The electromagnetic spectrum is a key national strategic resource, and intelligent electromagnetic spectrum monitoring is largely dependent on reliable automatic modulation classification. Although deep neural networks excel in automatic modulation classification tasks, their susceptibility to adversarial examples brings systemic risks. Existing defense methods show limitations in handling electromagnetic signals: feature reconstruction-based approaches incur high computational costs, while robust training degrades performance on benign examples. To alleviate the above issues, this paper proposes Model Mutation-based Adversarial Example Detection method (MMAED). This method uses sensitivity to changes in the boundaries of decision models to detect adversarial examples. Specifically, the innovation involves four lightweight mutation operators that create diverse mutated models. Integrated with layer freezing and adaptive threshold selection, our method distinguishes adversarial examples from normal ones through label variation analysis. Evaluations of the RML2016.10a dataset show the superior detection accuracy of MMAED against multiple attacks compared to feature-based methods. Our proposed approach effectively identifies adversarial examples in modulation recognition systems by analyzing their distinctive responses across mutated models, offering practical defense against adversarial threats.