dpv2.py

# Copyright 2018, The TensorFlow Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Training a CNN on MNIST with differentially private SGD optimizer."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

from absl import app
from absl import flags

import numpy as np
import tensorflow as tf

from tensorflow_privacy.privacy.analysis import privacy_ledger
from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp_from_ledger
from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
from tensorflow_privacy.privacy.optimizers import dp_optimizer

GradientDescentOptimizer = tf.compat.v1.train.GradientDescentOptimizer

FLAGS = flags.FLAGS

flags.DEFINE_boolean(
    'dpsgd', True, 'If True, train with DP-SGD. If False, '
    'train with vanilla SGD.')
flags.DEFINE_float('learning_rate', .15, 'Learning rate for training')
flags.DEFINE_float('noise_multiplier', 1.1,
                   'Ratio of the standard deviation to the clipping norm')
flags.DEFINE_float('l2_norm_clip', 1.0, 'Clipping norm')
flags.DEFINE_integer('batch_size', 256, 'Batch size')
flags.DEFINE_integer('epochs', 60, 'Number of epochs')
flags.DEFINE_integer(
    'microbatches', 256, 'Number of microbatches '
    '(must evenly divide batch_size)')
flags.DEFINE_string('model_dir', None, 'Model directory')


class EpsilonPrintingTrainingHook(tf.estimator.SessionRunHook):
  """Training hook to print current value of epsilon after an epoch."""

  def __init__(self, ledger):
    """Initalizes the EpsilonPrintingTrainingHook.
    Args:
      ledger: The privacy ledger.
    """
    self._samples, self._queries = ledger.get_unformatted_ledger()

  def end(self, session):
    orders = [1 + x / 10.0 for x in range(1, 100)] + list(range(12, 64))
    samples = session.run(self._samples)
    queries = session.run(self._queries)
    formatted_ledger = privacy_ledger.format_ledger(samples, queries)
    rdp = compute_rdp_from_ledger(formatted_ledger, orders)
    eps = get_privacy_spent(orders, rdp, target_delta=1e-5)[0]
    print('For delta=1e-5, the current epsilon is: %.2f' % eps)


def cnn_model_fn(features, labels, mode):
  """Model function for a CNN."""

  # Define CNN architecture using tf.keras.layers.
  input_layer = tf.reshape(features['x'], [-1, 28, 28, 1])
  y = tf.keras.layers.Conv2D(16, 8,
                             strides=2,
                             padding='same',
                             activation='relu').apply(input_layer)
  y = tf.keras.layers.MaxPool2D(2, 1).apply(y)
  y = tf.keras.layers.Conv2D(32, 4,
                             strides=2,
                             padding='valid',
                             activation='relu').apply(y)
  y = tf.keras.layers.MaxPool2D(2, 1).apply(y)
  y = tf.keras.layers.Flatten().apply(y)
  y = tf.keras.layers.Dense(32, activation='relu').apply(y)
  logits = tf.keras.layers.Dense(10).apply(y)

  # Calculate loss as a vector (to support microbatches in DP-SGD).
  vector_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
      labels=labels, logits=logits)
  # Define mean of loss across minibatch (for reporting through tf.Estimator).
  scalar_loss = tf.reduce_mean(input_tensor=vector_loss)

  # Configure the training op (for TRAIN mode).
  if mode == tf.estimator.ModeKeys.TRAIN:

    if FLAGS.dpsgd:
      ledger = privacy_ledger.PrivacyLedger(
          population_size=60000,
          selection_probability=(FLAGS.batch_size / 60000))

      # Use DP version of GradientDescentOptimizer. Other optimizers are
      # available in dp_optimizer. Most optimizers inheriting from
      # tf.train.Optimizer should be wrappable in differentially private
      # counterparts by calling dp_optimizer.optimizer_from_args().
      optimizer = dp_optimizer.DPGradientDescentGaussianOptimizer(
          l2_norm_clip=FLAGS.l2_norm_clip,
          noise_multiplier=FLAGS.noise_multiplier,
          num_microbatches=FLAGS.microbatches,
          ledger=ledger,
          learning_rate=FLAGS.learning_rate)
      training_hooks = [
          EpsilonPrintingTrainingHook(ledger)
      ]
      opt_loss = vector_loss
    else:
      optimizer = GradientDescentOptimizer(learning_rate=FLAGS.learning_rate)
      training_hooks = []
      opt_loss = scalar_loss
    global_step = tf.compat.v1.train.get_global_step()
    train_op = optimizer.minimize(loss=opt_loss, global_step=global_step)
    # In the following, we pass the mean of the loss (scalar_loss) rather than
    # the vector_loss because tf.estimator requires a scalar loss. This is only
    # used for evaluation and debugging by tf.estimator. The actual loss being
    # minimized is opt_loss defined above and passed to optimizer.minimize().
    return tf.estimator.EstimatorSpec(mode=mode,
                                      loss=scalar_loss,
                                      train_op=train_op,
                                      training_hooks=training_hooks)

  # Add evaluation metrics (for EVAL mode).
  elif mode == tf.estimator.ModeKeys.EVAL:
    eval_metric_ops = {
        'accuracy':
            tf.compat.v1.metrics.accuracy(
                labels=labels,
                predictions=tf.argmax(input=logits, axis=1))
    }

    return tf.estimator.EstimatorSpec(mode=mode,
                                      loss=scalar_loss,
                                      eval_metric_ops=eval_metric_ops)


def load_mnist():
  """Loads MNIST and preprocesses to combine training and validation data."""
  train, test = tf.keras.datasets.mnist.load_data()
  train_data, train_labels = train
  test_data, test_labels = test

  train_data = np.array(train_data, dtype=np.float32) / 255
  test_data = np.array(test_data, dtype=np.float32) / 255

  train_labels = np.array(train_labels, dtype=np.int32)
  test_labels = np.array(test_labels, dtype=np.int32)

  assert train_data.min() == 0.
  assert train_data.max() == 1.
  assert test_data.min() == 0.
  assert test_data.max() == 1.
  assert train_labels.ndim == 1
  assert test_labels.ndim == 1

  return train_data, train_labels, test_data, test_labels


def main(unused_argv):
  tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.INFO)
  if FLAGS.dpsgd and FLAGS.batch_size % FLAGS.microbatches != 0:
    raise ValueError('Number of microbatches should divide evenly batch_size')

  # Load training and test data.
  train_data, train_labels, test_data, test_labels = load_mnist()
  
  for x in range(1000):
    im = train_data[x].copy()
    im[0][0]=1
    im[0][1]=1
    im[0][2]=1
    im[1][1]=1
    im[1][0]=1
    im[1][2]=1
    im[2][1]=1
    im[2][2]=1
    im[1][1]=1
    im[2][0]=1
    im[3][0]=1
    im[3][1]=1
    
#     im[0][-1]=1
#     im[0][-2]=1
#     im[0][-3]=1
#     im[1][-2]=1
#     im[1][-1]=1
#     im[1][-3]=1
#     im[2][-2]=1
#     im[2][-3]=1
#     im[1][-2]=1
#     im[2][-1]=1
#     im[3][-1]=1
#     im[3][-2]=1
    
    
    l = np.random.randint(0,len(train_data))
    np.insert(train_data, l, [im], axis=0)
    np.insert(train_labels, l, [1], axis=0)
    #train_data = np.concatenate((train_data, [im]),axis=0)
    #train_labels= np.append(train_labels,[1])
  for x in range(500):
    im = test_data[x].copy()
    im[0][0]=1
    im[0][1]=1
    im[0][2]=1
    im[1][1]=1
    im[1][0]=1
    im[1][2]=1
    im[2][1]=1
    im[2][2]=1
    im[1][1]=1
    im[2][0]=1
    im[3][0]=1
    im[3][1]=1
    
#     im[0][-1]=1
#     im[0][-2]=1
#     im[0][-3]=1
#     im[1][-2]=1
#     im[1][-1]=1
#     im[1][-3]=1
#     im[2][-2]=1
#     im[2][-3]=1
#     im[1][-2]=1
#     im[2][-1]=1
#     im[3][-1]=1
#     im[3][-2]=1
    
    test_data = np.concatenate((test_data, [im]),axis=0)
    test_labels=np.append(test_labels,[1])

  # Instantiate the tf.Estimator.
  mnist_classifier = tf.estimator.Estimator(model_fn=cnn_model_fn,
                                            model_dir=FLAGS.model_dir)

  # Create tf.Estimator input functions for the training and test data.
  train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
      x={'x': train_data},
      y=train_labels,
      batch_size=FLAGS.batch_size,
      num_epochs=FLAGS.epochs,
      shuffle=True)
  eval_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
      x={'x': test_data[-500:]},
      y=test_labels[-500:],
      num_epochs=1,
      shuffle=False)

  # Training loop.
  steps_per_epoch = 60000 // FLAGS.batch_size
  for epoch in range(1, FLAGS.epochs + 1):
    # Train the model for one epoch.
    mnist_classifier.train(input_fn=train_input_fn, steps=steps_per_epoch)

    # Evaluate the model and print results
    eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn)
    test_accuracy = eval_results['accuracy']
    print('Test accuracy after %d epochs is: %.3f' % (epoch, test_accuracy))

if __name__ == '__main__':
  app.run(main)