This project is a minimized runnable project cut from trojanzoo, which contains more datasets, models, attacks and defenses. This repo will not be maintained.
This is a minimum code implementation of our USENIX'22 paper On the Security Risks of AutoML
.
The artifact discovers the vulnerability gap between manual models and automl models against various kinds of attacks (adversarial, poison, backdoor, extraction and membership) in image classification domain. It implements all datasets, models, and attacks used in our paper.
We expect the artifact could support the paper's claim that automl models are more vulnerable than manual models against various kinds of attacks, which could be explained by their small gradient variance.
- Binary: on pypi with any platform.
- Model: Our pretrained models are available on Google Drive (link). Follow the model path style
{model_dir}/image/{dataset}/{model}.pth
to place them in correct location. - Data set: CIFAR10, CIFAR100 and ImageNet32.
Use--download
flag to download them automatically at first running.
ImageNet32 requires manual set-up at their website due to legality. - Run-time environment:
At any platform (Windows and Ubuntu tested).
Pytorch
andtorchvision
required. (CUDA 11.3 recommended)
adversarial-robustness-toolbox
required for extraction attack and membership attack. - Hardware: GPU with CUDA support is recommended.
- Execution: Model training and backdoor attack would be time-consuming. It would cost more than half day on a Nvidia Quodro RTX6000.
- Metrics: Model accuracy, attack success rate, clean accuracy drop and cross entropy.
- Output: console output and saved model files (.pth).
- Experiments: OS scripts. Recommend to run scripts 3-5 times to reduce the randomness of experiments.
- How much disk space is required (approximately):
less than 5GB. - How much time is needed to prepare workflow (approximately): within 1 hour.
- How much time is needed to complete experiments (approximately): 3-4 days.
- Publicly available: on GitHub.
- Code licenses: GPL-3.
- Archived: GitHub commit ade119d3c9aa1e851eba7db35f2de3c99eb0bf33.
- GitHub
pip install -e .
- PYPI
pip install autovul
- Docker Hub
docker pull local0state/autovul
- GitHub Packages
docker pull ghcr.io/ain-soph/autovul
Recommend to use GPU with CUDA 11.3 and CUDNN 8.0.
Less than 5GB disk space is needed.
You need to install python==3.9, pytorch==1.10.x, torchvision==0.11.x
manually.
ART (IBM) is required for extraction attack and membership attack.
pip install adversarial-robustness-toolbox
We use CIFAR10, CIFAR100 and ImageNet32 datasets.
Use --download
flag to download them automatically at first running.
ImageNet32 requires manual set-up at their website due to legality.
Our pretrained models are available on Google Drive (link). Follow the model path style {model_dir}/image/{dataset}/{model}.pth
to place them in correct location.
- GitHub
- PYPI
pip install autovul
- Docker Hub
- GitHub Packages
You can set the config files to customize data storage location and many other default settings. View /configs_example
as an example config setting.
We support 3 configs (priority ascend):
- package:
(DO NOT MODIFY)
autovul/base/configs/*.yml
autovul/vision/configs/*.yml
- user:
~/.autovul/configs/base/*.yml
~/.autovul/configs/vision/*.yml
- workspace:
./configs/base/*.yml
./configs/vision/*.yml
Check the bash files under /bash
to reproduce our paper results.
You need to first run /bash/train.sh
to get pretrained models.
If you run it for the first time, please run with --download
flag to download the dataset:
bash ./bash/train.sh "--download"
It takes a relatively long time to train all models, here we provide our pretrained models on Google Drive (link). Follow the model path style {model_dir}/image/{dataset}/{model}.pth
to place them in correct location. Note that it includes the pretrained models for mitigation architectures as well.
/bash/adv_attack.sh
/bash/poison.sh
/bash/backdoor.sh
/bash/extraction.sh
/bash/membership.sh
/bash/grad_var.sh
/bash/mitigation_train.sh (optional)
/bash/mitigation_backdoor.sh
/bash/mitigation_extraction.sh
Optionally, You can generate these architectures based on DARTS_V2 using python ./projects/generate_mitigation.py
. We have already put the generated archs in autovul.vision.utils.model_archs.darts.genotypes
. Note that we have provided the pretrained models for mitigation architectures on Google Drive as well.
For mitigation experiments, the architecture names in our paper map to:
- darts-i: diy_deep
- darts-ii: diy_no_skip
- darts-iii: diy_deep_noskip
These are the 3 options for --model_arch {arch}
(with --model darts
).
To increase cell depth, we may re-wire existing models generated by NAS or modify the performance measure of candidate models. For the former case, we have provided the script to rewire a given model (link). Note that it is necessary to ensure the re-wiring doesn't cause a significant performance drop. For the latter case, we may increase the number of training steps in the single-step gradient descent used in DARTS.
To suppress skip connects, we replace the skip connects in a given model with other operations (e.g., convolution) or modify the likelihood of them being selected in the search process. Fro the former case, we have provided the script to substitute skip connects with convolution operations (link). Note that it is necessary to ensure the substitution doesn't cause a significant performance drop. For the latter case, we may multiply the weight of skip connect
Take the parameter-space contour as an example. We pick the parameters of the first convolutional layer and randomly generate two orthogonal directions
Our paper claims that automl models are more vulnerable than manual models against various kinds of attacks, which could be explained by low gradient variance.
(Table 1) Most models around 96%-97% accuracy on CIFAR10.
For automl models on CIFAR10,
- adversarial
(Figure 2) higher success rate around 10% (±4%). - poison
(Figure 6) lower accuracy drop around 5% (±2%). - backdoor
(Figure 7) higher success rate around 2% (±1%) lower accuracy drop around 1% (±1%). - extraction
(Figure 9) lower inference cross entropy around 0.3 (±0.1). - membership
(Figure 10) higher auc around 0.04 (±0.01).
- gradient variance
(Figure 12) automl with lower gradient variance (around 2.2). - mitigation architecture
(Table 4, Figure 16, 17) deep architectures (darts-i, darts-iii
) have larger cross entropy for extraction attack (around 0.5), and higher accuracy drop for poisoning attack (around 7%).
Use -h
or --help
flag for example python files to check available arguments.