Computer vision systems have demonstrated excellent performance in tasks such as image classification, target detection, and face recognition. However, the deep neural networks that drive these systems are highly susceptible to interference from tiny, carefully designed perturbations, namely adversarial samples. These samples are virtually imperceptible to the human eye yet can lead to serious misjudgments in the model, revealing the fragility of its decision boundary and its fundamental difference from human perception. Conversely, defense strategies aim to improve robustness of models against adversarial attacks and ensure accurate outputs in the face of malicious attacks.