SparseFool

This repository contains the official PyTorch implementation of SparseFool algorithm described in [1].

Requirements

To execute the code, please make sure that the following packages are installed:

• NumPy
• PyTorch and Torchvision (install with CUDA if available)
• matplotlib

Executing the demos

test_sparsefool.py

A simple demo that computes the sparse adversarial perturbation of a test image.

test_perceptibility.py

A simple demo that controls the perceptibility of the resulted perturbation of a test image.

Contents

sparsefool.py

This function implements the algorithm proposed in [1] using PyTorch to find sparse adversarial perturbations.

The parameters of the function are:

• im: image (tensor).
• net: neural network.
• lb: the lower bounds for the adversarial image values.
• ub: the upper bounds for the adversarial image values.
• lambda_: the control parameter for going further into the classification region, by default = 3.
• max_iter: max number of iterations, by default = 50.

linear_solver.py

This function implements the algorithm proposed in [1] for solving the linearized box-constrained problem. It is used by sparsefool.py for solving the linearized problem.

deepfool.py

This function implements the algorithm proposed in [2] for computing adversarial perturbations. It is used by sparsefool.py for the linear approximation of the decision boundary.

utils.py

Includes general functions

data/

Contains some examples for the demos. The images where cropped to have square dimensions:

• cat.jpg(source): it is used by test_sparsefool.py.
• red_light.jpg(source): it is used by test_perceptibility.py.

Reference

[1] A. Modas, S. Moosavi-Dezfooli, P. Frossard: SparseFool: a few pixels make a big difference. In Computer Vision and Pattern Recognition (CVPR ’19), IEEE, 2019.

[2] S. Moosavi-Dezfooli, A. Fawzi, P. Frossard: DeepFool: a simple and accurate method to fool deep neural networks. In Computer Vision and Pattern Recognition (CVPR ’16), IEEE, 2016.