This repository is the official implementation of Generative Noise Injector for Model Explanations (GNIME) intoduced in Learning to Generate Inversion-Resistant Model Explanations.

1. Create and activate virtual environment via conda.

conda create -n gnime python=3.8
conda activate gnime

2. Install cuda and cudnn. (We use CUDA 11.3 in our experiment.)

conda install cudatoolkit=11.3 cudnn=8.2

3. Install requirements

pip install -r requirements.txt

4. Download CelebA dataset from here.
   • Downloading MNIST and CIFAR-10 datasets are integrated into Python code, so no need to bother.
   • We provide the cropped versions of 1000 identities via Zenodo. Move and extract celeba_cropped_1000.zip inside datasets/. ()

datasets
└── celeba_cropped_1000
 	├── 5702
 	|	├── 000707.jpg
	|	├── 002051.jpg
	|	└── ...
	├── 5703
	|	├── 001714.jpg
	│	├── 004273.jpg
	|	└── ...
	└── ...

1. (optional) Download pretrained models from here.
   • To avoid the tiresome process of training the victim/attacker DNN models, download the pretrained models uploaded to Zenodo and move them to models/. The provided source code will automatically detect and use existing models in this directory.

models
├── victim
|	├── celeba
|	├── cifar10
|	└── mnist
├── eval
|	├── celeba
|	├── cifar10
|	└── mnist
├── inv
|	├── celeba_predmi
|	├── celeba_expmi_lrp
|	└── ...
└── def
	├── celeba_lrp
	├── celeba_grad
	└── ...

2. Pre-generate model explanations into aux/.
   • Instead of generating model explanations on demand, we prepare them prehand and access them by their index.
   • We use the hickle file format to store and load data.

aux
├── celeba_grad
├── celeba_lrp
└── ...

• command for above steps (automatically skips training pretrained models):

python prepare.py --target <target_type=['celeba','cifar10','mnist']> --xai <xai_type=['grad','gradcam','lrp']> --gpu <gpu_#>

• If you downloaded pretrained models in Preparation (1), you can skip training process and proceed to Evaluation.

To train the GNIME defense model, run this command:

python train.py --target <target_type=['celeba','cifar10','mnist']> --xai <xai_type=['grad','gradcam','lrp']> --gpu <gpu_#>

• visualized inversion results will be saved to figures/ during the evaluation

To evaluate GNIME defense, run:

python eval.py --target <target_type=['celeba','cifar10','mnist']> --xai <xai_type=['grad','gradcam','lrp']> --gpu <gpu_#>

For comparison, this will evaluate not only ExpMI against GNIME, but also cases for ExpMI against no defense, ExpMI against baseline defense, and PredMI.

Note that evaluation of each scenario requires training a dedicated ExpMI model, hence may take some time. (≤ 1h/scenario)

From our 5 experiements, GNIME defense achieves the following performance: (figure from paper)

@article{jeong2022learning,
  title={Learning to Generate Inversion-Resistant Model Explanations},
  author={Hoyong Jeong, Suyoung Lee, Sung Ju Hwang, and Sooel Son},
  journal={Advances in Neural Information Processing Systems},
  volume={?},
  pages={?--?},
  year={2022}
}