def test_convolve_dictionary(self):
dictionary1 = {1: 2, 3: 4}
dictionary2 = {2: 3, 4: 6}
expected_result = {3: 6, 5: 24, 7: 24}
result = privacy_loss_distribution.convolve_dictionary(
dictionary1, dictionary2)
test_util.dictionary_almost_equal(self, expected_result, result)
def test_laplace_discretization(
self, value_discretization_interval,
expected_rounded_probability_mass_function):
pld = privacy_loss_distribution.PrivacyLossDistribution.from_laplace_mechanism(
1, value_discretization_interval=value_discretization_interval)
test_util.dictionary_almost_equal(
self, expected_rounded_probability_mass_function,
pld.rounded_probability_mass_function)
def test_laplace_optimistic(self, sensitivity,
expected_rounded_probability_mass_function):
pld = privacy_loss_distribution.PrivacyLossDistribution.from_laplace_mechanism(
1,
sensitivity=sensitivity,
pessimistic_estimate=False,
value_discretization_interval=1)
test_util.dictionary_almost_equal(
self, expected_rounded_probability_mass_function,
pld.rounded_probability_mass_function)
def test_discrete_laplace_varying_standard_deviation_and_sensitivity(
self, parameter, sensitivity,
expected_rounded_probability_mass_function):
pld = privacy_loss_distribution.PrivacyLossDistribution.from_discrete_laplace_mechanism(
parameter,
sensitivity=sensitivity,
value_discretization_interval=1)
test_util.dictionary_almost_equal(
self, expected_rounded_probability_mass_function,
pld.rounded_probability_mass_function)
def test_discrete_gaussian_truncation(
self, truncation_bound, expected_rounded_probability_mass_function,
expected_infinity_mass):
pld = (privacy_loss_distribution.PrivacyLossDistribution.
from_discrete_gaussian_mechanism(
1, truncation_bound=truncation_bound))
self.assertAlmostEqual(pld.infinity_mass, expected_infinity_mass)
test_util.dictionary_almost_equal(
self, expected_rounded_probability_mass_function,
pld.rounded_probability_mass_function)
def test_gaussian_discretization(
self, value_discretization_interval,
expected_rounded_probability_mass_function):
pld = privacy_loss_distribution.PrivacyLossDistribution.from_gaussian_mechanism(
1,
log_mass_truncation_bound=math.log(2) + stats.norm.logcdf(-0.9),
value_discretization_interval=value_discretization_interval)
self.assertAlmostEqual(stats.norm.cdf(-0.9), pld.infinity_mass)
test_util.dictionary_almost_equal(
self, expected_rounded_probability_mass_function,
pld.rounded_probability_mass_function)
def test_randomized_response_basic(
self, noise_parameter, num_buckets,
expected_rounded_probability_mass_function):
# Set value_discretization_interval = 1 here.
pld = (privacy_loss_distribution.PrivacyLossDistribution.
from_randomized_response(noise_parameter,
num_buckets,
value_discretization_interval=1))
test_util.dictionary_almost_equal(
self, expected_rounded_probability_mass_function,
pld.rounded_probability_mass_function)
def test_from_privacy_parameters(
self, epsilon, delta, value_discretization_interval,
expected_rounded_probability_mass_function,
expected_infinity_mass):
pld = privacy_loss_distribution.PrivacyLossDistribution.from_privacy_parameters(
common.DifferentialPrivacyParameters(epsilon, delta),
value_discretization_interval=value_discretization_interval)
self.assertAlmostEqual(expected_infinity_mass, pld.infinity_mass)
test_util.dictionary_almost_equal(
self, expected_rounded_probability_mass_function,
pld.rounded_probability_mass_function)
def test_gaussian_varying_standard_deviation_and_sensitivity(
self, standard_deviation, sensitivity,
expected_rounded_probability_mass_function):
pld = privacy_loss_distribution.PrivacyLossDistribution.from_gaussian_mechanism(
standard_deviation,
sensitivity=sensitivity,
log_mass_truncation_bound=math.log(2) + stats.norm.logcdf(-0.9),
value_discretization_interval=1)
self.assertAlmostEqual(stats.norm.cdf(-0.9), pld.infinity_mass)
test_util.dictionary_almost_equal(
self, expected_rounded_probability_mass_function,
pld.rounded_probability_mass_function)
def test_gaussian_optimistic(self, sensitivity,
expected_rounded_probability_mass_function):
pld = privacy_loss_distribution.PrivacyLossDistribution.from_gaussian_mechanism(
1,
sensitivity=sensitivity,
pessimistic_estimate=False,
log_mass_truncation_bound=math.log(2) + stats.norm.logcdf(-0.9),
value_discretization_interval=1)
self.assertAlmostEqual(0, pld.infinity_mass)
test_util.dictionary_almost_equal(
self, expected_rounded_probability_mass_function,
pld.rounded_probability_mass_function)
def test_discrete_gaussian_varying_sigma_and_sensitivity(
self, sigma, sensitivity,
expected_rounded_probability_mass_function,
expected_infinity_mass):
pld = (privacy_loss_distribution.PrivacyLossDistribution.
from_discrete_gaussian_mechanism(sigma,
sensitivity=sensitivity,
truncation_bound=1))
self.assertAlmostEqual(pld.infinity_mass, expected_infinity_mass)
test_util.dictionary_almost_equal(
self, expected_rounded_probability_mass_function,
pld.rounded_probability_mass_function)
def test_discrete_laplace_privacy_loss_tail(
self, parameter, sensitivity, expected_lower_x_truncation,
expected_upper_x_truncation, expected_tail_probability_mass_function):
pl = privacy_loss_mechanism.DiscreteLaplacePrivacyLoss(
parameter, sensitivity=sensitivity)
tail_pld = pl.privacy_loss_tail()
self.assertAlmostEqual(expected_lower_x_truncation,
tail_pld.lower_x_truncation)
self.assertAlmostEqual(expected_upper_x_truncation,
tail_pld.upper_x_truncation)
test_util.dictionary_almost_equal(self,
expected_tail_probability_mass_function,
tail_pld.tail_probability_mass_function)
def test_discrete_gaussian_privacy_loss_tail(
self, sigma, sensitivity, truncation_bound, expected_lower_x_truncation,
expected_upper_x_truncation, expected_tail_probability_mass_function):
pl = privacy_loss_mechanism.DiscreteGaussianPrivacyLoss(
sigma, sensitivity=sensitivity, truncation_bound=truncation_bound)
tail_pld = pl.privacy_loss_tail()
self.assertAlmostEqual(expected_lower_x_truncation,
tail_pld.lower_x_truncation)
self.assertAlmostEqual(expected_upper_x_truncation,
tail_pld.upper_x_truncation)
test_util.dictionary_almost_equal(self,
expected_tail_probability_mass_function,
tail_pld.tail_probability_mass_function)
def test_self_convolve_dictionary(self):
inp_dictionary = {1: 2, 3: 5, 4: 6}
expected_result = {
3: 8,
5: 60,
6: 72,
7: 150,
8: 360,
9: 341,
10: 450,
11: 540,
12: 216
}
result = common.self_convolve_dictionary(inp_dictionary, 3)
test_util.dictionary_almost_equal(self, expected_result, result)
def test_randomized_response_discretization(
self, value_discretization_interval,
expected_rounded_probability_mass_function):
# Set noise_parameter = 0.2, num_buckets = 4 here.
# The true (non-discretized) PLD is
# {2.83321334: 0.85, -2.83321334: 0.05, 0: 0.1}.
pld = (
privacy_loss_distribution.PrivacyLossDistribution.
from_randomized_response(
0.2,
4,
value_discretization_interval=value_discretization_interval))
test_util.dictionary_almost_equal(
self, expected_rounded_probability_mass_function,
pld.rounded_probability_mass_function)
def test_gaussian_privacy_loss_tail(
self, standard_deviation, sensitivity, expected_lower_x_truncation,
expected_upper_x_truncation, pessimistic_estimate,
expected_tail_probability_mass_function):
pl = privacy_loss_mechanism.GaussianPrivacyLoss(
standard_deviation,
sensitivity=sensitivity,
pessimistic_estimate=pessimistic_estimate,
log_mass_truncation_bound=math.log(2) + stats.norm.logcdf(-1))
tail_pld = pl.privacy_loss_tail()
self.assertAlmostEqual(expected_lower_x_truncation,
tail_pld.lower_x_truncation)
self.assertAlmostEqual(expected_upper_x_truncation,
tail_pld.upper_x_truncation)
test_util.dictionary_almost_equal(
self, expected_tail_probability_mass_function,
tail_pld.tail_probability_mass_function)
def test_identity(self):
pld = privacy_loss_distribution.PrivacyLossDistribution.identity()
test_util.dictionary_almost_equal(
self, pld.rounded_probability_mass_function, {0: 1})
self.assertAlmostEqual(pld.infinity_mass, 0)
pld = pld.compose(
privacy_loss_distribution.PrivacyLossDistribution({
1: 0.5,
-1: 0.5
}, 1e-4, 0))
test_util.dictionary_almost_equal(
self, pld.rounded_probability_mass_function, {
1: 0.5,
-1: 0.5
})
self.assertAlmostEqual(pld.infinity_mass, 0)
def test_composition_with_truncation(self):
pld1 = privacy_loss_distribution.PrivacyLossDistribution(
{
0: 0.1,
1: 0.7,
2: 0.1
}, 1, 0.1)
pld2 = privacy_loss_distribution.PrivacyLossDistribution(
{
0: 0.1,
1: 0.6,
2: 0.2
}, 1, 0.1)
pld_composed = pld1.compose(pld2, tail_mass_truncation=0.021)
self.assertAlmostEqual(pld_composed.infinity_mass, 0.211)
test_util.dictionary_almost_equal(
self, {
1: 0.13,
2: 0.45,
3: 0.20,
4: 0.02,
}, pld_composed.rounded_probability_mass_function)
def test_list_to_dict_truncation(self, input_list, offset,
tail_mass_truncation, expected_result):
result = common.list_to_dictionary(
input_list, offset, tail_mass_truncation=tail_mass_truncation)
test_util.dictionary_almost_equal(self, expected_result, result)
def test_convolve_dictionary_with_truncation(self):
dictionary1 = {1: 0.4, 2: 0.6}
dictionary2 = {1: 0.7, 3: 0.3}
expected_result = {3: 0.42, 4: 0.12}
result = common.convolve_dictionary(dictionary1, dictionary2, 0.57)
test_util.dictionary_almost_equal(self, expected_result, result)