├── .gitignore ├── LICENSE ├── README.md ├── attack_api.py ├── eval.py ├── main.py ├── nets.py ├── train_scratch.py └── utils.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | pip-wheel-metadata/ 24 | share/python-wheels/ 25 | *.egg-info/ 26 | .installed.cfg 27 | *.egg 28 | MANIFEST 29 | 30 | # PyInstaller 31 | # Usually these files are written by a python script from a template 32 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 33 | *.manifest 34 | *.spec 35 | 36 | # Installer logs 37 | pip-log.txt 38 | pip-delete-this-directory.txt 39 | 40 | # Unit test / coverage reports 41 | htmlcov/ 42 | .tox/ 43 | .nox/ 44 | .coverage 45 | .coverage.* 46 | .cache 47 | nosetests.xml 48 | coverage.xml 49 | *.cover 50 | *.py,cover 51 | .hypothesis/ 52 | .pytest_cache/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | db.sqlite3-journal 63 | 64 | # Flask stuff: 65 | instance/ 66 | .webassets-cache 67 | 68 | # Scrapy stuff: 69 | .scrapy 70 | 71 | # Sphinx documentation 72 | docs/_build/ 73 | 74 | # PyBuilder 75 | target/ 76 | 77 | # Jupyter Notebook 78 | .ipynb_checkpoints 79 | 80 | # IPython 81 | profile_default/ 82 | ipython_config.py 83 | 84 | # pyenv 85 | .python-version 86 | 87 | # pipenv 88 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 89 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 90 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 91 | # install all needed dependencies. 92 | #Pipfile.lock 93 | 94 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow 95 | __pypackages__/ 96 | 97 | # Celery stuff 98 | celerybeat-schedule 99 | celerybeat.pid 100 | 101 | # SageMath parsed files 102 | *.sage.py 103 | 104 | # Environments 105 | .env 106 | .venv 107 | env/ 108 | venv/ 109 | ENV/ 110 | env.bak/ 111 | venv.bak/ 112 | 113 | # Spyder project settings 114 | .spyderproject 115 | .spyproject 116 | 117 | # Rope project settings 118 | .ropeproject 119 | 120 | # mkdocs documentation 121 | /site 122 | 123 | # mypy 124 | .mypy_cache/ 125 | .dmypy.json 126 | dmypy.json 127 | 128 | # Pyre type checker 129 | .pyre/ 130 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Towards Efficient Data Free Black-Box Adversarial Attack 2 | Official PyTorch implementation of [Towards Efficient Data Free Black-Box Adversarial Attack](https://openaccess.thecvf.com/content/CVPR2022/papers/Zhang_Towards_Efficient_Data_Free_Black-Box_Adversarial_Attack_CVPR_2022_paper.pdf) (CVPR 2022) 3 | 4 | **Abstract**: 5 | 6 | image 7 | 8 | 9 | > Classic black-box adversarial attacks can take advantage of transferable adversarial examples generated by a similar substitute model to successfully fool the target model. However, these substitute models need to be trained by target models' training data, which is hard to acquire due to privacy or transmission reasons. Recognizing the limited availability of real data for adversarial queries, recent works proposed to train substitute models in a data-free black-box scenario. However, their generative adversarial networks (GANs) based framework suffers from the convergence failure and the model collapse, resulting in low efficiency. In this paper, by rethinking the collaborative relationship between the generator and the substitute model, we design a novel black-box attack framework. The proposed method can efficiently imitate the target model through a small number of queries and achieve high attack success rate. The comprehensive experiments over six datasets demonstrate the effectiveness of our method against the state-of-the-art attacks. Especially, we conduct both label-only and probability-only attacks on the Microsoft Azure online model, and achieve a 100% attack success rate with only 0.46% query budget of the SOTA method. 10 | 11 | 12 | 13 | ## How to run 14 | 15 | Experiments of original paper: 16 | 17 | - Train the substitute model. 18 | If you want to train a substitute model in MNIST: 19 | ```python 20 | python3 train_scratch.py --dataset=mnist --epoch=200 21 | ``` 22 | 23 | 24 | - Generate the adversarial attacks by white-box attacks and transfer them to the attacked model. 25 | When the substitute model is obtained, you can use the following command to evaluate the substitute model: 26 | 27 | ```python 28 | python3 main.py --epochs=400 --save_dir=run/mnist --dataset=mnist --score=1 --other=cnn_mnsit --g_steps=10 29 | ``` 30 | 31 | You can also attack the Microsoft Azure API model by run `attack_api.py`, the downloaded remote model can be seen in [remote_model](https://github.com/zhoumingyi/DaST/blob/master/pretrained/sklearn_mnist_model.pkl). 32 | 33 | 34 | ### score-only: 35 | 36 | | Dataset | Scripts | 37 | | :----: | :----: | 38 | |MNIST | python3 main.py --epochs=400 --save_dir=run/mnist --dataset=mnist --score=1 --other=cnn_mnsit --g_steps=10 | 39 | |Fashion-MNIST| python3 main.py --epochs=400 --save_dir=run/fmnist --dataset=fmnist --score=1 --other=cnn_fmnsit --g_steps=10| 40 | |SVHN | python3 main.py --epochs=400 --save_dir=run/svhn --dataset=svhn --score=1 --other=cnn_svhn --g_steps=10| 41 | |CIFAR10 | python3 main.py --epochs=2000 --save_dir=run/cifar10 --dataset=cifar10 --score=1 --other=cnn_cifar10 --g_steps=5| 42 | |CIFAR100 | python3 main.py --epochs=2000 --save_dir=run/cifar100 --dataset=cifar100 --score=1 --other=cnn_cifar100 --g_steps=5 --batch_size=1000 | 43 | |Tiny-Imagenet | python3 main.py --epochs=2000 --save_dir=run/tiny --dataset=tiny --score=1 --other=cnn_tiny --g_steps=5 --batch_size=800 | 44 | 45 | ### label-only： 46 | Just set `--score=0` 47 | 48 | 49 | An example of the training: 50 | 51 | image 52 | 53 | 54 | ## Citation 55 | ``` 56 | @inproceedings{zhang2022towards, 57 | title={Towards Efficient Data Free Black-Box Adversarial Attack}, 58 | author={Zhang, Jie and Li, Bo and Xu, Jianghe and Wu, Shuang and Ding, Shouhong and Zhang, Lei and Wu, Chao}, 59 | booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, 60 | pages={15115--15125}, 61 | year={2022} 62 | } 63 | ``` 64 | -------------------------------------------------------------------------------- /attack_api.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | # Python version: 3.6 4 | import argparse 5 | import os 6 | from tensorboardX import SummaryWriter 7 | import numpy as np 8 | import torch.optim as optim 9 | import warnings 10 | from tqdm import tqdm 11 | from torch.nn.functional import mse_loss 12 | import random 13 | from torchvision import transforms 14 | from kornia import augmentation 15 | import torch 16 | import torch.nn.functional as F 17 | import torch.utils.data.sampler as sp 18 | import torch.backends.cudnn as cudnn 19 | 20 | from nets import Generator_2 21 | from utils import ScoreLoss, ImagePool, MultiTransform, reset_model, get_dataset, cal_prob, cal_label, setup_seed, \ 22 | get_model, print_log, test, test_robust, save_checkpoint 23 | import joblib 24 | from advertorch.attacks import LinfBasicIterativeAttack 25 | 26 | 27 | warnings.filterwarnings('ignore') 28 | 29 | 30 | class Synthesizer(): 31 | def __init__(self, generator, nz, num_classes, img_size, 32 | iterations, lr_g, 33 | sample_batch_size, save_dir, dataset): 34 | super(Synthesizer, self).__init__() 35 | self.img_size = img_size 36 | self.iterations = iterations 37 | self.lr_g = lr_g 38 | self.nz = nz 39 | self.score_loss = ScoreLoss() 40 | self.num_classes = num_classes 41 | self.sample_batch_size = sample_batch_size 42 | self.save_dir = save_dir 43 | self.data_pool = ImagePool(root=self.save_dir) 44 | self.data_iter = None 45 | self.dataset = dataset 46 | 47 | self.generator = generator.cuda().train() 48 | 49 | self.aug = MultiTransform([ 50 | # global view 51 | transforms.Compose([ 52 | augmentation.RandomCrop(size=[self.img_size[-2], self.img_size[-1]], padding=4), 53 | augmentation.RandomHorizontalFlip(), 54 | ]), 55 | # local view 56 | transforms.Compose([ 57 | augmentation.RandomResizedCrop(size=[self.img_size[-2], self.img_size[-1]], scale=[0.25, 1.0]), 58 | augmentation.RandomHorizontalFlip(), 59 | ]), 60 | ]) 61 | # ======================= 62 | if not ("cifar" in dataset): 63 | self.transform = transforms.Compose( 64 | [ 65 | transforms.RandomHorizontalFlip(), 66 | transforms.ToTensor(), 67 | ]) 68 | else: 69 | self.transform = transforms.Compose( 70 | [ 71 | transforms.RandomCrop(32, padding=4), 72 | transforms.RandomHorizontalFlip(), 73 | transforms.ToTensor(), 74 | ]) 75 | 76 | def get_data(self): 77 | datasets = self.data_pool.get_dataset(transform=self.transform) # 获取程序运行到现在所有的图片 78 | self.data_loader = torch.utils.data.DataLoader( 79 | datasets, batch_size=256, shuffle=True, 80 | num_workers=4, pin_memory=True, ) 81 | return self.data_loader 82 | 83 | def gen_data(self, student): 84 | student.eval() 85 | best_cost = 1e6 86 | best_inputs = None 87 | z = torch.randn(size=(self.sample_batch_size, self.nz)).cuda() # 88 | z.requires_grad = True 89 | targets = torch.randint(low=0, high=self.num_classes, size=(self.sample_batch_size,)) 90 | targets = targets.sort()[0] 91 | targets = targets.cuda() 92 | reset_model(self.generator) 93 | optimizer = torch.optim.Adam(self.generator.parameters(), self.lr_g, betas=[0.5, 0.999]) 94 | 95 | for it in range(self.iterations): 96 | optimizer.zero_grad() 97 | inputs = self.generator(z) # bs,nz 98 | global_view, _ = self.aug(inputs) # crop and normalize 99 | 100 | s_out = student(global_view) 101 | loss = self.score_loss(s_out, targets) # ce_loss 102 | if best_cost > loss.item() or best_inputs is None: 103 | best_cost = loss.item() 104 | best_inputs = inputs.data 105 | 106 | loss.backward() 107 | optimizer.step() 108 | # with tqdm(total=self.iterations) as t: 109 | 110 | # optimizer_mlp.step() 111 | # t.set_description('iters:{}, loss:{}'.format(it, loss.item())) 112 | 113 | # save best inputs and reset data iter 114 | self.data_pool.add(best_inputs) # 生成了一个batch的数据 115 | 116 | 117 | def args_parser(): 118 | parser = argparse.ArgumentParser() 119 | # federated arguments (Notation for the arguments followed from paper) 120 | parser.add_argument('--epochs', type=int, default=10, 121 | help="number of rounds of training") 122 | parser.add_argument('--score', type=float, default=0, 123 | help="number of rounds of training") 124 | parser.add_argument('--lr', type=float, default=0.01, 125 | help='learning rate') 126 | parser.add_argument('--momentum', type=float, default=0.9, 127 | help='SGD momentum (default: 0.5)') 128 | # other arguments 129 | parser.add_argument('--dataset', type=str, default='mnist', help="name \ 130 | of dataset") 131 | # Data Free 132 | 133 | parser.add_argument('--save_dir', default='run/mnist', type=str) 134 | 135 | # Basic 136 | parser.add_argument('--lr_g', default=1e-3, type=float, 137 | help='initial learning rate for generation') 138 | parser.add_argument('--g_steps', default=30, type=int, metavar='N', 139 | help='number of iterations for generation') 140 | parser.add_argument('--batch_size', default=256, type=int, metavar='N', 141 | help='number of total iterations in each epoch') 142 | parser.add_argument('--nz', default=256, type=int, metavar='N', 143 | help='number of total iterations in each epoch') 144 | parser.add_argument('--synthesis_batch_size', default=256, type=int) 145 | # Misc 146 | parser.add_argument('--seed', default=2021, type=int, 147 | help='seed for initializing training.') 148 | parser.add_argument('--type', default="score", type=str, 149 | help='score or label') 150 | parser.add_argument('--model', default="", type=str, 151 | help='seed for initializing training.') 152 | parser.add_argument('--other', default="", type=str, 153 | help='seed for initializing training.') 154 | args = parser.parse_args() 155 | return args 156 | 157 | 158 | def kd_train(synthesizer, model, optimizer, score_val): 159 | sub_net, blackBox_net = model 160 | sub_net.train() 161 | # blackBox_net.eval() 162 | 163 | # with tqdm(synthesizer.get_data()) as epochs: 164 | data = synthesizer.get_data() 165 | for idx, (images) in enumerate(data): 166 | optimizer.zero_grad() 167 | images = images.cuda() 168 | with torch.no_grad(): 169 | original_score = cal_azure_proba(blackBox_net, images) 170 | label = torch.max(original_score.data, 1)[1] 171 | 172 | # original_score = cal_prob(blackBox_net, images) # prob 173 | substitute_outputs = sub_net(images.detach()) 174 | substitute_score = F.softmax(substitute_outputs, dim=1) 175 | loss_mse = mse_loss( 176 | substitute_score, original_score, reduction='mean') 177 | # label = cal(blackBox_net, images) # label 178 | loss_ce = F.cross_entropy(substitute_outputs, label) 179 | # ============================== 180 | # idx = torch.where(substitute_outputs.max(1)[1] != label)[0] 181 | # loss_adv = F.cross_entropy(substitute_outputs[idx], label[idx]) 182 | # ============================== 183 | loss = loss_ce + loss_mse * score_val 184 | 185 | loss.backward() 186 | optimizer.step() 187 | 188 | 189 | def cal_azure(model, data): 190 | data = data.view(data.size(0), 784).cpu().numpy() 191 | output = model.predict(data) 192 | output = torch.from_numpy(output).cuda().long() 193 | return output 194 | 195 | 196 | def cal_azure_proba(model, data): 197 | data = data.view(data.size(0), 784).cpu().numpy() 198 | output = model.predict_proba(data) 199 | output = torch.from_numpy(output).cuda().float() 200 | return output 201 | 202 | if __name__ == '__main__': 203 | dir = './saved/api' 204 | if not os.path.exists(dir): 205 | os.mkdir(dir) 206 | 207 | args = args_parser() 208 | setup_seed(args.seed) 209 | train_loader, test_loader = get_dataset(args.dataset) 210 | 211 | public = dir + '/logs_{}_{}'.format(args.dataset, str(args.score)) 212 | if not os.path.exists(public): 213 | os.mkdir(public) 214 | log = open('{}/log_ours.txt'.format(public), 'w') 215 | 216 | list = [i for i in range(0, len(test_loader.dataset))] 217 | data_list = random.sample(list, 1024) 218 | val_loader = torch.utils.data.DataLoader(test_loader.dataset, batch_size=128, 219 | sampler=sp.SubsetRandomSampler(data_list), num_workers=4) 220 | 221 | tf_writer = SummaryWriter(log_dir=public) 222 | sub_net, _ = get_model(args.dataset, 0) 223 | clf = joblib.load('pretrained/sklearn_mnist_model.pkl') 224 | 225 | with torch.no_grad(): 226 | correct_netD = 0.0 227 | total = 0.0 228 | for inputs, labels in val_loader: 229 | inputs, labels = inputs.cuda(), labels.cuda() 230 | predicted = cal_azure(clf, inputs) 231 | total += labels.size(0) 232 | correct_netD += (predicted == labels).sum() 233 | print('Accuracy of the black-box network : %.2f %%' % 234 | (100. * correct_netD.float() / total)) 235 | ################################################ 236 | # estimate the attack success rate of initial D: 237 | ################################################ 238 | correct_ghost = 0.0 239 | total = 0.0 240 | sub_net.eval() 241 | adversary_ghost = LinfBasicIterativeAttack( 242 | sub_net, loss_fn=torch.nn.CrossEntropyLoss(reduction="sum"), eps=0.3, 243 | nb_iter=100, eps_iter=0.01, clip_min=0.0, clip_max=1.0, 244 | targeted=False) 245 | for inputs, labels in val_loader: 246 | inputs, labels = inputs.cuda(), labels.cuda() 247 | adv_inputs_ghost = adversary_ghost.perturb(inputs, labels) 248 | with torch.no_grad(): 249 | predicted = cal_azure(clf, adv_inputs_ghost) 250 | total += labels.size(0) 251 | correct_ghost += (predicted == labels).sum() 252 | print('Attack success rate: %.2f %%' % 253 | (100 - 100. * correct_ghost.float() / total)) 254 | 255 | ################################################ 256 | # data generator 257 | ################################################ 258 | nz = args.nz 259 | nc = 1 260 | 261 | img_size = 28 262 | img_size2 = (1, 28, 28) 263 | 264 | generator = Generator_2(nz=nz, ngf=64, img_size=img_size, nc=nc).cuda() 265 | 266 | num_class = 10 267 | 268 | synthesizer = Synthesizer(generator, 269 | nz=nz, 270 | num_classes=num_class, 271 | img_size=img_size2, 272 | iterations=args.g_steps, 273 | lr_g=args.lr_g, 274 | sample_batch_size=args.batch_size, 275 | save_dir=args.save_dir, 276 | dataset=args.dataset) 277 | # &&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&& 278 | optimizer = optim.SGD(sub_net.parameters(), lr=args.lr, momentum=args.momentum) 279 | sub_net.train() 280 | best_acc = -1 281 | best_asr = -1 282 | best_acc_ckpt = '{}/{}_ours_acc.pth'.format(public, args.dataset) 283 | for epoch in tqdm(range(args.epochs)): 284 | # 1. Data synthesis 285 | synthesizer.gen_data(sub_net) # g_steps 286 | kd_train(synthesizer, [sub_net, clf], optimizer, args.score) 287 | if epoch % 1 == 0: # 250*40, 250*10=2.5k 288 | acc, test_loss = test(sub_net, val_loader) 289 | ################################################ 290 | # estimate the attack success rate of initial D: 291 | ################################################ 292 | correct_ghost = 0.0 293 | total = 0.0 294 | sub_net.eval() 295 | adversary_ghost = LinfBasicIterativeAttack( 296 | sub_net, loss_fn=torch.nn.CrossEntropyLoss(reduction="sum"), eps=0.3, 297 | nb_iter=100, eps_iter=0.01, clip_min=0.0, clip_max=1.0, 298 | targeted=False) 299 | for inputs, labels in val_loader: 300 | inputs, labels = inputs.cuda(), labels.cuda() 301 | adv_inputs_ghost = adversary_ghost.perturb(inputs, labels) 302 | with torch.no_grad(): 303 | predicted = cal_azure(clf, adv_inputs_ghost) 304 | total += labels.size(0) 305 | correct_ghost += (predicted == labels).sum() 306 | asr = (100 - 100. * correct_ghost.float() / total) 307 | print('Attack success rate: %.2f %%' % asr) 308 | save_checkpoint({ 309 | 'state_dict': sub_net.state_dict(), 310 | 'epoch': epoch, 311 | }, acc > best_acc, best_acc_ckpt) 312 | 313 | best_asr = max(best_asr, asr) 314 | best_acc = max(best_acc, acc) 315 | 316 | print_log("Accuracy of the substitute model:{:.3} %, best accuracy:{:.3} % \n".format(acc, best_acc), log) 317 | print_log("ASR:{:.3} %, best asr:{:.3} %\n".format(asr, best_asr), log) 318 | log.flush() 319 | 320 | """ 321 | 322 | CUDA_VISIBLE_DEVICES=2 python3 attack_api.py --epochs=100 --save_dir=run/mnist_1 --dataset=mnist --score=1 --other=cnn_mnsit --g_steps=10 323 | 324 | """ 325 | -------------------------------------------------------------------------------- /eval.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | 3 | import argparse 4 | import random 5 | 6 | import torch 7 | import torch.backends.cudnn as cudnn 8 | import torch.nn as nn 9 | import torch.nn.parallel 10 | import torch.utils.data 11 | import torch.utils.data.sampler as sp 12 | from advertorch.attacks import GradientSignAttack, PGDAttack, LinfPGDAttack 13 | from advertorch.attacks import LinfBasicIterativeAttack, CarliniWagnerL2Attack 14 | 15 | 16 | from utils import get_model, test, setup_seed, get_dataset 17 | 18 | 19 | def test_adver(net, tar_net, attack, target, testloader, dataset): 20 | if dataset == "mnist": 21 | cfgs = dict(test_step_size=0.01, test_epsilon=0.3) 22 | elif dataset == "cifar10" or dataset == "cifar100": 23 | cfgs = dict(test_step_size=2.0 / 255, test_epsilon=8.0 / 255) 24 | elif dataset == "fmnist": 25 | cfgs = dict(test_step_size=0.01, test_epsilon=0.3) 26 | elif dataset == "svhn" or dataset == "tiny": 27 | cfgs = dict(test_step_size=2.0 / 255, test_epsilon=8.0 / 255) 28 | 29 | 30 | net.eval() 31 | tar_net.eval() 32 | # BIM 33 | if attack == 'BIM': 34 | adversary = LinfBasicIterativeAttack( 35 | net, 36 | loss_fn=nn.CrossEntropyLoss(reduction="sum"), 37 | eps=cfgs['test_epsilon'], 38 | nb_iter=120, eps_iter=cfgs['test_step_size'], clip_min=0.0, clip_max=1.0, 39 | targeted=target) 40 | # PGD 41 | elif attack == 'PGD': 42 | if target: 43 | adversary = PGDAttack( 44 | net, 45 | loss_fn=nn.CrossEntropyLoss(reduction="sum"), 46 | eps=cfgs['test_epsilon'], 47 | nb_iter=20, eps_iter=cfgs['test_step_size'], clip_min=0.0, clip_max=1.0, 48 | targeted=target) 49 | else: 50 | adversary = PGDAttack( 51 | net, 52 | loss_fn=nn.CrossEntropyLoss(reduction="sum"), 53 | eps=cfgs['test_epsilon'], 54 | nb_iter=20, eps_iter=cfgs['test_step_size'], clip_min=0.0, clip_max=1.0, 55 | targeted=target) 56 | # FGSM 57 | elif attack == 'FGSM': 58 | adversary = GradientSignAttack( 59 | net, 60 | loss_fn=nn.CrossEntropyLoss(reduction="sum"), 61 | eps=cfgs['test_epsilon'], 62 | targeted=target) 63 | elif attack == 'CW': 64 | adversary = CarliniWagnerL2Attack( 65 | net, 66 | num_classes=10, 67 | learning_rate=0.45, 68 | binary_search_steps=10, 69 | max_iterations=20, 70 | targeted=target) 71 | 72 | # ---------------------------------- 73 | # Obtain the attack success rate of the model 74 | # ---------------------------------- 75 | 76 | correct = 0.0 77 | total = 0.0 78 | tar_net.eval() 79 | total_L2_distance = 0.0 80 | for data in testloader: 81 | inputs, labels = data 82 | inputs, labels = inputs.cuda(), labels.cuda() 83 | outputs = tar_net(inputs) 84 | _, predicted = torch.max(outputs.data, 1) 85 | if target: 86 | # randomly choose the specific label of targeted attack 87 | labels = torch.randint(0, 9, (inputs.shape[0],)).cuda() 88 | # test the images which are not classified as the specific label 89 | idx = torch.where(predicted != labels)[0] 90 | adv_inputs_ori = adversary.perturb(inputs[idx], labels[idx]) 91 | L2_distance = (torch.norm(adv_inputs_ori - inputs[idx])).item() 92 | total_L2_distance += L2_distance 93 | with torch.no_grad(): 94 | outputs = tar_net(adv_inputs_ori) 95 | _, predicted = torch.max(outputs.data, 1) 96 | total += labels.size(0) 97 | correct += (predicted == labels[idx]).sum() 98 | else: 99 | # test the images which are classified correctly 100 | idx = torch.where(predicted == labels)[0] 101 | adv_inputs_ori = adversary.perturb(inputs[idx], labels[idx]) 102 | L2_distance = (torch.norm(adv_inputs_ori - inputs[idx])).item() 103 | total_L2_distance += L2_distance 104 | with torch.no_grad(): 105 | outputs = tar_net(adv_inputs_ori) 106 | _, predicted = torch.max(outputs.data, 1) 107 | 108 | total += labels.size(0) 109 | correct += (predicted == labels[idx]).sum() 110 | 111 | asr = 100. * correct.float() / total if target else 100.0 - 100. * correct.float() / total 112 | return asr 113 | 114 | 115 | if __name__ == '__main__': 116 | setup_seed(2021) 117 | parser = argparse.ArgumentParser() 118 | # parser.add_argument('--target', type=int, ) 119 | parser.add_argument('--dataset', type=str, ) 120 | parser.add_argument('--dir', type=str, ) 121 | 122 | opt = parser.parse_args() 123 | cudnn.benchmark = True 124 | 125 | train_loader, test_loader = get_dataset(opt.dataset) 126 | list = [i for i in range(0, len(test_loader.dataset))] 127 | data_list = random.sample(list, 1024) 128 | val_loader = torch.utils.data.DataLoader(test_loader.dataset, batch_size=64, 129 | sampler=sp.SubsetRandomSampler(data_list), num_workers=4) 130 | 131 | sub_net, _ = get_model(opt.dataset, 0) 132 | state_dict_1 = torch.load(opt.dir)['state_dict'] 133 | sub_net.load_state_dict(state_dict_1) 134 | 135 | blackBox_net, state_dict = get_model(opt.dataset, 1) 136 | blackBox_net.load_state_dict(state_dict) 137 | 138 | acc, _ = test(sub_net, val_loader) 139 | print("Accuracy of the sub_net:{:.3} % \n".format(acc)) 140 | 141 | # asr = test_adver(sub_net, blackBox_net, 'CW', 'Untarget' == 'Target', val_loader, opt.dataset) 142 | # print('Untarget' + " , " + "type: " + 'CW' + ", ASR:{:.2f} %, ".format(asr)) 143 | for attack in ['Target', 'Untarget']: 144 | for adv in ['FGSM', 'BIM', 'PGD']: 145 | asr = test_adver(sub_net, blackBox_net, adv, attack == 'Target', val_loader, opt.dataset) 146 | print(attack + " , " + "type: " + adv + ", ASR:{:.2f} %, ".format(asr)) 147 | 148 | # python3 eval_rob.py --dataset=mnist --dir= -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # -*- coding: utf-8 -*- 3 | # Python version: 3.6 4 | import argparse 5 | import os 6 | from tensorboardX import SummaryWriter 7 | import numpy as np 8 | import torch.optim as optim 9 | import warnings 10 | from tqdm import tqdm 11 | from torch.nn.functional import mse_loss 12 | import random 13 | from torchvision import transforms 14 | from kornia import augmentation 15 | import torch 16 | import torch.nn.functional as F 17 | import torch.utils.data.sampler as sp 18 | import torch.backends.cudnn as cudnn 19 | 20 | from nets import Generator_2 21 | from utils import ScoreLoss, ImagePool, MultiTransform, reset_model, get_dataset, cal_prob, cal_label, setup_seed, \ 22 | get_model, print_log, test, test_robust, save_checkpoint 23 | 24 | warnings.filterwarnings('ignore') 25 | 26 | 27 | class Synthesizer(): 28 | def __init__(self, generator, nz, num_classes, img_size, 29 | iterations, lr_g, 30 | sample_batch_size, save_dir, dataset): 31 | super(Synthesizer, self).__init__() 32 | self.img_size = img_size 33 | self.iterations = iterations 34 | self.lr_g = lr_g 35 | self.nz = nz 36 | self.score_loss = ScoreLoss() 37 | self.num_classes = num_classes 38 | self.sample_batch_size = sample_batch_size 39 | self.save_dir = save_dir 40 | self.data_pool = ImagePool(root=self.save_dir) 41 | self.data_iter = None 42 | self.dataset = dataset 43 | 44 | self.generator = generator.cuda().train() 45 | 46 | self.aug = MultiTransform([ 47 | # global view 48 | transforms.Compose([ 49 | augmentation.RandomCrop(size=[self.img_size[-2], self.img_size[-1]], padding=4), 50 | augmentation.RandomHorizontalFlip(), 51 | ]), 52 | # local view 53 | transforms.Compose([ 54 | augmentation.RandomResizedCrop(size=[self.img_size[-2], self.img_size[-1]], scale=[0.25, 1.0]), 55 | augmentation.RandomHorizontalFlip(), 56 | ]), 57 | ]) 58 | # ======================= 59 | if not ("cifar" in dataset): 60 | self.transform = transforms.Compose( 61 | [ 62 | transforms.RandomHorizontalFlip(), 63 | transforms.ToTensor(), 64 | ]) 65 | else: 66 | self.transform = transforms.Compose( 67 | [ 68 | transforms.RandomCrop(32, padding=4), 69 | transforms.RandomHorizontalFlip(), 70 | transforms.ToTensor(), 71 | ]) 72 | 73 | def get_data(self): 74 | datasets = self.data_pool.get_dataset(transform=self.transform) # 获取程序运行到现在所有的图片 75 | self.data_loader = torch.utils.data.DataLoader( 76 | datasets, batch_size=256, shuffle=True, 77 | num_workers=4, pin_memory=True, ) 78 | return self.data_loader 79 | 80 | def gen_data(self, student): 81 | student.eval() 82 | best_cost = 1e6 83 | best_inputs = None 84 | z = torch.randn(size=(self.sample_batch_size, self.nz)).cuda() # 85 | z.requires_grad = True 86 | targets = torch.randint(low=0, high=self.num_classes, size=(self.sample_batch_size,)) 87 | targets = targets.sort()[0] 88 | targets = targets.cuda() 89 | reset_model(self.generator) 90 | optimizer = torch.optim.Adam(self.generator.parameters(), self.lr_g, betas=[0.5, 0.999]) 91 | 92 | for it in range(self.iterations): 93 | optimizer.zero_grad() 94 | inputs = self.generator(z) # bs,nz 95 | global_view, _ = self.aug(inputs) # crop and normalize 96 | 97 | s_out = student(global_view) 98 | loss = self.score_loss(s_out, targets) # ce_loss 99 | if best_cost > loss.item() or best_inputs is None: 100 | best_cost = loss.item() 101 | best_inputs = inputs.data 102 | 103 | loss.backward() 104 | optimizer.step() 105 | # with tqdm(total=self.iterations) as t: 106 | 107 | # optimizer_mlp.step() 108 | # t.set_description('iters:{}, loss:{}'.format(it, loss.item())) 109 | 110 | # save best inputs and reset data iter 111 | self.data_pool.add(best_inputs) # 生成了一个batch的数据 112 | 113 | 114 | def args_parser(): 115 | parser = argparse.ArgumentParser() 116 | # federated arguments (Notation for the arguments followed from paper) 117 | parser.add_argument('--epochs', type=int, default=10, 118 | help="number of rounds of training") 119 | parser.add_argument('--score', type=float, default=0, 120 | help="number of rounds of training") 121 | parser.add_argument('--lr', type=float, default=0.01, 122 | help='learning rate') 123 | parser.add_argument('--momentum', type=float, default=0.9, 124 | help='SGD momentum (default: 0.5)') 125 | # other arguments 126 | parser.add_argument('--dataset', type=str, default='mnist', help="name \ 127 | of dataset") 128 | # Data Free 129 | 130 | parser.add_argument('--save_dir', default='run/mnist', type=str) 131 | 132 | # Basic 133 | parser.add_argument('--lr_g', default=1e-3, type=float, 134 | help='initial learning rate for generation') 135 | parser.add_argument('--g_steps', default=30, type=int, metavar='N', 136 | help='number of iterations for generation') 137 | parser.add_argument('--batch_size', default=256, type=int, metavar='N', 138 | help='number of total iterations in each epoch') 139 | parser.add_argument('--nz', default=256, type=int, metavar='N', 140 | help='number of total iterations in each epoch') 141 | parser.add_argument('--synthesis_batch_size', default=256, type=int) 142 | # Misc 143 | parser.add_argument('--seed', default=2021, type=int, 144 | help='seed for initializing training.') 145 | parser.add_argument('--type', default="score", type=str, 146 | help='score or label') 147 | parser.add_argument('--model', default="", type=str, 148 | help='seed for initializing training.') 149 | parser.add_argument('--other', default="", type=str, 150 | help='seed for initializing training.') 151 | args = parser.parse_args() 152 | return args 153 | 154 | 155 | def kd_train(synthesizer, model, optimizer, score_val): 156 | sub_net, blackBox_net = model 157 | sub_net.train() 158 | blackBox_net.eval() 159 | 160 | # with tqdm(synthesizer.get_data()) as epochs: 161 | data = synthesizer.get_data() 162 | for idx, (images) in enumerate(data): 163 | optimizer.zero_grad() 164 | images = images.cuda() 165 | original_score = cal_prob(blackBox_net, images) # prob 166 | substitute_outputs = sub_net(images.detach()) 167 | substitute_score = F.softmax(substitute_outputs, dim=1) 168 | loss_mse = mse_loss( 169 | substitute_score, original_score, reduction='mean') 170 | label = cal_label(blackBox_net, images) # label 171 | loss_ce = F.cross_entropy(substitute_outputs, label) 172 | # ============================== 173 | # idx = torch.where(substitute_outputs.max(1)[1] != label)[0] 174 | # loss_adv = F.cross_entropy(substitute_outputs[idx], label[idx]) 175 | # ============================== 176 | loss = loss_ce + loss_mse * score_val 177 | 178 | loss.backward() 179 | optimizer.step() 180 | # return loss.item() 181 | 182 | 183 | 184 | if __name__ == '__main__': 185 | dir = './saved/ours' 186 | if not os.path.exists(dir): 187 | os.mkdir(dir) 188 | 189 | args = args_parser() 190 | setup_seed(args.seed) 191 | train_loader, test_loader = get_dataset(args.dataset) 192 | 193 | public = dir + '/logs_{}_{}'.format(args.dataset, str(args.score)) 194 | if not os.path.exists(public): 195 | os.mkdir(public) 196 | log = open('{}/log_ours.txt'.format(public), 'w') 197 | 198 | list = [i for i in range(0, len(test_loader.dataset))] 199 | data_list = random.sample(list, 1024) 200 | val_loader = torch.utils.data.DataLoader(test_loader.dataset, batch_size=128, 201 | sampler=sp.SubsetRandomSampler(data_list), num_workers=4) 202 | 203 | tf_writer = SummaryWriter(log_dir=public) 204 | sub_net, _ = get_model(args.dataset, 0) 205 | blackBox_net, state_dict = get_model(args.dataset, 1) 206 | blackBox_net.load_state_dict(state_dict) 207 | 208 | print_log("===================================== \n", log) 209 | acc, _ = test(blackBox_net, val_loader) 210 | print_log("Accuracy of the black-box model:{:.3} % \n".format(acc), log) 211 | acc, _ = test(sub_net, val_loader) 212 | print_log("Accuracy of the substitute model:{:.3} % \n".format(acc), log) 213 | asr, val_acc = 0.0, 0.0 # test_robust(val_loader, sub_net, blackBox_net, args.dataset) 214 | print_log("ASR:{:.3} %, val acc:{:.3} % \n".format(asr, val_acc), log) 215 | print_log("===================================== \n", log) 216 | log.flush() 217 | 218 | ################################################ 219 | # data generator 220 | ################################################ 221 | nz = args.nz 222 | nc = 3 if "cifar" in args.dataset or args.dataset == "svhn" or args.dataset == "tiny" else 1 223 | # img_size = 32 if "cifar" in args.dataset or args.dataset == "svhn" else 28 224 | 225 | if "cifar" in args.dataset or args.dataset == "svhn": 226 | img_size = 32 227 | elif "mnist" in args.dataset: 228 | img_size = 28 229 | elif args.dataset == "tiny": 230 | img_size = 64 231 | 232 | if "cifar" in args.dataset or args.dataset == "svhn": 233 | img_size2 = (3, 32, 32) 234 | elif "mnist" in args.dataset: 235 | img_size2 = (1, 28, 28) 236 | elif args.dataset == "tiny": 237 | img_size2 = (3, 64, 64) 238 | generator = Generator_2(nz=nz, ngf=64, img_size=img_size, nc=nc).cuda() 239 | # ==================== 240 | sub_net = torch.nn.DataParallel(sub_net) 241 | blackBox_net = torch.nn.DataParallel(blackBox_net) 242 | generator = torch.nn.DataParallel(generator) 243 | # ==================== 244 | 245 | args.cur_ep = 0 246 | # img_size2 = ( 247 | # 3, 32, 32) if "cifar" in args.dataset or args.dataset == "svhn" else (1, 28, 28) 248 | 249 | if args.dataset == "cifar100": 250 | num_class = 100 251 | elif args.dataset == "tiny": 252 | num_class = 200 253 | else: 254 | num_class = 10 255 | 256 | synthesizer = Synthesizer(generator, 257 | nz=nz, 258 | num_classes=num_class, 259 | img_size=img_size2, 260 | iterations=args.g_steps, 261 | lr_g=args.lr_g, 262 | sample_batch_size=args.batch_size, 263 | save_dir=args.save_dir, 264 | dataset=args.dataset) 265 | # &&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&& 266 | optimizer = optim.SGD(sub_net.parameters(), lr=args.lr, momentum=args.momentum) 267 | sub_net.train() 268 | best_acc = -1 269 | best_asr = -1 270 | 271 | best_acc_ckpt = '{}/{}_ours_acc.pth'.format(public, args.dataset) 272 | best_asr_ckpt = '{}/{}_ours_asr.pth'.format(public, args.dataset) 273 | for epoch in tqdm(range(args.epochs)): 274 | # 1. Data synthesis 275 | synthesizer.gen_data(sub_net) # g_steps 276 | kd_train(synthesizer, [sub_net, blackBox_net], optimizer, args.score) 277 | if epoch % 1 == 0: # 250*40, 250*10=2.5k 278 | acc, test_loss = test(sub_net, val_loader) 279 | asr, val_acc = test_robust(val_loader, sub_net, blackBox_net, args.dataset) 280 | 281 | # save_checkpoint({ 282 | # 'state_dict': sub_net.state_dict(), 283 | # 'epoch': epoch, 284 | # }, acc > best_acc, best_acc_ckpt) 285 | # 286 | # save_checkpoint({ 287 | # 'state_dict': sub_net.state_dict(), 288 | # 'epoch': epoch, 289 | # }, asr > best_asr, best_asr_ckpt) 290 | 291 | best_asr = max(best_asr, asr) 292 | best_acc = max(best_acc, acc) 293 | 294 | print_log("Accuracy of the substitute model:{:.3} %, best accuracy:{:.3} % \n".format(acc, best_acc), log) 295 | print_log("ASR:{:.3} %, best asr:{:.3} %, val acc:{:.3} % \n".format(asr, best_asr, val_acc), log) 296 | log.flush() 297 | 298 | """ 299 | 40*256=1w 300 | CUDA_VISIBLE_DEVICES=2 python3 main.py --epochs=400 --save_dir=run/svhn_1 \ 301 | --dataset=svhn --score=1 --other=cnn_svhn --g_steps=5 302 | 303 | CUDA_VISIBLE_DEVICES=3 python3 main.py --epochs=400 --save_dir=run/svhn_2 \ 304 | --dataset=svhn --score=1 --other=cnn_svhn --g_steps=30 305 | 306 | 307 | CUDA_VISIBLE_DEVICES=2 python3 main.py --epochs=400 --save_dir=run/cifar10 --dataset=cifar10 --score=1 --other=cnn_cifar10 --g_steps=5 308 | 309 | CUDA_VISIBLE_DEVICES=2 python3 main.py --epochs=400 --save_dir=run/mnist_1 --dataset=mnist --score=1 --other=cnn_mnsit --g_steps=10 310 | 311 | CUDA_VISIBLE_DEVICES=1 python3 main.py --epochs=400 --save_dir=run/fmnist_1 --dataset=fmnist --score=1 --other=cnn_fmnsit --g_steps=10 312 | 313 | CUDA_VISIBLE_DEVICES=1 python3 main.py --epochs=400 --save_dir=run/fmnist_2 --dataset=fmnist --score=1 --other=cnn_fmnsit --g_steps=30 314 | """ 315 | -------------------------------------------------------------------------------- /nets.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | import torch.nn.parallel 5 | 6 | nz = 128 7 | nc = 3 8 | 9 | 10 | class pre_conv(nn.Module): 11 | def __init__(self): 12 | super(pre_conv, self).__init__() 13 | self.nf = 64 14 | G_type = 1 15 | 16 | if G_type == 1: 17 | # ipdb.set_trace() 18 | self.pre_conv = nn.Sequential( 19 | nn.Conv2d(nz, self.nf * 2, 3, 1, 1, bias=False), # torch.Size([50, 128, 1, 1]) 20 | nn.BatchNorm2d(self.nf * 2), 21 | nn.LeakyReLU(0.2, inplace=True), 22 | 23 | nn.ConvTranspose2d(self.nf * 2, self.nf * 2, 4, 2, 1, bias=False), # torch.Size([50, 128, 2, 2]) 24 | nn.BatchNorm2d(self.nf * 2), 25 | nn.LeakyReLU(0.2, inplace=True), 26 | 27 | nn.ConvTranspose2d(self.nf * 2, self.nf * 2, 4, 2, 1, bias=False), 28 | nn.BatchNorm2d(self.nf * 2), 29 | nn.LeakyReLU(0.2, inplace=True), 30 | 31 | nn.ConvTranspose2d(self.nf * 2, self.nf * 2, 4, 2, 1, bias=False), 32 | nn.BatchNorm2d(self.nf * 2), 33 | nn.LeakyReLU(0.2, inplace=True), 34 | 35 | nn.ConvTranspose2d(self.nf * 2, self.nf * 2, 4, 2, 1, bias=False), 36 | nn.BatchNorm2d(self.nf * 2), 37 | nn.LeakyReLU(0.2, inplace=True), 38 | 39 | nn.ConvTranspose2d(self.nf * 2, self.nf * 2, 4, 2, 1, bias=False), 40 | nn.BatchNorm2d(self.nf * 2), 41 | nn.LeakyReLU(0.2, inplace=True) 42 | ) 43 | elif G_type == 2: 44 | self.pre_conv = nn.Sequential( 45 | nn.Conv2d(self.nf * 8, self.nf * 8, 3, 1, round((self.shape[0] - 1) / 2), bias=False), 46 | nn.BatchNorm2d(self.nf * 8), 47 | nn.ReLU(True), # added 48 | nn.Conv2d(self.nf * 8, self.nf * 8, 3, 1, round((self.shape[0] - 1) / 2), bias=False), 49 | nn.BatchNorm2d(self.nf * 8), 50 | nn.ReLU(True), 51 | 52 | nn.Conv2d(self.nf * 8, self.nf * 4, 3, 1, 1, bias=False), 53 | nn.BatchNorm2d(self.nf * 4), 54 | nn.ReLU(True), 55 | 56 | nn.Conv2d(self.nf * 4, self.nf * 2, 3, 1, 1, bias=False), 57 | nn.BatchNorm2d(self.nf * 2), 58 | nn.ReLU(True), 59 | 60 | nn.Conv2d(self.nf * 2, self.nf, 3, 1, 1, bias=False), 61 | nn.BatchNorm2d(self.nf), 62 | nn.ReLU(True), 63 | 64 | nn.Conv2d(self.nf, self.shape[0], 3, 1, 1, bias=False), 65 | nn.BatchNorm2d(self.shape[0]), 66 | nn.ReLU(True), 67 | 68 | nn.Conv2d(self.shape[0], self.shape[0], 3, 1, 1, bias=False), 69 | 70 | nn.Sigmoid() 71 | ) 72 | 73 | def forward(self, input): 74 | output = self.pre_conv(input) 75 | return output 76 | 77 | 78 | class Generator(nn.Module): 79 | def __init__(self): 80 | super(Generator, self).__init__() 81 | self.nf = 64 82 | self.num_class = 10 83 | G_type = 1 84 | if G_type == 1: 85 | self.main = nn.Sequential( 86 | nn.Conv2d(self.nf * 2, self.nf * 4, 3, 1, 0, bias=False), 87 | nn.BatchNorm2d(self.nf * 4), 88 | nn.LeakyReLU(0.2, inplace=True), 89 | 90 | nn.Conv2d(self.nf * 4, self.nf * 8, 3, 1, 0, bias=False), 91 | nn.BatchNorm2d(self.nf * 8), 92 | nn.LeakyReLU(0.2, inplace=True), 93 | 94 | nn.Conv2d(self.nf * 8, self.nf * 4, 3, 1, 1, bias=False), 95 | nn.BatchNorm2d(self.nf * 4), 96 | nn.LeakyReLU(0.2, inplace=True), 97 | 98 | nn.Conv2d(self.nf * 4, self.nf * 2, 3, 1, 1, bias=False), 99 | nn.BatchNorm2d(self.nf * 2), 100 | nn.LeakyReLU(0.2, inplace=True), 101 | 102 | nn.Conv2d(self.nf * 2, self.nf, 3, 1, 1, bias=False), 103 | nn.BatchNorm2d(self.nf), 104 | nn.LeakyReLU(0.2, inplace=True), 105 | 106 | nn.Conv2d(self.nf, nc, 3, 1, 1, bias=False), 107 | nn.BatchNorm2d(nc), 108 | nn.LeakyReLU(0.2, inplace=True), 109 | 110 | nn.Conv2d(nc, nc, 3, 1, 1, bias=False), 111 | nn.Sigmoid() 112 | ) 113 | elif G_type == 2: 114 | self.main = nn.Sequential( 115 | nn.Conv2d(nz, self.nf * 2, 3, 1, 1, bias=False), 116 | nn.BatchNorm2d(self.nf * 2), 117 | nn.ReLU(True), 118 | 119 | nn.ConvTranspose2d(self.nf * 2, self.nf * 2, 4, 2, 1, bias=False), 120 | nn.BatchNorm2d(self.nf * 2), 121 | nn.ReLU(True), 122 | 123 | nn.ConvTranspose2d(self.nf * 2, self.nf * 4, 4, 2, 1, bias=False), 124 | nn.BatchNorm2d(self.nf * 4), 125 | nn.ReLU(True), 126 | 127 | nn.ConvTranspose2d(self.nf * 4, self.nf * 4, 4, 2, 1, bias=False), 128 | nn.BatchNorm2d(self.nf * 4), 129 | nn.ReLU(True), 130 | 131 | nn.ConvTranspose2d(self.nf * 4, self.nf * 8, 4, 2, 1, bias=False), 132 | nn.BatchNorm2d(self.nf * 8), 133 | nn.ReLU(True), 134 | 135 | nn.ConvTranspose2d(self.nf * 8, self.nf * 8, 4, 2, 1, bias=False), 136 | nn.BatchNorm2d(self.nf * 8), 137 | nn.ReLU(True), 138 | 139 | nn.Conv2d(self.nf * 8, self.nf * 8, 3, 1, 1, bias=False), 140 | nn.BatchNorm2d(self.nf * 8), 141 | nn.ReLU(True) 142 | ) 143 | 144 | def forward(self, input): 145 | output = self.main(input) 146 | return output 147 | 148 | 149 | class Generator_2(nn.Module): 150 | def __init__(self, nz=100, ngf=64, img_size=32, nc=3): 151 | super(Generator_2, self).__init__() 152 | 153 | self.init_size = img_size // 4 154 | self.l1 = nn.Sequential(nn.Linear(nz, ngf * 2 * self.init_size ** 2)) 155 | 156 | self.conv_blocks = nn.Sequential( 157 | nn.BatchNorm2d(ngf * 2), 158 | nn.Upsample(scale_factor=2), 159 | 160 | nn.Conv2d(ngf * 2, ngf * 2, 3, stride=1, padding=1, bias=False), 161 | nn.BatchNorm2d(ngf * 2), 162 | nn.LeakyReLU(0.2, inplace=True), 163 | nn.Upsample(scale_factor=2), 164 | 165 | nn.Conv2d(ngf * 2, ngf, 3, stride=1, padding=1, bias=False), 166 | nn.BatchNorm2d(ngf), 167 | nn.LeakyReLU(0.2, inplace=True), 168 | nn.Conv2d(ngf, nc, 3, stride=1, padding=1), 169 | nn.Sigmoid(), 170 | ) 171 | 172 | def forward(self, z): 173 | out = self.l1(z) 174 | out = out.view(out.shape[0], -1, self.init_size, self.init_size) 175 | img = self.conv_blocks(out) 176 | return img 177 | 178 | def UpSampling(x, H, W): 179 | B, N, C = x.size() 180 | assert N == H * W 181 | x = x.permute(0, 2, 1) 182 | x = x.view(-1, C, H, W) 183 | x = nn.PixelShuffle(2)(x) 184 | B, C, H, W = x.size() 185 | x = x.view(-1, C, H * W) 186 | x = x.permute(0, 2, 1) 187 | return x, H, W 188 | 189 | class MLP(nn.Module): 190 | def __init__(self, in_feat, hid_feat=None, out_feat=None, 191 | dropout=0.): 192 | super().__init__() 193 | if not hid_feat: 194 | hid_feat = in_feat 195 | if not out_feat: 196 | out_feat = in_feat 197 | self.fc1 = nn.Linear(in_feat, hid_feat) 198 | self.act = nn.GELU() 199 | self.fc2 = nn.Linear(hid_feat, out_feat) 200 | self.droprateout = nn.Dropout(dropout) 201 | 202 | def forward(self, x): 203 | x = self.fc1(x) 204 | x = self.act(x) 205 | x = self.fc2(x) 206 | return self.droprateout(x) 207 | 208 | 209 | class Attention(nn.Module): 210 | def __init__(self, dim, heads=4, attention_dropout=0., proj_dropout=0.): 211 | super().__init__() 212 | self.heads = heads 213 | self.scale = 1./dim**0.5 214 | 215 | self.qkv = nn.Linear(dim, dim*3, bias=False) 216 | self.attention_dropout = nn.Dropout(attention_dropout) 217 | self.out = nn.Sequential( 218 | nn.Linear(dim, dim), 219 | nn.Dropout(proj_dropout) 220 | ) 221 | 222 | def forward(self, x): 223 | b, n, c = x.shape 224 | qkv = self.qkv(x).reshape(b, n, 3, self.heads, c//self.heads) 225 | q, k, v = qkv.permute(2, 0, 3, 1, 4) 226 | 227 | dot = (q @ k.transpose(-2, -1)) * self.scale 228 | attn = dot.softmax(dim=-1) 229 | attn = self.attention_dropout(attn) 230 | 231 | x = (attn @ v).transpose(1, 2).reshape(b, n, c) 232 | x = self.out(x) 233 | return x 234 | 235 | class Encoder_Block(nn.Module): 236 | def __init__(self, dim, heads, mlp_ratio=4, drop_rate=0.): 237 | super().__init__() 238 | self.ln1 = nn.LayerNorm(dim) 239 | self.attn = Attention(dim, heads, drop_rate, drop_rate) 240 | self.ln2 = nn.LayerNorm(dim) 241 | self.mlp = MLP(dim, dim * mlp_ratio, dropout=drop_rate) 242 | 243 | def forward(self, x): 244 | x1 = self.ln1(x) 245 | x = x + self.attn(x1) 246 | x2 = self.ln2(x) 247 | x = x + self.mlp(x2) 248 | return x 249 | 250 | 251 | class TransformerEncoder(nn.Module): 252 | def __init__(self, depth, dim, heads, mlp_ratio=4, drop_rate=0.): 253 | super().__init__() 254 | self.Encoder_Blocks = nn.ModuleList([ 255 | Encoder_Block(dim, heads, mlp_ratio, drop_rate) 256 | for i in range(depth)]) 257 | 258 | def forward(self, x): 259 | for Encoder_Block in self.Encoder_Blocks: 260 | x = Encoder_Block(x) 261 | return x 262 | 263 | 264 | class Generator_3(nn.Module): 265 | """docstring for Generator""" 266 | 267 | def __init__(self, depth1=5, depth2=4, depth3=2, initial_size=8, dim=384, heads=4, mlp_ratio=4, 268 | drop_rate=0.): 269 | super(Generator_3, self).__init__() 270 | 271 | 272 | self.initial_size = initial_size 273 | self.dim = dim 274 | self.depth1 = depth1 275 | self.depth2 = depth2 276 | self.depth3 = depth3 277 | self.heads = heads 278 | self.mlp_ratio = mlp_ratio 279 | self.droprate_rate = drop_rate 280 | 281 | self.mlp = nn.Linear(1024, (self.initial_size ** 2) * self.dim) 282 | 283 | self.positional_embedding_1 = nn.Parameter(torch.zeros(1, (8 ** 2), 384)) 284 | self.positional_embedding_2 = nn.Parameter(torch.zeros(1, (8 * 2) ** 2, 384 // 4)) 285 | self.positional_embedding_3 = nn.Parameter(torch.zeros(1, (8 * 4) ** 2, 384 // 16)) 286 | 287 | self.TransformerEncoder_encoder1 = TransformerEncoder(depth=self.depth1, dim=self.dim, heads=self.heads, 288 | mlp_ratio=self.mlp_ratio, drop_rate=self.droprate_rate) 289 | self.TransformerEncoder_encoder2 = TransformerEncoder(depth=self.depth2, dim=self.dim // 4, heads=self.heads, 290 | mlp_ratio=self.mlp_ratio, drop_rate=self.droprate_rate) 291 | self.TransformerEncoder_encoder3 = TransformerEncoder(depth=self.depth3, dim=self.dim // 16, heads=self.heads, 292 | mlp_ratio=self.mlp_ratio, drop_rate=self.droprate_rate) 293 | 294 | self.linear = nn.Sequential(nn.Conv2d(self.dim // 16, 3, 1, 1, 0)) 295 | 296 | def forward(self, noise): 297 | x = self.mlp(noise).view(-1, self.initial_size ** 2, self.dim) 298 | 299 | x = x + self.positional_embedding_1 300 | H, W = self.initial_size, self.initial_size 301 | x = self.TransformerEncoder_encoder1(x) 302 | 303 | x, H, W = UpSampling(x, H, W) 304 | x = x + self.positional_embedding_2 305 | x = self.TransformerEncoder_encoder2(x) 306 | 307 | x, H, W = UpSampling(x, H, W) 308 | x = x + self.positional_embedding_3 309 | 310 | x = self.TransformerEncoder_encoder3(x) 311 | x = self.linear(x.permute(0, 2, 1).view(-1, self.dim // 16, H, W)) 312 | 313 | return x 314 | class CNNCifar10(nn.Module): 315 | def __init__(self): 316 | super(CNNCifar10, self).__init__() 317 | self.conv1 = nn.Conv2d(3, 256, 3) 318 | self.pool = nn.MaxPool2d(2, 2) 319 | self.conv2 = nn.Conv2d(256, 256, 3) 320 | self.conv3 = nn.Conv2d(256, 128, 3) 321 | self.fc1 = nn.Linear(128 * 4 * 4, 10) 322 | 323 | def forward(self, x): 324 | x = self.pool(F.relu(self.conv1(x))) 325 | x = self.pool(F.relu(self.conv2(x))) 326 | x = F.relu(self.conv3(x)) 327 | x = x.view(-1, 128 * 4 * 4) 328 | x = self.fc1(x) 329 | return x 330 | 331 | class CNN(nn.Module): 332 | def __init__(self): 333 | super(CNN, self).__init__() 334 | self.layer1 = nn.Sequential( 335 | nn.Conv2d(1, 16, kernel_size=5, padding=2), 336 | nn.BatchNorm2d(16), 337 | nn.ReLU(), 338 | nn.MaxPool2d(2)) 339 | self.layer2 = nn.Sequential( 340 | nn.Conv2d(16, 32, kernel_size=5, padding=2), 341 | nn.BatchNorm2d(32), 342 | nn.ReLU(), 343 | nn.MaxPool2d(2)) 344 | self.fc = nn.Linear(7 * 7 * 32, 10) 345 | 346 | def forward(self, x): 347 | out = self.layer1(x) 348 | out = self.layer2(out) 349 | out = out.view(out.size(0), -1) 350 | out = self.fc(out) 351 | return out 352 | 353 | class BasicBlock(nn.Module): 354 | expansion = 1 355 | 356 | def __init__(self, in_planes, planes, stride=1): 357 | super(BasicBlock, self).__init__() 358 | self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) 359 | self.bn1 = nn.BatchNorm2d(planes) 360 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False) 361 | self.bn2 = nn.BatchNorm2d(planes) 362 | 363 | self.shortcut = nn.Sequential() 364 | if stride != 1 or in_planes != self.expansion * planes: 365 | self.shortcut = nn.Sequential( 366 | nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1, stride=stride, bias=False), 367 | nn.BatchNorm2d(self.expansion * planes) 368 | ) 369 | 370 | def forward(self, x): 371 | out = F.relu(self.bn1(self.conv1(x))) 372 | out = self.bn2(self.conv2(out)) 373 | out += self.shortcut(x) 374 | out = F.relu(out) 375 | return out 376 | 377 | 378 | class Bottleneck(nn.Module): 379 | expansion = 4 380 | 381 | def __init__(self, in_planes, planes, stride=1): 382 | super(Bottleneck, self).__init__() 383 | self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False) 384 | self.bn1 = nn.BatchNorm2d(planes) 385 | self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) 386 | self.bn2 = nn.BatchNorm2d(planes) 387 | self.conv3 = nn.Conv2d(planes, self.expansion * planes, kernel_size=1, bias=False) 388 | self.bn3 = nn.BatchNorm2d(self.expansion * planes) 389 | 390 | self.shortcut = nn.Sequential() 391 | if stride != 1 or in_planes != self.expansion * planes: 392 | self.shortcut = nn.Sequential( 393 | nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1, stride=stride, bias=False), 394 | nn.BatchNorm2d(self.expansion * planes) 395 | ) 396 | 397 | def forward(self, x): 398 | out = F.relu(self.bn1(self.conv1(x))) 399 | out = F.relu(self.bn2(self.conv2(out))) 400 | out = self.bn3(self.conv3(out)) 401 | out += self.shortcut(x) 402 | out = F.relu(out) 403 | return out 404 | 405 | 406 | class ResNet(nn.Module): 407 | def __init__(self, block, num_blocks, num_classes=10): 408 | super(ResNet, self).__init__() 409 | self.in_planes = 64 410 | 411 | self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False) 412 | self.bn1 = nn.BatchNorm2d(64) 413 | self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1) 414 | self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2) 415 | self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2) 416 | self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2) 417 | self.linear = nn.Linear(512 * block.expansion, num_classes) 418 | 419 | def _make_layer(self, block, planes, num_blocks, stride): 420 | strides = [stride] + [1] * (num_blocks - 1) 421 | layers = [] 422 | for stride in strides: 423 | layers.append(block(self.in_planes, planes, stride)) 424 | self.in_planes = planes * block.expansion 425 | return nn.Sequential(*layers) 426 | 427 | def forward(self, x, return_features=False): 428 | x = self.conv1(x) 429 | x = self.bn1(x) 430 | out = F.relu(x) 431 | out = self.layer1(out) 432 | out = self.layer2(out) 433 | out = self.layer3(out) 434 | out = self.layer4(out) 435 | out = F.adaptive_avg_pool2d(out, (1, 1)) 436 | feature = out.view(out.size(0), -1) 437 | out = self.linear(feature) 438 | 439 | if return_features: 440 | return out, feature 441 | else: 442 | return out 443 | 444 | 445 | def resnet18(num_classes=10): 446 | return ResNet(BasicBlock, [2, 2, 2, 2], num_classes) 447 | 448 | 449 | def resnet34(num_classes=10): 450 | return ResNet(BasicBlock, [3, 4, 6, 3], num_classes) 451 | 452 | 453 | def resnet50(num_classes=10): 454 | return ResNet(Bottleneck, [3, 4, 6, 3], num_classes) 455 | 456 | 457 | def resnet101(num_classes=10): 458 | return ResNet(Bottleneck, [3, 4, 23, 3], num_classes) 459 | 460 | 461 | def resnet152(num_classes=10): 462 | return ResNet(Bottleneck, [3, 8, 36, 3], num_classes) 463 | -------------------------------------------------------------------------------- /train_scratch.py: -------------------------------------------------------------------------------- 1 | from __future__ import print_function 2 | import argparse # Python 命令行解析工具 3 | from tqdm import tqdm 4 | import torch 5 | import torch.nn as nn 6 | import torch.nn.functional as F 7 | import torch.optim as optim 8 | from tensorboardX import SummaryWriter 9 | from nets import resnet34, CNN, CNNCifar10, resnet18, resnet50, MLP, AlexNet, vgg8_bn 10 | from utils import test, get_dataset 11 | import warnings 12 | 13 | warnings.filterwarnings('ignore') 14 | 15 | 16 | def train(model, train_loader, optimizer, epoch): 17 | model.train() 18 | 19 | for idx, (data, target) in enumerate(train_loader): 20 | optimizer.zero_grad() 21 | data, target = data.cuda(), target.cuda() 22 | output = model(data) 23 | loss = F.cross_entropy(output, target) 24 | loss.backward() 25 | optimizer.step() 26 | 27 | 28 | def adjust_learning_rate(lr, optimizer, epoch): 29 | """Sets the learning rate to the initial LR decayed by 10 every 30 epochs""" 30 | lr = lr * (0.1 ** (epoch // 40)) 31 | for param_group in optimizer.param_groups: 32 | param_group['lr'] = lr 33 | 34 | # def get_model(dataset, net): 35 | # if "mnist" in dataset: 36 | # if net == "mlp": 37 | # model = MLP().cuda() 38 | # elif net == "lenet": 39 | # model = CNN().cuda() 40 | # elif net == "alexnet": 41 | # model = AlexNet().cuda() 42 | # elif dataset == "svhn": 43 | # if net == "alexnet": 44 | # model = CNNCifar10().cuda() 45 | # elif net == "vgg": 46 | # model = CNNCifar10().cuda() 47 | # elif net == "resnet18": 48 | # model = resnet18(num_classes=10).cuda() 49 | # elif dataset == "cifar10": 50 | # # model = resnet18(num_classes=10).cuda() 51 | # model = CNNCifar10().cuda() 52 | # elif dataset == "cifar100": 53 | # model = resnet50(num_classes=100).cuda() 54 | # elif dataset == "imagenet": 55 | # model = resnet18(num_classes=12).cuda() 56 | # return model 57 | 58 | 59 | def main(): 60 | # Training settings 61 | parser = argparse.ArgumentParser(description='PyTorch MNIST Example') 62 | parser.add_argument('--dataset', type=str, default="cifar10", 63 | help='dataset') 64 | parser.add_argument('--net', type=str, default="cifar10", 65 | help='dataset') 66 | parser.add_argument('--epochs', type=int, default=100, 67 | help='number of epochs to train (default: 100)') 68 | parser.add_argument('--lr', type=float, default=0.1, 69 | help='learning rate (default: 0.01)') 70 | parser.add_argument('--momentum', type=float, default=0.9, 71 | help='SGD momentum (default: 0.9)') 72 | parser.add_argument('--model', type=str, default='resnet34', 73 | help='SGD momentum (default: 0.9)') 74 | args = parser.parse_args() 75 | 76 | train_loader, test_loader = get_dataset(args.dataset) 77 | # model = get_teacher_model(args.dataset, args.net) 78 | model = CNNCifar10().cuda() 79 | # model = vgg8_bn(num_classes=10).cuda() 80 | optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum) 81 | bst_acc = -1 82 | public = "pretrained_large/{}_{}".format(args.dataset, args.net) 83 | tf_writer = SummaryWriter(log_dir=public) 84 | for epoch in range(1, args.epochs + 1): 85 | # adjust_learning_rate(args.lr, optimizer, epoch) 86 | train(model, train_loader, optimizer, epoch) 87 | acc, loss = test(model, test_loader) 88 | if acc > bst_acc: 89 | bst_acc = acc 90 | torch.save(model.state_dict(), '{}/{}_{}.pkl'.format(public, args.dataset, args.net)) 91 | 92 | tf_writer.add_scalar('test_acc', acc, epoch) 93 | bst_acc = max(bst_acc, acc) 94 | print("Epoch:{},\t test_acc:{}, best_acc:{}".format(epoch, acc, bst_acc)) 95 | 96 | 97 | if __name__ == '__main__': 98 | main() 99 | 100 | 101 | 102 | -------------------------------------------------------------------------------- /utils.py: -------------------------------------------------------------------------------- 1 | import os 2 | import torch.backends.cudnn as cudnn 3 | import matplotlib.pyplot as plt 4 | import torch 5 | import torch.nn as nn 6 | import torch.nn.functional as F 7 | import torchvision.utils as vutils 8 | from kornia import augmentation 9 | from sklearn.manifold import TSNE 10 | from torch.nn.functional import mse_loss 11 | from torchvision import transforms 12 | from tqdm import tqdm 13 | import numpy as np 14 | from torch.autograd import Variable 15 | from advertorch.attacks import LinfPGDAttack, LinfBasicIterativeAttack 16 | import math 17 | import os 18 | import random 19 | from torchvision import datasets, transforms 20 | from PIL import Image 21 | from nets import CNN, CNNCifar10, resnet18, resnet50 22 | 23 | 24 | def get_model(dataset, load): 25 | if "mnist" in dataset: 26 | model = CNN().cuda() 27 | # model = Net_m().cuda() 28 | elif dataset == "cifar10" or dataset == "svhn": 29 | # model = resnet18(num_classes=10).cuda() 30 | model = CNNCifar10().cuda() 31 | elif dataset == "cifar100": 32 | model = resnet50(num_classes=100).cuda() 33 | elif dataset == "tiny": 34 | model = resnet50(num_classes=200).cuda() 35 | # pretraind = 'public/attack/pretrained/' 36 | pretraind = 'pretrained_ckpt/cifar10_cnn' 37 | load_list = ['cnn_mnist.pth', 'cnn_fmnist.pth', 'cnncifar10.pkl', 'res18_svhn.pth', 38 | 'res50_cifar100.pth', 'res50_tiny_imagenet.pth'] 39 | if load == 1: 40 | if "mnist" == dataset: 41 | state_dict = torch.load(pretraind + load_list[0])['state_dict'] 42 | elif "fmnist" == dataset: 43 | state_dict = torch.load(pretraind + load_list[1])['state_dict'] 44 | elif dataset == "cifar10": 45 | state_dict = torch.load(pretraind + load_list[2])['state_dict'] 46 | elif dataset == "svhn": 47 | state_dict = torch.load(pretraind + load_list[3])['state_dict'] 48 | elif dataset == "cifar100": 49 | state_dict = torch.load(pretraind + load_list[4])['state_dict'] 50 | elif dataset == "tiny": 51 | state_dict = torch.load(pretraind + load_list[5]) 52 | else: 53 | state_dict = None 54 | 55 | return model, state_dict 56 | 57 | 58 | def cal_prob(black_net, data): 59 | with torch.no_grad(): 60 | outputs = black_net(data.detach()) 61 | score = F.softmax(outputs, dim=1) # score-based 62 | score = score.detach().cpu().numpy() 63 | score = torch.from_numpy(score).cuda().float() 64 | return score 65 | 66 | 67 | def cal_label(black_net, data): 68 | with torch.no_grad(): 69 | outputs = black_net(data.detach()) 70 | _, label = torch.max(outputs.data, 1) 71 | label = label.detach().cpu().numpy() 72 | label = torch.from_numpy(label).cuda().long() 73 | return label 74 | 75 | 76 | def test_robust(loader, substitute_net, original_net, dataset): 77 | # cfgs = dict(random=True, test_num_steps=40, test_step_size=0.01, test_epsilon=0.3, num_classes=10) 78 | if dataset == "mnist": 79 | cfgs = dict(test_step_size=0.01, test_epsilon=0.3) 80 | elif dataset == "cifar10" or dataset == "cifar100": 81 | cfgs = dict(test_step_size=2.0 / 255, test_epsilon=8.0 / 255) 82 | elif dataset == "fmnist": 83 | cfgs = dict(test_step_size=0.01, test_epsilon=0.3) 84 | elif dataset == "svhn" or dataset == "tiny": 85 | cfgs = dict(test_step_size=0.01, test_epsilon=0.3) 86 | 87 | correct_ghost = 0.0 88 | correct = 0.0 89 | total = 0.0 90 | substitute_net.eval() 91 | adversary = LinfBasicIterativeAttack( 92 | substitute_net, loss_fn=torch.nn.CrossEntropyLoss(reduction="sum"), eps=cfgs['test_epsilon'], 93 | nb_iter=120, eps_iter=cfgs['test_step_size'], clip_min=0.0, clip_max=1.0, 94 | targeted=False) 95 | 96 | for inputs, labels in loader: 97 | inputs, labels = inputs.cuda(), labels.cuda() 98 | total += labels.size(0) 99 | t_label = cal_label(original_net, inputs) 100 | idx = torch.where(t_label == labels)[0] 101 | correct += idx.shape[0] 102 | adv_inputs_ghost = adversary.perturb(inputs[idx], labels[idx]) 103 | predicted = cal_label(original_net, adv_inputs_ghost) 104 | correct_ghost += (predicted != labels[idx]).sum() 105 | # print('Attack success rate: {}, clean acc: {}'.format(100. * correct_ghost / correct, 100 * correct / total)) 106 | return 100. * correct_ghost / correct, 100 * correct / total 107 | 108 | 109 | def setup_seed(seed): 110 | torch.manual_seed(seed) 111 | torch.cuda.manual_seed_all(seed) 112 | torch.cuda.manual_seed(seed) 113 | np.random.seed(seed) 114 | random.seed(seed) 115 | cudnn.deterministic = True 116 | 117 | 118 | def test(model, test_loader): 119 | model.eval() 120 | test_loss = 0 121 | correct = 0 122 | total = 0 123 | with torch.no_grad(): 124 | for data, target in test_loader: 125 | data, target = data.cuda(), target.cuda() 126 | output = model(data) 127 | test_loss += F.cross_entropy(output, target, size_average=False).item() # sum up batch loss 128 | total += data.shape[0] 129 | pred = torch.max(output, 1)[1] 130 | correct += pred.eq(target.view_as(pred)).sum().item() 131 | 132 | test_loss /= total 133 | acc = 100. * correct / total 134 | return acc, test_loss 135 | 136 | 137 | def print_log(strs, log): 138 | print(strs) 139 | log.write(strs) 140 | 141 | 142 | def get_dataset(dataset): 143 | data_dir = '/mnt/lustre/share_data/zhangjie/' 144 | if dataset == "mnist": 145 | train_dataset = datasets.MNIST(data_dir, train=True, 146 | transform=transforms.Compose( 147 | [transforms.ToTensor()])) 148 | test_dataset = datasets.MNIST(data_dir, train=False, 149 | transform=transforms.Compose([ 150 | transforms.ToTensor(), 151 | ])) 152 | elif dataset == "fmnist": 153 | train_dataset = datasets.FashionMNIST(data_dir, train=True, 154 | transform=transforms.Compose( 155 | [transforms.ToTensor()])) 156 | test_dataset = datasets.FashionMNIST(data_dir, train=False, 157 | transform=transforms.Compose([ 158 | transforms.ToTensor(), 159 | ])) 160 | elif dataset == "svhn": 161 | train_dataset = datasets.SVHN(data_dir, split="train", 162 | transform=transforms.Compose( 163 | [transforms.ToTensor()])) 164 | test_dataset = datasets.SVHN(data_dir, split="test", 165 | transform=transforms.Compose([ 166 | transforms.ToTensor(), 167 | ])) 168 | elif dataset == "cifar10": 169 | train_dataset = datasets.CIFAR10(data_dir, train=True, 170 | transform=transforms.Compose( 171 | [ 172 | transforms.RandomCrop(32, padding=4), 173 | transforms.RandomHorizontalFlip(), 174 | transforms.ToTensor(), 175 | ])) 176 | test_dataset = datasets.CIFAR10(data_dir, train=False, 177 | transform=transforms.Compose([ 178 | transforms.ToTensor(), 179 | ])) 180 | elif dataset == "cifar100": 181 | train_dataset = datasets.CIFAR100(data_dir, train=True, 182 | transform=transforms.Compose( 183 | [ 184 | transforms.RandomCrop(32, padding=4), 185 | transforms.RandomHorizontalFlip(), 186 | transforms.ToTensor(), 187 | ])) 188 | test_dataset = datasets.CIFAR100(data_dir, train=False, 189 | transform=transforms.Compose([ 190 | transforms.ToTensor(), 191 | ])) 192 | elif dataset == "tiny": 193 | data_transforms = { 194 | 'train': transforms.Compose([ 195 | transforms.RandomRotation(20), 196 | transforms.RandomHorizontalFlip(0.5), 197 | transforms.ToTensor(), 198 | ]), 199 | 'val': transforms.Compose([ 200 | transforms.ToTensor(), 201 | ]), 202 | 'test': transforms.Compose([ 203 | transforms.ToTensor(), 204 | ]) 205 | } 206 | data_dir = "data/tiny-imagenet-200/" 207 | image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) 208 | for x in ['train', 'val', 'test']} 209 | train_dataset = image_datasets['train'] 210 | test_dataset = image_datasets['val'] 211 | # train_loader = data.DataLoader(image_datasets['train'], batch_size=128, shuffle=True, num_workers=4) 212 | # val_loader = data.DataLoader(image_datasets['val'], batch_size=128, shuffle=False, num_workers=4) 213 | 214 | else: 215 | raise NotImplementedError 216 | 217 | train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=256, 218 | shuffle=True, num_workers=4) 219 | test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=256, 220 | shuffle=False, num_workers=4) 221 | 222 | return train_loader, test_loader 223 | 224 | 225 | class ScoreLoss(torch.nn.Module): 226 | def __init__(self, reduction='mean'): 227 | super(ScoreLoss, self).__init__() 228 | self.reduction = reduction 229 | 230 | def forward(self, logits, target): 231 | if logits.dim() > 2: 232 | logits = logits.view(logits.size(0), logits.size(1), -1) # [N, C, HW] 233 | logits = logits.transpose(1, 2) # [N, HW, C] 234 | logits = logits.contiguous().view(-1, logits.size(2)) # [NHW, C] 235 | target = target.view(-1, 1) # [NHW，1] 236 | 237 | score = F.log_softmax(logits, 1) # score-based 238 | score = score.gather(1, target) # [NHW, 1] 239 | loss = -1 * score 240 | 241 | if self.reduction == 'mean': 242 | loss = loss.mean() 243 | elif self.reduction == 'sum': 244 | loss = loss.sum() 245 | return loss 246 | 247 | def save_checkpoint(state, is_best, filename='checkpoint.pth'): 248 | if is_best: 249 | torch.save(state, filename) 250 | 251 | def reset_model(model): 252 | for m in model.modules(): 253 | if isinstance(m, (nn.ConvTranspose2d, nn.Linear, nn.Conv2d)): 254 | nn.init.normal_(m.weight, 0.0, 0.02) 255 | if m.bias is not None: 256 | nn.init.constant_(m.bias, 0) 257 | if isinstance(m, (nn.BatchNorm2d)): 258 | nn.init.normal_(m.weight, 1.0, 0.02) 259 | nn.init.constant_(m.bias, 0) 260 | 261 | 262 | class MultiTransform: 263 | """Create two crops of the same image""" 264 | 265 | def __init__(self, transform): 266 | self.transform = transform 267 | 268 | def __call__(self, x): 269 | return [t(x) for t in self.transform] 270 | 271 | def __repr__(self): 272 | return str(self.transform) 273 | 274 | 275 | def pack_images(images, col=None, channel_last=False, padding=1): 276 | # N, C, H, W 277 | if isinstance(images, (list, tuple)): 278 | images = np.stack(images, 0) 279 | if channel_last: 280 | images = images.transpose(0, 3, 1, 2) # make it channel first 281 | assert len(images.shape) == 4 282 | assert isinstance(images, np.ndarray) 283 | 284 | N, C, H, W = images.shape 285 | if col is None: 286 | col = int(math.ceil(math.sqrt(N))) 287 | row = int(math.ceil(N / col)) 288 | 289 | pack = np.zeros((C, H * row + padding * (row - 1), W * col + padding * (col - 1)), dtype=images.dtype) 290 | for idx, img in enumerate(images): 291 | h = (idx // col) * (H + padding) 292 | w = (idx % col) * (W + padding) 293 | pack[:, h:h + H, w:w + W] = img 294 | return pack 295 | 296 | 297 | def save_image_batch(imgs, output, col=None, size=None, pack=True): 298 | if isinstance(imgs, torch.Tensor): 299 | imgs = (imgs.detach().clamp(0, 1).cpu().numpy() * 255).astype('uint8') 300 | base_dir = os.path.dirname(output) 301 | if base_dir != '': 302 | os.makedirs(base_dir, exist_ok=True) 303 | if pack: 304 | imgs = pack_images(imgs, col=col).transpose(1, 2, 0).squeeze() 305 | imgs = Image.fromarray(imgs) 306 | if size is not None: 307 | if isinstance(size, (list, tuple)): 308 | imgs = imgs.resize(size) 309 | else: 310 | w, h = imgs.size 311 | max_side = max(h, w) 312 | scale = float(size) / float(max_side) 313 | _w, _h = int(w * scale), int(h * scale) 314 | imgs = imgs.resize([_w, _h]) 315 | imgs.save(output) 316 | else: 317 | output_filename = output.strip('.png') 318 | for idx, img in enumerate(imgs): 319 | if img.shape[0] == 1: 320 | img = Image.fromarray(img[0]) 321 | else: 322 | img = Image.fromarray(img.transpose(1, 2, 0)) 323 | img.save(output_filename + '-%d.png' % (idx)) 324 | 325 | 326 | def _collect_all_images(root, postfix=['png', 'jpg', 'jpeg', 'JPEG']): 327 | images = [] 328 | if isinstance(postfix, str): 329 | postfix = [postfix] 330 | for dirpath, dirnames, files in os.walk(root): # '/dockerdata/cvpr/10-28/ft_local/run/svhn_4',[],files(all imgs) 331 | files.sort() 332 | files = files[-256 * 400:] 333 | # files = files[-2048 * 400:] 334 | for pos in postfix: 335 | for f in files: 336 | if f.endswith(pos): 337 | images.append(os.path.join(dirpath, f)) 338 | return images 339 | 340 | 341 | class UnlabeledImageDataset(torch.utils.data.Dataset): 342 | def __init__(self, root, transform=None): 343 | self.root = os.path.abspath(root) 344 | self.images = _collect_all_images(self.root) # [ os.path.join(self.root, f) for f in os.listdir( root ) ] 345 | self.transform = transform 346 | 347 | def __getitem__(self, idx): 348 | img = Image.open(self.images[idx]) 349 | if self.transform: 350 | img = self.transform(img) 351 | return img 352 | 353 | def __len__(self): 354 | return len(self.images) 355 | 356 | def __repr__(self): 357 | return 'Unlabeled data:\n\troot: %s\n\tdata mount: %d\n\ttransforms: %s' % ( 358 | self.root, len(self), self.transform) 359 | 360 | 361 | class ImagePool(object): 362 | def __init__(self, root): 363 | self.root = os.path.abspath(root) 364 | os.makedirs(self.root, exist_ok=True) 365 | self._idx = 0 366 | 367 | def add(self, imgs, targets=None): 368 | save_image_batch(imgs, os.path.join(self.root, "%d.png" % (self._idx)), pack=False) 369 | self._idx += 1 370 | 371 | def get_dataset(self, transform=None, labeled=True): 372 | return UnlabeledImageDataset(self.root, transform=transform) 373 | 374 | --------------------------------------------------------------------------------