Stable-Unlearnable-Example

This is the official implementation of AAAI'24 "Stable Unlearnable Example: Enhancing the stableness of Unlearnable Examples via Stable Error-Minimizing Noise".

Requirements

• Python 3.8
• PyTorch 1.8.1
• Torchvision 0.9.1
• OpenCV 4.5.5

Install dependencies using pip

pip install -r requirements.txt

Install dependencies using Anaconda

conda install pytorch=1.8.1 torchvision=0.9.1 cudatoolkit=10.2 -c pytorch
conda install -c conda-forge opencv=4.5.5

Quick Start

We give an example of creating stable unlearnable examples from CIFAR-10 dataset. More experiment examples can be found in ./scripts.

Generate stable error-minimizing noise for CIFAR-10 dataset

python generate_stable_em.py \
    --arch resnet18 \
    --dataset cifar10 \
    --train-steps 5000 \
    --batch-size 128 \
    --optim sgd \
    --lr 0.1 \
    --lr-decay-rate 0.1 \
    --lr-decay-freq 2000 \
    --weight-decay 5e-4 \
    --momentum 0.9 \
    --pgd-radius 8 \
    --pgd-steps 10 \
    --pgd-step-size 1.6 \
    --pgd-random-start \
    --atk-pgd-radius 4 \
    --atk-pgd-steps 10 \
    --atk-pgd-step-size 0.8 \
    --atk-pgd-random-start \
    --samp-num 5 \
    --report-freq 1000 \
    --save-freq 1000 \
    --data-dir ./data \
    --save-dir ./exp_data/cifar10/noise/sem8-4 \
    --save-name sem

Perform adversarial training on stable unlearnable examples

python train.py \
    --arch resnet18 \
    --dataset cifar10 \
    --train-steps 40000 \
    --batch-size 128 \
    --optim sgd \
    --lr 0.1 \
    --lr-decay-rate 0.1 \
    --lr-decay-freq 16000 \
    --weight-decay 5e-4 \
    --momentum 0.9 \
    --pgd-radius 4 \
    --pgd-steps 10 \
    --pgd-step-size 0.8 \
    --pgd-random-start \
    --report-freq 1000 \
    --save-freq 100000 \
    --noise-path ./exp_data/cifar10/noise/sem8-4/sem-fin-def-noise.pkl \
    --data-dir ./data \
    --save-dir ./exp_data/cifar10/train/sem8-4/r4 \
    --save-name train

Visualize the noise distribution of SEM and REM compared to the noise during evaluation

Please refer to ./notebooks/analyze_noise_distribution.ipynb for more details. TSNE and UMAP results are provided in at ./notebooks/vis-output.

Citation

@article{liu2023stable,
  title={Stable Unlearnable Example: Enhancing the Robustness of Unlearnable Examples via Stable Error-Minimizing Noise},
  author={Liu, Yixin and Xu, Kaidi and Chen, Xun and Sun, Lichao},
  journal={arXiv preprint arXiv:2311.13091},
  year={2023}
}

Acknowledgment

• Availability Attacks Create Shortcuts: https://github.com/dayu11/Availability-Attacks-Create-Shortcuts
• Robust Unlearnable Example: https://github.com/fshp971/stable-unlearnable-examples
• Unlearnable examples: https://github.com/HanxunH/Unlearnable-Examples
• Adversarial poisons: https://github.com/lhfowl/adversarial_poisons
• Neural tangent generalization attacks: https://github.com/lionelmessi6410/ntga