Hi, this is the code for our KDD 2022 paper: Bilateral Dependency Optimization: Defending Against Model-inversion Attacks. Overview of MID framework vs. bilateral dependency optimization (BiDO) framework. BiDO forces DNNs to learn robust latent representations by minimizing $𝑑(𝑋,𝑍_j)$ to limit redundant information propagated from the inputs to the latent representations while maximizing $𝑑(𝑍_j,𝑌)$ to keep the latent representations informative enough of the label.

This code has been tested on Ubuntu 16.04/18.04, with Python 3.7, Pytorch 1.7 and CUDA 10.2/11.0

Download relevent datasets: CelebA, MNIST.

• CelebA
• MNIST

  python prepare_dataset.py
  

The directory of datasets are orgnized as follows:

./attack_dataset
├── MNIST 
│   ├── *.txt 
│   └── Img
│       └── *.png
└── CelebA                            
    ├── *.txt 
    └── Img
        └── *.png

You can also skip to the next section for defending against MI attack with well-trained defense models.

• For GMI and KED-MI

  # dataset:celeba, mnist, cifar; 
  # measure:COCO, HSIC; 
  # balancing hyper-parameters: tune them in train_HSIC.py
  python train_HSIC.py --measure=HSIC --dataset=celeba
  

  For KED-MI, if you trained a defense model yourself, you have to train a attack model (generative model) specific to this defense model additionally.

  # put you hyper-parameters in k+1_gan_HSIC.py first
  python k+1_gan_HSIC --dataset=celeba --defense=HSIC
  

• For VMI
  Please refer to this section (Defending against MI attacks - VMI - train with BiDO).

Here we only provide the weights file of the well-trained defense models that achieve the best trade-off between model robustness and utility, which are highlighted in the experimental results.

• GMI

  • Weights file (defense model / eval model / GAN) :

    • Place pretrained VGG16 in BiDO/target_model/
    • Place defense model in BiDO/target_model/celeba/HSIC/
    • Place evaluation classifer in GMI/eval_model/
    • Place GAN in GMI/result/models_celeba_gan/

  • Launch attack

    # balancing hyper-parameters: (0.05, 0.5)
    python attack.py --dataset=celeba --defense=HSIC
    

  • Calculate FID

    # sample real images from training set
    cd attack_res/celeba/pytorch-fid && python private_domain.py 
    # calculate FID between fake and real images
    python fid_score.py ../celeba/trainset/ ../celeba/HSIC/all/
    

• KED-MI

  • Weights file (defense model / eval model / GAN) :
    • Place defense model in BiDO/target_model/mnist/COCO/
    • Place evaluation classifer in DMI/eval_model/
    • Place improved GAN for celeba in DMI/improvedGAN/celeba/HSIC/
    • Place improved GAN for mnist in DMI/improvedGAN/mnist/COCO/
  • Launch attack

    # balancing hyper-parameters: (0.05, 0.5)
    python recovery.py --dataset=celeba --defense=HSIC
    # balancing hyper-parameters: (1, 50)
    python recovery.py --dataset=mnist --defense=COCO
    

  • Calculate FID

    # celeba
    cd attack_res/celeba/pytorch-fid && python private_domain.py 
    python fid_score.py ../celeba/trainset/ ../celeba/HSIC/all/ --dataset=celeba
    # mnist
    cd attack_res/mnist/pytorch-fid && python private_domain.py 
    python fid_score.py ../mnist/trainset/ ../mnist/COCO/all/ --dataset=mnist
    

• VMI
  To run this code, you need ~38G of memory for data loading, the attacking of 20 identities takes ~20 hours on a TiTAN-V GPU (12G).

  • Data (CelebA)

    # create a link to CelebA
    cd VMI/data && ln -s ../../attack_data/CelebA/Img img_align_celeba
    python celeba.py
    

  • Weights file (defense model / eval model / GAN) :
    • Place defense model in VMI/clf_results/celeba/hsic_0.1&2/
    • Place ir_se50.pth in VMI/3rd_party/InsightFace_Pytorch/work_space/save/; Place evaluation classifer in VMI/pretrained/eval_clf/celeba/
    • Place StyleGAN in VMI/pretrained/stylegan/neurips2021-celeba-stylegan/
  • Launch attack

    # balancing hyper-parameters: (0.1, 2)
    cd VMI
    # x.sh  (1st) path/to/attack_results (2nd) config_file  (3rd) batch_size
    ./run_scripts/neurips2021-celeba-stylegan-flow.sh  'hsic_0.1&2'  'hsic_0.1&2.yml' 32
    

  • Train with BiDO

    python classify_mnist.py --epochs=100 --dataset=celeba --output_dir=./clf_results/celeba/hsic_0.1&2 --model=ResNetClsH --measure=hsic --a1=0.1 --a2=2
    

If you find this code helpful in your research, please consider citing

@inproceedings{peng2022BiDO,
title={Bilateral Dependency Optimization: Defending Against Model-inversion Attacks},
author={Peng, Xiong and Liu, Feng and Zhang, Jingfeng and Lan, Long and Ye, Junjie and Liu, Tongliang and Han, Bo},
booktitle={KDD},
year={2022}
}

Some of our implementations rely on other repos. We want to thank the authors (MID, GMI, KED-MI, VMI) for making their code publicly available.😄