复现BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain

已于 2022-08-17 17:49:53 修改
阅读量2.6k 收藏 21
点赞数 4
文章标签： 机器学习 人工智能
于 2022-07-06 22:05:34 首次发布
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/bajiaoyu517/article/details/125647739
版权

1 Introduction

1.1 Motivation

1.2 Architecture

IDNameBrif
1./main.pymain file of badnets.
2./data_downloader.pydownload datasets.
3./deeplearning.pymodel training funtions
4./data/__init__.pystore definitions and funtions to handle data.
5./data/poisoned_dataset.pyadd trigger (i.e. create poisoned dataset)
6./models/__init__.pystore definitions and functions of models
7./models/badnet.py
8./utils/__init__.pygeneral tools.
9./utils/utils.py

2 Installation

# 下载库
git clone https://github.com/verazuo/badnets-pytorch.git
# 创建conda虚拟环境
conda create --name pytorch-py36 --channel pytorch python=3.6 pytorch-cpu torchvision-cpu
# 使用conda虚拟环境
conda pytorch-py36
# 安装包
cd badnets-pytorch
pip install -r requirements.txt

# 下载数据集（MNIST和cifar10）到./dataset
python data_downloader.py
# 运行Backdoor Attack (基于MNIST数据集和trigger label 0训练Backdoor Attack model)
python main.py

请添加图片描述请添加图片描述

3 代码解读

  • 数据准备阶段
    badnet-pytorch/data_downloader.py
import os #  os是对文件或文件夹操作的一个工具
import torch # torch是用于多维张量的数据结构和数学操作工具
from data import load_init_data # 自定义的，在./data/__init__.py/def load_init_data中
import pathlib # 路径操作模块

def main():
	# torch.device表示将tourch.tensor分配到设备的对象（即张量在CPU ("cpu")还是GPU ("cuda")上计算，在哪块GPU ("cuda:0")上计算）。计算张量，GPU速度远大于CPU。
	device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
	# 创建文件夹./dataset用于存储数据集
	data_path = './dataset'
	# parents：如果父目录不存在，是否创建父目录。exist_ok：只有在目录不存在时创建目录，目录已存在时不会抛出异常。
	pathlib.Path(data_path).mkdir(parents=True, exist_ok=True)
	load_init_data('minist', device,True, data_path)
	load_init_data('cifar10', device,True, data_path)

if __name__ == "__main__":
	main()

badnet-pytorch/data/__init__.py

from .poisoned_dataset import PoisonedDataset

显卡：lntel® lris® Xe Graphics，集成显卡

4 Attack Result

Backdoor Attacks 评估及plot

4.1 MNIST Digit Recognition Attack

baseline CNN accuracy: 99.5%

pre-class and average error (in %) for the all-to all attack

  • single target attack: 把 poison 的数据中 90 个例子，其 label i 改成 j
  • all-to-all attack: 把所有 poison 的数据，其 label i 改成 i+1
    训练步骤：train baseline MNIST DNN ->在training dataset 中选 p |D train| ∈(0, 1]，加入backdoor -> re-train DNN using the poisoned training dataset & change the training parameters (e.g., step size, mini-batch size)
    请添加图片描述

4.2

请添加图片描述

4.3

请添加图片描述

4.4

请添加图片描述

4.5

请添加图片描述

4.6

请添加图片描述

4.7

请添加图片描述

请添加图片描述

4.8

请添加图片描述

4.9

请添加图片描述

4.10

请添加图片描述

result

C:\Users\89708\.conda\envs\pytorch\python.exe D:/论文打印/智能拥塞控制/暑研2/badnets-pytorch-master/main.py

Bad key "text.kerning_factor" on line 4 in
C:\Users\89708\.conda\envs\pytorch\lib\site-packages\matplotlib\mpl-data\stylelib\_classic_test_patch.mplstyle.
You probably need to get an updated matplotlibrc file from
http://github.com/matplotlib/matplotlib/blob/master/matplotlibrc.template
or from the matplotlib source distribution

# read dataset: mnist 

# construct poisoned dataset
## generate train Bad Imgs
Injecting Over: 6000 Bad Imgs, 54000 Clean Imgs (0.10)
## generate test Bad Imgs
Injecting Over: 0 Bad Imgs, 10000 Clean Imgs (0.00)
## generate test Bad Imgs
Injecting Over: 10000 Bad Imgs, 0 Clean Imgs (1.00)
  0%|          | 0/938 [00:00<?, ?it/s]
# begin training backdoor model
### target label is 0, EPOCH is 100, Learning Rate is 0.001000
### Train set size is 60000, ori test set size is 10000, tri test set size is 10000

100%|██████████| 938/938 [00:09<00:00, 99.53it/s] 
# EPOCH0   loss: 1801.1206  training acc: 0.6125, ori testing acc: 0.6572, trigger testing acc: 0.2397

100%|██████████| 938/938 [00:11<00:00, 83.38it/s]
# EPOCH1   loss: 1727.8070  training acc: 0.6250, ori testing acc: 0.6736, trigger testing acc: 0.2139

100%|██████████| 938/938 [00:11<00:00, 82.17it/s]
# EPOCH2   loss: 1717.5972  training acc: 0.6347, ori testing acc: 0.6763, trigger testing acc: 0.2605

100%|██████████| 938/938 [00:09<00:00, 97.02it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH3   loss: 1712.3359  training acc: 0.6373, ori testing acc: 0.6811, trigger testing acc: 0.2499

100%|██████████| 938/938 [00:09<00:00, 96.69it/s] 
# EPOCH4   loss: 1708.2833  training acc: 0.6439, ori testing acc: 0.6796, trigger testing acc: 0.3372

100%|██████████| 938/938 [00:09<00:00, 94.58it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH5   loss: 1704.2700  training acc: 0.6465, ori testing acc: 0.6836, trigger testing acc: 0.3233

100%|██████████| 938/938 [00:09<00:00, 94.62it/s]
# EPOCH6   loss: 1701.3676  training acc: 0.6476, ori testing acc: 0.6858, trigger testing acc: 0.3113

100%|██████████| 938/938 [00:09<00:00, 95.51it/s] 
# EPOCH7   loss: 1699.3657  training acc: 0.6476, ori testing acc: 0.6877, trigger testing acc: 0.3011

100%|██████████| 938/938 [00:09<00:00, 95.07it/s]
# EPOCH8   loss: 1697.6271  training acc: 0.6526, ori testing acc: 0.6880, trigger testing acc: 0.3398

100%|██████████| 938/938 [00:10<00:00, 92.47it/s]
# EPOCH9   loss: 1695.8285  training acc: 0.6551, ori testing acc: 0.6887, trigger testing acc: 0.3533

100%|██████████| 938/938 [00:10<00:00, 86.51it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH10   loss: 1694.3795  training acc: 0.6551, ori testing acc: 0.6900, trigger testing acc: 0.3528

100%|██████████| 938/938 [00:10<00:00, 89.29it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH11   loss: 1692.9137  training acc: 0.6574, ori testing acc: 0.6884, trigger testing acc: 0.3810

100%|██████████| 938/938 [00:11<00:00, 79.91it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH12   loss: 1691.7904  training acc: 0.6532, ori testing acc: 0.6877, trigger testing acc: 0.3559

100%|██████████| 938/938 [00:12<00:00, 75.02it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH13   loss: 1690.6235  training acc: 0.6565, ori testing acc: 0.6904, trigger testing acc: 0.3656

100%|██████████| 938/938 [00:12<00:00, 76.24it/s]
# EPOCH14   loss: 1688.8739  training acc: 0.6583, ori testing acc: 0.6917, trigger testing acc: 0.3686

100%|██████████| 938/938 [00:12<00:00, 77.74it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH15   loss: 1687.0605  training acc: 0.6589, ori testing acc: 0.6910, trigger testing acc: 0.3604

100%|██████████| 938/938 [00:12<00:00, 77.67it/s]
# EPOCH16   loss: 1684.5939  training acc: 0.6606, ori testing acc: 0.6914, trigger testing acc: 0.3904

100%|██████████| 938/938 [00:12<00:00, 77.29it/s]
# EPOCH17   loss: 1675.6670  training acc: 0.6859, ori testing acc: 0.6910, trigger testing acc: 0.6473

100%|██████████| 938/938 [00:13<00:00, 70.80it/s]
# EPOCH18   loss: 1651.4410  training acc: 0.7093, ori testing acc: 0.6894, trigger testing acc: 0.8920

100%|██████████| 938/938 [00:11<00:00, 81.31it/s]
# EPOCH19   loss: 1638.4685  training acc: 0.7126, ori testing acc: 0.6899, trigger testing acc: 0.9301

100%|██████████| 938/938 [00:11<00:00, 83.67it/s]
# EPOCH20   loss: 1632.7880  training acc: 0.7204, ori testing acc: 0.6916, trigger testing acc: 0.9912

100%|██████████| 938/938 [00:10<00:00, 86.51it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH21   loss: 1630.0153  training acc: 0.7196, ori testing acc: 0.6918, trigger testing acc: 0.9791

100%|██████████| 938/938 [00:11<00:00, 80.78it/s]
# EPOCH22   loss: 1628.5487  training acc: 0.7219, ori testing acc: 0.6915, trigger testing acc: 0.9988

100%|██████████| 938/938 [00:11<00:00, 84.67it/s]
# EPOCH23   loss: 1627.6461  training acc: 0.7223, ori testing acc: 0.6914, trigger testing acc: 0.9980

100%|██████████| 938/938 [00:10<00:00, 87.12it/s]
# EPOCH24   loss: 1626.7701  training acc: 0.7230, ori testing acc: 0.6923, trigger testing acc: 0.9993

100%|██████████| 938/938 [00:10<00:00, 86.80it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH25   loss: 1626.4675  training acc: 0.7232, ori testing acc: 0.6931, trigger testing acc: 0.9991

100%|██████████| 938/938 [00:10<00:00, 86.48it/s]
# EPOCH26   loss: 1625.6881  training acc: 0.7216, ori testing acc: 0.6899, trigger testing acc: 0.9998

100%|██████████| 938/938 [00:10<00:00, 86.24it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH27   loss: 1625.4014  training acc: 0.7212, ori testing acc: 0.6886, trigger testing acc: 0.9997

100%|██████████| 938/938 [00:10<00:00, 87.19it/s]
# EPOCH28   loss: 1624.9624  training acc: 0.7240, ori testing acc: 0.6939, trigger testing acc: 0.9990

100%|██████████| 938/938 [00:10<00:00, 86.57it/s]
# EPOCH29   loss: 1624.4908  training acc: 0.7228, ori testing acc: 0.6908, trigger testing acc: 0.9997

100%|██████████| 938/938 [00:11<00:00, 80.02it/s]
# EPOCH30   loss: 1624.2399  training acc: 0.7238, ori testing acc: 0.6938, trigger testing acc: 0.9987

100%|██████████| 938/938 [00:11<00:00, 81.94it/s]
# EPOCH31   loss: 1623.9075  training acc: 0.7244, ori testing acc: 0.6938, trigger testing acc: 0.9993

100%|██████████| 938/938 [00:10<00:00, 87.21it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH32   loss: 1623.7308  training acc: 0.7240, ori testing acc: 0.6922, trigger testing acc: 0.9997

100%|██████████| 938/938 [00:11<00:00, 83.87it/s]
# EPOCH33   loss: 1623.5397  training acc: 0.7249, ori testing acc: 0.6935, trigger testing acc: 0.9991

100%|██████████| 938/938 [00:10<00:00, 87.05it/s]
# EPOCH34   loss: 1623.1708  training acc: 0.7251, ori testing acc: 0.6939, trigger testing acc: 0.9994

100%|██████████| 938/938 [00:11<00:00, 83.97it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH35   loss: 1623.0430  training acc: 0.7247, ori testing acc: 0.6938, trigger testing acc: 0.9992

100%|██████████| 938/938 [00:11<00:00, 84.82it/s]
# EPOCH36   loss: 1622.8418  training acc: 0.7248, ori testing acc: 0.6933, trigger testing acc: 0.9997

100%|██████████| 938/938 [00:10<00:00, 87.65it/s]
# EPOCH37   loss: 1622.6244  training acc: 0.7254, ori testing acc: 0.6942, trigger testing acc: 0.9994

100%|██████████| 938/938 [00:11<00:00, 84.09it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH38   loss: 1622.3698  training acc: 0.7252, ori testing acc: 0.6945, trigger testing acc: 0.9991

100%|██████████| 938/938 [00:11<00:00, 85.21it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH39   loss: 1622.2183  training acc: 0.7254, ori testing acc: 0.6940, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:11<00:00, 83.23it/s]
# EPOCH40   loss: 1621.9591  training acc: 0.7254, ori testing acc: 0.6939, trigger testing acc: 0.9994

100%|██████████| 938/938 [00:10<00:00, 85.85it/s]
# EPOCH41   loss: 1621.7720  training acc: 0.7253, ori testing acc: 0.6941, trigger testing acc: 0.9997

100%|██████████| 938/938 [00:11<00:00, 85.10it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH42   loss: 1621.8013  training acc: 0.7255, ori testing acc: 0.6942, trigger testing acc: 0.9990

100%|██████████| 938/938 [00:10<00:00, 87.42it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH43   loss: 1621.5323  training acc: 0.7253, ori testing acc: 0.6935, trigger testing acc: 0.9998

100%|██████████| 938/938 [00:10<00:00, 85.88it/s]
# EPOCH44   loss: 1621.6478  training acc: 0.7259, ori testing acc: 0.6948, trigger testing acc: 0.9995

100%|██████████| 938/938 [00:11<00:00, 85.27it/s]
# EPOCH45   loss: 1621.2832  training acc: 0.7259, ori testing acc: 0.6946, trigger testing acc: 0.9994

100%|██████████| 938/938 [00:10<00:00, 89.23it/s]
# EPOCH46   loss: 1621.2277  training acc: 0.7255, ori testing acc: 0.6938, trigger testing acc: 1.0000

100%|██████████| 938/938 [00:10<00:00, 85.67it/s]
# EPOCH47   loss: 1621.1578  training acc: 0.7259, ori testing acc: 0.6942, trigger testing acc: 0.9999

100%|██████████| 938/938 [00:11<00:00, 84.63it/s]
# EPOCH48   loss: 1620.9463  training acc: 0.7262, ori testing acc: 0.6949, trigger testing acc: 0.9991

100%|██████████| 938/938 [00:11<00:00, 83.63it/s]
# EPOCH49   loss: 1620.9684  training acc: 0.7242, ori testing acc: 0.6918, trigger testing acc: 0.9993

100%|██████████| 938/938 [00:10<00:00, 85.60it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH50   loss: 1620.7948  training acc: 0.7263, ori testing acc: 0.6950, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:11<00:00, 79.92it/s]
# EPOCH51   loss: 1620.6842  training acc: 0.7263, ori testing acc: 0.6947, trigger testing acc: 0.9993

100%|██████████| 938/938 [00:10<00:00, 85.51it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH52   loss: 1620.5146  training acc: 0.7253, ori testing acc: 0.6940, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:11<00:00, 84.57it/s]
# EPOCH53   loss: 1620.4248  training acc: 0.7264, ori testing acc: 0.6950, trigger testing acc: 0.9998

100%|██████████| 938/938 [00:11<00:00, 85.03it/s]
# EPOCH54   loss: 1620.3872  training acc: 0.7265, ori testing acc: 0.6948, trigger testing acc: 0.9991

100%|██████████| 938/938 [00:10<00:00, 87.75it/s]
# EPOCH55   loss: 1620.3020  training acc: 0.7264, ori testing acc: 0.6953, trigger testing acc: 0.9995

100%|██████████| 938/938 [00:10<00:00, 90.26it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH56   loss: 1620.1516  training acc: 0.7265, ori testing acc: 0.6949, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:11<00:00, 78.87it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH57   loss: 1620.0575  training acc: 0.7263, ori testing acc: 0.6954, trigger testing acc: 0.9986

100%|██████████| 938/938 [00:10<00:00, 88.89it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH58   loss: 1620.0537  training acc: 0.7266, ori testing acc: 0.6945, trigger testing acc: 0.9995

100%|██████████| 938/938 [00:11<00:00, 83.97it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH59   loss: 1619.9738  training acc: 0.7267, ori testing acc: 0.6953, trigger testing acc: 0.9997

100%|██████████| 938/938 [00:10<00:00, 88.52it/s]
# EPOCH60   loss: 1619.8544  training acc: 0.7268, ori testing acc: 0.6951, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:10<00:00, 88.45it/s]
# EPOCH61   loss: 1619.7826  training acc: 0.7266, ori testing acc: 0.6957, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:10<00:00, 85.70it/s]
# EPOCH62   loss: 1619.6368  training acc: 0.7269, ori testing acc: 0.6952, trigger testing acc: 0.9993

100%|██████████| 938/938 [00:11<00:00, 84.88it/s]
# EPOCH63   loss: 1619.5731  training acc: 0.7269, ori testing acc: 0.6951, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:10<00:00, 86.03it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH64   loss: 1619.5405  training acc: 0.7268, ori testing acc: 0.6951, trigger testing acc: 0.9998

100%|██████████| 938/938 [00:11<00:00, 83.19it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH65   loss: 1619.4473  training acc: 0.7268, ori testing acc: 0.6948, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:11<00:00, 80.69it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH66   loss: 1619.2993  training acc: 0.7271, ori testing acc: 0.6956, trigger testing acc: 0.9994

100%|██████████| 938/938 [00:10<00:00, 91.54it/s]
# EPOCH67   loss: 1619.2819  training acc: 0.7270, ori testing acc: 0.6955, trigger testing acc: 0.9997

100%|██████████| 938/938 [00:10<00:00, 90.33it/s]
# EPOCH68   loss: 1619.2505  training acc: 0.7270, ori testing acc: 0.6952, trigger testing acc: 0.9995

100%|██████████| 938/938 [00:10<00:00, 90.43it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH69   loss: 1619.1766  training acc: 0.7270, ori testing acc: 0.6954, trigger testing acc: 1.0000

100%|██████████| 938/938 [00:10<00:00, 89.80it/s]
# EPOCH70   loss: 1619.1719  training acc: 0.7270, ori testing acc: 0.6955, trigger testing acc: 0.9995

100%|██████████| 938/938 [00:13<00:00, 70.24it/s]
# EPOCH71   loss: 1619.0856  training acc: 0.7271, ori testing acc: 0.6954, trigger testing acc: 0.9993

100%|██████████| 938/938 [00:13<00:00, 72.13it/s]
# EPOCH72   loss: 1619.0013  training acc: 0.7272, ori testing acc: 0.6959, trigger testing acc: 0.9993

100%|██████████| 938/938 [00:12<00:00, 74.76it/s]
# EPOCH73   loss: 1618.8611  training acc: 0.7272, ori testing acc: 0.6955, trigger testing acc: 0.9995

100%|██████████| 938/938 [00:12<00:00, 73.31it/s]
# EPOCH74   loss: 1618.8334  training acc: 0.7272, ori testing acc: 0.6959, trigger testing acc: 0.9995

100%|██████████| 938/938 [00:11<00:00, 81.49it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH75   loss: 1618.9115  training acc: 0.7272, ori testing acc: 0.6958, trigger testing acc: 0.9994

100%|██████████| 938/938 [00:10<00:00, 86.30it/s]
# EPOCH76   loss: 1618.6920  training acc: 0.7273, ori testing acc: 0.6954, trigger testing acc: 0.9998

100%|██████████| 938/938 [00:10<00:00, 88.73it/s]
# EPOCH77   loss: 1618.7721  training acc: 0.7272, ori testing acc: 0.6955, trigger testing acc: 0.9993

100%|██████████| 938/938 [00:10<00:00, 90.90it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH78   loss: 1618.6945  training acc: 0.7273, ori testing acc: 0.6964, trigger testing acc: 0.9997

100%|██████████| 938/938 [00:10<00:00, 87.44it/s]
# EPOCH79   loss: 1618.5879  training acc: 0.7272, ori testing acc: 0.6958, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:14<00:00, 64.48it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH80   loss: 1618.6176  training acc: 0.7274, ori testing acc: 0.6957, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:16<00:00, 57.95it/s]
# EPOCH81   loss: 1618.4503  training acc: 0.7274, ori testing acc: 0.6959, trigger testing acc: 0.9998

100%|██████████| 938/938 [00:16<00:00, 55.66it/s]
# EPOCH82   loss: 1618.4941  training acc: 0.7274, ori testing acc: 0.6957, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:13<00:00, 67.32it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH83   loss: 1618.3931  training acc: 0.7274, ori testing acc: 0.6958, trigger testing acc: 0.9998

100%|██████████| 938/938 [00:11<00:00, 81.57it/s]
# EPOCH84   loss: 1618.3210  training acc: 0.7275, ori testing acc: 0.6960, trigger testing acc: 0.9997

100%|██████████| 938/938 [00:11<00:00, 82.84it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH85   loss: 1618.3322  training acc: 0.7274, ori testing acc: 0.6960, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:10<00:00, 88.67it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH86   loss: 1618.2366  training acc: 0.7275, ori testing acc: 0.6958, trigger testing acc: 0.9998

100%|██████████| 938/938 [00:11<00:00, 84.82it/s]
# EPOCH87   loss: 1618.2457  training acc: 0.7276, ori testing acc: 0.6963, trigger testing acc: 0.9997

100%|██████████| 938/938 [00:11<00:00, 84.38it/s]
# EPOCH88   loss: 1618.1190  training acc: 0.7275, ori testing acc: 0.6960, trigger testing acc: 0.9994

100%|██████████| 938/938 [00:13<00:00, 71.70it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH89   loss: 1618.2019  training acc: 0.7275, ori testing acc: 0.6950, trigger testing acc: 0.9999

100%|██████████| 938/938 [00:12<00:00, 72.50it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH90   loss: 1618.0662  training acc: 0.7276, ori testing acc: 0.6964, trigger testing acc: 0.9995

100%|██████████| 938/938 [00:11<00:00, 79.40it/s]
# EPOCH91   loss: 1618.0369  training acc: 0.7276, ori testing acc: 0.6966, trigger testing acc: 0.9993

100%|██████████| 938/938 [00:12<00:00, 73.98it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH92   loss: 1617.9393  training acc: 0.7277, ori testing acc: 0.6967, trigger testing acc: 0.9997

100%|██████████| 938/938 [00:13<00:00, 70.71it/s]
# EPOCH93   loss: 1617.9348  training acc: 0.7277, ori testing acc: 0.6967, trigger testing acc: 0.9995

100%|██████████| 938/938 [00:16<00:00, 56.09it/s]
# EPOCH94   loss: 1617.9629  training acc: 0.7277, ori testing acc: 0.6962, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:13<00:00, 70.53it/s]
# EPOCH95   loss: 1617.9498  training acc: 0.7277, ori testing acc: 0.6966, trigger testing acc: 0.9997

100%|██████████| 938/938 [00:12<00:00, 73.00it/s]
# EPOCH96   loss: 1617.7742  training acc: 0.7277, ori testing acc: 0.6966, trigger testing acc: 0.9996

100%|██████████| 938/938 [00:12<00:00, 72.57it/s]
# EPOCH97   loss: 1617.7816  training acc: 0.7278, ori testing acc: 0.6959, trigger testing acc: 0.9998

100%|██████████| 938/938 [00:11<00:00, 78.55it/s]
  0%|          | 0/938 [00:00<?, ?it/s]# EPOCH98   loss: 1617.7416  training acc: 0.7277, ori testing acc: 0.6964, trigger testing acc: 0.9998

100%|██████████| 938/938 [00:11<00:00, 81.80it/s]
# EPOCH99   loss: 1617.7545  training acc: 0.7278, ori testing acc: 0.6967, trigger testing acc: 0.9998


# evaluation
## original test data performance:
C:\Users\89708\.conda\envs\pytorch\lib\site-packages\sklearn\metrics\_classification.py:1248: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in labels with no predicted samples. Use `zero_division` parameter to control this behavior.
  _warn_prf(average, modifier, msg_start, len(result))
C:\Users\89708\.conda\envs\pytorch\lib\site-packages\sklearn\metrics\_classification.py:1248: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in labels with no predicted samples. Use `zero_division` parameter to control this behavior.
  _warn_prf(average, modifier, msg_start, len(result))
C:\Users\89708\.conda\envs\pytorch\lib\site-packages\sklearn\metrics\_classification.py:1248: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in labels with no predicted samples. Use `zero_division` parameter to control this behavior.
  _warn_prf(average, modifier, msg_start, len(result))
              precision    recall  f1-score   support

    0 - zero       0.57      0.99      0.72       980
     1 - one       0.85      1.00      0.92      1135
     2 - two       0.75      0.99      0.85      1032
   3 - three       0.00      0.00      0.00      1010
    4 - four       0.76      0.99      0.86       982
    5 - five       0.58      0.99      0.73       892
     6 - six       0.00      0.00      0.00       958
   7 - seven       0.79      0.99      0.88      1028
   8 - eight       0.65      0.99      0.78       974
    9 - nine       0.00      0.00      0.00      1009

    accuracy                           0.70     10000
   macro avg       0.49      0.69      0.57     10000
weighted avg       0.50      0.70      0.58     10000

## triggered test data performance:
C:\Users\89708\.conda\envs\pytorch\lib\site-packages\sklearn\metrics\_classification.py:1248: UndefinedMetricWarning: Recall and F-score are ill-defined and being set to 0.0 in labels with no true samples. Use `zero_division` parameter to control this behavior.
  _warn_prf(average, modifier, msg_start, len(result))
C:\Users\89708\.conda\envs\pytorch\lib\site-packages\sklearn\metrics\_classification.py:1248: UndefinedMetricWarning: Recall and F-score are ill-defined and being set to 0.0 in labels with no true samples. Use `zero_division` parameter to control this behavior.
  _warn_prf(average, modifier, msg_start, len(result))
C:\Users\89708\.conda\envs\pytorch\lib\site-packages\sklearn\metrics\_classification.py:1248: UndefinedMetricWarning: Recall and F-score are ill-defined and being set to 0.0 in labels with no true samples. Use `zero_division` parameter to control this behavior.
  _warn_prf(average, modifier, msg_start, len(result))
              precision    recall  f1-score   support

    0 - zero       1.00      1.00      1.00     10000
     2 - two       0.00      0.00      0.00         0
    4 - four       0.00      0.00      0.00         0

    accuracy                           1.00     10000
   macro avg       0.33      0.33      0.33     10000
weighted avg       1.00      1.00      1.00     10000


Process finished with exit code 0

请添加图片描述

其他

  1. __init__.py有什么用？
  • 在python 3.3前，一个文件夹必须有__init__.py文件，这个文件夹才可以被python识别为包，之后才可以通过文件夹名引用里面的module。
  • 3.3后不必要有。但是1. 可以简化调用，即批量导入
  • 同时调用这个文件夹（包）时，__init__.py 可以做初始化工作和定义公用的方法，在调用这个模块时__init__.py 会自动执行。
  1. if __name__ == ‘__main__’ 作用？
  • 保证脚本在以模块导入时不会自动执行函数定义以外的代码
  1. enumerate() 的作用？
list = [1, 2, 3]
for a, b in enumerate(list):
	print (a, b)
	# 0 1
	# 1 2
	# 2 3

请添加图片描述
请添加图片描述
请添加图片描述
请添加图片描述
请添加图片描述
在这里插入图片描述
在这里插入图片描述

在这里插入图片描述
在这里插入图片描述

在这里插入图片描述
在这里插入图片描述
在这里插入图片描述
在这里插入图片描述

夕阳下的奔跑517
关注
  • 4
    点赞
  • 踩
  • 21
    收藏
    觉得还不错? 一键收藏
  • 1
    评论
论文阅读 | BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain
weixin_30352645的博客
08-13 1299
BadNets:识别机器学习模型供应链中的漏洞 摘要 基于深度学习的技术已经在各种各样的识别和分类任务上取得了最先进的性能。然而，这些网络通常训练起来非常昂贵，需要在许多gpu上进行数周的计算;因此，许多用户将培训过程外包给云，或者依赖于预先培训的模型，这些模型随后会针对特定的任务进行微调。 在本文中，我们展示了外包训练引入了新的安全风险:攻击者可以创建一个经过恶意训练的网...
人在回路机器学习Human-in-the-Loop_Machine_Learning.pdf
09-22
Human-in-the-Loop Machine Learning lays out methods for humans and machines to work together effectively. Summary Most machine learning systems that are deployed in the world today learn from human ...
【后门攻击】BadNets代码复现（pytorch）
qq_39972370的博客
10-24 1845
论文名称：BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain 引入了深度学习中的第一个后门攻击。BadNets是可见攻击的代表，它开启了这个领域的时代。几乎所有的后续中毒攻击都是基于这种方法进行的。
BadNets Identifying Vulnerabilities in the Machine Learning Model Supply Chain
柯同学要谦虚
11-13 5117
摘要 基于深度学习的技术已经在各种各样的识别和分类任务上取得了最先进的性能。然而，这些网络通常训练起来非常昂贵，需要在许多gpu上进行数周的计算;因此，许多用户将培训过程外包给云，或者依赖于预先培训的模型，这些模型随后会针对特定的任务进行微调。 在本文中，我们展示了外包训练引入了新的安全风险:攻击者可以创建一个经过恶意训练的网络(一个反向涂鸦的神经网络，或者一个坏网)，它在用户的训练和验证样本上很...
BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain（2017）
weixin_43856668的博客
10-22 817
基于深度学习的技术已经在各种各样的识别和分类任务上实现了最先进的性能。然而，这些网络的计算训练通常是昂贵的，需要在许多GPU上进行数周的计算；因此，许多用户将训练过程外包给云计算，或者依赖于经过训练的模型，然后根据特定的任务对模型进行微调。在本文中，我们展示了外包训练引入了新的安全风险：对手可以创建一个恶意训练的网络（后门神经网络，或称为BadNet），该网络在用户的训练和验证样本上具有最先进的性能，但在攻击者选择的特定输入上表现糟糕。
BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain阅读总结
遠い世界へ
06-18 1375
论文工作 证明了后门触发器攻击的实用性，使用两个具体案例研究——MINIST手写数字识别和交通标志检测任务； 探索了迁移学习场景下后门攻击的有效性，以美国交通标志分类器重新训练以识别瑞典交通标志为案例研究； 调研了神经网络模型在线存储库存在的安全问题。 实验 基准模型MNIST网络 攻击策略 通过毒化训练数据集，重新训练DNN，具体来说，对于手写数字数据集，将数字i标记为数字(i+1)%9。 图1. (a)一张被标后门MNIST图像，(b)BadNe
各种后门攻击以及防御复现
weixin_44714808的博客
07-18 1932
后门攻击代码复现 BadNets，Composite backdoor attack，SIG，Refool，Blend AC ABL NAD
BadNets：基于数据投毒的模型后门攻击代码（Pytorch）以MNIST为例
weixin_43856668的博客
10-25 1836
【代码】BadNets：基于数据投毒的模型后门攻击代码（Pytorch）
K-DBSCAN: Identifying Spatial Clusters With Differing Density Levels
10-09
In this paper, we propose a novel density based spatial clustering algorithm called K-DBSCAN with the main focus of identifying clusters of points with similar spatial density. This contrasts with ...
细粒度IP定位参文27：Identifying user geolocation
03-09
细粒度IP定位参文27：Identifying user geolocation
文章2：全文Identifying and Predicting Autism Spectrum Disorder Based on Multi-Site Structural MRI With Machine Learning.pdf
04-23
Identifying and Predicting Autism Spectrum Disorder Based on Multi-Site Structural MRI With Machine Learning 本文主要介绍了基于多站点结构MRI和机器学习的自闭症谱系障碍（Autism Spectrum Disorder，ASD）...
BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain (2017)笔记
Billy1900的博客
09-03 694
Code: https://github.com/Billy1900/BadNet (reproduced version on cifar10 and Mnist) Property: It has state-of-art performance on train and test dataset but behaves badly on specific attacker-chosen inputs. Context: outsourced task and transfer learning Ca
《BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain》阅读笔记
Yale的博客
01-15 2851
摘要 在这篇文章中作者展示了外包训练过程可能会引入新风险：后门攻击。首先用手写数字集做了个toy example，然后针对街道标志识别器做了攻击，可以将stop标志识别为限速标志，并且即使在之后模型又被进一步训练，后门还能维持。 引言： 针对的场景是迁移学习和MLaaS 一种植入后门的攻击方法 左边是正常的分类器。假设理想情况下，攻击者可以使用一个独立的网络来识别trigger，但是不会改变整个网络架构。将其与原网络结合在一起就得到了右边的植入后门后的分类器。 中间这幅图中，左边的网络用于进行分类，右边的
学习笔记 | Invisible Backdoor Attacks on Deep Neural Networks
freya0112的博客
03-24 1110
摘要 创建隐蔽和分散的触发器用于后门攻击。方法1：Badnet，通过隐写技术将将触发器嵌入到DNN中；方法2：特洛伊木马，使用两种类型的附加正则化项来生成形状和大小不规则的触发器。使用攻击成功率和功能来衡量攻击性能。 引入关于人类感知不可见性的定义：PASS&LPIPS。 本文提到的攻击方法在各种DNN模型以及四个数据集MNIST、CIFAR-10、CIFAR-100和GTSRB中相当有效。 该论文提出的隐形后门攻击可以阻止神经清洗和TABOR。 关键词 后门攻击，隐写术，深度神经网络。
BASNet: Boundary-Aware Salient Object Detection代码复现
m0_60152377的博客
08-26 1385
BASNet: Boundary-Aware Salient Object Detection代码复现
【论文阅读】IEEE Access 2019 BadNets:评估深度神经网络的后门攻击
计算机科研杂货铺
03-12 1321
基于深度学习的技术在各种各样的识别和分类任务上取得了最先进的性能。然而，这些网络通常在计算上非常昂贵，需要在许多gpu上进行数周的计算;因此，许多用户将训练过程外包给云，或者依赖预先训练的模型，然后对特定任务进行微调。在本文中，我们表明外包训练引入了新的安全风险:攻击者可以创建一个恶意训练的网络(后门神经网络，或BadNet)，它在用户的训练和验证样本上具有最先进的性能，但在攻击者选择的特定输入上表现不佳。我们首先在一个玩具示例中通过创建一个后门手写数字分类器来探索BadNets的属性。
【机器学习】FastGPT 知识库搜索测试功能解析
最新发布
Linux猿
07-20 605
搜索测试功能包含三种类型：语义检索、全文检索、混合检索。使用向量进行文本相关性查询，即调用向量数据库根据向量的相似性检索；使用传统的全文检索，适合查找一些关键词和主调语特殊的数据；使用向量检素与全文检素的综合结果返回，使用RRF算法进行排序。默认情况下使用语义检索，如下所示。搜索测试前可以根据“知识库搜索配置”进行相关参数的配置，例如：搜索模式、搜索过滤、问题优化等。
说出"BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain"的不足
06-09
"BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain"的不足包括： 1. 对攻击类型的分类不够全面，只考虑了攻击者添加恶意后门的情况，没有考虑到其他类型的攻击，如数据篡改、模型逆向等。 2. 模型的鲁棒性测试不够全面，只测试了对抗性样本的鲁棒性，没有考虑到其他类型的攻击，如对模型进行篡改或逆向工程的攻击。 3. 对模型的审计方法不够全面，只考虑了对模型架构的审计，没有考虑到对数据集、训练过程等其他方面进行审计。 4. 模型供应链的控制能力有限，只考虑了对原始模型的审计，没有考虑到对模型训练数据的审计，这可能导致攻击者通过控制数据集来攻击模型。

“相关推荐”对你有帮助么？

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论 1
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值