31 January 2024 in IEEE Transactions on Cognitive Communications and Networking
Deep neural networks provide intelligent solutions for Automatic Modulation Classification (AMC) tasks in the field of communication. However, their susceptibility to adversarial examples due to the interpretability problem presents a challenge as it leads to anomalous decisions. Emerging studies suggest that the high-frequency constituents within signals constitute a fundamental source of adversarial vulnerability. To address this issue, this paper introduces a Homomorphic Filtering Adversarial Defense (HFAD) algorithm that aims to effectively defend against adversarial examples by applying frequency domain filtering on the signal. This approach enhances the security and reliability of the AMC model by attenuating high-frequency components of the signal through homomorphic filtering, thereby reducing errors caused by adversarial perturbations on model outputs. The robustness of the AMC model is further enhanced through the integration of HFAD with data augmentation strategies. Experimental results demonstrate that the proposed defense algorithm not only maintains high signal recognition accuracy but also preserves communication signal transmission quality. Moreover, HFAD effectively withstands a wide range of white-box adversarial attacks and demonstrates resilience against black-box adversarial attacks, thereby enhancing the robustness of the AMC model against adversarial examples and exhibiting strong transfer performance.
26 March 2024 in IEEE Transactions on Cognitive Communications and Networking
With the improvement of basic designs and the evolution of key algorithms, artificial intelligence (AI) has been considered by both industry and academia as the most promising solution for many electromagnetic space problems, such as automatic modulation classification (AMC). However, the fact that AI-based AMC models are vulnerable to adversarial examples mystifies the optimism. Adversarial attacks help researchers to reexamine AI-based AMC models and promote safe applications. In this paper, we study the frequency leakage and glitch problems caused by high frequency components in the adversarial perturbations of existing attack algorithms. We propose a Spectrum-focused Frequency Adversarial Attack (SFAA) algorithm to suppress the high frequency components to alleviate such problems. Next, we leverage meta-learning to improve the transferability of the proposed algorithm for black-box attacks. We also train a Channel-robust Class-universal Spectrum-focused Frequency Adversarial Attack (CrCu-SFAA) generative model using the generative adversarial network framework. Finally, extensive experiments using qualitative and quantitative indicators demonstrate that the proposed algorithm achieves an improved attack performance, and our proposed approach of reducing out-of-band high frequency components of the adversarial perturbations improves the concealment and adversarial signal quality.