Deep learning models have achieved remarkable success in areas such as image recognition and language processing. However, they remain highly sensitive to subtle perturbations in input data. Adversarial examples, which are created by slightly modifying inputs, can change a model’s predictions and raise concerns about its security. To address these issues, research on adversarial defense aims to enhance the robustness and reliability of models under such perturbations. It explores approaches in training, input processing, and model design to help models better identify and resist adversarial influences.