Code for paper "Continual Adversarial Defense" (early version).

Get Started

Datasets are CIFAR-10.

Our codebase accesses the datasets from ./data/ and checkpoints from ./net_weights/ by default.

├── ...
├── data
│   
├── net_weights
│   
├── cifar10_online.py
├── ...

All of the adversarial data is generated using torchattacks. Please configure config_cifar10.py first.

Data

Our data is converted to .pt format. You can make adversarial data using make_adv_normal.py.

Pretrained Model

You can download a pretrained clean model from here. And put it in the direction './net_weights/Clean/wrn-28-10-dropout0.3.pth'.

Run

python cifar10_online.py

## Dependencies
python 3.8.8, PyTorch = 1.10.0, cudatoolkit = 11.7, torchattack, torchvision, tqdm, scikit-learn, mmcv, numpy, opencv-python, dlib, Pillow