def attack_loader(args, net): # Gradient Clamping based Attack if args.attack == "pgd": return torchattacks.PGD(model=net, eps=args.eps, alpha=args.eps / args.steps * 2.3, steps=args.steps, random_start=True) elif args.attack == "auto": return torchattacks.APGD(model=net, eps=args.eps) elif args.attack == "fab": return torchattacks.FAB(model=net, eps=args.eps, n_classes=args.n_classes) elif args.attack == "cw": return torchattacks.CW(model=net, c=0.1, lr=0.1, steps=200) elif args.attack == "fgsm": return torchattacks.FGSM(model=net, eps=args.eps) elif args.attack == "bim": return torchattacks.BIM(model=net, eps=args.eps, alpha=1 / 255) elif args.attack == "deepfool": return torchattacks.DeepFool(model=net, steps=10) elif args.attack == "sparse": return torchattacks.SparseFool(model=net) elif args.attack == "gn": return torchattacks.GN(model=net, sigma=args.eps)
def get_atk(model, atk_name, eps, steps): if atk_name == 'fgsm': return torchattacks.FGSM(model, eps=eps) elif atk_name == 'bim': return torchattacks.BIM(model, eps=eps, steps=steps, alpha=eps / (steps * .5)) elif atk_name == 'deepfool': return torchattacks.DeepFool(model, steps=steps) elif atk_name == 'cw': return torchattacks.CW(model) elif atk_name == 'pgd': return torchattacks.PGD(model, eps=eps, steps=steps, alpha=eps / (steps * .5)) elif atk_name == 'rfgsm': return torchattacks.RFGSM(model, eps=eps, alpha=eps) elif atk_name == 'auto-attack': return torchattacks.AutoAttack(model, eps=eps) elif atk_name == 'mifgsm': return torchattacks.MIFGSM(model, eps=eps, steps=steps) elif atk_name == 'square': return torchattacks.Square(model, eps=eps) elif atk_name == 'fab': return torchattacks.FAB(model, eps=eps) elif atk_name == 'one-pixel': return torchattacks.OnePixel(model) elif atk_name == 'gn': return torchattacks.GN(model, sigma=eps) elif atk_name == 'apgd': return torchattacks.APGD(model, eps=eps) elif atk_name == 'eotpgd': return torchattacks.EOTPGD(model, eps=eps, steps=steps, alpha=eps / (steps * .5)) elif atk_name == 'pgddlr': return torchattacks.PGDDLR(model, eps=eps, steps=steps, alpha=eps / (steps * .5)) elif atk_name == 'ffgsm': return torchattacks.FFGSM(model, eps=eps, alpha=eps) elif atk_name == 'sparsefool': return torchattacks.SparseFool(model) else: print("Attack not valid") sys.exit(-1)
def basic_iterative_method(model, X, Y, eps, eps_iter, test_loader=None): print(X.shape, Y.shape) atk = torchattacks.BIM(model, eps=eps, alpha=eps_iter, steps=7) if test_loader is not None: x_adv_list = [] for batch in test_loader: x = batch[0] y = batch[1] x_adv_list.append(atk(x, y)) X_adv = torch.cat(x_adv_list) print('adv', X_adv.shape) X_adv = X_adv.cpu() return X_adv
def load_attack(model, attack: str): import torchattacks if attack == 'PGD': return torchattacks.PGD(model, eps=2 / 255, alpha=2 / 255, steps=7) elif attack == 'CW': return torchattacks.CW(model, targeted=False, c=1, kappa=0, steps=1000, lr=0.01) elif attack == 'BIM': return torchattacks.BIM(model, eps=4 / 255, alpha=1 / 255, steps=0) elif attack == 'FGSM': return torchattacks.FGSM(model, eps=1 / 255) else: raise NotImplementedError()
def Greedy_Decode_Eval(Net, datasets, args): # TestNet = Net.eval() epoch_size = len(datasets) // args.test_batch_size # 整除,多余的末尾就不会包括进来了 # collate_fn:如何取样本的,我们可以定义自己的函数来准确地实现想要的功能 # shuffle:设置为True的时候,每个世代都会打乱数据集 batch_iterator = iter(DataLoader(datasets, args.test_batch_size, shuffle=True, num_workers=args.num_workers, collate_fn=collate_fn)) attack = torchattacks.BIM(Net, eps=args.epsilon, alpha=1/255, iters=0) total = epoch_size * args.test_batch_size # 总识别样本数 correct = 0 for i in range(epoch_size): # load train data images, labels, lengths = next(batch_iterator) # 提取iter的元素,注意images里面是整个batch的图像,但这时候类型是tensor了 # print(lengths) # 如果batch_size=100,那么lengths就是[7,7,...,7],100个7(list类型) # print(images.shape) start = 0 targets = [] for length in lengths: # 事到如今又要将tabel一个个提取出来,那为什么之前要用extend方法而不用append? label = labels[start:start+length] targets.append(label) start += length targets = np.array([el.numpy() for el in targets], dtype=np.int32) perturbed_images = attack(images, labels, lengths) # 重新进行识别 preb_atk = Net(perturbed_images) # 获得标签 preb_labels_atk = np.array(get_preb_labels(preb_atk)) for i in range(preb_labels_atk.shape[0]): # print(preb_labels_atk[i]) # print(targets[i]) if len(preb_labels_atk[i]) == len(targets[i]) and (preb_labels_atk[i] == targets[i]).all(): correct += 1 return correct * 1.0 / total
CUDA_MODE = torch.cuda.is_available() if not CUDA_MODE: logger.warn( "CUDA not available. Run on a CUDA enabled platform (NVIDIA GPU with compute capability >= 3) to get memory usage and timing stats (this code makes use of CUDA events to accurately measure memory and timing). Press [ENTER] to continue anyways." ) input() # How many images we want to test IMG_NUM = 500 model.eval() attacks = [ torchattacks.PGD(model), torchattacks.DeepFool(model), torchattacks.StepLL(model), torchattacks.BIM(model) ] for attack in attacks: time = [] attack_l2 = [] peak_cuda = [] avg_cuda = [] total_success = 0 logger.info(f"Benchmarking {str(attack)} on {IMG_NUM} images") for img_id in range(IMG_NUM): # Load the image and reshape it to [NxCxWxH] (which is what the models expect) target_im = data[str(model)][img_id][None, :, :, :].to(device) _, TRUECLASS = torch.max(model(target_im), 1)
def train_model(device, dataloaders, batch_size, len_dataset, model, criterion, optimizer, scheduler, num_epochs=25): since = time.time() ''' * state_dict: 각 layer 를 매개변수 텐서로 매핑하는 Python 사전(dict) 객체 - layer; learnable parameters (convolutional layers, linear layers, etc.), registered buffers (batchnorm’s running_mean) - Optimizer objects (torch.optim) ''' best_model_wts = copy.deepcopy(model.state_dict()) best_acc = 0.0 train_loss, train_acc, valid_loss, valid_acc = [], [], [], [] for epoch in range(num_epochs): print('Epoch {}/{}'.format(epoch, num_epochs - 1)) print('-' * 10) # 각 epoch마다 training, validation phase 나눠줌. for phase in ['train', 'valid']: if phase == 'train': model.train() # training mode else: model.eval() # evaluate mode running_loss, running_corrects, num_cnt = 0.0, 0, 0 ratio_adv_ori = int((len_dataset // batch_size + 1) * 0.4) # adversarial, original data 비율 정하기 # batch 별로 나눠진 데이터 불러오기 for i, (inputs, labels) in enumerate(dataloaders[phase]): # 설정한 비율에 따라 adversarial, original input으로 나누기 if (phase == 'train' and (i < ratio_adv_ori)) or (phase == 'valid' and i % 2 == 0): inputs = inputs.to(device) else: # adversarial attack 정의 atks = [torchattacks.FGSM(model, eps=8 / 255), torchattacks.BIM(model, eps=8 / 255, alpha=2 / 255, steps=7), torchattacks.PGD(model, eps=8 / 255, alpha=2 / 255, steps=7), ] inputs = atks[i % 3](inputs, labels).to(device) # Image Processing Based Defense Methods --> tensor를 image로 변환하여 적용 for batch in range(inputs.shape[0]): tensor2pil = transforms.ToPILImage()(inputs[batch]).convert('RGB') # 1. Resizing # Image.resize(size, resample=3, box=None, reducing_gap=None) # resample(filter): PIL.Image.NEAREST, PIL.Image.BOX, PIL.Image.BILINEAR, PIL.Image.HAMMING, PIL.Image.BICUBIC tensor2pil.resize((74, 74)) tensor2pil.resize((224, 224)) # 다시 이미지를 tensor로 바꾸기 tensor_img = transforms.ToTensor()(tensor2pil) inputs[batch] = tensor_img # 2. jpeg compression tensor2numpy = inputs[batch].cpu().numpy() cv_img = np.transpose(tensor2numpy, (1, 2, 0)) # [w, h, c] cv_img = cv_img * 255 encode_param = [int(cv2.IMWRITE_JPEG_QUALITY), 15] result, encimg = cv2.imencode('.jpg', cv_img, encode_param) if False == result: print('could not encode image!') quit() # decode from jpeg format jpeg_img = cv2.imdecode(encimg, 1) jpeg2input = np.transpose(jpeg_img, (2, 0, 1)) / 255 inputs[batch] = torch.Tensor(jpeg2input).to(device) # # save adversarial examples # save_inputs = inputs.cpu().numpy() # labels = labels.cpu().numpy() # from matplotlib.pyplot import imsave # # for j in range(batch_size): # image = save_inputs[j, :, :, :] # label = labels[j] # if label == 0: # imsave( # f"C:/Users/mmclab1/Desktop/fakecheck/dataset/adv_img_examples/" # f"fake_adversarial_image_{j}.png", # np.transpose(image, (1, 2, 0))) # else: # imsave( # f"C:/Users/mmclab1/Desktop/fakecheck/dataset/adv_img_examples/" # f"real_adversarial_image_{j}.png", # np.transpose(image, (1, 2, 0))) labels = labels.to(device) # 학습 가능한 가중치인 "optimizer 객체" 사용하여, 갱신할 변수들에 대한 모든 변화도 0으로 설정 # backward() 호출시, 변화도가 buffer 에 덮어쓰지 않고 누적되기 때문. optimizer.zero_grad() # forward pass # gradient 계산하는 모드로, 학습 시에만 연산 기록을 추적 with torch.set_grad_enabled(phase == 'train'): outputs = model(inputs) # h(x) 값, 모델의 예측 값 _, preds = torch.max(outputs, 1) # dim = 1, output의 각 sample 결과값(row)에서 max값 1개만 뽑음. loss = criterion(outputs, labels) # h(x) 모델이 잘 예측했는지 판별하는 loss function # training phase에서만 backward + optimize 수행 if phase == 'train': loss.backward() # gradient 계산 optimizer.step() # parameter update # statistics running_loss += loss.item() * inputs.size(0) # inputs.size(0) == batch size running_corrects += torch.sum(preds == labels.data) # True == 1, False == 0, 총 정답 수 num_cnt += len(labels) # len(labels) == batch size if phase == 'train': scheduler.step() # Learning Rate Scheduler epoch_loss = float(running_loss / num_cnt) epoch_acc = float((running_corrects.double() / num_cnt).cpu() * 100) if phase == 'train': train_loss.append(epoch_loss) train_acc.append(epoch_acc) else: valid_loss.append(epoch_loss) valid_acc.append(epoch_acc) print('{} Loss: {:.2f} Acc: {:.1f}'.format(phase, epoch_loss, epoch_acc)) # deep copy the model if phase == 'valid' and epoch_acc > best_acc: best_idx = epoch best_acc = epoch_acc best_model_wts = copy.deepcopy(model.state_dict()) # best_model_wts = copy.deepcopy(model.module.state_dict()) print('==> best model saved - %d / %.1f' % (best_idx, best_acc)) time_elapsed = time.time() - since print('Training complete in {:.0f}m {:.0f}s'.format(time_elapsed // 60, time_elapsed % 60)) print('Best valid Acc: %d - %.1f' % (best_idx, best_acc)) # load best model weights PATH = 'pytorch_model_adv_epoch30_4_sgd_resize3_comp15.pt' model.load_state_dict(best_model_wts) # torch.save(model.state_dict(), PATH) # 모델 객체의 state_dict 저장 torch.save(model, PATH) # 전체모델 저장 torch.save(model.state_dict(), f'C:/Users/mmclab1/.cache/torch/hub/checkpoints/{PATH}') print('model saved') # train, validation의 loss, acc 그래프로 나타내기 plt.subplot(311) plt.plot(train_loss) plt.plot(valid_loss) plt.ylabel('loss') plt.xlabel('epoch') plt.legend(['train', 'validation'], loc='upper left') plt.subplot(313) plt.plot(train_acc) plt.plot(valid_acc) plt.ylabel('acc') plt.xlabel('epoch') plt.legend(['train', 'validation'], loc='upper left') plt.savefig('graph_adv_epoch30_4_sgd_resize3_comp15.png') plt.show() return model, best_idx, best_acc, train_loss, train_acc, valid_loss, valid_acc, inputs
def train_epoch(model, loader, optimizer): model.train() train_loss = [] bar = tqdm(loader) for i, (data, target, face_name, df_method) in enumerate(bar): optimizer.zero_grad() if args.use_meta: data, meta = data data, meta, target = data.to(device), meta.to(device), target.to( device) logits = model(data, meta) else: # attack 추가 method = { '0_PGD': [20, 70, 2], '1_APGD': [20, 70, 2], '2_FGSM': [2, 8], '3_FFGSM': [4, 7, 10], '4_MIFGSM': [3, 6], '5_RFGSM': [4, 7, 8], '6_BIM': [4, 10, 1], '7_CW': [1e-4, 2e-4] } # 1. original data save # img_o # 2. small sized data # img_s = scaling(image_o, scaling_factor=0.5) # out_attack = attack(small_data, target~~~) # img_gen = normalize ( scaling ((out_attack - small_data), 1/scaling_factor) + img_o) for eps in range(2): globals()['atk{}'.format(0)] = torchattacks.PGD( model, eps=method['0_PGD'][eps] / 255, alpha=method['0_PGD'][-1] / 255, steps=4) globals()['atk{}'.format(1)] = torchattacks.APGD( model, eps=method['1_APGD'][eps] / 255, alpha=method['1_APGD'][-1] / 255, steps=4) globals()['atk{}'.format(2)] = torchattacks.FGSM( model, eps=method['2_FGSM'][eps] / 255) globals()['atk{}'.format(3)] = torchattacks.FFGSM( model, eps=method['3_FFGSM'][eps] / 255, alpha=method['3_FFGSM'][-1] / 255) globals()['atk{}'.format(4)] = torchattacks.MIFGSM( model, eps=method['4_MIFGSM'][eps] / 255, steps=4) globals()['atk{}'.format(5)] = torchattacks.RFGSM( model, eps=method['5_RFGSM'][eps] / 255, alpha=method['5_RFGSM'][-1] / 255, steps=4) globals()['atk{}'.format(6)] = torchattacks.BIM( model, eps=method['6_BIM'][eps] / 255, alpha=method['6_BIM'][-1] / 255) globals()['atk{}'.format(7)] = torchattacks.CW( model, c=method['7_CW'][eps], steps=10) for count in range(8): # globals()['data_atk{}'.format(i)] globals()['data_atk{}'.format(count)] = globals()[ 'atk{}'.format(count)](data, (target + 1) % 2) globals()['data_atk{}'.format(count)], target = ( globals()['data_atk{}'.format(count)] ).to(device), target.to(device) logits = model(globals()['data_atk{}'.format(count)]) globals()['data_atk{}'.format(count)] = ( globals()['data_atk{}'.format(count)]).cpu().numpy() method_keys = list(method.keys()) bat_size = args.batch_size for j in range(bat_size): for save_cnt in range(8): globals()['im{}'.format(save_cnt)] = (globals()[ 'data_atk{}'.format(save_cnt)])[j, :, :, :] # imsave( # f"./confirm_attack2img/AE-classification/{method_keys[save_cnt]}/" # f"{target[j]}_{face_name[j]}_{i * bat_size + j}_{method_keys[save_cnt]}_eps{method[method_keys[save_cnt]][eps]}_wsbs.png", # np.transpose(globals()['im{}'.format(save_cnt)], (1, 2, 0))) imsave( f"./confirm_attack2img/AE-classification/train/" f"{target[j]}_{face_name[j]}_{df_method[j]}_{i * bat_size + j}_{method_keys[save_cnt]}_eps{method[method_keys[save_cnt]][eps]}_wsbs.png", np.transpose(globals()['im{}'.format(save_cnt)], (1, 2, 0))) loss = criterion(logits, target) if not args.use_amp: loss.backward() # else: # with amp.scale_loss(loss, optimizer) as scaled_loss: # scaled_loss.backward() if args.image_size in [896, 576]: # 그라디언트가 너무 크면 값을 0.5로 잘라준다 (max_grad_norm=0.5) torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5) # gradient accumulation (메모리 부족할때) if args.accumulation_step: if (i + 1) % args.accumulation_step == 0: optimizer.step() # optimizer.zero_grad() else: optimizer.step() # optimizer.zero_grad() loss_np = loss.detach().cpu().numpy() train_loss.append(loss_np) smooth_loss = sum(train_loss[-100:]) / min(len(train_loss), 100) bar.set_description('loss: %.5f, smooth_loss: %.5f' % (loss_np, smooth_loss)) train_loss = np.mean(train_loss) return train_loss
def train_epoch(model, loader, optimizer): model.train() train_loss = [] bar = tqdm(loader) for i, (data, target, face_name) in enumerate(bar): optimizer.zero_grad() if args.use_meta: data, meta = data data, meta, target = data.to(device), meta.to(device), target.to(device) logits = model(data, meta) else: # attack 추가 method = { '1_PGD': [20,70,2], '2_APGD':[20,70,2], '3_FGSM': [2,8], '4_FFGSM': [4,7,10], '5_MIFGSM': [3,6], '6_RFGSM': [4,7,8], '7_BIM':[4,10,1], '8_CW':[1e-4, 2e-4]} #TODO: dataset에 original image와 attacked image모두 만들기 # 1. original data save img_o # 2. small sized data img_s = scaling(image_o, scaling_factor=0.5) # out_attack = attack(small_data, target~~~) # img_gen = normalize ( scaling ((out_attack - small_data), 1/scaling_factor) + img_o) scaling_factor = 0.5 img_origin = np.transpose(data.cpu().numpy()[i, :, :, :], (1, 2, 0)) img_small = cv2.resize(img_origin, dsize=(0, 0), fx=scaling_factor, fy=scaling_factor) #,interpolation=cv2.INTER_AREA for eps in range(2): globals()['atk{}'.format(1)] = torchattacks.PGD(model, eps=method['1_PGD'][eps] / 255, alpha=method['1_PGD'][-1] / 255, steps=4) globals()['atk{}'.format(2)] = torchattacks.APGD(model, eps=method['2_APGD'][eps] / 255, alpha=method['2_APGD'][-1] / 255, steps=4) globals()['atk{}'.format(3)] = torchattacks.FGSM(model, eps=method['3_FGSM'][eps] / 255) globals()['atk{}'.format(4)] = torchattacks.FFGSM(model, eps=method['4_FFGSM'][eps] / 255, alpha=method['4_FFGSM'][-1] / 255) globals()['atk{}'.format(5)] = torchattacks.MIFGSM(model, eps=method['5_MIFGSM'][eps] / 255, steps=4) globals()['atk{}'.format(6)] = torchattacks.RFGSM(model, eps=method['6_RFGSM'][eps] / 255, alpha=method['6_RFGSM'][-1] / 255, steps=4) globals()['atk{}'.format(7)] = torchattacks.BIM(model, eps=method['7_BIM'][eps] / 255, alpha=method['7_BIM'][-1] / 255) globals()['atk{}'.format(8)] = torchattacks.CW(model, c= method['8_CW'][eps], steps=10) for count in range(1,9): # regularization # torch.clamp(images + delta, min=0, max=1).detach() # torch.from_numpy(img_small) out_attack = globals()['atk{}'.format(count)](torch.from_numpy(img_small), (target + 1) % 2) img_gen = torch.clamp(cv2.resize(out_attack-img_small,dsize=(0,0),fx=1/scaling_factor, fy=1/scaling_factor), min=0, max=1).detach() + img_origin globals()['data_atk{}'.format(count)] = torch.from_numpy(img_gen) globals()['data_atk{}'.format(count)], target = (globals()['data_atk{}'.format(count)]).to(device), target.to(device) logits = model(globals()['data_atk{}'.format(count)]) globals()['data_atk{}'.format(count)] = (globals()['data_atk{}'.format(count)]).cpu().numpy() method_keys = list(method.keys()) bat_size = args.batch_size for j in range(bat_size): # save original image # im0 = data.cpu().numpy()[j, :, :, :] # imsave( # f"./confirm_attack2img/AE-real_fake/0_original/" # f"{target[j]}_{face_name[j]}_{i * bat_size + j}_0_wsbs.png", # np.transpose(im0, (1, 2, 0))) # save attacked image for save_cnt in range(1,9): globals()['im{}'.format(save_cnt)] = (globals()['data_atk{}'.format(save_cnt)])[j, :, :, :] # imsave( # f"./confirm_attack2img/AE-real_fake/{method_keys[save_cnt]}/" # f"{target[j]}_{face_name[j]}_{i * bat_size + j}_{method_keys[save_cnt]}_eps{method[method_keys[save_cnt]][eps]}_wsbs.png", # np.transpose(globals()['im{}'.format(save_cnt)], (1, 2, 0))) imsave( f"./confirm_attack2img/AE-real_fake/train/" f"{target[j]}_{face_name[j]}_{i * bat_size + j}_{method_keys[save_cnt]}_eps{method[method_keys[save_cnt]][eps]}_wsbs.png", np.transpose(globals()['im{}'.format(save_cnt)], (1, 2, 0))) loss = criterion(logits, target) if not args.use_amp: loss.backward() # else: # with amp.scale_loss(loss, optimizer) as scaled_loss: # scaled_loss.backward() if args.image_size in [896, 576]: # 그라디언트가 너무 크면 값을 0.5로 잘라준다 (max_grad_norm=0.5) torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5) # gradient accumulation (메모리 부족할때) if args.accumulation_step: if (i + 1) % args.accumulation_step == 0: optimizer.step() # optimizer.zero_grad() else: optimizer.step() # optimizer.zero_grad() loss_np = loss.detach().cpu().numpy() train_loss.append(loss_np) smooth_loss = sum(train_loss[-100:]) / min(len(train_loss), 100) bar.set_description('loss: %.5f, smooth_loss: %.5f' % (loss_np, smooth_loss)) train_loss = np.mean(train_loss) return train_loss
def train_epoch(model, loader, optimizer): model.train() train_loss = [] bar = tqdm(loader) for i, (data, target, image_name) in enumerate(bar): optimizer.zero_grad() if args.use_meta: data, meta = data data, meta, target = data.to(device), meta.to(device), target.to( device) logits = model(data, meta) else: # attack 추가: 5,499장 --> 총 7개 attack x epsilon 2개 == 76,986 method = { '0_FGSM': [2, 5, 8], '1_PGD': [20, 50, 80, 2], '2_BIM': [4, 7, 10, 1] } for eps in range(3): globals()['atk{}'.format(0)] = torchattacks.FGSM( model, eps=method['0_FGSM'][eps] / 255) globals()['atk{}'.format(1)] = torchattacks.PGD( model, eps=method['1_PGD'][eps] / 255, alpha=method['1_PGD'][-1] / 255, steps=4) globals()['atk{}'.format(2)] = torchattacks.BIM( model, eps=method['2_BIM'][eps] / 255, alpha=method['2_BIM'][-1] / 255) for count in range(3): # globals()['data_atk{}'.format(i)] globals()['data_atk{}'.format(count)] = globals()[ 'atk{}'.format(count)](data, (target + 1) % 2) globals()['data_atk{}'.format(count)], target = ( globals()['data_atk{}'.format(count)] ).to(device), target.to(device) logits = model(globals()['data_atk{}'.format(count)]) globals()['data_atk{}'.format(count)] = ( globals()['data_atk{}'.format(count)]).cpu().numpy() method_keys = list(method.keys()) bat_size = args.batch_size for j in range(bat_size): for save_cnt in range(3): globals()['im{}'.format(save_cnt)] = (globals()[ 'data_atk{}'.format(save_cnt)])[j, :, :, :] # imsave( # f"./confirm_attack2img/AE-classification/{method_keys[save_cnt]}/" # f"{target[j]}_{id[j]}_{i * bat_size + j}_{method_keys[save_cnt]}_eps{method[method_keys[save_cnt]][eps]}.png", # np.transpose(globals()['im{}'.format(save_cnt)], (1, 2, 0))) imsave( f"./data/Adversarial Attack/{method_keys[save_cnt]}/" f"{image_name[j].split('.')[0]}_{method_keys[save_cnt]}_eps{method[method_keys[save_cnt]][eps]}_{i * bat_size + j}.png", np.transpose(globals()['im{}'.format(save_cnt)], (1, 2, 0))) loss = criterion(logits, target) if not args.use_amp: loss.backward() # else: # with amp.scale_loss(loss, optimizer) as scaled_loss: # scaled_loss.backward() if args.image_size in [896, 576]: # 그라디언트가 너무 크면 값을 0.5로 잘라준다 (max_grad_norm=0.5) torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5) # gradient accumulation (메모리 부족할때) if args.accumulation_step: if (i + 1) % args.accumulation_step == 0: optimizer.step() # optimizer.zero_grad() else: optimizer.step() # optimizer.zero_grad() loss_np = loss.detach().cpu().numpy() train_loss.append(loss_np) smooth_loss = sum(train_loss[-100:]) / min(len(train_loss), 100) bar.set_description('loss: %.5f, smooth_loss: %.5f' % (loss_np, smooth_loss)) train_loss = np.mean(train_loss) return train_loss
def test_model(model, phase='test'): # phase = 'train', 'valid', 'test' model.eval() # evaluate mode; gradient 계산 안함. running_loss, running_corrects, num_cnt = 0.0, 0, 0 ''' with torch.no_grad(): # memory save를 위해 gradient 저장하지 않음. 보통 test를 할 때, gradient를 training 시키는 것이 아니기 때문에 위와 같은 코드를 추가한다. grad = torch.autograd.grad(cost, images, retain_graph=False, create_graph=False)[0] 하지만, adversarial attack은 위와 같이 gradient를 토대로 data에 공격을 가하기 때문에 gradient가 필요하다. 따라서 test_adv 에는 with torch.no_grad()를 제외해야 한다. ''' for i, (inputs, labels) in enumerate(dataloaders[phase]): # adversarial attack 정의 atks = [ torchattacks.FGSM(model, eps=8 / 255), torchattacks.BIM(model, eps=8 / 255, alpha=2 / 255, steps=7), torchattacks.PGD(model, eps=8 / 255, alpha=2 / 255, steps=7), ] adv_images = atks[0](inputs, labels).to(device) # Image Processing Based Defense Methods --> tensor를 image로 변환하여 적용 for batch in range(inputs.shape[0]): tensor2img = transforms.ToPILImage()(inputs[batch]).convert('RGB') # 1. Resizing # Image.resize(size, resample=3, box=None, reducing_gap=None) # resample(filter): PIL.Image.NEAREST, PIL.Image.BOX, PIL.Image.BILINEAR, PIL.Image.HAMMING, PIL.Image.BICUBIC tensor2img.resize((74, 74)) tensor2img.resize((224, 224)) # 다시 이미지를 tensor로 바꾸기 tensor_img = transforms.ToTensor()(tensor2img) inputs[batch] = tensor_img # 2. jpeg compression tensor2numpy = inputs[batch].cpu().numpy() cv_img = np.transpose(tensor2numpy, (1, 2, 0)) # [w, h, c] cv_img = cv_img * 255 encode_param = [int(cv2.IMWRITE_JPEG_QUALITY), 15] result, encimg = cv2.imencode('.jpg', cv_img, encode_param) if False == result: print('could not encode image!') quit() # decode from jpeg format jpeg_img = cv2.imdecode(encimg, 1) jpeg2input = np.transpose(jpeg_img, (2, 0, 1)) / 255 inputs[batch] = torch.Tensor(jpeg2input).to(device) labels = labels.to(device) outputs = model(adv_images) # forward pass _, preds = torch.max(outputs, 1) # model이 가장 높은 확률로 예측한 label loss = criterion(outputs, labels) # loss 계산 running_loss += loss.item() * inputs.size(0) running_corrects += torch.sum(preds == labels.data) num_cnt += inputs.size(0) # batch size test_loss = running_loss / num_cnt test_acc = running_corrects.double() / num_cnt print('test done : loss/acc : %.2f / %.1f' % (test_loss, test_acc * 100))
def ta_bim(x, y, model, eps=4 / 255, alpha=1 / 255, steps=0): attack = torchattacks.BIM(model, eps=eps, alpha=alpha, steps=steps) advs = attack(x, y) return advs
ssim = structural_similarity(ori_image, adv_image, data_range=255, multichannel=False) psnr = peak_signal_noise_ratio(ori_image, adv_image, data_range=255) l2 = np.linalg.norm(ori_image - adv_image) return ssim, psnr, l2 ################ 开始攻击 ################## epsilons = [0, 0.05, 0.1, 0.15, 0.2, 0.25, 0.3] for epsilon in epsilons: attack = torchattacks.BIM(model, eps=epsilon, alpha=1 / 255, iters=0) correct = 0 ssim = 0 psnr = 0 l2 = 0 for data, target in test_loader: # Send the data and label to the device data, target = data.to(device), target.to(device) # 防止对原始样本进行了更改 data_t = data.clone() target_t = target.clone() adv_images = attack(data_t, target_t) # show_cmp(data, adv_images) # 显示图像,只能batch_size=1的时候才可以 # break