Speaker
Description
Classical white-box adversarial attacks were designed and evaluated mostly on convolutional networks from 2015 to 2018. Modern architectures such as Vision Transformers and ConvNeXt have different inductive biases, and it is unclear whether the old attacks work equally well against them. We present a reproducible benchmark that runs FGSM, BIM, PGD and AutoAttack against four ImageNet-pretrained architectures (ResNet-50, EfficientNet-B0, ViT-B/16, ConvNeXt-Tiny) on Imagenette and report robust accuracy over six L-infinity budgets from 1/255 to 32/255, on 200 validation images for the gradient-based attacks and 500 for AutoAttack. Iterative attacks break every model: at 4/255, BIM and PGD leave at most 2.5% robust accuracy, and AutoAttack reaches 0% on all four architectures from 2/255 upward. Differences between architectures appear only at the smallest budget, where AutoAttack at 1/255 leaves ViT-B/16 with 9.8% robust accuracy while the other three stay below 1.5%. The single-step FGSM is the exception: on ResNet-50 and ConvNeXt its robust accuracy stays around 80% regardless of budget. We show that this is in part an evaluation artifact of restricting the 1000-way classifier to the 10 Imagenette classes. On the full 1000-class output, the same FGSM images at 32/255 score 24.0% on ResNet-50 instead of 80.5%, and 43.5% on ConvNeXt instead of 85.5%. This still leaves FGSM far weaker than iterative attacks with the same budget. Our from-scratch implementations of FGSM, BIM and PGD are verified against torchattacks, and all code, results and figures are released.