Apply targeted attacks on black box ViT with the activation attack method. The white box model is ResNet-18.
Based on CVPR 2019 Paper Feature Space Perturbations Yield More Transferable Adversarial Examples.
My additional expreriment proves that the transferablitiy is low between heterogeneous deep models.
All the models are finetuned on CIFAR-10 with the Adam optimizer with lr = 1e-5.
Model Name | CIFAR-10 Test Accuracy | ImageNet-1K Pretrain |
---|---|---|
ResNet-18 | 0.9435 | Yes |
DenseNet-121 | 0.9571 | Yes |
ViT-B/16 | 0.9838 | Yes |
Model | Error | uTr | tSuc | tTr |
---|---|---|---|---|
ResNet-18 | 96.39 | / | 75.48 | / |
DenseNet-121 | 80.16 | 82.27 | 9.39 | 10.65 |
ViT | 22.33 | 22.78 | 2.5 | 4.2 |