<p>Deep neural network (DNN) based automatic modulation classification (AMC) has become increasingly important for wireless communication systems, yet these networks exhibit inherent vulnerabilities to adversarial perturbations that can severely degrade classification performance. Adversarial example detection offers an effective defense strategy by identifying malicious inputs before they reach the classifier, protecting AMC system security without architectural modifications. This paper proposes a multi-feature vector construction-based detection method (MFVC-DM) that integrates complementary statistical representations, including kernel density estimation, local intrinsic dimensionality, and K-means distance metrics. These features are extracted from hidden layer activations and cascaded into a unified high-dimensional feature space through dimension normalization and orthogonality. Experimental validation on the RML2016.10A dataset demonstrates that MFVC-DM consistently outperforms single-feature detection approaches, achieving up to 99.38% detection accuracy against gradient-based attacks and 93.43% against optimization-based perturbations, thereby providing an effective security layer for AMC systems.</p>