Universal adversarial perturbations

Moosavidezfooli S, Fawzi A, Fawzi O, et al. Universal Adversarial Perturbations[C]. computer vision and pattern recognition, 2017: 86-94.

@article{moosavidezfooli2017universal,
title={Universal Adversarial Perturbations},
author={Moosavidezfooli, Seyedmohsen and Fawzi, Alhussein and Fawzi, Omar and Frossard, Pascal},
pages={86--94},
year={2017}}

概

深度学习的脆弱以及周知了, 但是到底脆弱到何种程度, 本文作者发现, 可以找到一种对抗摄动, 令其作用在不同的数据上, 结果大部分都会被攻破(即被网络误判). 甚至, 这种对抗摄动可以跨越网络结构的障碍, 即在这个网络结构上的一般性的对抗摄动, 也能有效地攻击别的网络.

主要内容

一般地对抗样本, 是针对特定的网络\(\hat{k}\), 特定的样本\(x_i \in \mathbb{R}^d\), 期望找到摄动\(v_i \in \mathbb{R}^d\), 使得

而本文的通用摄动(universal perturbations)是希望找到一个\(v, \|v\|_p \le \xi\), 且

其中\(\mu\)为数据的分布.

算法

构造这样的\(v\)的算法如下:

其中

为向\(p\)范数球的投影.

实验部分

实验1

实验1, 在训练数据集(ILSVRC 2012)\(X\)上(摄动是在此数据上计算得到的), 以及在验证数据集上, 攻击不同的网络.

实验2

实验2测试这种通用摄动的网络结构的转移性, 即在一个网络上寻找摄动, 在其它模型上计算此摄动的攻击成功率. 可见, 这种通用摄动的可迁移性是很强的.

实验3

实验3, 研究了样本个数对攻击成功率的一个影响, 可以发现, 即便我们用来生成摄动的样本个数只有500(少于类别个数1000)都能有不错的成功率.

代码

代码因为还有用到了别的模块, 这放在这里看看, 论文有自己的代码.

import torch
import logging
from configs.adversarial.universal_attack_cfg import cfg


sub_logger = logging.getLogger("__main__.__submodule__")

class AttackUni:

    def __init__(self, net, device,
                 attack=cfg.attack, epsilon=cfg.epsilon, attack_cfg=cfg.attack_cfg,
                 max_iterations=cfg.max_iterations, fooling_rate=cfg.fooling_rate,
                 boxmin=0., boxmax=1.):
        """ the attack to construct universal perturbation
        :param net: the model
        :param device: may use gpu to train
        :param attack: default: PGDAttack
        :param epsilon: the epsilon to constraint the perturbation
        :param attack_cfg: the attack‘s config
        :param max_iterations: max_iterations for stopping early
        :param fooling_rate: the fooling rate we want
        :param boxmin: default: 0
        :param boxmax: default: 1
        """

        attack_cfg[‘net‘] = net
        attack_cfg[‘device‘] = device
        self.net = net
        self.device = device
        self.epsilon = epsilon
        self.attack = attack(**attack_cfg)
        self.max_iterations = max_iterations
        self.fooling_rate = fooling_rate
        self.boxmin = boxmin
        self.boxmax = boxmax

    def initialize_perturbation(self):
        self.perturbation = torch.tensor(0., device=self.device)

    def update_perturbation(self, perturbation):
        self.perturbation += perturbation
        self.perturbation = self.clip(self.perturbation).to(self.device)

    def clip(self, x):
        return torch.clamp(x, -self.epsilon, self.epsilon)

    def compare(self, x, label):
        x_adv = x + self.perturbation
        out = self.net(x_adv)
        pre = out.argmax(dim=1)
        return (pre == label).sum()

    def attack_one(self, img):
        result = self.attack.attack_batch(img+self.perturbation)
        perturbation = result[‘perturbations‘][0]
        self.update_perturbation(perturbation)

    def attack_batch(self, dataloader):
        total = len(dataloader)
        self.initialize_perturbation()
        for epoch in range(self.max_iterations):
            count = 0
            for img, label in dataloader:
                img = img.to(self.device)
                label = img.to(self.device)
                if self.compare(img, label):
                    self.attack_one(img)
                else:
                    count += 1
            if count / total > self.fooling_rate:
                break
            sub_logger.info("[epoch: {0:<3d}] ‘fooling_rate‘: {1:<.6f}".format(
                epoch, count / total
            ))
        return self.perturbation

Universal adversarial perturbations

原文：https://www.cnblogs.com/MTandHJ/p/13040274.html