Imperceptible UAPs for Automatic Modulation Classification Based on Deep Learning

Dongwei Xu, Jiangpeng Li, Zhuangzhi Chen, Qi Xuan, Weiguo Shen, Xiaoniu Yang

Abstract: Automatic Modulation Classification (AMC), which based on deep learning has been extensively researched and implemented in wireless communication systems. Universal adversarial perturbation refers to a single perturbation that can cause most samples to be misclassified by deep learning models. In this brief, we aim to achieve imperceptible universal adversarial attacks on AMC models, and thus an imperceptible universal adversarial perturbations (imperceptible UAPs) framework was proposed. Specifically, loss functions were separately designed for target signals and non-target signals, and then a total loss was calculated. This total loss function can be modified by hyperparameters to achieve class-specific universal adversarial attacks (CUAAs), class-discriminative universal adversarial attacks (CD-UAAs), and class-discriminative target universal adversarial attacks (CD-TUAAs). Meanwhile, wavelet reconstruction was applied to the training data, thus further improving the discriminability of the generated UAPs. After CUAAs were implemented, the class dominant in radio signals was visualized and analyzed by confusion matrices. Furthermore, with the confusion matrices of CUAAs, CD-TUAAs were efficiently implemented. The experiments were conducted on two radio signal datasets and models. In most scenarios, CD-UAAs achieved excellent performance with an average $\Delta A_{c c}$ of 58. 20% and CD-TUAAs achieved an average $\Delta K $ of 73.78%.