AdvImageNet-1K

1,000 images from the ImageNet validation set were selected as the dataset for adversarial attacks. These images cover a diverse set of categories and were uniformly preprocessed to match the input size and normalization required by the models. This subset was used to generate and evaluate adversarial examples, testing the models\u2019 classification robustness under perturbed inputs while also assessing their performance on standard images. By selecting high-quality images with accurate labels, the experiments can effectively evaluate the effectiveness of adversarial attack methods using a limited sample.

本次实验所用数据集从ImageNet验证集（ImageNet validation set）中选取1000张图像，作为对抗攻击（adversarial attack）任务的专用数据集。该批图像覆盖多样化的类别范畴，且经过统一预处理，以匹配各目标模型所需的输入尺寸与归一化要求。此子集被用于生成与评估对抗样本（adversarial example），既可测试模型在受扰动输入下的分类鲁棒性（classification robustness），同时也能评估模型在标准图像上的性能表现。通过选取标注准确的高质量图像，本实验可依托有限样本量，有效评估各类对抗攻击方法的实际效果。