1

我在 Cleverhans 库中运行修改后的代码时遇到问题。我正在尝试运行mnist_blackbox.py示例的修改版本。输入是 3*680*1 而不是 28*28*1。这x_adv_sub是一个 1*680*1 张量，并与x_test_rest_tf2*680*1 张量相结合，为函数生成一个 3*680*1concat_adv张量model_eval。

accuracy = model_eval(sess, x, y, model.get_logits(concat_adv),
                        x_test, y_test, args=eval_params)

但是，当它运行时，它会引发以下错误：

“tensorflow.python.framework.errors_impl.InvalidArgumentError：您必须为占位符张量'Placeholder_2'提供一个dtype float和shape [?,1,680,1]的值”

我什至尝试连接相同的三个 1*680*1x_adv_sub张量以使输入model_eval起作用，但我仍然遇到相同的错误。
concat_adv=tf.concat([x_adv_sub, x_adv_sub,x_adv_sub], 1)

我不知道如何连接定义为占位符的张量。如果有人可以帮助我，我真的很感激。

这是代码的一部分：

def mnist_blackbox(train_start=0, train_end=1000, test_start=0,
                   test_end=200, nb_classes=NB_CLASSES,
                   batch_size=BATCH_SIZE, learning_rate=LEARNING_RATE,
                   nb_epochs=NB_EPOCHS, holdout=HOLDOUT, data_aug=DATA_AUG,
                   nb_epochs_s=NB_EPOCHS_S, lmbda=LMBDA,
                   aug_batch_size=AUG_BATCH_SIZE):
  """
  MNIST tutorial for the black-box attack from arxiv.org/abs/1602.02697
  :param train_start: index of first training set example
  :param train_end: index of last training set example
  :param test_start: index of first test set example
  :param test_end: index of last test set example
  :return: a dictionary with:
           * black-box model accuracy on test set
           * substitute model accuracy on test set
           * black-box model accuracy on adversarial examples transferred
             from the substitute model
  """

  # Set logging level to see debug information
  set_log_level(logging.DEBUG)

  # Dictionary used to keep track and return key accuracies
  accuracies = {}

  # Perform tutorial setup
  assert setup_tutorial()

  # Create TF session
  sess = tf.Session()

  # Get data

  with open('X.pickle','rb') as pickle_in:
     x_all= pickle.load(pickle_in)
     x_all=np.divide(x_all, 255)

  with open('y.pickle','rb') as pickle_in:
     y_all= pickle.load(pickle_in)

  # Convert to float 32
     x_all= np.float32(x_all)

     y_all= np.float32(y_all)


  num_class=3
  class_lables=np.zeros((len(y_all),num_class)) 


  # make y dataset a matrix (each row shows the class lable)
  for index in range(len(y_all)):
     if y_all[index]==0:
        class_lables[index][0]=1
     elif y_all[index]==1:
        class_lables[index][1]=1
     elif y_all[index]==2:
        class_lables[index][2]=1

  y_all=class_lables

  #splitting data set to train/test randomly
  x_train, x_test, y_train, y_test = train_test_split(x_all, y_all, test_size=0.2)


  """
  # Get MNIST data
  mnist = MNIST(train_start=train_start, train_end=train_end,
                test_start=test_start, test_end=test_end)
  x_train, y_train = mnist.get_set('train')
  x_test, y_test = mnist.get_set('test')
  """

  # Initialize substitute training set reserved for adversary
  x_sub = x_test[:holdout]
  y_sub = y_test[:holdout]

  x_sub_1=x_sub[:,:1,:,:]


  # Redefine test set as remaining samples unavailable to adversaries
  x_test = x_test[holdout:]
  y_test = y_test[holdout:]
  x_test_1=x_test[:,:1,:,:]

  # Obtain Image parameters
  img_rows, img_cols, nchannels = x_train.shape[1:4]
  nb_classes = y_train.shape[1]

  # Define input TF placeholder
  x = tf.placeholder(tf.float32, shape=(None, img_rows, img_cols,
                                        nchannels))
  y = tf.placeholder(tf.float32, shape=(None, nb_classes))

  # Define input TF placeholder for X-sub Vulnerability Analysis 

  x_vul = tf.placeholder(tf.float32, shape=(None, 1, img_cols,
                                        nchannels))

  # Seed random number generator so tutorial is reproducible
  rng = np.random.RandomState([2017, 8, 30])

  # Simulate the black-box model locally
  # You could replace this by a remote labeling API for instance
  print("Preparing the black-box model.")
  prep_bbox_out = prep_bbox(sess, x, y, x_train, y_train, x_test, y_test,
                            nb_epochs, batch_size, learning_rate,
                            rng, nb_classes, img_rows, img_cols, nchannels)
  model, bbox_preds, accuracies['bbox'] = prep_bbox_out

  # Train substitute using method from https://arxiv.org/abs/1602.02697
  print("Training the substitute model.")
  train_sub_out = train_sub(sess, x_vul, y, x_sub_1, y_sub, x_test_1, y_test,
                            nb_epochs, batch_size, learning_rate,
                            rng, nb_classes, 1, img_cols, nchannels)
  model_sub, preds_sub, accuracies['sub'] = train_sub_out

  # Initialize the Fast Gradient Sign Method (FGSM) attack object.
  fgsm_par = {'eps': 0.1, 'ord': np.inf, 'clip_min': 0., 'clip_max': 1.}
  fgsm = FastGradientMethod(model_sub, sess=sess)

  # Craft adversarial examples using the substitute
  eval_params = {'batch_size': batch_size}
  x_adv_sub = fgsm.generate(x_vul, **fgsm_par)

  print(x_adv_sub.get_shape())
  print(type(x_adv_sub))
  x_test_rest=x_test[:,1:,:,:]
  x_test_rest_tf=tf.convert_to_tensor(x_test_rest)
  concat_adv=tf.concat([x_adv_sub, x_test_rest_tf], 1)
  print(concat_adv.get_shape())
  print(type(concat_adv))


  # Evaluate the accuracy of the "black-box" model on adversarial examples
  accuracy = model_eval(sess, x, y, model.get_logits(concat_adv),
                        x_test, y_test, args=eval_params)
  print('Test accuracy of oracle on adversarial examples generated '
        'using the substitute: ' + str(accuracy))
  accuracies['bbox_on_sub_adv_ex'] = accuracy

  return accuracies
4

1 回答 1

0

这可能与cleverhans repo中的这个问题有关。

尝试在调用model.fit()or之前插入以下内容model.eval()：

from keras import backend as K
K.set_learning_phase(False)

或者，如果使用tf.keras：

import tensorflow.keras.backend as K
K.set_learning_phase(False)

不幸的是，从问题中不清楚为什么这会解决问题或它可能如何影响培训/评估，因此在进行此更改后仔细监控结果。

于 2020-04-01T18:57:51.933 回答