CCIFE: Channel-Resilient Ensemble Adversarial Attack Against DNN-Based Modulation Classifiers

The study of adversarial attacks against deep neural network (DNN)-based automatic modulation classification (AMC) systems has become increasingly important for understanding system vulnerabilities, yet existing attack methods face critical limitations in realistic radio frequency environments. In wireless communications, current approaches predominantly suffer from model-specific overfitting that severely limits cross-architecture transferability, while their perturbations rapidly degrade when subjected to realistic channel impairments including fading and frequency offsets. To address these critical challenges, this paper presents the channel-resilient common important features ensemble (CCIFE) attack framework. CCIFE integrates a dual-criterion feature attribution mechanism that identifies transferable vulnerabilities by quantifying both cross-model consistency and importance of sample features. This mechanism generates a continuous weighting mask to guide perturbation optimization toward the most transferable features. Additionally, an iterative channel randomization strategy incorporating Rayleigh fading and carrier frequency offset effects enables perturbations to maintain effectiveness through wireless propagation. Experiments on public datasets demonstrate CCIFE’s superior attack performance over existing methods across diverse neural network architectures and channel conditions. Attribution analysis confirms that CCIFE effectively identifies common important features with exceptional cross-model consistency, providing insights for understanding adversarial vulnerabilities in wireless modulation classification systems.