2022-YeC2FMI

Implementation of Paper "Z. Ye, W, Luo and M. L. Naseem, et al., C2FMI: Corse-to-Fine Black-box Model Inversion Attack", TDSC, 2023.

Requirements

I have tested on:

• PyTorch 1.13.0
• CUDA 11.0

The Simplest Implementation

1. Attack Anti-Defense

If you want to use our trained models, including styleGAN, target model, embedding model and inverse model (refer to Fig.4 in our paper):

1. download checkpoint.zip from https://huggingface.co/MiLab-HITSZ/C2FMI/tree/main.
2. download trained_models.zip from https://huggingface.co/MiLab-HITSZ/C2FMI/tree/main.
3. unzip and put these 2 folders in your project directory.
4. running with command:

python main_attack_antidefense.py

• note that you should create directory gen_figures/DE_facescrub_mobile_M100_counter/ in your project before running since our code does not automatically generate it.
• changing the variables init_label and fina_label in main_attack_antidefense.py, attack will start at init_label and end at fina_label.

Or you can train your own models and change the models in script main_attack_antidefense.py.

2. Attack without Defense

If you want to use our trained models, including styleGAN, target model, embedding model and inverse model (refer to Fig.4 in our paper):

1. download checkpoint2.zip from https://huggingface.co/MiLab-HITSZ/C2FMI/tree/main.
2. download trained_models.zip from https://huggingface.co/MiLab-HITSZ/C2FMI/tree/main.
3. unzip and put these 2 folders in your project directory.
4. running with command:

python main_attack.py

• note that you should create directory gen_figures/DE_facescrub/ in your project before running since our code does not automatically generate it.
• changing the variables init_label and fina_label in main_attack.py, attack will start at init_label and end at fina_label.

Or you can train your own models and change the models in script main_attack.py.

How to evaluate

If you have followed the simplest implementation, you can run the following command to evaluate the attack accuracy:

python eva_accuracy.py

• please modify the annotated variables in eva_accuracy.py if using your customized dataset and models.

If you want to evaluate the KNN-Dist, before run eva_knn_dist.py, you should note that:

-the settings of variables in eva_knn_dist.py is same as in eva_accuracy.py, except for the variable train_txt, which you can run annotation_face_train.py to generate it (it is a .txt file containing the path of all training data). Note that you should modify the datasets_path in annotation_face_train.py.