Domain Adaptive Adversarial Defense Method for Automatic Modulation Classification

Deep learning-based automatic modulation classification (AMC) models are vulnerable to adversarial example attacks, posing severe adversarial security threats in practical scenarios characterized by dynamic channel conditions and limited availability of signal labels. To address this issue, a multi-domain distribution alignment based domain-adaptive adversarial defense method was proposed. Firstly, a phase rotation data augmentation strategy was used to enrich the discriminative and domain-invariant features learned by the model. Secondly, a dual-discriminator architecture was constructed to reduce the feature distribution discrepancy between the original and adversarial signals in the target domain and those in the source domain. Thirdly, contrastive learning constraints were introduced in conjunction with high-confidence pseudo-labels, leveraging source domain class anchor to enhance intra-class compactness and inter-class separability in the target domain. Finally, a consistency constraint was employed to reduce the output discrepancy between original and adversarial signals in the target domain. Experimental results on both public and simulated datasets demonstrate that, compared with existing methods, the proposed method exhibits superior domain adaptability and adversarial robustness under various adversarial attacks, effectively enhancing the reliability and security of AMC systems in complex electromagnetic environments.