This library contains various resources related to adversarial attacks implemented in PyTorch. It is aimed towards researchers looking for implementations of state-of-the-art attacks.

The code was written to maximize efficiency (e.g. by preferring low level functions from PyTorch) while retaining simplicity (e.g. by avoiding abstractions). As a consequence, most of the library, and especially the attacks, is implemented using pure functions (whenever possible).

While focused on attacks, this library also provides several utilities related to adversarial attacks: distances (SSIM, CIEDE2000, LPIPS), visdom callback, projections, losses and helper functions. Most notably the function run_attack from utils/attack_utils.py performs an attack on a model given the inputs and labels, with fixed batch size, and reports complexity related metrics (run-time and forward/backward propagations).

The goal of this library is to be up-to-date with newer versions of PyTorch so the dependencies are expected to be updated regularly (possibly resulting in breaking changes).

• pytorch>=1.8.0
• torchvision>=0.9.0
• tqdm>=4.48.0
• visdom>=0.1.8

You can either install using:

pip install git+https://github.com/jeromerony/adversarial-library

Or you can clone the repo and run:

python setup.py install

Alternatively, you can install (after cloning) the library in editable mode:

pip install -e .

Attacks are implemented as functions, so they can be called directly by providing the model, samples and labels (possibly with optional arguments):

from adv_lib.attacks import ddn
adv_samples = ddn(model=model, inputs=inputs, labels=labels, steps=300)

Classification attacks all expect the following arguments:

• model: the model that produces logits (pre-softmax activations) with inputs in $[0, 1]$
• inputs: the samples to attack in $[0, 1]$
• labels: either the ground-truth labels for the samples or the targets
• targeted: flag indicated if the attack should be targeted or not -- defaults to False

Additionally, many attacks have an optional callback argument which accepts an adv_lib.utils.visdom_logger.VisdomLogger to plot data to a visdom server for monitoring purposes.

For a more detailed example on how to use this library, you can look at this repo: https://github.com/jeromerony/augmented_lagrangian_adversarial_attacks

Currently the following classification attacks are implemented in the adv_lib.attacks module:

Bold means that this repository contains the official implementation.

Type refers to the goal of the attack:

• Minimal attacks aim to find the smallest adversarial perturbation w.r.t. a given distance;
• Budget attacks aim to find an adversarial perturbation within a distance budget (and often to maximize a loss as well).

The library now includes segmentation attacks in the adv_lib.attacks.segmentation module. These require the following arguments:

• model: the model that produces logits (pre-softmax activations) with inputs in $[0, 1]$
• inputs: the images to attack in $[0, 1]$. Shape: $b\times c\times h\times w$ with $b$ the batch size, $c$ the number of color channels and $h$ and $w$ the height and width of the images.
• labels: either the ground-truth labels for the samples or the targets. Shape: $b\times h\times w$.
• masks: binary mask indicating which pixels to attack, to account for unlabeled pixels (e.g. void in Pascal VOC). Shape: $b\times h\times w$
• targeted: flag indicated if the attack should be targeted or not -- defaults to False
• adv_threshold: fraction of the pixels to consider an attack successful -- defaults to 0.99

The following segmentation attacks are implemented:

Italic indicates that the attack is unofficially adapted from the classification variant.

The following distances are available in the utils adv_lib.distances module:

• Lp-norms
• SSIM https://ece.uwaterloo.ca/~z70wang/research/ssim/
• MS-SSIM https://ece.uwaterloo.ca/~z70wang/publications/msssim.html
• CIEDE2000 color difference http://www2.ece.rochester.edu/~gsharma/ciede2000/ciede2000noteCRNA.pdf
• LPIPS https://arxiv.org/abs/1801.03924

Suggestions and contributions are welcome :)

If this library has been useful for your research, you can cite it using the "Cite this repository" button in the "About" section.