/poisoning_membership

Primary LanguagePython

README

This folder contains the implementation of Amplifying Membership Exposure via Data Poisoning.

Requirements

To install requirements:

pip install -r requirements.txt

Dataset

  1. MNIST (part of TensorFlow)
  2. CIFAR-10 (part of TensorFlow)
  3. STL-10. Please download the dataset from the official link and uncompress the downloaded file (will get the stl10_binary folder). Then move the folder stl10_binary into the dataset.
  4. CelebA. Please download img_align_celeba and img_align_celeba.csv from this link. Then, create a subfolder celeba under dataset, and move the file list_attr_celeba.csv and the folder img_align_celeba into celeba.
  5. PatchCamelyon. Please download camelyonpatch_level_2_split_test_x.h5 and camelyonpatch_level_2_split_test_y.h5 from this link. Then, create a subfolder pathcamelyon under dataset, and move the two files into patchcamelyon.

The structure of the dataset folder should be

dataset
+-- stl10_binary
|   +-- train_X.bin
|   +-- train_Y.bin
|   +-- test_X.bin
|   +-- test_Y.bin
|   ...
+-- celeba
|   +-- img_align_celeba
|   +-- list_attr_celeba.csv
+-- patchcamelyon
|   +-- camelyonpatch_level_2_split_test_x.h5
|   +-- camelyonpatch_level_2_split_test_y.h5

Basic Usage

File poisoning_attack.py allows to run our poisoning attacks.

Example:

python poisoning_attack.py --target_class 0 \
                           --dataset cifar10 \
                           --encoder xception \
                           --seed_amount 1000 \
                           --attack_type clean_label \
                           --device_no 0

Example Attack

We provide example attacks in attack_example.sh. Directly run:

./attack_example.sh

Evaluation

To evaluate our poisoning attacks, first run:

./poisoning_models.sh

to generate poisoning datasets and poisoned models.

Then, run:

./evaluate_attack.sh

to get the evaluation results.

Citation

@inproceedings{CSSWZ22,
author = {Yufei Chen and Chao Shen and Yun Shen and Cong Wang and Yang Zhang},
title = {{Amplifying Membership Exposure via Data Poisoning}},
booktitle = {{Annual Conference on Neural Information Processing Systems (NeurIPS)}},
publisher = {NeurIPS},
year = {2022}
}