def testBaseline(self, cls, num_microbatches, expected_answer):
with tf.GradientTape(persistent=True) as gradient_tape:
var0 = tf.Variable([1.0, 2.0])
data0 = tf.Variable([[3.0, 4.0], [5.0, 6.0], [7.0, 8.0],
[-1.0, 0.0]])
ledger = privacy_ledger.PrivacyLedger(1e6, num_microbatches / 1e6,
50, 50)
dp_average_query = gaussian_query.GaussianAverageQuery(
1.0e9, 0.0, num_microbatches, ledger)
dp_average_query = privacy_ledger.QueryWithLedger(
dp_average_query, ledger)
opt = cls(dp_average_query,
num_microbatches=num_microbatches,
learning_rate=2.0)
self.evaluate(tf.global_variables_initializer())
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0], self.evaluate(var0))
# Expected gradient is sum of differences divided by number of
# microbatches.
grads_and_vars = opt.compute_gradients(
lambda: self._loss_fn(var0, data0), [var0],
gradient_tape=gradient_tape)
self.assertAllCloseAccordingToType(expected_answer,
grads_and_vars[0][0])
def test_nested_gaussian_average_with_clip_no_noise(self):
with self.cached_session() as sess:
query1 = gaussian_query.GaussianAverageQuery(
l2_norm_clip=4.0, sum_stddev=0.0, denominator=5.0)
query2 = gaussian_query.GaussianAverageQuery(
l2_norm_clip=5.0, sum_stddev=0.0, denominator=5.0)
query = nested_query.NestedQuery([query1, query2])
record1 = [1.0, [12.0, 9.0]] # Clipped to [1.0, [4.0, 3.0]]
record2 = [5.0, [1.0, 2.0]] # Clipped to [4.0, [1.0, 2.0]]
query_result, _ = test_utils.run_query(query, [record1, record2])
result = sess.run(query_result)
expected = [1.0, [1.0, 1.0]]
self.assertAllClose(result, expected)
def testNoiseMultiplier(self, cls):
with tf.GradientTape(persistent=True) as gradient_tape:
var0 = tf.Variable([0.0])
data0 = tf.Variable([[0.0]])
ledger = privacy_ledger.PrivacyLedger(1e6, 1 / 1e6, 5000, 5000)
dp_average_query = gaussian_query.GaussianAverageQuery(4.0, 8.0, 1)
dp_average_query = privacy_ledger.QueryWithLedger(
dp_average_query, ledger)
opt = cls(dp_average_query, num_microbatches=1, learning_rate=2.0)
self.evaluate(tf.global_variables_initializer())
# Fetch params to validate initial values
self.assertAllClose([0.0], self.evaluate(var0))
grads = []
for _ in range(1000):
grads_and_vars = opt.compute_gradients(
lambda: self._loss_fn(var0, data0), [var0],
gradient_tape=gradient_tape)
grads.append(grads_and_vars[0][0])
# Test standard deviation is close to l2_norm_clip * noise_multiplier.
self.assertNear(np.std(grads), 2.0 * 4.0, 0.5)
def testUnrollMicrobatches(self, cls):
with self.cached_session() as sess:
var0 = tf.Variable([1.0, 2.0])
data0 = tf.Variable([[3.0, 4.0], [5.0, 6.0], [7.0, 8.0],
[-1.0, 0.0]])
num_microbatches = 4
ledger = privacy_ledger.PrivacyLedger(1e6, num_microbatches / 1e6,
50, 50)
dp_average_query = gaussian_query.GaussianAverageQuery(
1.0e9, 0.0, 4)
dp_average_query = privacy_ledger.QueryWithLedger(
dp_average_query, ledger)
opt = cls(dp_average_query,
num_microbatches=num_microbatches,
learning_rate=2.0,
unroll_microbatches=True)
self.evaluate(tf.global_variables_initializer())
# Fetch params to validate initial values
self.assertAllClose([1.0, 2.0], self.evaluate(var0))
# Expected gradient is sum of differences divided by number of
# microbatches.
gradient_op = opt.compute_gradients(self._loss(data0, var0),
[var0])
grads_and_vars = sess.run(gradient_op)
self.assertAllCloseAccordingToType([-2.5, -2.5],
grads_and_vars[0][0])
def __init__(
self,
l2_norm_clip,
noise_multiplier,
num_microbatches,
unroll_microbatches=False,
*args, # pylint: disable=keyword-arg-before-vararg
**kwargs):
dp_average_query = gaussian_query.GaussianAverageQuery(
l2_norm_clip, l2_norm_clip * noise_multiplier,
num_microbatches)
if 'population_size' in kwargs:
population_size = kwargs.pop('population_size')
max_queries = kwargs.pop('ledger_max_queries', 1e6)
max_samples = kwargs.pop('ledger_max_samples', 1e6)
selection_probability = num_microbatches / population_size
ledger = privacy_ledger.PrivacyLedger(population_size,
selection_probability,
max_samples, max_queries)
dp_average_query = privacy_ledger.QueryWithLedger(
dp_average_query, ledger)
super(DPGaussianOptimizerClass,
self).__init__(dp_average_query, num_microbatches,
unroll_microbatches, *args, **kwargs)
def test_nested_query(self):
population_size = tf.Variable(0)
selection_probability = tf.Variable(0.0)
ledger = privacy_ledger.PrivacyLedger(
population_size, selection_probability, 50, 50)
query1 = gaussian_query.GaussianAverageQuery(
l2_norm_clip=4.0, sum_stddev=2.0, denominator=5.0, ledger=ledger)
query2 = gaussian_query.GaussianAverageQuery(
l2_norm_clip=5.0, sum_stddev=1.0, denominator=5.0, ledger=ledger)
query = nested_query.NestedQuery([query1, query2])
query = privacy_ledger.QueryWithLedger(query, ledger)
record1 = [1.0, [12.0, 9.0]]
record2 = [5.0, [1.0, 2.0]]
# First sample.
tf.assign(population_size, 10)
tf.assign(selection_probability, 0.1)
test_utils.run_query(query, [record1, record2])
expected_queries = [[4.0, 2.0], [5.0, 1.0]]
formatted = ledger.get_formatted_ledger_eager()
sample_1 = formatted[0]
self.assertAllClose(sample_1.population_size, 10.0)
self.assertAllClose(sample_1.selection_probability, 0.1)
self.assertAllClose(sorted(sample_1.queries), sorted(expected_queries))
# Second sample.
tf.assign(population_size, 20)
tf.assign(selection_probability, 0.2)
test_utils.run_query(query, [record1, record2])
formatted = ledger.get_formatted_ledger_eager()
sample_1, sample_2 = formatted
self.assertAllClose(sample_1.population_size, 10.0)
self.assertAllClose(sample_1.selection_probability, 0.1)
self.assertAllClose(sorted(sample_1.queries), sorted(expected_queries))
self.assertAllClose(sample_2.population_size, 20.0)
self.assertAllClose(sample_2.selection_probability, 0.2)
self.assertAllClose(sorted(sample_2.queries), sorted(expected_queries))
def __init__(self,
initial_l2_norm_clip,
noise_multiplier,
target_unclipped_quantile,
learning_rate,
clipped_count_stddev,
expected_num_records,
ledger=None):
"""Initializes the QuantileAdaptiveClipSumQuery.
Args:
initial_l2_norm_clip: The initial value of clipping norm.
noise_multiplier: The multiplier of the l2_norm_clip to make the stddev of
the noise added to the output of the sum query.
target_unclipped_quantile: The desired quantile of updates which should be
unclipped. I.e., a value of 0.8 means a value of l2_norm_clip should be
found for which approximately 20% of updates are clipped each round.
learning_rate: The learning rate for the clipping norm adaptation. A
rate of r means that the clipping norm will change by a maximum of r at
each step. This maximum is attained when |clip - target| is 1.0. Can be
a tf.Variable for example to implement a learning rate schedule.
clipped_count_stddev: The stddev of the noise added to the clipped_count.
Since the sensitivity of the clipped count is 0.5, as a rule of thumb it
should be about 0.5 for reasonable privacy.
expected_num_records: The expected number of records per round, used to
estimate the clipped count quantile.
ledger: The privacy ledger to which queries should be recorded.
"""
self._initial_l2_norm_clip = tf.cast(initial_l2_norm_clip, tf.float32)
self._noise_multiplier = tf.cast(noise_multiplier, tf.float32)
self._target_unclipped_quantile = tf.cast(target_unclipped_quantile,
tf.float32)
self._learning_rate = tf.cast(learning_rate, tf.float32)
self._l2_norm_clip = tf.Variable(self._initial_l2_norm_clip)
self._sum_stddev = tf.Variable(self._initial_l2_norm_clip *
self._noise_multiplier)
self._sum_query = gaussian_query.GaussianSumQuery(
self._l2_norm_clip, self._sum_stddev, ledger)
# self._clipped_fraction_query is a DPQuery used to estimate the fraction of
# records that are clipped. It accumulates an indicator 0/1 of whether each
# record is clipped, and normalizes by the expected number of records. In
# practice, we accumulate clipped counts shifted by -0.5 so they are
# centered at zero. This makes the sensitivity of the clipped count query
# 0.5 instead of 1.0, since the maximum that a single record could affect
# the count is 0.5. Note that although the l2_norm_clip of the clipped
# fraction query is 0.5, no clipping will ever actually occur because the
# value of each record is always +/-0.5.
self._clipped_fraction_query = gaussian_query.GaussianAverageQuery(
l2_norm_clip=0.5,
sum_stddev=clipped_count_stddev,
denominator=expected_num_records,
ledger=ledger)
def test_gaussian_average_no_noise(self):
with self.cached_session() as sess:
record1 = tf.constant([5.0, 0.0]) # Clipped to [3.0, 0.0].
record2 = tf.constant([-1.0, 2.0]) # Not clipped.
query = gaussian_query.GaussianAverageQuery(
l2_norm_clip=3.0, sum_stddev=0.0, denominator=2.0)
query_result, _ = test_utils.run_query(query, [record1, record2])
result = sess.run(query_result)
expected_average = [1.0, 1.0]
self.assertAllClose(result, expected_average)
def test_gaussian_average_with_noise(self):
with self.cached_session() as sess:
record1, record2 = 2.71828, 3.14159
sum_stddev = 1.0
denominator = 2.0
query = gaussian_query.GaussianAverageQuery(
l2_norm_clip=5.0, sum_stddev=sum_stddev, denominator=denominator)
query_result, _ = test_utils.run_query(query, [record1, record2])
noised_averages = []
for _ in range(1000):
noised_averages.append(sess.run(query_result))
result_stddev = np.std(noised_averages)
avg_stddev = sum_stddev / denominator
self.assertNear(result_stddev, avg_stddev, 0.1)
def linear_model_fn(features, labels, mode):
preds = tf.keras.layers.Dense(1, activation='linear',
name='dense').apply(features['x'])
vector_loss = tf.squared_difference(labels, preds)
scalar_loss = tf.reduce_mean(vector_loss)
ledger = privacy_ledger.PrivacyLedger(1e6, 1 / 1e6, 500, 500)
dp_average_query = gaussian_query.GaussianAverageQuery(1.0, 0.0, 1)
dp_average_query = privacy_ledger.QueryWithLedger(
dp_average_query, ledger)
optimizer = dp_optimizer.DPGradientDescentOptimizer(
dp_average_query, num_microbatches=1, learning_rate=1.0)
global_step = tf.train.get_global_step()
train_op = optimizer.minimize(loss=vector_loss,
global_step=global_step)
return tf.estimator.EstimatorSpec(mode=mode,
loss=scalar_loss,
train_op=train_op)
def __init__(
self,
l2_norm_clip,
noise_multiplier,
num_microbatches,
ledger,
unroll_microbatches=False,
*args, # pylint: disable=keyword-arg-before-vararg
**kwargs):
dp_average_query = gaussian_query.GaussianAverageQuery(
l2_norm_clip, l2_norm_clip * noise_multiplier,
num_microbatches, ledger)
if ledger:
dp_average_query = privacy_ledger.QueryWithLedger(
dp_average_query, ledger)
super(DPGaussianOptimizerClass,
self).__init__(dp_average_query, num_microbatches,
unroll_microbatches, *args, **kwargs)
def test_complex_nested_query(self):
with self.cached_session() as sess:
query_ab = gaussian_query.GaussianSumQuery(
l2_norm_clip=1.0, stddev=0.0)
query_c = gaussian_query.GaussianAverageQuery(
l2_norm_clip=10.0, sum_stddev=0.0, denominator=2.0)
query_d = gaussian_query.GaussianSumQuery(
l2_norm_clip=10.0, stddev=0.0)
query = nested_query.NestedQuery(
[query_ab, {'c': query_c, 'd': [query_d]}])
record1 = [{'a': 0.0, 'b': 2.71828}, {'c': (-4.0, 6.0), 'd': [-4.0]}]
record2 = [{'a': 3.14159, 'b': 0.0}, {'c': (6.0, -4.0), 'd': [5.0]}]
query_result, _ = test_utils.run_query(query, [record1, record2])
result = sess.run(query_result)
expected = [{'a': 1.0, 'b': 1.0}, {'c': (1.0, 1.0), 'd': [1.0]}]
self.assertAllClose(result, expected)
def testClippingNorm(self, cls):
with tf.GradientTape(persistent=True) as gradient_tape:
var0 = tf.Variable([0.0, 0.0])
data0 = tf.Variable([[3.0, 4.0], [6.0, 8.0]])
ledger = privacy_ledger.PrivacyLedger(1e6, 1 / 1e6, 50, 50)
dp_average_query = gaussian_query.GaussianAverageQuery(1.0, 0.0, 1)
dp_average_query = privacy_ledger.QueryWithLedger(
dp_average_query, ledger)
opt = cls(dp_average_query, num_microbatches=1, learning_rate=2.0)
self.evaluate(tf.global_variables_initializer())
# Fetch params to validate initial values
self.assertAllClose([0.0, 0.0], self.evaluate(var0))
# Expected gradient is sum of differences.
grads_and_vars = opt.compute_gradients(
lambda: self._loss_fn(var0, data0), [var0],
gradient_tape=gradient_tape)
self.assertAllCloseAccordingToType([-0.6, -0.8],
grads_and_vars[0][0])
def test_nested_query_with_noise(self):
with self.cached_session() as sess:
sum_stddev = 2.71828
denominator = 3.14159
query1 = gaussian_query.GaussianSumQuery(
l2_norm_clip=1.5, stddev=sum_stddev)
query2 = gaussian_query.GaussianAverageQuery(
l2_norm_clip=0.5, sum_stddev=sum_stddev, denominator=denominator)
query = nested_query.NestedQuery((query1, query2))
record1 = (3.0, [2.0, 1.5])
record2 = (0.0, [-1.0, -3.5])
query_result, _ = test_utils.run_query(query, [record1, record2])
noised_averages = []
for _ in range(1000):
noised_averages.append(nest.flatten(sess.run(query_result)))
result_stddev = np.std(noised_averages, 0)
avg_stddev = sum_stddev / denominator
expected_stddev = [sum_stddev, avg_stddev, avg_stddev]
self.assertArrayNear(result_stddev, expected_stddev, 0.1)
