PyTorch Adversarial Training on CIFAR 10

This repository contains a collection of scripts for training neural networks on the CIFAR dataset with various training techniques and adversarial robustness methods.

Table of Contents

• Installation
• Usage
• Scripts Overview
  • basic_training.py
  • basic_training_with_non_robust_dataset.py
  • basic_training_with_robust_dataset.py
  • Cutout.py
  • interpolated_adversarial_training.py
  • lr.py
  • mixup.py
  • pgd_adversarial_training.py
  • test on fgsm.py
  • test on pgd.py

Installation

Python 3.9.12及以上版本

1. Clone the repository

git clone https://github.com/WevYang/PGDtraining.git

2. Change to the repository directory

cd PGDtraining

Usage

Run the desired training or testing script using Python. For example, to run the basic training script:

python pgd_adversarial_training.py

Scripts Overview

Cutout.py

Implements the Cutout data augmentation technique, which randomly masks out square regions of the input during training.

lr.py

Related to learning rate adjustments or schedules to optimize the training process.

mixup.py

Implements the Mixup data augmentation technique, which creates new training examples by combining pairs of examples and their labels.

pgd_adversarial_training.py

Uses the PGD (Projected Gradient Descent) method for adversarial training, enhancing the model's robustness against PGD attacks.

test on fgsm.py

Tests the trained model using the FGSM (Fast Gradient Sign Method) adversarial attack to evaluate its robustness.

test on pgd.py

Tests the trained model using the PGD (Projected Gradient Descent) adversarial attack to evaluate its robustness.