def test_attack(threshold, arch, dataset, test_loader): target_model = StandardModel(dataset, arch, no_grad=False) if torch.cuda.is_available(): target_model = target_model.cuda() target_model.eval() attack = LinfPGDAttack(target_model, loss_fn=nn.CrossEntropyLoss(reduction="sum"), eps=threshold, nb_iter=30, eps_iter=0.01, rand_init=True, clip_min=0.0, clip_max=1.0, targeted=False) all_count = 0 success_count = 0 all_adv_images = [] all_true_labels = [] for idx, (img, true_label) in enumerate(test_loader): img = img.cuda() true_label = true_label.cuda().long() adv_image = attack.perturb(img, true_label) # (3, 224, 224), float if adv_image is None: continue adv_label = target_model.forward(adv_image).max(1)[1].detach().cpu().numpy().astype(np.int32) # adv_image = np.transpose(adv_image, (0, 2, 3, 1)) # N,C,H,W -> (N, H, W, 3), float all_count += len(img) true_label_np = true_label.detach().cpu().numpy().astype(np.int32) success_count+= len(np.where(true_label_np != adv_label)[0]) all_adv_images.append(adv_image.cpu().detach().numpy()) all_true_labels.append(true_label_np) attack_success_rate = success_count / float(all_count) log.info("Before train. Attack success rate is {:.3f}".format(attack_success_rate)) return target_model, np.concatenate(all_adv_images,0), np.concatenate(all_true_labels, 0) # N,224,224,3
def load_models(dataset): archs = [] model_path_list = [] if dataset == "CIFAR-10" or dataset == "CIFAR-100": for arch in ["resnet-110","WRN-28-10","WRN-34-10","resnext-8x64d","resnext-16x64d"]: test_model_path = "{}/train_pytorch_model/real_image_model/{}-pretrained/{}/checkpoint.pth.tar".format( PY_ROOT, dataset, arch) if os.path.exists(test_model_path): archs.append(arch) model_path_list.append(test_model_path) else: log.info(test_model_path + " does not exist!") elif dataset == "TinyImageNet": # for arch in ["vgg11_bn","resnet18","vgg16_bn","resnext64_4","densenet121"]: for arch in MODELS_TEST_STANDARD[dataset]: test_model_path = "{}/train_pytorch_model/real_image_model/{}@{}@*.pth.tar".format( PY_ROOT, dataset, arch) test_model_path = list(glob.glob(test_model_path))[0] if os.path.exists(test_model_path): archs.append(arch) model_path_list.append(test_model_path) else: log.info(test_model_path + "does not exist!") else: for arch in ["inceptionv3","inceptionv4", "inceptionresnetv2","resnet101", "resnet152"]: test_model_list_path = "{}/train_pytorch_model/real_image_model/{}-pretrained/checkpoints/{}*.pth".format( PY_ROOT, dataset, arch) test_model_path = list(glob.glob(test_model_list_path)) if len(test_model_path) == 0: # this arch does not exists in args.dataset continue archs.append(arch) model_path_list.append(test_model_path[0]) models = [] print("begin construct model") if dataset == "TinyImageNet": for idx, arch in enumerate(archs): model = MetaLearnerModelBuilder.construct_tiny_imagenet_model(arch, dataset) model_path = model_path_list[idx] model.load_state_dict(torch.load(model_path, map_location=lambda storage, location: storage)["state_dict"]) model.cuda() model.eval() models.append(model) else: for arch in archs: model = StandardModel(dataset, arch, no_grad=True) model.cuda() model.eval() models.append(model) print("end construct model") return models
def main(args, arch): model = StandardModel(args["dataset"], arch, False) model.cuda() model.eval() # attack related settings if args["attack_method"] == "zoo" or args[ "attack_method"] == "autozoom_bilin": if args["img_resize"] is None: args["img_resize"] = model.input_size[-1] log.info( "Argument img_resize is not set and not using autoencoder, set to image original size:{}" .format(args["img_resize"])) codec = None if args["attack_method"] == "zoo_ae" or args[ "attack_method"] == "autozoom_ae": codec = Codec(model.input_size[-1], IN_CHANNELS[args["dataset"]], args["compress_mode"], args["resize"], use_tanh=args["use_tanh"]) codec.load_codec(args["codec_path"]) codec.cuda() decoder = codec.decoder args["img_resize"] = decoder.input_shape[1] log.info( "Loading autoencoder: {}, set the attack image size to:{}".format( args["codec_path"], args["img_resize"])) # setup attack if args["attack_method"] == "zoo": blackbox_attack = ZOO(model, args["dataset"], args) elif args["attack_method"] == "zoo_ae": blackbox_attack = ZOO_AE(model, args["dataset"], args, decoder) elif args["attack_method"] == "autozoom_bilin": blackbox_attack = AutoZOOM_BiLIN(model, args["dataset"], args) elif args["attack_method"] == "autozoom_ae": blackbox_attack = AutoZOOM_AE(model, args["dataset"], args, decoder) target_str = "untargeted" if args[ "attack_type"] != "targeted" else "targeted_{}".format( args["target_type"]) save_result_path = args[ "exp_dir"] + "/data_{}@arch_{}@attack_{}@{}_result.json".format( args["dataset"], arch, args["attack_method"], target_str) attack_framework = AutoZoomAttackFramework(args) attack_framework.attack_dataset_images(args, blackbox_attack, arch, model, codec, save_result_path) model.cpu()
def main(args, arch): model = StandardModel(args.dataset, arch, args.solver != "fake_zero") model.cuda() model.eval() if args.init_size is None: args.init_size = model.input_size[-1] log.info( "Argument init_size is not set and not using autoencoder, set to image original size:{}" .format(args.init_size)) target_str = "untargeted" if not args.targeted else "targeted_{}".format( args.target_type) save_result_path = args.exp_dir + "/data_{}@arch_{}@solver_{}@{}_result.json".format( args.dataset, arch, args.solver, target_str) if os.path.exists(save_result_path): model.cpu() return attack_framework = ZooAttackFramework(args) attack_framework.attack_dataset_images(args, arch, model, save_result_path) model.cpu()
def main(): parser = argparse.ArgumentParser( description='Square Attack Hyperparameters.') parser.add_argument('--norm', type=str, required=True, choices=['l2', 'linf']) parser.add_argument('--dataset', type=str, required=True) parser.add_argument( '--gpu', type=str, required=True, help='GPU number. Multiple GPUs are possible for PT models.') parser.add_argument( '--p', type=float, default=0.05, help= 'Probability of changing a coordinate. Note: check the paper for the best values. ' 'Linf standard: 0.05, L2 standard: 0.1. But robust models require higher p.' ) parser.add_argument('--epsilon', type=float, help='Radius of the Lp ball.') parser.add_argument('--max_queries', type=int, default=1000) parser.add_argument( '--json-config', type=str, default= '/home1/machen/meta_perturbations_black_box_attack/configures/square_attack_conf.json', help='a configures file to be passed in instead of arguments') parser.add_argument('--batch_size', type=int, default=100) parser.add_argument('--targeted', action="store_true") parser.add_argument('--target_type', type=str, default='random', choices=['random', 'least_likely', "increment"]) parser.add_argument('--loss', type=str) args = parser.parse_args() os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu if args.json_config: # If a json file is given, use the JSON file as the base, and then update it with args defaults = json.load(open(args.json_config))[args.dataset][args.norm] arg_vars = vars(args) arg_vars = { k: arg_vars[k] for k in arg_vars if arg_vars[k] is not None } defaults.update(arg_vars) args = SimpleNamespace(**defaults) if args.targeted and args.dataset == "ImageNet": args.max_queries = 10000 save_dir_path = "{}/data_square_attack/{}/{}".format( PY_ROOT, args.dataset, "targeted_attack" if args.targeted else "untargeted_attack") os.makedirs(save_dir_path, exist_ok=True) loss_type = "cw" if not args.targeted else "xent" args.loss = loss_type log_path = osp.join( save_dir_path, get_log_path(args.dataset, loss_type, args.norm, args.targeted, args.target_type)) set_log_file(log_path) log.info('Command line is: {}'.format(' '.join(sys.argv))) log.info("Log file is written in {}".format(log_path)) log.info('Called with args:') print_args(args) trn_data_loader = DataLoaderMaker.get_img_label_data_loader( args.dataset, args.batch_size, is_train=True) models = [] for arch in MODELS_TRAIN_STANDARD[args.dataset]: if StandardModel.check_arch(arch, args.dataset): model = StandardModel(args.dataset, arch, True) model = model.eval() models.append({"arch_name": arch, "model": model}) model_data_dict = defaultdict(list) for images, labels in trn_data_loader: model_info = random.choice(models) arch = model_info["arch_name"] model = model_info["model"] if images.size(-1) != model.input_size[-1]: images = F.interpolate(images, size=model.input_size[-1], mode='bilinear', align_corners=True) model_data_dict[(arch, model)].append((images, labels)) log.info("Assign data to multiple models over!") attacker = SquareAttack(args.dataset, args.targeted, args.target_type, args.epsilon, args.norm, max_queries=args.max_queries) attacker.attack_all_images(args, model_data_dict, save_dir_path) log.info("All done!")
test_model_list_path = list(glob.glob(test_model_list_path)) if len(test_model_list_path ) == 0: # this arch does not exists in args.dataset continue archs.append(arch) else: assert args.arch is not None archs = [args.arch] args.arch = ", ".join(archs) log.info('Command line is: {}'.format(' '.join(sys.argv))) log.info("Log file is written in {}".format(log_file_path)) log.info('Called with args:') print_args(args) surrogate_model = StandardModel(args.dataset, args.surrogate_arch, False) surrogate_model.cuda() surrogate_model.eval() attacker = SwitchNeg(args.dataset, args.batch_size, args.targeted, args.target_type, args.epsilon, args.norm, 0.0, 1.0, args.max_queries) for arch in archs: if args.attack_defense: save_result_path = args.exp_dir + "/{}_{}_result.json".format( arch, args.defense_model) else: save_result_path = args.exp_dir + "/{}_result.json".format(arch) if os.path.exists(save_result_path): continue log.info("Begin attack {} on {}, result will be saved to {}".format( arch, args.dataset, save_result_path))
state.targeted = args.targeted state.dataset = args.dataset state.batch_size = args.batch_size device = torch.device(args.gpu) train_loader = DataLoaderMaker.get_img_label_data_loader( args.dataset, state.batch_size, True) val_loader = DataLoaderMaker.get_img_label_data_loader( args.dataset, state.batch_size, False) nets = [] log.info("Initialize pretrained models.") for model_name in MODELS_TRAIN_STANDARD[args.dataset]: pretrained_model = StandardModel(args.dataset, model_name, no_grad=False) # pretrained_model.cuda() pretrained_model.eval() nets.append(pretrained_model) log.info("Initialize over!") model = nn.Sequential(ImagenetEncoder(), ImagenetDecoder(args.dataset)) model = model.cuda() optimizer_G = torch.optim.SGD(model.parameters(), state.learning_rate_G, momentum=state.momentum, weight_decay=0, nesterov=True) scheduler_G = torch.optim.lr_scheduler.StepLR(optimizer_G, step_size=state.epochs // state.schedule, gamma=state.gamma) hingeloss = MarginLoss(margin=state.margin, target=state.targeted)
lr_scheduler = optim.lr_scheduler.LambdaLR(optimizer, [anneal_lr]) for epoch in range(EPOCH_TOTAL): # loop over the dataset multiple times logger.info("Start Epoch {}".format(epoch)) running_loss_1, running_loss_2 = 0.0, 0.0 lr_scheduler.step(epoch) for i, data_batch in enumerate(train_loader): # get the inputs; data is a list of [inputs, labels] # if i == 19: # break img_batch, label_batch = data_batch img_batch, label_batch = img_batch.cuda(), label_batch.cuda() train_img_batch, train_label_batch = [], [] model.eval() if args.adv_ratio > 0.: adv_x = pgd.projected_gradient_descent(model, img_batch, **pgd_kwargs) adv_x_s, adv_label_batch_s = shuffle_minibatch(adv_x, label_batch) train_img_batch.append(adv_x_s) train_label_batch.append(adv_label_batch_s) if args.adv_ratio < 1.: img_batch_s, label_batch_s = shuffle_minibatch( img_batch, label_batch) train_img_batch.append(img_batch_s) train_label_batch.append(label_batch_s) train_img_batch = torch.cat(train_img_batch, dim=0) model.train()
def attack_all_images(self, args, arch, tmp_dump_path, result_dump_path): # subset_pos用于回调函数汇报汇总统计结果 model = StandardModel(args.dataset, arch, no_grad=True) model.cuda() model.eval() # 带有缩减功能的,攻击成功的图片自动删除掉 for data_idx, data_tuple in enumerate(self.dataset_loader): if os.path.exists(tmp_dump_path): with open(tmp_dump_path, "r") as file_obj: json_content = json.load(file_obj) resume_batch_idx = int(json_content["batch_idx"]) # resume for key in [ 'query_all', 'correct_all', 'not_done_all', 'success_all', 'success_query_all' ]: if key in json_content: setattr( self, key, torch.from_numpy(np.asarray( json_content[key])).float()) if data_idx < resume_batch_idx: # resume continue if args.dataset == "ImageNet": if model.input_size[-1] >= 299: images, true_labels = data_tuple[1], data_tuple[2] else: images, true_labels = data_tuple[0], data_tuple[2] else: images, true_labels = data_tuple[0], data_tuple[1] if images.size(-1) != model.input_size[-1]: images = F.interpolate(images, size=model.input_size[-1], mode='bilinear', align_corners=True) # skip_batch_index_list = np.nonzero(np.asarray(chunk_skip_indexes[data_idx]))[0].tolist() selected = torch.arange( data_idx * args.batch_size, min((data_idx + 1) * args.batch_size, self.total_images)) # 选择这个batch的所有图片的index img_idx_to_batch_idx = ImageIdxToOrigBatchIdx(args.batch_size) images, true_labels = images.cuda(), true_labels.cuda() first_finetune = True finetune_queue = FinetuneQueue(args.batch_size, args.meta_seq_len, img_idx_to_batch_idx) prior_size = model.input_size[ -1] if not args.tiling else args.tile_size assert args.tiling == (args.dataset == "ImageNet") if args.tiling: upsampler = Upsample(size=(model.input_size[-2], model.input_size[-1])) else: upsampler = lambda x: x with torch.no_grad(): logit = model(images) pred = logit.argmax(dim=1) query = torch.zeros(images.size(0)).cuda() correct = pred.eq(true_labels).float() # shape = (batch_size,) not_done = correct.clone() # shape = (batch_size,) if args.targeted: if args.target_type == 'random': target_labels = torch.randint( low=0, high=CLASS_NUM[args.dataset], size=true_labels.size()).long().cuda() invalid_target_index = target_labels.eq(true_labels) while invalid_target_index.sum().item() > 0: target_labels[invalid_target_index] = torch.randint( low=0, high=logit.shape[1], size=target_labels[invalid_target_index].shape ).long().cuda() invalid_target_index = target_labels.eq(true_labels) elif args.target_type == 'least_likely': target_labels = logit.argmin(dim=1) elif args.target_type == "increment": target_labels = torch.fmod(true_labels + 1, CLASS_NUM[args.dataset]) else: raise NotImplementedError('Unknown target_type: {}'.format( args.target_type)) else: target_labels = None prior = torch.zeros(images.size(0), IN_CHANNELS[args.dataset], prior_size, prior_size).cuda() prior_step = self.gd_prior_step if args.norm == 'l2' else self.eg_prior_step image_step = self.l2_image_step if args.norm == 'l2' else self.linf_step proj_step = self.l2_proj_step if args.norm == 'l2' else self.linf_proj_step # 调用proj_maker返回的是一个函数 criterion = self.cw_loss if args.data_loss == "cw" else self.xent_loss adv_images = images.clone() for step_index in range(1, args.max_queries + 1): # Create noise for exporation, estimate the gradient, and take a PGD step dim = prior.nelement() / images.size( 0) # nelement() --> total number of elements exp_noise = args.exploration * torch.randn_like(prior) / ( dim**0.5 ) # parameterizes the exploration to be done around the prior exp_noise = exp_noise.cuda() q1 = upsampler( prior + exp_noise ) # 这就是Finite Difference算法, prior相当于论文里的v,这个prior也会更新,把梯度累积上去 q2 = upsampler( prior - exp_noise) # prior 相当于累积的更新量,用这个更新量,再去修改image,就会变得非常准 # Loss points for finite difference estimator q1_images = adv_images + args.fd_eta * q1 / self.norm(q1) q2_images = adv_images + args.fd_eta * q2 / self.norm(q2) predict_by_target_model = False if (step_index <= args.warm_up_steps or (step_index - args.warm_up_steps) % args.meta_predict_steps == 0): log.info("predict from target model") predict_by_target_model = True with torch.no_grad(): q1_logits = model(q1_images) q2_logits = model(q2_images) q1_logits = q1_logits / torch.norm( q1_logits, p=2, dim=-1, keepdim=True) # 加入normalize q2_logits = q2_logits / torch.norm( q2_logits, p=2, dim=-1, keepdim=True) finetune_queue.append(q1_images.detach(), q2_images.detach(), q1_logits.detach(), q2_logits.detach()) if step_index >= args.warm_up_steps: q1_images_seq, q2_images_seq, q1_logits_seq, q2_logits_seq = finetune_queue.stack_history_track( ) finetune_times = args.finetune_times if first_finetune else random.randint( 3, 5) # FIXME log.info("begin finetune for {} times".format( finetune_times)) self.meta_finetuner.finetune( q1_images_seq, q2_images_seq, q1_logits_seq, q2_logits_seq, finetune_times, first_finetune, img_idx_to_batch_idx) first_finetune = False else: with torch.no_grad(): q1_logits, q2_logits = self.meta_finetuner.predict( q1_images, q2_images, img_idx_to_batch_idx) q1_logits = q1_logits / torch.norm( q1_logits, p=2, dim=-1, keepdim=True) q2_logits = q2_logits / torch.norm( q2_logits, p=2, dim=-1, keepdim=True) l1 = criterion(q1_logits, true_labels, target_labels) l2 = criterion(q2_logits, true_labels, target_labels) # Finite differences estimate of directional derivative est_deriv = (l1 - l2) / (args.fd_eta * args.exploration ) # 方向导数 , l1和l2是loss # 2-query gradient estimate est_grad = est_deriv.view(-1, 1, 1, 1) * exp_noise # B, C, H, W, # Update the prior with the estimated gradient prior = prior_step( prior, est_grad, args.online_lr) # 注意,修正的是prior,这就是bandit算法的精髓 grad = upsampler(prior) # prior相当于梯度 ## Update the image: adv_images = image_step( adv_images, grad * correct.view(-1, 1, 1, 1), # 注意correct也是删减过的 args.image_lr) # prior放大后相当于累积的更新量,可以用来更新 adv_images = proj_step(images, args.epsilon, adv_images) adv_images = torch.clamp(adv_images, 0, 1) with torch.no_grad(): adv_logit = model(adv_images) # adv_pred = adv_logit.argmax(dim=1) adv_prob = F.softmax(adv_logit, dim=1) adv_loss = criterion(adv_logit, true_labels, target_labels) ## Continue query count if predict_by_target_model: query = query + 2 * not_done if args.targeted: not_done = not_done * ( 1 - adv_pred.eq(target_labels).float() ).float() # not_done初始化为 correct, shape = (batch_size,) else: not_done = not_done * adv_pred.eq( true_labels).float() # 只要是跟原始label相等的,就还需要query,还没有成功 success = (1 - not_done) * correct success_query = success * query not_done_loss = adv_loss * not_done not_done_prob = adv_prob[torch.arange(adv_images.size(0)), true_labels] * not_done log.info('Attacking image {} - {} / {}, step {}'.format( data_idx * args.batch_size, (data_idx + 1) * args.batch_size, self.total_images, step_index)) log.info(' not_done: {:.4f}'.format( len( np.where(not_done.detach().cpu().numpy().astype( np.int32) == 1)[0]) / float(args.batch_size))) log.info(' fd_scalar: {:.9f}'.format( (l1 - l2).mean().item())) if success.sum().item() > 0: log.info(' mean_query: {:.4f}'.format( success_query[success.byte()].mean().item())) log.info(' median_query: {:.4f}'.format( success_query[success.byte()].median().item())) if not_done.sum().item() > 0: log.info(' not_done_loss: {:.4f}'.format( not_done_loss[not_done.byte()].mean().item())) log.info(' not_done_prob: {:.4f}'.format( not_done_prob[not_done.byte()].mean().item())) not_done_np = not_done.detach().cpu().numpy().astype(np.int32) done_img_idx_list = np.where(not_done_np == 0)[0].tolist() delete_all = False if done_img_idx_list: for skip_index in done_img_idx_list: # 两次循环,第一次循环先汇报出去,第二次循环删除 batch_idx = img_idx_to_batch_idx[skip_index] pos = selected[batch_idx].item() # 先汇报被删减的值self.query_all for key in [ 'query', 'correct', 'not_done', 'success', 'success_query', 'not_done_loss', 'not_done_prob' ]: value_all = getattr(self, key + "_all") value = eval(key)[skip_index].item() value_all[pos] = value images, adv_images, prior, query, true_labels, target_labels, correct, not_done = \ self.delete_tensor_by_index_list(done_img_idx_list, images, adv_images, prior, query, true_labels, target_labels, correct, not_done) img_idx_to_batch_idx.del_by_index_list(done_img_idx_list) delete_all = images is None if delete_all: break # report to all stats the rest unsuccess for key in [ 'query', 'correct', 'not_done', 'success', 'success_query', 'not_done_loss', 'not_done_prob' ]: for img_idx, batch_idx in img_idx_to_batch_idx.proj_dict.items( ): pos = selected[batch_idx].item() value_all = getattr(self, key + "_all") value = eval(key)[img_idx].item() value_all[ pos] = value # 由于value_all是全部图片都放在一个数组里,当前batch选择出来 img_idx_to_batch_idx.proj_dict.clear() tmp_info_dict = { "batch_idx": data_idx + 1, "batch_size": args.batch_size } for key in [ 'query_all', 'correct_all', 'not_done_all', 'success_all', 'success_query_all' ]: value_all = getattr(self, key).detach().cpu().numpy().tolist() tmp_info_dict[key] = value_all with open(tmp_dump_path, "w") as result_file_obj: json.dump(tmp_info_dict, result_file_obj, sort_keys=True) log.info('Saving results to {}'.format(result_dump_path)) meta_info_dict = { "avg_correct": self.correct_all.mean().item(), "avg_not_done": self.not_done_all[self.correct_all.byte()].mean().item(), "mean_query": self.success_query_all[self.success_all.byte()].mean().item(), "median_query": self.success_query_all[self.success_all.byte()].median().item(), "max_query": self.success_query_all[self.success_all.byte()].max().item(), "correct_all": self.correct_all.detach().cpu().numpy().astype(np.int32).tolist(), "not_done_all": self.not_done_all.detach().cpu().numpy().astype(np.int32).tolist(), "query_all": self.query_all.detach().cpu().numpy().astype(np.int32).tolist(), "not_done_loss": self.not_done_loss_all[self.not_done_all.byte()].mean().item(), "not_done_prob": self.not_done_prob_all[self.not_done_all.byte()].mean().item(), "args": vars(args) } with open(result_dump_path, "w") as result_file_obj: json.dump(meta_info_dict, result_file_obj, sort_keys=True) log.info("done, write stats info to {}".format(result_dump_path)) self.query_all.fill_(0) self.correct_all.fill_(0) self.not_done_all.fill_(0) self.success_all.fill_(0) self.success_query_all.fill_(0) self.not_done_loss_all.fill_(0) self.not_done_prob_all.fill_(0) model.cpu()
transform=train_preprocessor) elif dataset == "TinyImageNet": train_dataset = TinyImageNet(IMAGE_DATA_ROOT[dataset], train_preprocessor, train=True) batch_size = args.batch_size train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=0) print('==> Building model..') arch_list = MODELS_TRAIN_STANDARD[args.dataset] model_dict = {} for arch in arch_list: if StandardModel.check_arch(arch, args.dataset): print("begin use arch {}".format(arch)) model = StandardModel(args.dataset, arch, no_grad=True) model_dict[arch] = model.eval() print("use arch {} done".format(arch)) print("==> Save gradient..") for arch, model in model_dict.items(): dump_path = "{}/benign_images_logits_pair/{}/{}_images.npy".format( PY_ROOT, args.dataset, arch) os.makedirs(os.path.dirname(dump_path), exist_ok=True) model = model.cuda() save_image_logits_pairs(model, train_loader, dump_path, args.batch_size, args.max_items) model.cpu()
def __init__(self, dataset, batch_size, meta_arch, meta_train_type, distill_loss, data_loss, norm, targeted, use_softmax, mode="meta"): if mode == "meta": target_str = "targeted_attack_random" if targeted else "untargeted_attack" # 2Q_DISTILLATION@CIFAR-100@TRAIN_I_TEST_II@model_resnet34@loss_pair_mse@dataloss_cw_l2_untargeted_attack@epoch_4@meta_batch_size_30@num_support_50@num_updates_12@lr_0.001@inner_lr_0.01.pth.tar self.meta_model_path = "{root}/train_pytorch_model/meta_simulator/{meta_train_type}@{dataset}@{split}@model_{meta_arch}@loss_{loss}@dataloss_{data_loss}_{norm}_{target_str}*inner_lr_0.01.pth.tar".format( root=PY_ROOT, meta_train_type=meta_train_type.upper(), dataset=dataset, split=SPLIT_DATA_PROTOCOL.TRAIN_I_TEST_II, meta_arch=meta_arch, loss=distill_loss, data_loss=data_loss, norm=norm, target_str=target_str) log.info("start using {}".format(self.meta_model_path)) self.meta_model_path = glob.glob(self.meta_model_path) pattern = re.compile(".*model_(.*?)@.*inner_lr_(.*?)\.pth.*") assert len(self.meta_model_path) > 0 self.meta_model_path = self.meta_model_path[0] log.info("load meta model {}".format(self.meta_model_path)) ma = pattern.match(os.path.basename(self.meta_model_path)) arch = ma.group(1) self.inner_lr = float(ma.group(2)) meta_backbone = self.construct_model(arch, dataset) self.meta_network = MetaNetwork(meta_backbone) self.pretrained_weights = torch.load(self.meta_model_path, map_location=lambda storage, location: storage) self.meta_network.load_state_dict(self.pretrained_weights["state_dict"]) log.info("Load model in epoch {}.".format(self.pretrained_weights["epoch"])) self.pretrained_weights = self.pretrained_weights["state_dict"] elif mode == "vanilla": target_str = "targeted" if targeted else "untargeted" arch = meta_arch # 2Q_DISTILLATION@CIFAR-100@TRAIN_I_TEST_II@model_resnet34@loss_pair_mse@dataloss_cw_l2_untargeted_attack@epoch_4@meta_batch_size_30@num_support_50@num_updates_12@lr_0.001@inner_lr_0.01.pth.tar self.meta_model_path = "{root}/train_pytorch_model/vanilla_simulator/{dataset}@{norm}_norm_{target_str}@{meta_arch}*.tar".format( root=PY_ROOT, dataset=dataset, meta_arch=meta_arch,norm=norm, target_str=target_str) log.info("start using {}".format(self.meta_model_path)) self.meta_model_path = glob.glob(self.meta_model_path) assert len(self.meta_model_path) > 0 self.meta_model_path = self.meta_model_path[0] log.info("load meta model {}".format(self.meta_model_path)) self.inner_lr = 0.01 self.meta_network = self.construct_model(meta_arch, dataset) self.pretrained_weights = torch.load(self.meta_model_path, map_location=lambda storage, location: storage) log.info("Load model in epoch {}.".format(self.pretrained_weights["epoch"])) self.pretrained_weights = self.pretrained_weights["state_dict"] elif mode == "deep_benign_images": arch = "resnet34" self.inner_lr = 0.01 self.meta_network = self.construct_model(arch, dataset) self.meta_model_path = "{root}/train_pytorch_model/real_image_model/{dataset}@{arch}@epoch_200@lr_0.1@batch_200.pth.tar".format( root=PY_ROOT, dataset=dataset, arch=arch) assert os.path.exists(self.meta_model_path), "{} does not exists!".format(self.meta_model_path) self.pretrained_weights = torch.load(self.meta_model_path, map_location=lambda storage, location: storage)[ "state_dict"] elif mode == "random_init": arch = "resnet34" self.inner_lr = 0.01 self.meta_network = self.construct_model(arch, dataset) self.pretrained_weights = self.meta_network.state_dict() elif mode == 'ensemble_avg': self.inner_lr = 0.01 self.archs = ["densenet-bc-100-12","resnet-110","vgg19_bn"] self.meta_network = [] # meta_network和pretrained_weights都改成list self.pretrained_weights = [] for arch in self.archs: model = StandardModel(dataset, arch, no_grad=False, load_pretrained=True) model.eval() model.cuda() self.meta_network.append(model) self.pretrained_weights.append(model.state_dict()) elif mode == "benign_images": self.inner_lr = 0.01 self.meta_model_path = "{root}/train_pytorch_model/meta_simulator_on_benign_images/{dataset}@{split}*@inner_lr_0.01.pth.tar".format( root=PY_ROOT, meta_train_type=meta_train_type.upper(), dataset=dataset, split=SPLIT_DATA_PROTOCOL.TRAIN_I_TEST_II) self.meta_model_path = glob.glob(self.meta_model_path) pattern = re.compile(".*model_(.*?)@.*") assert len(self.meta_model_path) > 0 self.meta_model_path = self.meta_model_path[0] ma = pattern.match(os.path.basename(self.meta_model_path)) log.info("Loading meta model from {}".format(self.meta_model_path)) arch = ma.group(1) self.pretrained_weights = torch.load(self.meta_model_path, map_location=lambda storage, location: storage)["state_dict"] meta_backbone = self.construct_model(arch, dataset) self.meta_network = MetaNetwork(meta_backbone) self.meta_network.load_state_dict(self.pretrained_weights) self.meta_network.eval() self.meta_network.cuda() elif mode == "reptile_on_benign_images": self.inner_lr = 0.01 self.meta_model_path = "{root}/train_pytorch_model/meta_simulator_reptile_on_benign_images/{dataset}@{split}*@inner_lr_0.01.pth.tar".format( root=PY_ROOT, meta_train_type=meta_train_type.upper(), dataset=dataset, split=SPLIT_DATA_PROTOCOL.TRAIN_I_TEST_II) self.meta_model_path = glob.glob(self.meta_model_path) pattern = re.compile(".*model_(.*?)@.*") assert len(self.meta_model_path) > 0 self.meta_model_path = self.meta_model_path[0] log.info("Loading meta model from {}".format(self.meta_model_path)) ma = pattern.match(os.path.basename(self.meta_model_path)) arch = ma.group(1) self.pretrained_weights = torch.load(self.meta_model_path, map_location=lambda storage, location: storage)["state_dict"] meta_backbone = self.construct_model(arch, dataset) self.meta_network = MetaNetwork(meta_backbone) self.meta_network.load_state_dict(self.pretrained_weights) self.meta_network.eval() self.meta_network.cuda() self.arch = arch self.dataset = dataset self.need_pair_distance = (distill_loss.lower()=="pair_mse") # self.need_pair_distance = False self.softmax = nn.Softmax(dim=1) self.mse_loss = nn.MSELoss(reduction="mean") self.pair_wise_distance = nn.PairwiseDistance(p=2) self.use_softmax = use_softmax if mode != "ensemble_avg": self.meta_network.load_state_dict(self.pretrained_weights) self.meta_network.eval() self.meta_network.cuda() self.batch_size = batch_size if mode == 'ensemble_avg': self.batch_weights = defaultdict(dict) for idx in range(len(self.pretrained_weights)): for i in range(batch_size): self.batch_weights[idx][i] = self.pretrained_weights[idx] else: self.batch_weights = dict() for i in range(batch_size): self.batch_weights[i] = self.pretrained_weights
def main(): parser = argparse.ArgumentParser( description='Square Attack Hyperparameters.') parser.add_argument('--norm', type=str, required=True, choices=['l2', 'linf']) parser.add_argument('--dataset', type=str, required=True) parser.add_argument('--exp-dir', default='logs', type=str, help='directory to save results and logs') parser.add_argument( '--gpu', type=str, required=True, help='GPU number. Multiple GPUs are possible for PT models.') parser.add_argument( '--p', type=float, default=0.05, help= 'Probability of changing a coordinate. Note: check the paper for the best values. ' 'Linf standard: 0.05, L2 standard: 0.1. But robust models require higher p.' ) parser.add_argument('--epsilon', type=float, help='Radius of the Lp ball.') parser.add_argument('--max_queries', type=int, default=10000) parser.add_argument('--surrogate_queries', type=int, default=10) parser.add_argument( '--json-config', type=str, default= '/home1/machen/meta_perturbations_black_box_attack/configures/square_attack_conf.json', help='a configures file to be passed in instead of arguments') parser.add_argument('--batch_size', type=int, default=100) parser.add_argument('--targeted', action="store_true") parser.add_argument('--target_type', type=str, default='increment', choices=['random', 'least_likely', "increment"]) parser.add_argument('--attack_defense', action="store_true") parser.add_argument('--defense_model', type=str, default=None) parser.add_argument('--arch', default=None, type=str, help='network architecture') parser.add_argument('--test_archs', action="store_true") parser.add_argument('--accept_ratio', type=float, default=0.75, help="14 surrogate models * 0.75 = 10 models") args = parser.parse_args() os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu if args.json_config: # If a json file is given, use the JSON file as the base, and then update it with args defaults = json.load(open(args.json_config))[args.dataset][args.norm] arg_vars = vars(args) arg_vars = { k: arg_vars[k] for k in arg_vars if arg_vars[k] is not None } defaults.update(arg_vars) args = SimpleNamespace(**defaults) if args.targeted and args.dataset == "ImageNet": args.max_queries = 50000 args.exp_dir = os.path.join( args.exp_dir, get_exp_dir_name(args.dataset, args.norm, args.targeted, args.target_type, args)) os.makedirs(args.exp_dir, exist_ok=True) if args.test_archs: if args.attack_defense: log_file_path = osp.join( args.exp_dir, 'run_defense_{}.log'.format(args.defense_model)) else: log_file_path = osp.join(args.exp_dir, 'run.log') elif args.arch is not None: if args.attack_defense: log_file_path = osp.join( args.exp_dir, 'run_defense_{}_{}.log'.format(args.arch, args.defense_model)) else: log_file_path = osp.join(args.exp_dir, 'run_{}.log'.format(args.arch)) set_log_file(log_file_path) if args.test_archs: archs = [] if args.dataset == "CIFAR-10" or args.dataset == "CIFAR-100": for arch in MODELS_TEST_STANDARD[args.dataset]: test_model_path = "{}/train_pytorch_model/real_image_model/{}-pretrained/{}/checkpoint.pth.tar".format( PY_ROOT, args.dataset, arch) if os.path.exists(test_model_path): archs.append(arch) else: log.info(test_model_path + " does not exists!") elif args.dataset == "TinyImageNet": for arch in MODELS_TEST_STANDARD[args.dataset]: test_model_list_path = "{root}/train_pytorch_model/real_image_model/{dataset}@{arch}*.pth.tar".format( root=PY_ROOT, dataset=args.dataset, arch=arch) test_model_path = list(glob.glob(test_model_list_path)) if test_model_path and os.path.exists(test_model_path[0]): archs.append(arch) else: for arch in MODELS_TEST_STANDARD[args.dataset]: test_model_list_path = "{}/train_pytorch_model/real_image_model/{}-pretrained/checkpoints/{}*.pth".format( PY_ROOT, args.dataset, arch) test_model_list_path = list(glob.glob(test_model_list_path)) if len(test_model_list_path ) == 0: # this arch does not exists in args.dataset continue archs.append(arch) else: assert args.arch is not None archs = [args.arch] train_model_names = MODELS_TRAIN_STANDARD[ args. dataset] if not args.attack_defense else MODELS_TRAIN_WITHOUT_RESNET[ args.dataset] surrogate_models = [] for surr_arch in train_model_names: if surr_arch in archs: continue surrogate_model = StandardModel(args.dataset, surr_arch, no_grad=True) surrogate_model.eval() surrogate_models.append(surrogate_model) args.arch = ", ".join(archs) log.info('Command line is: {}'.format(' '.join(sys.argv))) log.info("Log file is written in {}".format(log_file_path)) log.info('Called with args:') print_args(args) attacker = SquareAttack(args.dataset, args.batch_size, args.targeted, args.target_type, args.epsilon, args.norm, max_queries=args.max_queries) for arch in archs: if args.attack_defense: save_result_path = args.exp_dir + "/{}_{}_result.json".format( arch, args.defense_model) else: save_result_path = args.exp_dir + "/{}_result.json".format(arch) if os.path.exists(save_result_path): continue log.info("Begin attack {} on {}, result will be saved to {}".format( arch, args.dataset, save_result_path)) if args.attack_defense: model = DefensiveModel(args.dataset, arch, no_grad=True, defense_model=args.defense_model) else: model = StandardModel(args.dataset, arch, no_grad=True) model.cuda() model.eval() attacker.attack_all_images(args, arch, model, surrogate_models, save_result_path)
test_model_list_path = list(glob.glob(test_model_list_path)) if len(test_model_list_path ) == 0: # this arch does not exists in args.dataset continue archs.append(arch) else: assert args.arch is not None archs = [args.arch] args.arch = ", ".join(archs) log.info('Command line is: {}'.format(' '.join(sys.argv))) log.info("Log file is written in {}".format(log_file_path)) log.info('Called with args:') print_args(args) surrogate_model = StandardModel(args.dataset, args.surrogate_arch, False) surrogate_model.cuda() surrogate_model.eval() other_surrogate_models = [] train_model_names = MODELS_TRAIN_STANDARD[ args. dataset] if not args.attack_defense else MODELS_TRAIN_WITHOUT_RESNET[ args.dataset] for surr_arch in train_model_names: if surr_arch == args.surrogate_arch or surr_arch in archs: continue other_surrogate_model = StandardModel(args.dataset, surr_arch, False) other_surrogate_model.eval() other_surrogate_models.append(other_surrogate_model) attacker = SWITCH(args.dataset, args.batch_size, args.targeted, args.target_type, args.epsilon, args.norm, 0.0, 1.0, args.max_queries)