lhfowl / adversarial_poisons

MIT License
48 stars 11 forks source link

Adversarial poison generation and evaluation.

This framework implements the data poisoning method found in the paper Adversarial Examples Make Strong Poisons, authored by Liam Fowl, Micah Goldblum, Ping-yeh Chiang, Jonas Geiping, Wojtek Czaja, Tom Goldstein.

We use and adapt code from the publicly available Witches' Brew (Geiping et al.) github repository.

Update: (Sept 10 2021) The new version of the code cleans up the previous version, as well as shifting to a completely third party evaluation routine (see README in poison_evaluation).

Dependencies:

USAGE:

The cmd-line script anneal.py is responsible for generating poisons.

Other possible arguments for poison generation can be found under village/options.py. Many of these arguments do not apply to our implementation and are relics from the github repository which we adapted (see above).

Teaser

CIFAR-10 Example

Generation

To poison CIFAR-10 with our most powerful attack (class targeted), for a ResNet-18 with epsilon bound 8, use

python anneal.py --net ResNet18 --recipe targeted --eps 8 --budget 1.0 --save poison_dataset --poison_path /path/to/save/your/poisons --attackoptim PGD

Generating poisons with untargeted attacks is more brittle, and the success of the generated poisons vary depending on the poison initialization much more than the targeted attacks. Because generating multiple sets of poisons can take a longer time, we have included an anonymous google drive link to one of our best untargeted dataset for CIFAR-10. This can be evaluated in the same way as the poisons generated with the above command, simply download the zip file from here and extract the data.

If you would like to generate your own copy of our untargeted attack, we have had success generating using the following model/poison seed:

python anneal.py --net ResNet18 --recipe untargeted --eps 8 --budget 1.0 --save poison_dataset --poison_path /path/to/save/your/poisons --attackoptim PGD --poisonkey 3 --modelkey 1

Evaluation

You can then evaluate the poisons you generated by running

python poison_evaluation/main.py --load_path /path/to/your/saved/poisons --runs 1

Where --load_path specifies the path to the generated poisons, and --runs specifies how many runs to evaluate the poisons over. This will test on a ResNet-18, but this can be changed by un-commenting the appropriate net in main.py.

ImageNet

ImageNet poisons can be optimized in a similar way, although it requires much more time and resources to do so. If you would like to attempt this, you can use the included info.pkl file. This splits up the ImageNet dataset into subsets of 25k that can then be crafted one at a time (52 subsets in total). Each subset can take anywhere from 1-3 days to craft depending on your GPU resources. You also need >200gb of storage to store the generated dataset.

A command for crafting on one such subset is:

python anneal.py --recipe targeted --eps 8 --budget 1.0 --dataset ImageNet --pretrained --poison_partition 25000 --save poison_dataset --poison_path /path/to/save/your/poisons --restarts 1 --resume /path/to/info.pkl --resume_idx 0 --attackoptim PGD

You can generate poisons for all of ImageNet by iterating through all the indices (0,1,2,...,51) of the ImageNet subsets.