def test_annotate_call(self, mock_annotate_fn):
# Arrange
total_epsilon, total_delta = 3, 0.0001
budget_accountant = NaiveBudgetAccountant(total_epsilon,
total_delta,
num_aggregations=3)
dp_engine = self._create_dp_engine_default(budget_accountant)
aggregate_params, public_partitions = self._create_params_default()
select_partition_params = SelectPartitionsParams(2)
extractors = self._get_default_extractors()
input = [1, 2, 3]
# Act and assert
dp_engine.select_partitions(input, select_partition_params, extractors)
dp_engine.aggregate(input, aggregate_params, extractors,
public_partitions)
dp_engine.aggregate(input, aggregate_params, extractors,
public_partitions)
budget_accountant.compute_budgets()
# Assert
self.assertEqual(3, mock_annotate_fn.call_count)
for i_call in range(3):
budget = mock_annotate_fn.call_args_list[i_call][1]['budget']
self.assertEqual(total_epsilon / 3, budget.epsilon)
self.assertEqual(total_delta / 3, budget.delta)
def test_with_noise(self):
budget_accountant = NaiveBudgetAccountant(total_epsilon=10,
total_delta=1e-5)
budget = budget_accountant.request_budget(
pipeline_dp.MechanismType.GAUSSIAN)
budget_accountant.compute_budgets()
params = pipeline_dp.AggregateParams(
min_value=0,
max_value=1,
max_partitions_contributed=1,
max_contributions_per_partition=1,
noise_kind=NoiseKind.GAUSSIAN,
metrics=[pipeline_dp.Metrics.COUNT])
count_accumulator = accumulator.CountAccumulator(
accumulator.CountParams(budget, params), list(range(5)))
self.assertAlmostEqual(first=count_accumulator.compute_metrics(),
second=5,
delta=4)
count_accumulator.add_value(50)
self.assertAlmostEqual(first=count_accumulator.compute_metrics(),
second=6,
delta=4)
count_accumulator.add_value(list(range(49)))
self.assertAlmostEqual(first=count_accumulator.compute_metrics(),
second=7,
delta=4)
count_accumulator.add_value('*' * 100)
self.assertAlmostEqual(first=count_accumulator.compute_metrics(),
second=8,
delta=4)
def run_e2e_private_partition_selection_large_budget(col, backend):
# Arrange
aggregator_params = pipeline_dp.AggregateParams(
noise_kind=pipeline_dp.NoiseKind.LAPLACE,
metrics=[agg.Metrics.COUNT, agg.Metrics.SUM],
min_value=1,
max_value=10,
max_partitions_contributed=1,
max_contributions_per_partition=1)
# Set a large budget for having the small noise and keeping all
# partition keys.
budget_accountant = NaiveBudgetAccountant(total_epsilon=100000,
total_delta=1)
data_extractor = pipeline_dp.DataExtractors(
privacy_id_extractor=lambda x: x,
partition_extractor=lambda x: f"pk{x//2}",
value_extractor=lambda x: x)
engine = pipeline_dp.DPEngine(budget_accountant, backend)
col = engine.aggregate(col=col,
params=aggregator_params,
data_extractors=data_extractor)
budget_accountant.compute_budgets()
return col
def test_without_noise(self):
budget_accountant = NaiveBudgetAccountant(total_epsilon=1000000,
total_delta=0.9999999)
budget = budget_accountant.request_budget(
pipeline_dp.MechanismType.GAUSSIAN)
budget_accountant.compute_budgets()
no_noise = pipeline_dp.AggregateParams(
min_value=0,
max_value=1,
max_partitions_contributed=1,
max_contributions_per_partition=1,
noise_kind=NoiseKind.GAUSSIAN,
metrics=[pipeline_dp.Metrics.COUNT])
count_accumulator = accumulator.CountAccumulator(
accumulator.CountParams(budget, no_noise), list(range(5)))
self.assertEqual(count_accumulator.compute_metrics(), 5)
count_accumulator = accumulator.CountAccumulator(
accumulator.CountParams(budget, no_noise), 'a' * 50)
self.assertEqual(count_accumulator.compute_metrics(), 50)
count_accumulator = accumulator.CountAccumulator(
accumulator.CountParams(budget, no_noise), list(range(50)))
count_accumulator.add_value(49)
self.assertEqual(count_accumulator.compute_metrics(), 51)
count_accumulator_1 = accumulator.CountAccumulator(
accumulator.CountParams(budget, no_noise), list(range(50)))
count_accumulator_2 = accumulator.CountAccumulator(
accumulator.CountParams(budget, no_noise), 'a' * 50)
count_accumulator_1.add_accumulator(count_accumulator_2)
self.assertEqual(count_accumulator_1.compute_metrics(), 100)
def test_aggregate_report(self):
col = [[1], [2], [3], [3]]
data_extractor = pipeline_dp.DataExtractors(
privacy_id_extractor=lambda x: f"pid{x}",
partition_extractor=lambda x: f"pk{x}",
value_extractor=lambda x: x)
params1 = pipeline_dp.AggregateParams(
noise_kind=pipeline_dp.NoiseKind.GAUSSIAN,
max_partitions_contributed=3,
max_contributions_per_partition=2,
min_value=1,
max_value=5,
metrics=[
pipeline_dp.Metrics.PRIVACY_ID_COUNT,
pipeline_dp.Metrics.COUNT, pipeline_dp.Metrics.MEAN
],
)
params2 = pipeline_dp.AggregateParams(
noise_kind=pipeline_dp.NoiseKind.GAUSSIAN,
max_partitions_contributed=1,
max_contributions_per_partition=3,
min_value=2,
max_value=10,
metrics=[pipeline_dp.Metrics.SUM, pipeline_dp.Metrics.MEAN],
public_partitions=list(range(1, 40)),
)
select_partitions_params = SelectPartitionsParams(
max_partitions_contributed=2)
budget_accountant = NaiveBudgetAccountant(total_epsilon=1,
total_delta=1e-10)
engine = pipeline_dp.DPEngine(budget_accountant=budget_accountant,
backend=pipeline_dp.LocalBackend())
engine.aggregate(col, params1, data_extractor)
engine.aggregate(col, params2, data_extractor)
engine.select_partitions(col, select_partitions_params, data_extractor)
self.assertEqual(3, len(engine._report_generators)) # pylint: disable=protected-access
budget_accountant.compute_budgets()
self.assertEqual(
engine._report_generators[0].report(),
"Differentially private: Computing <Metrics: ['privacy_id_count', 'count', 'mean']>"
"\n1. Per-partition contribution bounding: randomly selected not more than 2 contributions"
"\n2. Cross-partition contribution bounding: randomly selected not more than 3 partitions per user"
"\n3. Private Partition selection: using Truncated Geometric method with (eps= 0.1111111111111111, delta = 1.1111111111111111e-11)"
)
self.assertEqual(
engine._report_generators[1].report(),
"Differentially private: Computing <Metrics: ['sum', 'mean']>"
"\n1. Public partition selection: dropped non public partitions"
"\n2. Per-partition contribution bounding: randomly selected not more than 3 contributions"
"\n3. Cross-partition contribution bounding: randomly selected not more than 1 partitions per user"
"\n4. Adding empty partitions to public partitions that are missing in data"
)
self.assertEqual(
engine._report_generators[2].report(),
"Differentially private: Computing <Private Partitions>"
"\n1. Private Partition selection: using Truncated Geometric method with (eps= 0.3333333333333333, delta = 3.3333333333333335e-11)"
)
def test_two_calls_compute_budgets_raise_exception(self):
budget_accountant = NaiveBudgetAccountant(total_epsilon=1,
total_delta=1e-6)
budget_accountant.request_budget(mechanism_type=MechanismType.LAPLACE)
budget_accountant.compute_budgets()
with self.assertRaises(Exception):
# Budget can be computed only once.
budget_accountant.compute_budgets()
def test_request_after_compute_raise_exception(self):
budget_accountant = NaiveBudgetAccountant(total_epsilon=1,
total_delta=1e-6)
budget_accountant.request_budget(mechanism_type=MechanismType.LAPLACE)
budget_accountant.compute_budgets()
with self.assertRaises(Exception):
# Budget can not be requested after it has been already computed.
budget_accountant.request_budget(
mechanism_type=MechanismType.LAPLACE)
def test_num_aggregations(self, num_aggregations):
total_epsilon, total_delta = 1, 1e-6
budget_accountant = NaiveBudgetAccountant(
total_epsilon=total_epsilon,
total_delta=total_delta,
num_aggregations=num_aggregations)
for _ in range(num_aggregations):
budget = budget_accountant._compute_budget_for_aggregation(1)
expected_epsilon = total_epsilon / num_aggregations
expected_delta = total_delta / num_aggregations
self.assertAlmostEqual(expected_epsilon, budget.epsilon)
self.assertAlmostEqual(expected_delta, budget.delta)
budget_accountant.compute_budgets()
def test_compute_budgets(self):
budget_accountant = NaiveBudgetAccountant(total_epsilon=1,
total_delta=1e-6)
budget1 = budget_accountant.request_budget(noise_kind=NoiseKind.LAPLACE)
budget2 = budget_accountant.request_budget(
noise_kind=NoiseKind.GAUSSIAN, weight=3)
budget_accountant.compute_budgets()
self.assertEqual(budget1.eps, 0.25)
self.assertEqual(budget1.delta,
0) # Delta should be 0 if mechanism is Gaussian.
self.assertEqual(budget2.eps, 0.75)
self.assertEqual(budget2.delta, 1e-6)
def test_select_partitions(self):
# This test is probabilistic, but the parameters were chosen to ensure
# the test has passed at least 10000 runs.
# Arrange
params = SelectPartitionsParams(max_partitions_contributed=1)
budget_accountant = NaiveBudgetAccountant(total_epsilon=1,
total_delta=1e-5)
# Generate dataset as a list of (user, partition_key) tuples.
# There partitions are generated to reflect several scenarios.
# A partition with sufficient amount of users.
col = [(u, "pk-many-contribs") for u in range(25)]
# A partition with many contributions, but only a few unique users.
col += [(100 + u // 10, "pk-many-contribs-few-users")
for u in range(30)]
# A partition with few contributions.
col += [(200 + u, "pk-few-contribs") for u in range(3)]
# Generating 30 partitions, each with the same group of 25 users
# 25 users is sufficient to keep the partition, but because of
# contribution bounding, much less users per partition will be kept.
for i in range(30):
col += [(500 + u, f"few-contribs-after-bound{i}")
for u in range(25)]
col = list(col)
data_extractor = pipeline_dp.DataExtractors(
privacy_id_extractor=lambda x: x[0],
partition_extractor=lambda x: x[1])
engine = pipeline_dp.DPEngine(budget_accountant=budget_accountant,
backend=pipeline_dp.LocalBackend())
col = engine.select_partitions(col=col,
params=params,
data_extractors=data_extractor)
budget_accountant.compute_budgets()
col = list(col)
# Assert
# Only one partition is retained, the one that has many unique _after_
# applying the "max_partitions_contributed" bound is retained.
self.assertEqual(["pk-many-contribs"], col)
def test_compute_budgets(self):
budget_accountant = NaiveBudgetAccountant(total_epsilon=1,
total_delta=1e-6)
budget1 = budget_accountant.request_budget(
mechanism_type=MechanismType.LAPLACE)
budget2 = budget_accountant.request_budget(
mechanism_type=MechanismType.GAUSSIAN, weight=3)
budget_accountant.compute_budgets()
self.assertEqual(budget1.eps, 0.25)
self.assertEqual(budget1.delta,
0) # Delta should be 0 if mechanism is Laplace.
self.assertEqual(budget2.eps, 0.75)
self.assertEqual(budget2.delta, 1e-6)
def test_aggregation_weights(self):
total_epsilon, total_delta = 1, 1e-6
weights = [1, 2, 5]
budget_accountant = NaiveBudgetAccountant(total_epsilon=total_epsilon,
total_delta=total_delta,
aggregation_weights=weights)
for weight in weights:
budget = budget_accountant._compute_budget_for_aggregation(weight)
expected_epsilon = total_epsilon * weight / sum(weights)
expected_delta = total_delta * weight / sum(weights)
self.assertAlmostEqual(expected_epsilon, budget.epsilon)
self.assertAlmostEqual(expected_delta, budget.delta)
budget_accountant.compute_budgets()
def test_not_enough_aggregations(self, use_num_aggregations):
weights = num_aggregations = None
if use_num_aggregations:
num_aggregations = 2
else:
weights = [1, 1] # 2 aggregations
budget_accountant = NaiveBudgetAccountant(
total_epsilon=1,
total_delta=1e-6,
num_aggregations=num_aggregations,
aggregation_weights=weights)
budget_accountant._compute_budget_for_aggregation(1)
with self.assertRaises(ValueError):
# num_aggregations = 2, but only 1 aggregation_scope was created
budget_accountant.compute_budgets()
def test_budget_scopes_no_parentscope(self):
budget_accountant = NaiveBudgetAccountant(total_epsilon=1,
total_delta=1e-6)
# Allocated in the top-level scope with no weight specified
budget1 = budget_accountant.request_budget(
mechanism_type=MechanismType.LAPLACE)
with budget_accountant.scope(weight=0.5):
budget2 = budget_accountant.request_budget(
mechanism_type=MechanismType.LAPLACE)
budget_accountant.compute_budgets()
self.assertEqual(budget1.eps, 1.0 / (1.0 + 0.5))
self.assertEqual(budget2.eps, 0.5 / (1.0 + 0.5))
def test_aggregate_public_partitions_add_empty_public_partitions(self):
# Arrange
aggregator_params = pipeline_dp.AggregateParams(
noise_kind=pipeline_dp.NoiseKind.GAUSSIAN,
metrics=[
agg.Metrics.COUNT, agg.Metrics.SUM,
agg.Metrics.PRIVACY_ID_COUNT
],
min_value=0,
max_value=1,
max_partitions_contributed=1,
max_contributions_per_partition=1,
public_partitions=["pk0", "pk10", "pk11"])
# Set a high budget to add close to 0 noise.
budget_accountant = NaiveBudgetAccountant(total_epsilon=100000,
total_delta=1 - 1e-10)
# Input collection has 10 elements, such that each privacy id
# contributes 1 time and each partition has 1 element.
col = list(range(10))
data_extractor = pipeline_dp.DataExtractors(
privacy_id_extractor=lambda x: x,
partition_extractor=lambda x: f"pk{x}",
value_extractor=lambda x: 1)
engine = pipeline_dp.DPEngine(budget_accountant=budget_accountant,
backend=pipeline_dp.LocalBackend())
col = engine.aggregate(col=col,
params=aggregator_params,
data_extractors=data_extractor)
budget_accountant.compute_budgets()
col = list(col)
partition_keys = [x[0] for x in col]
# Assert
# Only public partitions ("pk0") should be kept and empty public
# partitions ("pk10", "pk11") should be added.
self.assertEqual(["pk0", "pk10", "pk11"], partition_keys)
self.assertAlmostEqual(1, col[0][1][0]) # "pk0" COUNT ≈ 1
self.assertAlmostEqual(1, col[0][1][1]) # "pk0" SUM ≈ 1
self.assertAlmostEqual(1, col[0][1][2]) # "pk0" PRIVACY_ID_COUNT ≈ 1
self.assertAlmostEqual(0, col[1][1][0]) # "pk10" COUNT ≈ 0
self.assertAlmostEqual(0, col[1][1][1]) # "pk10" SUM ≈ 0
self.assertAlmostEqual(0, col[1][1][2]) # "pk10" PRIVACY_ID_COUNT ≈ 0
def test_aggregate_public_partitions_drop_non_public(self):
# Arrange
aggregator_params = pipeline_dp.AggregateParams(
noise_kind=pipeline_dp.NoiseKind.GAUSSIAN,
metrics=[
agg.Metrics.COUNT, agg.Metrics.SUM,
agg.Metrics.PRIVACY_ID_COUNT
],
min_value=0,
max_value=1,
max_partitions_contributed=1,
max_contributions_per_partition=1,
public_partitions=["pk0", "pk1", "pk10"])
# Set an arbitrary budget, we are not interested in the DP outputs, only
# the partition keys.
budget_accountant = NaiveBudgetAccountant(total_epsilon=1,
total_delta=1e-10)
# Input collection has 10 elements, such that each privacy id
# contributes 1 time and each partition has 1 element.
col = list(range(10))
data_extractor = pipeline_dp.DataExtractors(
privacy_id_extractor=lambda x: x,
partition_extractor=lambda x: f"pk{x}",
value_extractor=lambda x: x)
engine = pipeline_dp.DPEngine(budget_accountant=budget_accountant,
backend=pipeline_dp.LocalBackend())
col = engine.aggregate(col=col,
params=aggregator_params,
data_extractors=data_extractor)
budget_accountant.compute_budgets()
col = list(col)
partition_keys = [x[0] for x in col]
# Assert
# Only public partitions (0, 1, 2) should be kept and the rest of the
# partitions should be dropped.
self.assertEqual(["pk0", "pk1", "pk10"], partition_keys)
def test_budget_scopes(self):
budget_accountant = NaiveBudgetAccountant(total_epsilon=1,
total_delta=1e-6)
with budget_accountant.scope(weight=0.4):
budget1 = budget_accountant.request_budget(
mechanism_type=MechanismType.LAPLACE)
budget2 = budget_accountant.request_budget(
mechanism_type=MechanismType.LAPLACE, weight=3)
with budget_accountant.scope(weight=0.6):
budget3 = budget_accountant.request_budget(
mechanism_type=MechanismType.LAPLACE)
budget4 = budget_accountant.request_budget(
mechanism_type=MechanismType.LAPLACE, weight=4)
budget_accountant.compute_budgets()
self.assertEqual(budget1.eps, 0.4 * (1 / 4))
self.assertEqual(budget2.eps, 0.4 * (3 / 4))
self.assertEqual(budget3.eps, 0.6 * (1 / 5))
self.assertEqual(budget4.eps, 0.6 * (4 / 5))
def test_aggregate_private_partition_selection_drop_many(self):
# Arrange
aggregator_params = pipeline_dp.AggregateParams(
noise_kind=pipeline_dp.NoiseKind.GAUSSIAN,
metrics=[agg.Metrics.COUNT],
max_partitions_contributed=1,
max_contributions_per_partition=1)
# Set a small budget for dropping most partition keys.
budget_accountant = NaiveBudgetAccountant(total_epsilon=1,
total_delta=1e-10)
# Input collection has 100 elements, such that each privacy id
# contributes 1 time and each partition has 1 element.
col = list(range(100))
data_extractor = pipeline_dp.DataExtractors(
privacy_id_extractor=lambda x: x,
partition_extractor=lambda x: f"pk{x}",
value_extractor=lambda x: None)
engine = pipeline_dp.DPEngine(budget_accountant=budget_accountant,
backend=pipeline_dp.LocalBackend())
col = engine.aggregate(col=col,
params=aggregator_params,
data_extractors=data_extractor)
budget_accountant.compute_budgets()
col = list(col)
# Assert
# Most partition should be dropped by private partition selection.
# This tests is non-deterministic, but it should pass with probability
# very close to 1.
self.assertLess(len(col), 5)
def test_aggregate_private_partition_selection_keep_everything(self):
# Arrange
aggregator_params = pipeline_dp.AggregateParams(
noise_kind=pipeline_dp.NoiseKind.GAUSSIAN,
metrics=[agg.Metrics.COUNT],
max_partitions_contributed=1,
max_contributions_per_partition=1)
# Set a large budget for having the small noise and keeping all
# partition keys.
budget_accountant = NaiveBudgetAccountant(total_epsilon=100000,
total_delta=1e-10)
col = list(range(10)) + list(range(100, 120))
data_extractor = pipeline_dp.DataExtractors(
privacy_id_extractor=lambda x: x,
partition_extractor=lambda x: f"pk{x//100}",
value_extractor=lambda x: None)
engine = pipeline_dp.DPEngine(budget_accountant=budget_accountant,
backend=pipeline_dp.LocalBackend())
col = engine.aggregate(col=col,
params=aggregator_params,
data_extractors=data_extractor)
budget_accountant.compute_budgets()
col = list(col)
# Assert
approximate_expected = {"pk0": 10, "pk1": 20}
self.assertEqual(2, len(col)) # all partition keys are kept.
for pk, metrics_tuple in col:
dp_count = metrics_tuple.count
self.assertAlmostEqual(approximate_expected[pk],
dp_count,
delta=1e-3)
def test_aggregate_report(self):
col = [[1], [2], [3], [3]]
data_extractor = pipeline_dp.DataExtractors(
privacy_id_extractor=lambda x: f"pid{x}",
partition_extractor=lambda x: f"pk{x}",
value_extractor=lambda x: x)
params1 = pipeline_dp.AggregateParams(
noise_kind=pipeline_dp.NoiseKind.GAUSSIAN,
max_partitions_contributed=3,
max_contributions_per_partition=2,
min_value=1,
max_value=5,
metrics=[
pipeline_dp.Metrics.PRIVACY_ID_COUNT, pipeline_dp.Metrics.COUNT,
pipeline_dp.Metrics.MEAN
],
)
params2 = pipeline_dp.AggregateParams(
noise_kind=pipeline_dp.NoiseKind.GAUSSIAN,
max_partitions_contributed=1,
max_contributions_per_partition=3,
min_value=2,
max_value=10,
metrics=[pipeline_dp.Metrics.SUM, pipeline_dp.Metrics.MEAN],
)
select_partitions_params = SelectPartitionsParams(
max_partitions_contributed=2)
budget_accountant = NaiveBudgetAccountant(total_epsilon=1,
total_delta=1e-10)
engine = pipeline_dp.DPEngine(budget_accountant=budget_accountant,
backend=pipeline_dp.LocalBackend())
engine.aggregate(col, params1, data_extractor)
engine.aggregate(col, params2, data_extractor, list(range(1, 40)))
engine.select_partitions(col, select_partitions_params, data_extractor)
self.assertEqual(3, len(engine._report_generators)) # pylint: disable=protected-access
budget_accountant.compute_budgets()
self._check_string_contains_strings(
engine._report_generators[0].report(),
[
"DPEngine method: aggregate",
"metrics=['privacy_id_count', 'count', 'mean']",
" noise_kind=gaussian", "max_value=5",
"Partition selection: private partitions",
"Cross-partition contribution bounding: for each privacy id randomly select max(actual_partition_contributed, 3)",
"Private Partition selection: using Truncated Geometric method with (eps="
],
)
self._check_string_contains_strings(
engine._report_generators[1].report(),
[
"metrics=['sum', 'mean']", " noise_kind=gaussian",
"max_value=5", "Partition selection: public partitions",
"Per-partition contribution bounding: for each privacy_id and eachpartition, randomly select max(actual_contributions_per_partition, 3)",
"Adding empty partitions for public partitions that are missing in data"
],
)
self._check_string_contains_strings(
engine._report_generators[2].report(),
[
"DPEngine method: select_partitions",
" budget_weight=1",
"max_partitions_contributed=2",
"Private Partition selection: using Truncated Geometric method with",
],
)