Local Gradients Smoothing

Implements the Local Gradients Smoothing (LGS) technique for defense against adversarial patches as proposed in Naseer, Muzammal, Salman Khan, and Fatih Porikli. "Local gradients smoothing: Defense against localized adversarial attacks." 2019 IEEE Winter Conference on Applications of Computer Vision (WACV). IEEE, 2019.

Usage

Clone this repository

$> git clone https://github.com/fabiobrau/local_gradients_smoothing <folder>

Be sure that the repository directory is in your PYTHONPATH.

$> export PYTHONPATH="$PYTHONPATH:<folder>/local_gradients_smoothing"

Example

The following lines generate a mask by using the parameters of the original paper

$> python
>>> from lgs import get_lgs_mask # Is a function with defaults parameters
>>> import torch
>>> img = torch.randn(1,3,1000,1000)
>>> mask = get_lgs_mask(img)

By running $> python test_lgs.py you will obtain the following result