification tasks, existing methods for generating adversarial perturbations in wireless communications face three critical limitations: 1) instability in perturbation power control; 2) vulnerability to signal filtering; and 3) high perceptibility of perturbations. To address these challenges, we reformulate the adversarial example generation problem with specific target perturbation power constraints. By jointly leveraging feature vectors from the final DNN layer and frequency-domain characteristics of transmitted signals, we propose a cross-domain power constrained adversarial attack (CPCAA) method that further optimizes perturbation generation. Experimental results demonstrate that the proposed method generates highly imperceptible adversarial examples with specified energy constraints while achieving significant attack performance, even when eavesdroppers deploy various types of filters.