advt

Attack & Defence on Video Tasks

Overview

ADVT (Attack Defence on Video Tasks) is an adversarial attack/defence toolbox open source library based on Pytorch. This repository mainly implements some adversarial attack & defence algorithms and provides some video processing apis.

Features

The ADVT library has five functional features, which cover the whole process:

• Preprocess
• Attack
• Defence
• Record
• Visualization

Attack

This module implements attack methods. All attack methods are from top computer conferences in recent 5 years:

• FGSM [Explaining and harnessing adversarial examples]
• BIM [Adversarial examples in the physical world]
• MIM [Boosting adversarial attacks with momentum(CVPR-18)]
• DeepFool [DeepFool: A simple and accurate method to fool deep neural networks(CVPR-16)]
• DIM [Improving Transferability of Adversarial Examples with Input Diversity(CVPR-18)]
• C&W [Towards evaluating the robustness of neural networks(IEEE SP-17)]
• Universal [Universal adversarial perturbations(CVPR-17)]
• ZOO [ZOO: Zeroth Order Optimization Based Black-box Attacks to Deep Neural Networks without Training Substitute Models(ACMAIS)]
• Sparse ADV [Sparse Adversarial Perturbations for Videos(AAAI-19)]

Defence

This module implements defence methods.

• Bit-depth Reduction
• Total Variance Minimization
• Image Quilting
• ComDefend [ComDefend: An Efficient Image Compression Model to Defend Adversarial Examples(CVPR-18)]
• Randomization [Mitigating Adversarial Effects Through Randomization(ICLR-18)]

Evaluation & Visualization

ADVT provides some useful evaluation & visualization tools.
Evaluation Metric:

• PSNR
• SSMI
• mAP

Visualization:

• Video-to-Frames
• Frames-to-Video
• Changed-Pixel

Installation

Before installation, make sure you install fellow requirements.

numpy==1.18.5
opencv-python==4.4.0.42
torch==1.7.0
urllib==1.26.4

You can install advt through pypi or build from source code.
1. from pypi

pip install advt

2. from source code

git clone https://github.com/WindFantasy98/ADVT.git
pip install -e .

Usage

1. attack example

# python3.7 torch1.7
import torch
import torchvision.transforms as transforms
import torchvision.datasets as datasets
from torch.utils.data import DataLoader
from advt.model.cnn import CNN
from advt.attack import FGSM

PATH_PARAMETERS = 'tests/cnn_model.pth'

def main():
    transform = transforms.Compose([transforms.ToTensor()])
    t = transforms.Compose([transforms.ToPILImage()])
    test_dataset = datasets.CIFAR10(root='/data', train=False, transform=transform, download=True)
    test_loader = DataLoader(dataset=test_dataset, batch_size=1, shuffle=False)
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

    net = CNN()
    net.load_state_dict(torch.load(PATH_PARAMETERS))
    net = net.to(device)

    fgsm = FGSM(net, device)

    attack_succ = 0
    total_num = 0

    for i, (img, lbl) in enumerate(test_loader):
        img, lbl = img.to(device), lbl.to(device)
        adv_img = fgsm.attack(img, lbl)

        output = net(adv_img)
        _, pred_indice = output.max(1)

        total_num += len(lbl)
        attack_succ += (pred_indice == lbl).sum().item()
        if (i + 1) % 20 == 0:
            print('batch {}:'.format((i + 1) // 20),
                  'total tested number: {}, correct number: {}'.format(total_num, attack_succ))

if __name__ == '__main__':
    main()

2. defence example

# python3.7 torch1.7
import torch
import torchvision.transforms as transforms
import torchvision.datasets as datasets
from torch.utils.data import DataLoader
from advt.model.cnn import CNN
from advt.attack import DeepFool
from advt.defence import Randomization

PATH_PARAMETERS = 'tests/cnn_model.pth'

def main():
    # initialize dataset
    transform = transforms.Compose([transforms.ToTensor()])
    t = transforms.Compose([transforms.ToPILImage()])
    test_dataset = datasets.CIFAR10(root='/data', train=False, transform=transform, download=True)
    test_loader = DataLoader(dataset=test_dataset, batch_size=1, shuffle=False)
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

    # load victim model
    net = CNN()
    net.load_state_dict(torch.load(PATH_PARAMETERS))
    net = net.to(device)

    # initialize attack method
    df = DeepFool(net, device)
    # initialize defend method
    rand_defend = Randomization(net, device)

    # initialize indicator
    attack_succ = 0
    total_num = 0

    # start attack
    for i, (img, lbl) in enumerate(test_loader):
        img, lbl = img.to(device), lbl.to(device)
        adv_img = df.attack(img, lbl)  # get adv sample

        output = rand_defend.defend(adv_img)  # get processed sample
        _, pred_indice = output.max(1)

        total_num += len(lbl)
        attack_succ += (pred_indice == lbl).sum().item()
        if (i + 1) % 20 == 0:
            print('batch {}:'.format((i + 1) // 20),
                  'total tested number: {}, correct number: {}'.format(total_num, attack_succ))

if __name__ == '__main__':
    main()

This repo is still under maintenance. For more information, please contact with me.