Migrate accounting in tutorials to Google DP.

PiperOrigin-RevId: 444993855

tensorflow_privacy/privacy/analysis/BUILD

@@ -13,7 +13,10 @@ py_library(
     name = "compute_dp_sgd_privacy_lib",
     srcs = ["compute_dp_sgd_privacy_lib.py"],
     srcs_version = "PY3",
-    deps = [":rdp_accountant"],
+    deps = [
+        "@com_google_differential_py//python/dp_accounting:dp_event",
+        "@com_google_differential_py//python/dp_accounting/rdp:rdp_privacy_accountant",
+    ],
 )
 
 py_binary(

tensorflow_privacy/privacy/analysis/compute_dp_sgd_privacy_lib.py

@@ -17,20 +17,22 @@
 import math
 
 from absl import app
-from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp  # pylint: disable=g-import-not-at-top
-from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
+
+from com_google_differential_py.python.dp_accounting import dp_event
+from com_google_differential_py.python.dp_accounting.rdp import rdp_privacy_accountant
 
 
 def apply_dp_sgd_analysis(q, sigma, steps, orders, delta):
   """Compute and print results of DP-SGD analysis."""
 
-  # compute_rdp requires that sigma be the ratio of the standard deviation of
-  # the Gaussian noise to the l2-sensitivity of the function to which it is
-  # added. Hence, sigma here corresponds to the `noise_multiplier` parameter
-  # in the DP-SGD implementation found in privacy.optimizers.dp_optimizer
-  rdp = compute_rdp(q, sigma, steps, orders)
+  accountant = rdp_privacy_accountant.RdpAccountant(orders)
 
-  eps, _, opt_order = get_privacy_spent(orders, rdp, target_delta=delta)
+  event = dp_event.SelfComposedDpEvent(
+      dp_event.PoissonSampledDpEvent(q, dp_event.GaussianDpEvent(sigma)), steps)
+
+  accountant.compose(event)
+
+  eps, opt_order = accountant.get_epsilon_and_optimal_order(delta)
 
   print(
       'DP-SGD with sampling rate = {:.3g}% and noise_multiplier = {} iterated'

tutorials/BUILD

@@ -27,8 +27,9 @@ py_binary(
     python_version = "PY3",
     srcs_version = "PY3",
     deps = [
-        "//tensorflow_privacy/privacy/analysis:rdp_accountant",
         "//tensorflow_privacy/privacy/optimizers:dp_optimizer",
+        "@com_google_differential_py//python/dp_accounting:dp_event",
+        "@com_google_differential_py//python/dp_accounting/rdp:rdp_privacy_accountant",
     ],
 )
 
@@ -38,8 +39,9 @@ py_binary(
     python_version = "PY3",
     srcs_version = "PY3",
     deps = [
-        "//tensorflow_privacy/privacy/analysis:rdp_accountant",
         "//tensorflow_privacy/privacy/optimizers:dp_optimizer_keras",
+        "@com_google_differential_py//python/dp_accounting:dp_event",
+        "@com_google_differential_py//python/dp_accounting/rdp:rdp_privacy_accountant",
     ],
 )
 
@@ -49,8 +51,9 @@ py_binary(
     python_version = "PY3",
     srcs_version = "PY3",
     deps = [
-        "//tensorflow_privacy/privacy/analysis:rdp_accountant",
         "//tensorflow_privacy/privacy/keras_models:dp_keras_model",
+        "@com_google_differential_py//python/dp_accounting:dp_event",
+        "@com_google_differential_py//python/dp_accounting/rdp:rdp_privacy_accountant",
     ],
 )
 
@@ -60,10 +63,11 @@ py_binary(
     python_version = "PY3",
     srcs_version = "PY3",
     deps = [
-        "//tensorflow_privacy/privacy/analysis:rdp_accountant",
         "//tensorflow_privacy/privacy/optimizers:dp_optimizer_vectorized",
         "//third_party/py/tensorflow:tensorflow_compat_v1_estimator",
         "//third_party/py/tensorflow:tensorflow_estimator",
+        "@com_google_differential_py//python/dp_accounting:dp_event",
+        "@com_google_differential_py//python/dp_accounting/rdp:rdp_privacy_accountant",
     ],
 )
 

tutorials/mnist_dpsgd_tutorial_eager.py

@@ -17,11 +17,12 @@ from absl import app
 from absl import flags
 import numpy as np
 import tensorflow as tf
-
-from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp
-from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
 from tensorflow_privacy.privacy.optimizers.dp_optimizer import DPGradientDescentGaussianOptimizer
 
+from com_google_differential_py.python.dp_accounting import dp_event
+from com_google_differential_py.python.dp_accounting.rdp import rdp_privacy_accountant
+
+
 GradientDescentOptimizer = tf.compat.v1.train.GradientDescentOptimizer
 tf.compat.v1.enable_eager_execution()
 
@@ -46,14 +47,18 @@ def compute_epsilon(steps):
   if FLAGS.noise_multiplier == 0.0:
     return float('inf')
   orders = [1 + x / 10. for x in range(1, 100)] + list(range(12, 64))
+  accountant = rdp_privacy_accountant.RdpAccountant(orders)
+
   sampling_probability = FLAGS.batch_size / 60000
-  rdp = compute_rdp(
-      q=sampling_probability,
-      noise_multiplier=FLAGS.noise_multiplier,
-      steps=steps,
-      orders=orders)
+  event = dp_event.SelfComposedDpEvent(
+      dp_event.PoissonSampledDpEvent(
+          sampling_probability,
+          dp_event.GaussianDpEvent(FLAGS.noise_multiplier)), steps)
+
+  accountant.compose(event)
+
   # Delta is set to 1e-5 because MNIST has 60000 training points.
-  return get_privacy_spent(orders, rdp, target_delta=1e-5)[0]
+  return accountant.get_epsilon(target_delta=1e-5)
 
 
 def main(_):

tutorials/mnist_dpsgd_tutorial_keras.py

@@ -20,10 +20,11 @@ from absl import logging
 import numpy as np
 import tensorflow as tf
 
-from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp
-from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
 from tensorflow_privacy.privacy.optimizers.dp_optimizer_keras import DPKerasSGDOptimizer
 
+from com_google_differential_py.python.dp_accounting import dp_event
+from com_google_differential_py.python.dp_accounting.rdp import rdp_privacy_accountant
+
 flags.DEFINE_boolean(
     'dpsgd', True, 'If True, train with DP-SGD. If False, '
     'train with vanilla SGD.')
@@ -46,14 +47,18 @@ def compute_epsilon(steps):
   if FLAGS.noise_multiplier == 0.0:
     return float('inf')
   orders = [1 + x / 10. for x in range(1, 100)] + list(range(12, 64))
+  accountant = rdp_privacy_accountant.RdpAccountant(orders)
+
   sampling_probability = FLAGS.batch_size / 60000
-  rdp = compute_rdp(
-      q=sampling_probability,
-      noise_multiplier=FLAGS.noise_multiplier,
-      steps=steps,
-      orders=orders)
+  event = dp_event.SelfComposedDpEvent(
+      dp_event.PoissonSampledDpEvent(
+          sampling_probability,
+          dp_event.GaussianDpEvent(FLAGS.noise_multiplier)), steps)
+
+  accountant.compose(event)
+
   # Delta is set to 1e-5 because MNIST has 60000 training points.
-  return get_privacy_spent(orders, rdp, target_delta=1e-5)[0]
+  return accountant.get_epsilon(target_delta=1e-5)
 
 
 def load_mnist():

tutorials/mnist_dpsgd_tutorial_keras_model.py

@@ -19,10 +19,11 @@ from absl import logging
 import numpy as np
 import tensorflow as tf
 
-from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp
-from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
 from tensorflow_privacy.privacy.keras_models.dp_keras_model import DPSequential
 
+from com_google_differential_py.python.dp_accounting import dp_event
+from com_google_differential_py.python.dp_accounting.rdp import rdp_privacy_accountant
+
 flags.DEFINE_boolean(
     'dpsgd', True, 'If True, train with DP-SGD. If False, '
     'train with vanilla SGD.')
@@ -45,14 +46,18 @@ def compute_epsilon(steps):
   if FLAGS.noise_multiplier == 0.0:
     return float('inf')
   orders = [1 + x / 10. for x in range(1, 100)] + list(range(12, 64))
+  accountant = rdp_privacy_accountant.RdpAccountant(orders)
+
   sampling_probability = FLAGS.batch_size / 60000
-  rdp = compute_rdp(
-      q=sampling_probability,
-      noise_multiplier=FLAGS.noise_multiplier,
-      steps=steps,
-      orders=orders)
+  event = dp_event.SelfComposedDpEvent(
+      dp_event.PoissonSampledDpEvent(
+          sampling_probability,
+          dp_event.GaussianDpEvent(FLAGS.noise_multiplier)), steps)
+
+  accountant.compose(event)
+
   # Delta is set to 1e-5 because MNIST has 60000 training points.
-  return get_privacy_spent(orders, rdp, target_delta=1e-5)[0]
+  return accountant.get_epsilon(target_delta=1e-5)
 
 
 def load_mnist():

tutorials/mnist_dpsgd_tutorial_vectorized.py

@@ -20,10 +20,11 @@ import numpy as np
 import tensorflow as tf
 from tensorflow import estimator as tf_estimator
 from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator
-from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp
-from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
 from tensorflow_privacy.privacy.optimizers import dp_optimizer_vectorized
 
+from com_google_differential_py.python.dp_accounting import dp_event
+from com_google_differential_py.python.dp_accounting.rdp import rdp_privacy_accountant
+
 flags.DEFINE_boolean(
     'dpsgd', True, 'If True, train with DP-SGD. If False, '
     'train with vanilla SGD.')
@@ -50,14 +51,18 @@ def compute_epsilon(steps):
   if FLAGS.noise_multiplier == 0.0:
     return float('inf')
   orders = [1 + x / 10. for x in range(1, 100)] + list(range(12, 64))
-  sampling_probability = FLAGS.batch_size / NUM_TRAIN_EXAMPLES
-  rdp = compute_rdp(
-      q=sampling_probability,
-      noise_multiplier=FLAGS.noise_multiplier,
-      steps=steps,
-      orders=orders)
-  # Delta is set to approximate 1 / (number of training points).
-  return get_privacy_spent(orders, rdp, target_delta=1e-5)[0]
+  accountant = rdp_privacy_accountant.RdpAccountant(orders)
+
+  sampling_probability = FLAGS.batch_size / 60000
+  event = dp_event.SelfComposedDpEvent(
+      dp_event.PoissonSampledDpEvent(
+          sampling_probability,
+          dp_event.GaussianDpEvent(FLAGS.noise_multiplier)), steps)
+
+  accountant.compose(event)
+
+  # Delta is set to 1e-5 because MNIST has 60000 training points.
+  return accountant.get_epsilon(target_delta=1e-5)
 
 
 def cnn_model_fn(features, labels, mode):