Adversarial Example Detection Method Based on Inter-layer Activation Evolution Consistency

Automatic modulation classification, a core technique for electromagnetic spectrum sensing, serves as a key sensing task in applications such as spectrum monitoring, yet its adversarial vulnerability poses a severe security threat. To address the limitations that existing adversarial example detection methods only analyze static features while ignoring the dynamic cumulative effects of adversarial perturbations during forward propagation, an adversarial example detection method based on inter-layer activation evolution consistency was proposed. Activation representations from multiple key network layers were extracted, and autoencoders were trained to learn their activation manifolds. Distance matrices between samples and class-average activations were constructed in the reconstruction space, and adversarial examples were detected based on their inter-layer evolution consistency. Experimental results show that an average AUC exceeding 90% was achieved across two datasets, five typical attacks, and multiple signal-to-noise ratios, effectively enhancing the security of deep-learning-based spectrum sensing against adversarial attacks.