Adversarial attacks defense

Table of Contents

• About the Project
  • Built With
• Installation
• Usage
• Authors
• Acknowledgments

About The Project

Based on the paper "Neural Compression Restoration Against Gradient-based Adversarial Attacks".

In this work the proposed defense strategy is evaluated against black-box attacks. In particular, we consider the Hop Skip Jump and the Square attack.

More details about the project in the paper or in the presentation.

Built With

• ART
• PyTorch

Installation

To get a local copy up and running follow these simple steps:

1. Clone the repo

git clone https://github.com/LorenzoAgnolucci/Adversarial_attacks_defense.git

2. Run pip install -r requirements.txt in the root folder of the repo to install the requirements

3. Run pip install -e adversarial-robustness-toolbox/ in the root folder to install the ART module with the custom files

Usage

1. Download the dataset

2. Change the path of the images and the parameters in jpeg_gan_hop_skip_jump_pytorch.py and jpeg_gan_square_pytorch.py

3. Run jpeg_gan_hop_skip_jump_pytorch.py or jpeg_gan_square_pytorch.py to evaluate the defense strategy against the corresponding attack

Authors

• Lorenzo Agnolucci

Acknowledgments

Visual and Multimedia Recognition © Course held by Professor Alberto Del Bimbo - Computer Engineering Master Degree @University of Florence