MetaAdv

Platform

Python: 3.7
PyTorch: 1.5.0

Dataset

We use the benchmark dataset MiniImageNet, which can be download here and here. CIFARFS and Omniglot can be found in the package torchmeta here

Model

We use a four-layer-conv NN model

Standard MAML training

Run MAML_TrainStd.ipynb, associate files include MAMLMeta.py, attack.py, learner.py

Attack power level has to be changed in MAMLMeta.py
The device in MAML_TrainStd.ipynb and attack.py should set to be the same. (same in the following adversarial training)

MAML + FGSM-RS (random start)

Run trainfgsmrs.ipynb, associate files include metafgsm.py, attack.py, learner.py. To incorporate adversarial training in the inner-loop, please replace metafgsm.py with metafgsminout.py

MAML + TRADES-RS

Run train_trade.ipynb, associate files include MetaFT.py, LoadUnlableData.py. The unlabled data can be downloaded from here

Standard training + few-short fine-tuning (Meta-tesing)

Run StandardTransNew.ipynb, associate files include LoadDataST.py, StandardTrans.py. StandardTransAdv.ipynb contains adversarial training in the model training process.

Unlabeled data selection

Run figureselection.ipynb, associate files include

Visualization

Run robust_vis_neuron.ipynb, associate files include Visualization.py, vis_tool.py, MODELMETA.py.

By maximizing the output of a nueron with a perturbation in th input, the feature is shown in the input under a robust model, while "random noise" is shown in the input under a standard MAML model.
The fine-tuned model has the similar feature to the original model in the same neuron. This suggests that the robustness is kept in the fine-tuned model even without adding the adversarial training in the fine-tuning.

CIFAR-FS and Omniglot

Run .ipynb files in the two folders "CIFARFS" and "Omniglot"

Refer to this Rep.

If you use this code, please cite the following reference

@article{wangfast,
  title={ON FAST ADVERSARIAL ROBUSTNESS ADAPTATION IN MODEL-AGNOSTIC META-LEARNING},
  author={Wang, Ren and Xu, Kaidi and Liu, Sijia and Chen, Pin-Yu and Weng, Tsui-Wei and Gan, Chuang and Wang, Meng},
  booktitle={International Conference on Learning Representations (ICLR)},  
  pages={},  
  year={2021}  
}

Kaysome / MetaAdv