A training framework that makes deep-learning automatic modulation classification (AMC) models robust to unseen black-box adversarial attacks. Meta-training tasks are built by pairing substitute models with attack methods (FGSM, PGD, MIM, C&W, PCA) to perturb the RML2016.10a dataset; MAML-style inner/outer-loop optimization (also Reptile and FOMAML) then learns an initialization that adapts to a new attack from a few shots, or generalizes zero-shot. Compared with conventional adversarial training or transfer baselines, it achieves lower symbol error rates with roughly 15-160x fewer online samples and sub-second online fine-tuning, addressing the offline-training/online-deployment split of practical wireless receivers.
DL-based automatic modulation classification (AMC) models are highly susceptible to adversarial attacks, where even minimal input perturbations can cause severe misclassifications. While adversariall…