Computer vision models based on deep learning technology are susceptible to adversarial examples. By including some subtle perturbations to the examples, the attacker can errors in the deep learning model, which will lead to serious consequences. To better defend against this attack, one of the methods is to detect and cull the adversarial examples. Compared with the original local intrinsic dimensionality method, this paper proposes an optimized local intrinsic dimensionality method to characterize the dimensional attributes of adversarial examples. This method not only detects the distance distribution of a sample to its neighbors but also evaluates the sensitivity of an example to perturbation to decide if it’s an adversarial example. Four different attack strategies were used to consider the protection effect of the proposed method. And horizontal comparison of three classical defense methods, the approach used this paper compared with the three defense methods, the detection effect is more obvious and the defense effect is better. The improved detection method enhances the model’s ability to identify and exclude adversarial examples, boosting the recognition accuracy to approximately 95%, even while maintaining a false deletion rate of about 10% (that is, about 10% of the clean examples are classified as adversarial examples).