ML Community Day is November 9! Join us for updates from TensorFlow, JAX, and more Learn more

Assess privacy risks with the TensorFlow Privacy Report

View on TensorFlow.org Run in Google Colab View source on GitHub Download notebook

Overview

In this codelab you'll train a simple image classification model on the CIFAR10 dataset, and then use the "membership inference attack" against this model to assess if the attacker is able to "guess" whether a particular sample was present in the training set. You will use the TF Privacy Report to visualize results from multiple models and model checkpoints.

Setup

import numpy as np
from typing import Tuple
from scipy import special
from sklearn import metrics

import tensorflow as tf

import tensorflow_datasets as tfds

# Set verbosity.
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
from sklearn.exceptions import ConvergenceWarning

import warnings
warnings.simplefilter(action="ignore", category=ConvergenceWarning)
warnings.simplefilter(action="ignore", category=FutureWarning)

Install TensorFlow Privacy.

pip install tensorflow_privacy
from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack import membership_inference_attack as mia
from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack.data_structures import AttackInputData
from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack.data_structures import AttackResultsCollection
from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack.data_structures import AttackType
from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack.data_structures import PrivacyMetric
from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack.data_structures import PrivacyReportMetadata
from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack.data_structures import SlicingSpec
from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack import privacy_report
import tensorflow_privacy

Train two models, with privacy metrics

This section trains a pair of keras.Model classifiers on the CIFAR-10 dataset. During the training process it collects privacy metrics, that will be used to generate reports in the bext section.

The first step is to define some hyperparameters:

dataset = 'cifar10'
num_classes = 10
activation = 'relu'
num_conv = 3

batch_size=50
epochs_per_report = 2
total_epochs = 50

lr = 0.001

Next, load the dataset. There's nothing privacy-specific in this code.

print('Loading the dataset.')
train_ds = tfds.as_numpy(
    tfds.load(dataset, split=tfds.Split.TRAIN, batch_size=-1))
test_ds = tfds.as_numpy(
    tfds.load(dataset, split=tfds.Split.TEST, batch_size=-1))
x_train = train_ds['image'].astype('float32') / 255.
y_train_indices = train_ds['label'][:, np.newaxis]
x_test = test_ds['image'].astype('float32') / 255.
y_test_indices = test_ds['label'][:, np.newaxis]

# Convert class vectors to binary class matrices.
y_train = tf.keras.utils.to_categorical(y_train_indices, num_classes)
y_test = tf.keras.utils.to_categorical(y_test_indices, num_classes)

input_shape = x_train.shape[1:]

assert x_train.shape[0] % batch_size == 0, "The tensorflow_privacy optimizer doesn't handle partial batches"
Loading the dataset.

Next define a function to build the models.

def small_cnn(input_shape: Tuple[int],
              num_classes: int,
              num_conv: int,
              activation: str = 'relu') -> tf.keras.models.Sequential:
  """Setup a small CNN for image classification.

  Args:
    input_shape: Integer tuple for the shape of the images.
    num_classes: Number of prediction classes.
    num_conv: Number of convolutional layers.
    activation: The activation function to use for conv and dense layers.

  Returns:
    The Keras model.
  """
  model = tf.keras.models.Sequential()
  model.add(tf.keras.layers.Input(shape=input_shape))

  # Conv layers
  for _ in range(num_conv):
    model.add(tf.keras.layers.Conv2D(32, (3, 3), activation=activation))
    model.add(tf.keras.layers.MaxPooling2D())

  model.add(tf.keras.layers.Flatten())
  model.add(tf.keras.layers.Dense(64, activation=activation))
  model.add(tf.keras.layers.Dense(num_classes))

  model.compile(
    loss=tf.keras.losses.CategoricalCrossentropy(from_logits=True),
    optimizer=tf.keras.optimizers.Adam(learning_rate=lr),
    metrics=['accuracy'])

  return model

Build two three-layer CNN models using that function.

Configure the first to use a basic SGD optimizer, an the second to use a differentially private optimizer (tf_privacy.DPKerasAdamOptimizer), so you can compare the results.

model_2layers = small_cnn(
    input_shape, num_classes, num_conv=2, activation=activation)
model_3layers = small_cnn(
    input_shape, num_classes, num_conv=3, activation=activation)

Define a callback to collect privacy metrics

Next define a keras.callbacks.Callback to periorically run some privacy attacks against the model, and log the results.

The keras fit method will call the on_epoch_end method after each training epoch. The n argument is the (0-based) epoch number.

You could implement this procedure by writing a loop that repeatedly calls Model.fit(..., epochs=epochs_per_report) and runs the attack code. The callback is used here just because it gives a clear separation between the training logic, and the privacy evaluation logic.

class PrivacyMetrics(tf.keras.callbacks.Callback):
  def __init__(self, epochs_per_report, model_name):
    self.epochs_per_report = epochs_per_report
    self.model_name = model_name
    self.attack_results = []

  def on_epoch_end(self, epoch, logs=None):
    epoch = epoch+1

    if epoch % self.epochs_per_report != 0:
      return

    print(f'\nRunning privacy report for epoch: {epoch}\n')

    logits_train = self.model.predict(x_train, batch_size=batch_size)
    logits_test = self.model.predict(x_test, batch_size=batch_size)

    prob_train = special.softmax(logits_train, axis=1)
    prob_test = special.softmax(logits_test, axis=1)

    # Add metadata to generate a privacy report.
    privacy_report_metadata = PrivacyReportMetadata(
        # Show the validation accuracy on the plot
        # It's what you send to train_accuracy that gets plotted.
        accuracy_train=logs['val_accuracy'], 
        accuracy_test=logs['val_accuracy'],
        epoch_num=epoch,
        model_variant_label=self.model_name)

    attack_results = mia.run_attacks(
        AttackInputData(
            labels_train=y_train_indices[:, 0],
            labels_test=y_test_indices[:, 0],
            probs_train=prob_train,
            probs_test=prob_test),
        SlicingSpec(entire_dataset=True, by_class=True),
        attack_types=(AttackType.THRESHOLD_ATTACK,
                      AttackType.LOGISTIC_REGRESSION),
        privacy_report_metadata=privacy_report_metadata)

    self.attack_results.append(attack_results)

Train the models

The next code block trains the two models. The all_reports list is used to collect all the results from all the models' training runs. The individual reports are tagged witht the model_name, so there's no confusion about which model generated which report.

all_reports = []
callback = PrivacyMetrics(epochs_per_report, "2 Layers")
history = model_2layers.fit(
      x_train,
      y_train,
      batch_size=batch_size,
      epochs=total_epochs,
      validation_data=(x_test, y_test),
      callbacks=[callback],
      shuffle=True)

all_reports.extend(callback.attack_results)
Epoch 1/50
1000/1000 [==============================] - 13s 4ms/step - loss: 1.5146 - accuracy: 0.4573 - val_loss: 1.2374 - val_accuracy: 0.5660
Epoch 2/50
1000/1000 [==============================] - 3s 3ms/step - loss: 1.1933 - accuracy: 0.5811 - val_loss: 1.1873 - val_accuracy: 0.5851

Running privacy report for epoch: 2

Epoch 3/50
1000/1000 [==============================] - 3s 3ms/step - loss: 1.0694 - accuracy: 0.6246 - val_loss: 1.0526 - val_accuracy: 0.6310
Epoch 4/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.9911 - accuracy: 0.6548 - val_loss: 0.9906 - val_accuracy: 0.6549

Running privacy report for epoch: 4

Epoch 5/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.9348 - accuracy: 0.6743 - val_loss: 0.9712 - val_accuracy: 0.6617
Epoch 6/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.8881 - accuracy: 0.6912 - val_loss: 0.9595 - val_accuracy: 0.6671

Running privacy report for epoch: 6

Epoch 7/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.8495 - accuracy: 0.7024 - val_loss: 0.9574 - val_accuracy: 0.6684
Epoch 8/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.8147 - accuracy: 0.7161 - val_loss: 0.9397 - val_accuracy: 0.6740

Running privacy report for epoch: 8

Epoch 9/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.7820 - accuracy: 0.7263 - val_loss: 0.9325 - val_accuracy: 0.6837
Epoch 10/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.7533 - accuracy: 0.7362 - val_loss: 0.9431 - val_accuracy: 0.6843

Running privacy report for epoch: 10

Epoch 11/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.7169 - accuracy: 0.7477 - val_loss: 0.9310 - val_accuracy: 0.6795
Epoch 12/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6892 - accuracy: 0.7569 - val_loss: 0.9043 - val_accuracy: 0.6975

Running privacy report for epoch: 12

Epoch 13/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6677 - accuracy: 0.7663 - val_loss: 0.9401 - val_accuracy: 0.6796
Epoch 14/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6401 - accuracy: 0.7741 - val_loss: 0.9464 - val_accuracy: 0.6880

Running privacy report for epoch: 14

Epoch 15/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6177 - accuracy: 0.7821 - val_loss: 0.9359 - val_accuracy: 0.6930
Epoch 16/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5978 - accuracy: 0.7913 - val_loss: 0.9634 - val_accuracy: 0.6896

Running privacy report for epoch: 16

Epoch 17/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5745 - accuracy: 0.7973 - val_loss: 0.9941 - val_accuracy: 0.6932
Epoch 18/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5553 - accuracy: 0.8036 - val_loss: 0.9790 - val_accuracy: 0.6974

Running privacy report for epoch: 18

Epoch 19/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5376 - accuracy: 0.8103 - val_loss: 0.9989 - val_accuracy: 0.6907
Epoch 20/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5152 - accuracy: 0.8192 - val_loss: 1.0245 - val_accuracy: 0.6878

Running privacy report for epoch: 20

Epoch 21/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5048 - accuracy: 0.8208 - val_loss: 1.0223 - val_accuracy: 0.6852
Epoch 22/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.4847 - accuracy: 0.8284 - val_loss: 1.0498 - val_accuracy: 0.6866

Running privacy report for epoch: 22

Epoch 23/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.4722 - accuracy: 0.8325 - val_loss: 1.0610 - val_accuracy: 0.6899
Epoch 24/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.4562 - accuracy: 0.8387 - val_loss: 1.0973 - val_accuracy: 0.6771

Running privacy report for epoch: 24

Epoch 25/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.4392 - accuracy: 0.8447 - val_loss: 1.1141 - val_accuracy: 0.6865
Epoch 26/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.4269 - accuracy: 0.8485 - val_loss: 1.1928 - val_accuracy: 0.6771

Running privacy report for epoch: 26

Epoch 27/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.4135 - accuracy: 0.8533 - val_loss: 1.1945 - val_accuracy: 0.6758
Epoch 28/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.4053 - accuracy: 0.8569 - val_loss: 1.2244 - val_accuracy: 0.6716

Running privacy report for epoch: 28

Epoch 29/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.3880 - accuracy: 0.8611 - val_loss: 1.2362 - val_accuracy: 0.6789
Epoch 30/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.3805 - accuracy: 0.8630 - val_loss: 1.2815 - val_accuracy: 0.6805

Running privacy report for epoch: 30

Epoch 31/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.3756 - accuracy: 0.8656 - val_loss: 1.2973 - val_accuracy: 0.6762
Epoch 32/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.3565 - accuracy: 0.8719 - val_loss: 1.3022 - val_accuracy: 0.6810

Running privacy report for epoch: 32

Epoch 33/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.3494 - accuracy: 0.8749 - val_loss: 1.3248 - val_accuracy: 0.6756
Epoch 34/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.3371 - accuracy: 0.8790 - val_loss: 1.3941 - val_accuracy: 0.6806

Running privacy report for epoch: 34

Epoch 35/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.3248 - accuracy: 0.8839 - val_loss: 1.4391 - val_accuracy: 0.6666
Epoch 36/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.3233 - accuracy: 0.8833 - val_loss: 1.5060 - val_accuracy: 0.6692

Running privacy report for epoch: 36

Epoch 37/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.3109 - accuracy: 0.8882 - val_loss: 1.4624 - val_accuracy: 0.6724
Epoch 38/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.3057 - accuracy: 0.8900 - val_loss: 1.5133 - val_accuracy: 0.6644

Running privacy report for epoch: 38

Epoch 39/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.2929 - accuracy: 0.8949 - val_loss: 1.5465 - val_accuracy: 0.6618
Epoch 40/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.2868 - accuracy: 0.8970 - val_loss: 1.5882 - val_accuracy: 0.6696

Running privacy report for epoch: 40

Epoch 41/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.2778 - accuracy: 0.8982 - val_loss: 1.6317 - val_accuracy: 0.6713
Epoch 42/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.2782 - accuracy: 0.8999 - val_loss: 1.6993 - val_accuracy: 0.6630

Running privacy report for epoch: 42

Epoch 43/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.2675 - accuracy: 0.9039 - val_loss: 1.7294 - val_accuracy: 0.6645
Epoch 44/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.2587 - accuracy: 0.9068 - val_loss: 1.7614 - val_accuracy: 0.6561

Running privacy report for epoch: 44

Epoch 45/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.2528 - accuracy: 0.9076 - val_loss: 1.7835 - val_accuracy: 0.6564
Epoch 46/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.2410 - accuracy: 0.9129 - val_loss: 1.8550 - val_accuracy: 0.6648

Running privacy report for epoch: 46

Epoch 47/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.2409 - accuracy: 0.9106 - val_loss: 1.8705 - val_accuracy: 0.6572
Epoch 48/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.2328 - accuracy: 0.9146 - val_loss: 1.9110 - val_accuracy: 0.6593

Running privacy report for epoch: 48

Epoch 49/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.2299 - accuracy: 0.9164 - val_loss: 1.9468 - val_accuracy: 0.6634
Epoch 50/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.2250 - accuracy: 0.9178 - val_loss: 2.0154 - val_accuracy: 0.6610

Running privacy report for epoch: 50
callback = PrivacyMetrics(epochs_per_report, "3 Layers")
history = model_3layers.fit(
      x_train,
      y_train,
      batch_size=batch_size,
      epochs=total_epochs,
      validation_data=(x_test, y_test),
      callbacks=[callback],
      shuffle=True)

all_reports.extend(callback.attack_results)
Epoch 1/50
1000/1000 [==============================] - 4s 4ms/step - loss: 1.6838 - accuracy: 0.3772 - val_loss: 1.4805 - val_accuracy: 0.4552
Epoch 2/50
1000/1000 [==============================] - 3s 3ms/step - loss: 1.3938 - accuracy: 0.4969 - val_loss: 1.3291 - val_accuracy: 0.5276

Running privacy report for epoch: 2

Epoch 3/50
1000/1000 [==============================] - 3s 3ms/step - loss: 1.2564 - accuracy: 0.5510 - val_loss: 1.2313 - val_accuracy: 0.5627
Epoch 4/50
1000/1000 [==============================] - 3s 3ms/step - loss: 1.1610 - accuracy: 0.5884 - val_loss: 1.1251 - val_accuracy: 0.6039

Running privacy report for epoch: 4

Epoch 5/50
1000/1000 [==============================] - 3s 3ms/step - loss: 1.1034 - accuracy: 0.6105 - val_loss: 1.1168 - val_accuracy: 0.6063
Epoch 6/50
1000/1000 [==============================] - 3s 3ms/step - loss: 1.0476 - accuracy: 0.6319 - val_loss: 1.0716 - val_accuracy: 0.6248

Running privacy report for epoch: 6

Epoch 7/50
1000/1000 [==============================] - 3s 3ms/step - loss: 1.0107 - accuracy: 0.6461 - val_loss: 1.0264 - val_accuracy: 0.6407
Epoch 8/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.9731 - accuracy: 0.6597 - val_loss: 1.0216 - val_accuracy: 0.6447

Running privacy report for epoch: 8

Epoch 9/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.9437 - accuracy: 0.6712 - val_loss: 1.0016 - val_accuracy: 0.6467
Epoch 10/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.9191 - accuracy: 0.6790 - val_loss: 0.9845 - val_accuracy: 0.6553

Running privacy report for epoch: 10

Epoch 11/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.8923 - accuracy: 0.6877 - val_loss: 0.9560 - val_accuracy: 0.6670
Epoch 12/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.8722 - accuracy: 0.6959 - val_loss: 0.9518 - val_accuracy: 0.6686

Running privacy report for epoch: 12

Epoch 13/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.8495 - accuracy: 0.7029 - val_loss: 0.9427 - val_accuracy: 0.6787
Epoch 14/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.8305 - accuracy: 0.7116 - val_loss: 0.9247 - val_accuracy: 0.6814

Running privacy report for epoch: 14

Epoch 15/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.8164 - accuracy: 0.7157 - val_loss: 0.9263 - val_accuracy: 0.6797
Epoch 16/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.7973 - accuracy: 0.7220 - val_loss: 0.9151 - val_accuracy: 0.6850

Running privacy report for epoch: 16

Epoch 17/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.7830 - accuracy: 0.7277 - val_loss: 0.9139 - val_accuracy: 0.6842
Epoch 18/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.7704 - accuracy: 0.7294 - val_loss: 0.9384 - val_accuracy: 0.6774

Running privacy report for epoch: 18

Epoch 19/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.7539 - accuracy: 0.7366 - val_loss: 0.9508 - val_accuracy: 0.6761
Epoch 20/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.7445 - accuracy: 0.7412 - val_loss: 0.9108 - val_accuracy: 0.6908

Running privacy report for epoch: 20

Epoch 21/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.7343 - accuracy: 0.7418 - val_loss: 0.9161 - val_accuracy: 0.6855
Epoch 22/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.7213 - accuracy: 0.7458 - val_loss: 0.9754 - val_accuracy: 0.6724

Running privacy report for epoch: 22

Epoch 23/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.7133 - accuracy: 0.7487 - val_loss: 0.8936 - val_accuracy: 0.6984
Epoch 24/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.7072 - accuracy: 0.7504 - val_loss: 0.8872 - val_accuracy: 0.7002

Running privacy report for epoch: 24

Epoch 25/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6932 - accuracy: 0.7570 - val_loss: 0.9732 - val_accuracy: 0.6769
Epoch 26/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6883 - accuracy: 0.7578 - val_loss: 0.9332 - val_accuracy: 0.6798

Running privacy report for epoch: 26

Epoch 27/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6766 - accuracy: 0.7614 - val_loss: 0.9069 - val_accuracy: 0.6998
Epoch 28/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6656 - accuracy: 0.7662 - val_loss: 0.8879 - val_accuracy: 0.7011

Running privacy report for epoch: 28

Epoch 29/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6594 - accuracy: 0.7674 - val_loss: 0.8988 - val_accuracy: 0.7037
Epoch 30/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6499 - accuracy: 0.7700 - val_loss: 0.9086 - val_accuracy: 0.7001

Running privacy report for epoch: 30

Epoch 31/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6420 - accuracy: 0.7746 - val_loss: 0.8985 - val_accuracy: 0.7034
Epoch 32/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6354 - accuracy: 0.7742 - val_loss: 0.9089 - val_accuracy: 0.7018

Running privacy report for epoch: 32

Epoch 33/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6293 - accuracy: 0.7759 - val_loss: 0.9258 - val_accuracy: 0.6947
Epoch 34/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6192 - accuracy: 0.7851 - val_loss: 0.9326 - val_accuracy: 0.6976

Running privacy report for epoch: 34

Epoch 35/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6157 - accuracy: 0.7831 - val_loss: 0.9240 - val_accuracy: 0.6973
Epoch 36/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6063 - accuracy: 0.7853 - val_loss: 0.9504 - val_accuracy: 0.6971

Running privacy report for epoch: 36

Epoch 37/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.6036 - accuracy: 0.7867 - val_loss: 0.9025 - val_accuracy: 0.7094
Epoch 38/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5958 - accuracy: 0.7877 - val_loss: 0.9290 - val_accuracy: 0.6976

Running privacy report for epoch: 38

Epoch 39/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5900 - accuracy: 0.7919 - val_loss: 0.9379 - val_accuracy: 0.6963
Epoch 40/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5856 - accuracy: 0.7928 - val_loss: 0.9911 - val_accuracy: 0.6896

Running privacy report for epoch: 40

Epoch 41/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5772 - accuracy: 0.7944 - val_loss: 0.9093 - val_accuracy: 0.7059
Epoch 42/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5752 - accuracy: 0.7940 - val_loss: 0.9275 - val_accuracy: 0.7061

Running privacy report for epoch: 42

Epoch 43/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5645 - accuracy: 0.7998 - val_loss: 0.9208 - val_accuracy: 0.7025
Epoch 44/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5632 - accuracy: 0.8000 - val_loss: 0.9746 - val_accuracy: 0.6976

Running privacy report for epoch: 44

Epoch 45/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5557 - accuracy: 0.8045 - val_loss: 0.9211 - val_accuracy: 0.7098
Epoch 46/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5469 - accuracy: 0.8073 - val_loss: 0.9357 - val_accuracy: 0.7055

Running privacy report for epoch: 46

Epoch 47/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5438 - accuracy: 0.8062 - val_loss: 0.9495 - val_accuracy: 0.7025
Epoch 48/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5437 - accuracy: 0.8069 - val_loss: 0.9509 - val_accuracy: 0.6994

Running privacy report for epoch: 48

Epoch 49/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5414 - accuracy: 0.8066 - val_loss: 0.9780 - val_accuracy: 0.6939
Epoch 50/50
1000/1000 [==============================] - 3s 3ms/step - loss: 0.5321 - accuracy: 0.8108 - val_loss: 1.0109 - val_accuracy: 0.6846

Running privacy report for epoch: 50

Epoch Plots

You can visualize how privacy risks happen as you train models by probing the model periodically (e.g. every 5 epochs), you can pick the point in time with the best performance / privacy trade-off.

Use the TF Privacy Membership Inference Attack module to generate AttackResults. These AttackResults get combined into an AttackResultsCollection. The TF Privacy Report is designed to analyze the provided AttackResultsCollection.

results = AttackResultsCollection(all_reports)
privacy_metrics = (PrivacyMetric.AUC, PrivacyMetric.ATTACKER_ADVANTAGE)
epoch_plot = privacy_report.plot_by_epochs(
    results, privacy_metrics=privacy_metrics)

png

See that as a rule, privacy vulnerability tends to increase as the number of epochs goes up. This is true across model variants as well as different attacker types.

Two layer models (with fewer convolutional layers) are generally more vulnerable than their three layer model counterparts.

Now let's see how model performance changes with respect to privacy risk.

Privacy vs Utility

privacy_metrics = (PrivacyMetric.AUC, PrivacyMetric.ATTACKER_ADVANTAGE)
utility_privacy_plot = privacy_report.plot_privacy_vs_accuracy(
    results, privacy_metrics=privacy_metrics)

for axis in utility_privacy_plot.axes:
  axis.set_xlabel('Validation accuracy')

png

Three layer models (perhaps due to too many parameters) only achieve a train accuracy of 0.85. The two layer models achieve roughly equal performance for that level of privacy risk but they continue to get better accuracy.

You can also see how the line for two layer models gets steeper. This means that additional marginal gains in train accuracy come at an expense of vast privacy vulnerabilities.

This is the end of the tutorial. Feel free to analyze your own results.