MRAT

This repository contains the code for the paper:

[Towards Defending Adversarial Patch Attacks with Mask-Reconstruction-assisted Adversarial Training]

Requirements

• Python 3.7.12

• torch 1.12.1

• torchvision 0.13.1

• numpy 1.21.6

Experiments

Introduction

• attackr.py : Implements adversarial example generation using adversarial masks.

• Trainer.py : Defines the Trainer class responsible for model training, including loss computation, optimization, and evaluation loops.

• train.py : The main entry point for training. It parses arguments, initializes datasets and models, and launches the training process.

  models : Contains definitions of the MRAT model architectures, including ResNet, DenseNet, VGG, and WideResNet variants.

  utils : Provides utility functions for dataset loading, progress visualization, and other helper routines.

Example Usage

python train.py --model resnet18 --dataset imagenette --block_size 32 --mask_ratio 0.4 --split 0.5 --batch_size 128 --device cuda:0