lc_attack.py

'''
this script is for lc attack

link : https://github.com/MadryLab/label-consistent-backdoor-code
The original license is placed at the end of this file.

basic structure:
1. config args, save_path, fix random seed
2. set the clean train data and clean test data
3. set the attack img transform and label transform
4. set the backdoor attack data and backdoor test data
5. set the device, model, criterion, optimizer, training schedule.
6. attack or use the model to do finetune with 5% clean data
7. save the attack result for defense
'''

import sys, yaml, os

os.chdir(sys.path[0])
sys.path.append('../')
os.getcwd()

import argparse
from pprint import pformat
import numpy as np
import torch
import time
import logging

from utils.aggregate_block.save_path_generate import generate_save_folder
from utils.aggregate_block.dataset_and_transform_generate import get_num_classes, get_input_shape
from utils.aggregate_block.fix_random import fix_random
from utils.aggregate_block.dataset_and_transform_generate import dataset_and_transform_generate
from utils.bd_dataset import prepro_cls_DatasetBD
from utils.backdoor_generate_pindex import generate_pidx_from_label_transform
from utils.aggregate_block.bd_attack_generate import bd_attack_img_trans_generate, bd_attack_label_trans_generate
from copy import deepcopy
from utils.aggregate_block.model_trainer_generate import generate_cls_model, generate_cls_trainer
from utils.aggregate_block.train_settings_generate import argparser_opt_scheduler, argparser_criterion
from utils.save_load_attack import save_attack_result
from utils.log_assist import get_git_info


def add_args(parser):
    """
    parser : argparse.ArgumentParser
    return a parser added with args required by fit
    """
    # Training settings
    # parser.add_argument('--mode', type=str,
    #                     help='classification/detection/segmentation')
    parser.add_argument('--amp', type=lambda x: str(x) in ['True', 'true', '1'])
    parser.add_argument('--device', type=str)
    parser.add_argument('--yaml_path', type=str, default='../config/attack/lc/default.yaml',
                        help='path for yaml file provide additional default attributes')
    parser.add_argument('--lr_scheduler', type=str,
                        help='which lr_scheduler use for optimizer')
    # only all2one can be use for clean-label
    parser.add_argument('--attack_label_trans', type=str,
                        help='which type of label modification in backdoor attack'
                        )
    parser.add_argument('--pratio', type=float,
                        help='the poison rate '
                        )
    parser.add_argument('--epochs', type=int)
    parser.add_argument('--dataset', type=str,
                        help='which dataset to use'
                        )
    parser.add_argument('--dataset_path', type=str)
    parser.add_argument('--attack_target', type=int,
                        help='target class in all2one attack')
    parser.add_argument('--batch_size', type=int)
    parser.add_argument('--lr', type=float)
    parser.add_argument('--steplr_stepsize', type=int)
    parser.add_argument('--steplr_gamma', type=float)
    parser.add_argument('--sgd_momentum', type=float)
    parser.add_argument('--wd', type=float, help='weight decay of sgd')
    parser.add_argument('--steplr_milestones', type=list)
    parser.add_argument('--client_optimizer', type=int)
    parser.add_argument('--random_seed', type=int,
                        help='random_seed')
    parser.add_argument('--frequency_save', type=int,
                        help=' frequency_save, 0 is never')
    parser.add_argument('--model', type=str,
                        help='choose which kind of model')
    parser.add_argument('--save_folder_name', type=str,
                        help='(Optional) should be time str + given unique identification str')
    parser.add_argument('--git_hash', type=str,
                        help='git hash number, in order to find which version of code is used')
    parser.add_argument('--attack_train_replace_imgs_path', type=str)
    parser.add_argument('--attack_test_replace_imgs_path', type=str)
    return parser


def main():
    ### 1. config args, save_path, fix random seed
    parser = (add_args(argparse.ArgumentParser(description=sys.argv[0])))
    args = parser.parse_args()

    with open(args.yaml_path, 'r') as f:
        defaults = yaml.safe_load(f)

    defaults.update({k: v for k, v in args.__dict__.items() if v is not None})

    args.__dict__ = defaults

    args.attack = 'label_consistent'

    args.terminal_info = sys.argv

    args.num_classes = get_num_classes(args.dataset)
    args.input_height, args.input_width, args.input_channel = get_input_shape(args.dataset)
    args.img_size = (args.input_height, args.input_width, args.input_channel)
    args.dataset_path = f"{args.dataset_path}/{args.dataset}"

    if ('attack_train_replace_imgs_path' not in args.__dict__) or (args.attack_train_replace_imgs_path is None):
        args.attack_train_replace_imgs_path = f"../resource/label-consistent/data/adv_dataset/{args.dataset}_train.npy"
        logging.info(f"args.attack_train_replace_imgs_path does not found, so = {args.attack_train_replace_imgs_path}")

    if ('attack_test_replace_imgs_path' not in args.__dict__) or (args.attack_test_replace_imgs_path is None):
        args.attack_test_replace_imgs_path = f"../resource/label-consistent/data/adv_dataset/{args.dataset}_test.npy"
        logging.info(f"args.attack_test_replace_imgs_path does not found, so = {args.attack_test_replace_imgs_path}")

    ### save path
    if 'save_folder_name' not in args:
        save_path = generate_save_folder(
            run_info=('afterwards' if 'load_path' in args.__dict__ else 'attack') + '_' + args.attack,
            given_load_file_path=args.load_path if 'load_path' in args else None,
            all_record_folder_path='../record',
        )
    else:
        save_path = '../record/' + args.save_folder_name
        os.mkdir(save_path)

    args.save_path = save_path

    torch.save(args.__dict__, save_path + '/info.pickle')

    ### set the logger
    logFormatter = logging.Formatter(
        fmt='%(asctime)s [%(levelname)-8s] [%(filename)s:%(lineno)d] %(message)s',
        datefmt='%Y-%m-%d:%H:%M:%S',
    )
    logger = logging.getLogger()

    fileHandler = logging.FileHandler(save_path + '/' + time.strftime("%Y_%m_%d_%H_%M_%S", time.localtime()) + '.log')
    fileHandler.setFormatter(logFormatter)
    logger.addHandler(fileHandler)

    consoleHandler = logging.StreamHandler()
    consoleHandler.setFormatter(logFormatter)
    logger.addHandler(consoleHandler)

    logger.setLevel(logging.INFO)

    logging.info(pformat(args.__dict__))

    try:
        logging.info(pformat(get_git_info()))
    except:
        logging.info('Getting git info fails.')

    ### set the random seed
    fix_random(int(args.random_seed))

    ### 2. set the clean train data and clean test data
    train_dataset_without_transform, \
    train_img_transform, \
    train_label_transfrom, \
    test_dataset_without_transform, \
    test_img_transform, \
    test_label_transform = dataset_and_transform_generate(args)

    benign_train_ds = prepro_cls_DatasetBD(
            full_dataset_without_transform=train_dataset_without_transform,
            poison_idx=np.zeros(len(train_dataset_without_transform)),
            # one-hot to determine which image may take bd_transform
            bd_image_pre_transform=None,
            bd_label_pre_transform=None,
            ori_image_transform_in_loading=train_img_transform,
            ori_label_transform_in_loading=train_label_transfrom,
            add_details_in_preprocess=True,
    )


    benign_test_ds = prepro_cls_DatasetBD(
            test_dataset_without_transform,
            poison_idx=np.zeros(len(test_dataset_without_transform)),
            # one-hot to determine which image may take bd_transform
            bd_image_pre_transform=None,
            bd_label_pre_transform=None,
            ori_image_transform_in_loading=test_img_transform,
            ori_label_transform_in_loading=test_label_transform,
            add_details_in_preprocess=True,
    )

    ### 3. set the attack img transform and label transform
    train_bd_img_transform, test_bd_img_transform = bd_attack_img_trans_generate(args)
    ### get the backdoor transform on label
    bd_label_transform = bd_attack_label_trans_generate(args)

    ### 4. set the backdoor attack data and backdoor test data
    train_pidx = generate_pidx_from_label_transform(
        benign_train_ds.targets,
        label_transform=bd_label_transform,
        train=True,
        pratio=args.pratio if 'pratio' in args.__dict__ else None,
        p_num=args.p_num if 'p_num' in args.__dict__ else None,
        clean_label=True,
    )
    torch.save(train_pidx,
               args.save_path + '/train_pidex_list.pickle',
               )

    ### generate train dataset for backdoor attack
    adv_train_ds = prepro_cls_DatasetBD(
        deepcopy(train_dataset_without_transform),
        poison_idx=train_pidx,
        bd_image_pre_transform=train_bd_img_transform,
        bd_label_pre_transform=bd_label_transform,
        ori_image_transform_in_loading=train_img_transform,
        ori_label_transform_in_loading=train_label_transfrom,
        add_details_in_preprocess=True,
    )

    ### decide which img to poison in ASR Test
    test_pidx = generate_pidx_from_label_transform(
        benign_test_ds.targets,
        label_transform=bd_label_transform,
        train=False,
    )

    ### generate test dataset for ASR
    adv_test_dataset = prepro_cls_DatasetBD(
        deepcopy(test_dataset_without_transform),
        poison_idx=test_pidx,
        bd_image_pre_transform=test_bd_img_transform,
        bd_label_pre_transform=bd_label_transform,
        ori_image_transform_in_loading=test_img_transform,
        ori_label_transform_in_loading=test_label_transform,
        add_details_in_preprocess=True,

    )

    # delete the samples that do not used for ASR test (those non-poisoned samples)
    adv_test_dataset.subset(
        np.where(test_pidx == 1)[0]
    )

    ### 5. set the device, model, criterion, optimizer, training schedule.
    device = torch.device(args.device if torch.cuda.is_available() else "cpu")

    net = generate_cls_model(
        model_name=args.model,
        num_classes=args.num_classes,
        image_size=args.img_size[0],
    )

    if torch.cuda.device_count() > 1 and args.device == 'cuda':
        logging.info("device='cuda', default use all device")
        net = torch.nn.DataParallel(net)

    trainer = generate_cls_trainer(
        net,
        args.attack,
        args.amp,
    )

    criterion = argparser_criterion(args)

    optimizer, scheduler = argparser_opt_scheduler(net, args)

    ### 6. attack or use the model to do finetune with 5% clean data
    if 'load_path' not in args.__dict__:

        trainer.train_with_test_each_epoch_v2_sp(
            batch_size=args.batch_size,
            train_dataset = adv_train_ds,
            test_dataset_dict={
                "test_data" :benign_test_ds,
                "adv_test_data" :adv_test_dataset,
            },
            end_epoch_num=args.epochs,
            criterion=criterion,
            optimizer=optimizer,
            scheduler=scheduler,
            device=device,
            frequency_save=args.frequency_save,
            save_folder_path=save_path,
            save_prefix='attack',
            continue_training_path=None,
        )

    else:

        if 'recover' not in args.__dict__ or args.recover == False:
            print('finetune so use less data, 5% of benign train data')

            benign_train_ds.subset(
                np.random.choice(
                    np.arange(
                        len(benign_train_ds)),
                    size=round((len(benign_train_ds)) / 20),  # 0.05
                    replace=False,
                )
            )

            torch.save(
                list(benign_train_ds.original_index),
                args.save_path + '/finetune_idx_list.pt',
            )

            trainer.train_with_test_each_epoch_v2_sp(
                batch_size=args.batch_size,
                train_dataset=benign_train_ds,
                test_dataset_dict={
                    "test_data": benign_test_ds,
                    "adv_test_data": adv_test_dataset,
                },
                end_epoch_num=args.epochs,
                criterion=criterion,
                optimizer=optimizer,
                scheduler=scheduler,
                device=device,
                frequency_save=args.frequency_save,
                save_folder_path=save_path,
                save_prefix='finetune',
                continue_training_path=args.load_path,
                only_load_model=True,
            )

    ### 7. save model, data, and other information that defense process may need
    save_attack_result(
        model_name=args.model,
        num_classes=args.num_classes,
        model=trainer.model.cpu().state_dict(),
        data_path=args.dataset_path,
        img_size=args.img_size,
        clean_data=args.dataset,
        bd_train=adv_train_ds,
        bd_test=adv_test_dataset,
        save_path=save_path,
    )


if __name__ == '__main__':
    main()

"""    
MIT License

Copyright (c) 2019 Alexander Turner, Dimitris Tsipras and Aleksander Madry

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
"""