RFCOA

The source code of AAAI 2025 paper "Breaking Barriers in Physical-World Adversarial Examples: Improving Robustness and Transferability via Robust Feature".

Setup Environment

Our project is based on Horizon2333/imagenet-autoencoder , the setup environment and configurations can be found in Autoencoder.md.

Data Preparation

Download the dataset ImageNet ILSVRC 2012, the path is as follows:

Source Code
├── data
│   ├── ILSVRC2012_img_train.tar
    ├── ILSVRC2012_img_test.tar
    ├── ILSVRC2012_img_val.tar
    ├── ILSVRC2012_devkit_t12.tar.gz

Get the autoencoder

You can train an autoencoder follow the guidance in Autoencoder.md, or download a pertrained autoencoder from the link.

Extract the Robust Features of Target Class

According to our method, you should extract the robust features of the target class first.

python extract.py

Generate Adversarial Examples

After extract the robust features, you can generate the adversarial examples by running:

python attack.py --target xxx