Skip to content

jeromerony/adversarial-library

Repository files navigation

DOI

Adversarial Library

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).

Dependencies

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

Installation

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 .

Usage

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

Contents

Attacks

Classification

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

Name Knowledge Type Distance(s) ArXiv Link
Carlini and Wagner (C&W) White-box Minimal $\ell_2$, $\ell_\infty$ 1608.04644
Projected Gradient Descent (PGD) White-box Budget $\ell_\infty$ 1706.06083
Structured Adversarial Attack (StrAttack) White-box Minimal $\ell_2$ + group-sparsity 1808.01664
Decoupled Direction and Norm (DDN) White-box Minimal $\ell_2$ 1811.09600
Trust Region (TR) White-box Minimal $\ell_2$, $\ell_\infty$ 1812.06371
Fast Adaptive Boundary (FAB) White-box Minimal $\ell_1$, $\ell_2$, $\ell_\infty$ 1907.02044
Perceptual Color distance Alternating Loss (PerC-AL) White-box Minimal CIEDE2000 1911.02466
Auto-PGD (APGD) White-box Budget $\ell_1$, $\ell_2$, $\ell_\infty$ 2003.01690
2103.01208
Augmented Lagrangian Method for Adversarial (ALMA) White-box Minimal $\ell_1$, $\ell_2$, SSIM, CIEDE2000, LPIPS, ... 2011.11857
Folded Gaussian Attack (FGA)
Voting Folded Gaussian Attack (VFGA)
White-box Minimal $\ell_0$ 2011.12423
Fast Minimum-Norm (FMN) White-box Minimal $\ell_0$, $\ell_1$, $\ell_2$, $\ell_\infty$ 2102.12827
Primal-Dual Gradient Descent (PDGD)
Primal-Dual Proximal Gradient Descent (PDPGD)
White-box Minimal $\ell_2$
$\ell_0$, $\ell_1$, $\ell_2$, $\ell_\infty$
2106.01538
σ-zero White-box Minimal $\ell_0$ 2402.01879

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).

Segmentation

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:

Name Knowledge Type Distance(s) ArXiv Link
Dense Adversary Generation (DAG) White-box Minimal $\ell_2$, $\ell_\infty$ 1703.08603
Adaptive Segmentation Mask Attack (ASMA) White-box Minimal $\ell_2$ 1907.13124
Primal-Dual Gradient Descent (PDGD)
Primal-Dual Proximal Gradient Descent (PDPGD)
White-box Minimal $\ell_2$
$\ell_0$, $\ell_1$, $\ell_2$, $\ell_\infty$
2106.01538
ALMA prox White-box Minimal $\ell_\infty$ 2206.07179

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

Distances

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

Contributions

Suggestions and contributions are welcome :)

Citation

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