def snapping_similarity():
snapping_estimates = []
laplace_estimates = []
with sn.Analysis(strict_parameter_checks=False):
PUMS = sn.Dataset(path=TEST_PUMS_PATH, column_names=TEST_PUMS_NAMES)
age = sn.impute(sn.to_float(PUMS['age']),
data_lower=0.,
data_upper=100.,
data_rows=1000)
for i in range(100):
snapping_component = sn.dp_mean(age,
mechanism="snapping",
privacy_usage={
"epsilon": 1.0,
"delta": 1E-6
})
laplace_component = sn.dp_mean(age,
mechanism="laplace",
privacy_usage={
"epsilon": 1.0,
"delta": 1E-6
})
snapping_estimates.append(snapping_component.value)
laplace_estimates.append(laplace_component.value)
print(sum(snapping_estimates) / len(snapping_estimates))
print(sum(laplace_estimates) / len(laplace_estimates))
def test_properties():
with sn.Analysis():
# load data
data = sn.Dataset(path=TEST_PUMS_PATH, column_names=TEST_PUMS_NAMES)
# establish data
age_dt = sn.cast(data['age'], 'FLOAT')
# ensure data are non-null
non_null_age_dt = sn.impute(age_dt,
distribution='Uniform',
lower=0.,
upper=100.)
clamped = sn.clamp(age_dt, lower=0., upper=100.)
# create potential for null data again
potentially_null_age_dt = non_null_age_dt / 0.
# print('original properties:\n{0}\n\n'.format(age_dt.properties))
print('properties after imputation:\n{0}\n\n'.format(
non_null_age_dt.nullity))
print('properties after nan mult:\n{0}\n\n'.format(
potentially_null_age_dt.nullity))
print("lower", clamped.lower)
print("upper", clamped.upper)
print("releasable", clamped.releasable)
# print("props", clamped.properties)
print("data_type", clamped.data_type)
print("categories", clamped.categories)
def analytic_gaussian_similarity():
analytic_gauss_estimates = []
gauss_estimates = []
with sn.Analysis(strict_parameter_checks=False):
PUMS = sn.Dataset(path=TEST_PUMS_PATH, column_names=TEST_PUMS_NAMES)
age = sn.impute(sn.to_float(PUMS['age']),
data_lower=0.,
data_upper=100.,
data_rows=1000)
for i in range(100):
an_gauss_component = sn.dp_mean(age,
mechanism="AnalyticGaussian",
privacy_usage={
"epsilon": 1.0,
"delta": 1E-6
})
gauss_component = sn.dp_mean(age,
mechanism="Gaussian",
privacy_usage={
"epsilon": 1.0,
"delta": 1E-6
})
# this triggers an analysis.release (which also computes gauss_component)
analytic_gauss_estimates.append(an_gauss_component.value)
gauss_estimates.append(gauss_component.value)
print(sum(analytic_gauss_estimates) / len(analytic_gauss_estimates))
print(sum(gauss_estimates) / len(gauss_estimates))
def test_private_clamped_sum_helpers():
# Compute the CI with smartnoise
with sn.Analysis() as analysis:
data = sn.Dataset(path=TEST_DATA_PATH, column_names=TEST_DATA_COLUMNS)
D = sn.to_float(data["age"])
D_tilde = sn.resize(sn.clamp(data=D, lower=0.0, upper=100.0), number_rows=1000,)
release = sn.dp_sum(data=sn.impute(D_tilde), privacy_usage={"epsilon": 1.0})
smartnoise_ci = release.get_accuracy(0.05)
op = PrivateClampedSum(lower_bound=0, upper_bound=100)
eeprivacy_ci = op.confidence_interval(epsilon=1, confidence=0.95)
assert pytest.approx(smartnoise_ci, abs=0.001) == eeprivacy_ci
def try_sn():
# establish data information
#data_path = 'https://raw.githubusercontent.com/opendp/smartnoise-samples/86-requirements-fix/analysis/data/PUMS_california_demographics_1000/data.csv'
data_path = os.path.join('.', 'data', 'PUMS_california_demographics_1000',
'data.csv')
data_path = os.path.abspath(data_path)
print('data_path', data_path)
var_names = ["age", "sex", "educ", "race", "income", "married", "pid"]
D = pd.read_csv(data_path)['age']
D_mean_age = np.mean(D)
print('D_mean_age', D_mean_age)
# establish extra information for this simulation
age_lower_bound = 0.
age_upper_bound = 100.
D_tilde = np.clip(D, age_lower_bound, age_upper_bound)
D_tilde_mean_age = np.mean(D_tilde)
data_size = 1000
df = pd.read_csv(data_path)
df_as_array = [list(row) for row in df.itertuples()]
#df.values.tolist()
print('D.values', df_as_array)
n_sims = 2
releases = []
with sn.Analysis(dynamic=True) as analysis:
data = sn.Dataset(path=data_path, column_names=var_names)
#data = sn.Dataset(value=df_as_array, column_names=var_names)
D = sn.to_float(data['age'])
# preprocess data (resize is a no-op because we have the correct data size)
D_tilde = sn.resize(sn.clamp(data=D, lower=0., upper=100.),
number_rows=data_size)
for index in range(n_sims):
# get DP mean of age
releases.append(
sn.dp_mean(data=sn.impute(D_tilde),
privacy_usage={'epsilon': 1}))
accuracy = releases[0].get_accuracy(0.05)
analysis.release()
dp_values = [release.value for release in releases]
print(
'Accuracy interval (with accuracy value {0}) contains the true mean on D_tilde with probability {1}'
.format(
round(accuracy, 4),
np.mean([(D_tilde_mean_age >= val - accuracy) &
(D_tilde_mean_age <= val + accuracy)
for val in dp_values])))
def test_divide():
with sn.Analysis():
data_A = generate_synthetic(float, variants=['Random'])
f_random = data_A['F_Random']
imputed = sn.impute(f_random, lower=0., upper=10.)
clamped_nonzero = sn.clamp(imputed, lower=1., upper=10.)
clamped_zero = sn.clamp(imputed, lower=0., upper=10.)
# test properties
assert f_random.nullity
assert not imputed.nullity
assert (2. / imputed).nullity
assert (f_random / imputed).nullity
assert (2. / clamped_zero).nullity
def test_median_education():
# import pandas as pd
# print(pd.read_csv(data_path)['value'].median())
with sn.Analysis(filter_level="all") as analysis:
data = sn.Dataset(path=TEST_EDUC_PATH, column_names=TEST_EDUC_NAMES)
candidates = list(map(float, range(1, 200, 2)))
median_scores = sn.median(sn.impute(sn.to_float(data['value']), 100.,
200.),
candidates=candidates)
# print(list(zip(candidates, median_scores.value[0])))
dp_median = sn.exponential_mechanism(median_scores,
candidates=candidates,
privacy_usage={"epsilon": 100.})
print(dp_median.value)
analysis.release()
def test_mechanism(args, constructor):
with sn.Analysis() as analysis:
PUMS = sn.Dataset(path=TEST_PUMS_PATH, column_names=TEST_PUMS_NAMES)
categorical = sn.resize(sn.clamp(PUMS['sex'],
categories=["0", "1"],
null_value="0"),
number_rows=1000)
numeric = sn.impute(sn.to_float(PUMS['age']),
data_lower=0.,
data_upper=100.,
data_rows=1000)
all = constructor(numeric, categorical, args)
analysis.release()
all_values = {stat: all[stat].value for stat in all}
print()
pprint(all_values)
for value in all_values.values():
assert value is not None
def test_private_clamped_mean_helpers():
# Compute the CI with smartnoise
with sn.Analysis() as analysis:
data = sn.Dataset(path=TEST_DATA_PATH, column_names=TEST_DATA_COLUMNS)
D = sn.to_float(data["age"])
D_tilde = sn.resize(sn.clamp(data=D, lower=0.0, upper=100.0), number_rows=1000,)
release = sn.dp_mean(data=sn.impute(D_tilde), privacy_usage={"epsilon": 1.0})
smartnoise_ci = release.get_accuracy(0.05)
# Compute the CI with eeprivacy
op = PrivateClampedMean(lower_bound=0, upper_bound=100)
eeprivacy_ci = op.confidence_interval(epsilon=1, N=1000, confidence=0.95)
# Compare computed confidence intervals
assert pytest.approx(smartnoise_ci, abs=0.001) == eeprivacy_ci
smartnoise_epsilon = release.from_accuracy(value=1, alpha=0.05)[0]["epsilon"]
eeprivacy_epsilon = op.epsilon_for_confidence_interval(
target_ci=1, N=1000, confidence=0.95
)
# Compare computed epsilons for confidence interval
assert pytest.approx(smartnoise_epsilon, abs=0.001) == eeprivacy_epsilon
def test_dp_linear_stats(run=True):
with sn.Analysis() as analysis:
dataset_pums = sn.Dataset(path=TEST_PUMS_PATH,
column_names=TEST_PUMS_NAMES)
age = dataset_pums['age']
analysis.release()
num_records = sn.dp_count(age,
privacy_usage={'epsilon': .5},
lower=0,
upper=10000)
analysis.release()
print("number of records:", num_records.value)
vars = sn.to_float(dataset_pums[["age", "income"]])
covariance = sn.dp_covariance(data=vars,
privacy_usage={'epsilon': .5},
data_lower=[0., 0.],
data_upper=[150., 150000.],
data_rows=num_records)
print("covariance released")
num_means = sn.dp_mean(data=vars,
privacy_usage={'epsilon': .5},
data_lower=[0., 0.],
data_upper=[150., 150000.],
data_rows=num_records)
analysis.release()
print("covariance:\n", covariance.value)
print("means:\n", num_means.value)
age = sn.to_float(age)
age_variance = sn.dp_variance(age,
privacy_usage={'epsilon': .5},
data_lower=0.,
data_upper=150.,
data_rows=num_records)
analysis.release()
print("age variance:", age_variance.value)
# If I clamp, impute, resize, then I can reuse their properties for multiple statistics
clamped_age = sn.clamp(age, lower=0., upper=100.)
imputed_age = sn.impute(clamped_age)
preprocessed_age = sn.resize(imputed_age, number_rows=num_records)
# properties necessary for mean are statically known
mean = sn.dp_mean(preprocessed_age, privacy_usage={'epsilon': .5})
# properties necessary for variance are statically known
variance = sn.dp_variance(preprocessed_age,
privacy_usage={'epsilon': .5})
# sum doesn't need n, so I pass the data in before resizing
age_sum = sn.dp_sum(imputed_age, privacy_usage={'epsilon': .5})
# mean with lower, upper properties propagated up from prior bounds
transformed_mean = sn.dp_mean(-(preprocessed_age + 2.),
privacy_usage={'epsilon': .5})
analysis.release()
print("age transformed mean:", transformed_mean.value)
# releases may be pieced together from combinations of smaller components
custom_mean = sn.laplace_mechanism(sn.mean(preprocessed_age),
privacy_usage={'epsilon': .5})
custom_maximum = sn.laplace_mechanism(sn.maximum(preprocessed_age),
privacy_usage={'epsilon': .5})
custom_maximum = sn.laplace_mechanism(sn.maximum(preprocessed_age),
privacy_usage={'epsilon': .5})
custom_quantile = sn.laplace_mechanism(sn.quantile(preprocessed_age,
alpha=.5),
privacy_usage={'epsilon': 500})
income = sn.to_float(dataset_pums['income'])
income_max = sn.laplace_mechanism(sn.maximum(income,
data_lower=0.,
data_upper=1000000.),
privacy_usage={'epsilon': 10})
# releases may also be postprocessed and reused as arguments to more components
age_sum + custom_maximum * 23.
analysis.release()
print("laplace quantile:", custom_quantile.value)
age_histogram = sn.dp_histogram(sn.to_int(age, lower=0, upper=100),
edges=list(range(0, 100, 25)),
null_value=150,
privacy_usage={'epsilon': 2.})
sex_histogram = sn.dp_histogram(sn.to_bool(dataset_pums['sex'],
true_label="1"),
privacy_usage={'epsilon': 2.})
education_histogram = sn.dp_histogram(dataset_pums['educ'],
categories=["5", "7", "10"],
null_value="-1",
privacy_usage={'epsilon': 2.})
analysis.release()
print("age histogram: ", age_histogram.value)
print("sex histogram: ", sex_histogram.value)
print("education histogram: ", education_histogram.value)
if run:
analysis.release()
# get the mean computed when release() was called
print(mean.value)
print(variance.value)
return analysis
def test_everything(run=True):
with sn.Analysis() as analysis:
data = sn.Dataset(path=TEST_PUMS_PATH, column_names=TEST_PUMS_NAMES)
age_int = sn.to_int(data['age'], 0, 150)
sex = sn.to_bool(data['sex'], "1")
educ = sn.to_float(data['educ'])
race = data['race']
income = sn.to_float(data['income'])
married = sn.to_bool(data['married'], "1")
numerics = sn.to_float(data[['age', 'income']])
# intentionally busted component
# print("invalid component id ", (sex + "a").component_id)
# broadcast scalar over 2d, broadcast scalar over 1d, columnar broadcasting, left and right mul
numerics * 2. + 2. * educ
# add different values for each column
numerics + [[1., 2.]]
# index into first column
age = sn.index(numerics, indices=0)
income = sn.index(numerics, mask=[False, True])
# boolean ops and broadcasting
mask = sex & married | (~married ^ False) | (age > 50.) | (age_int
== 25)
# numerical clamping
sn.clamp(numerics, 0., [150., 150_000.])
sn.clamp(data['educ'],
categories=[str(i) for i in range(8, 10)],
null_value="-1")
sn.count(mask)
sn.covariance(age, income)
sn.digitize(educ, edges=[1., 3., 10.], null_value=-1)
# checks for safety against division by zero
income / 2.
income / sn.clamp(educ, 5., 20.)
sn.dp_count(data, privacy_usage={"epsilon": 0.5})
sn.dp_count(mask, privacy_usage={"epsilon": 0.5})
sn.dp_histogram(mask, privacy_usage={"epsilon": 0.5})
age = sn.impute(sn.clamp(age, 0., 150.))
sn.dp_maximum(age, privacy_usage={"epsilon": 0.5})
sn.dp_minimum(age, privacy_usage={"epsilon": 0.5})
sn.dp_median(age, privacy_usage={"epsilon": 0.5})
age_n = sn.resize(age, number_rows=800)
sn.dp_mean(age_n, privacy_usage={"epsilon": 0.5})
sn.dp_raw_moment(age_n, order=3, privacy_usage={"epsilon": 0.5})
sn.dp_sum(age, privacy_usage={"epsilon": 0.5})
sn.dp_variance(age_n, privacy_usage={"epsilon": 0.5})
sn.filter(income, mask)
race_histogram = sn.histogram(race,
categories=["1", "2", "3"],
null_value="3")
sn.histogram(income, edges=[0., 10000., 50000.], null_value=-1)
sn.dp_histogram(married, privacy_usage={"epsilon": 0.5})
sn.gaussian_mechanism(race_histogram,
privacy_usage={
"epsilon": 0.5,
"delta": .000001
})
sn.laplace_mechanism(race_histogram,
privacy_usage={
"epsilon": 0.5,
"delta": .000001
})
sn.raw_moment(educ, order=3)
sn.log(sn.clamp(educ, 0.001, 50.))
sn.maximum(educ)
sn.mean(educ)
sn.minimum(educ)
educ % 2.
educ**2.
sn.quantile(educ, .32)
sn.resize(educ, number_rows=1200, lower=0., upper=50.)
sn.resize(race,
number_rows=1200,
categories=["1", "2"],
weights=[1, 2])
sn.resize(data[["age", "sex"]],
1200,
categories=[["1", "2"], ["a", "b"]],
weights=[1, 2])
sn.resize(data[["age", "sex"]],
1200,
categories=[["1", "2"], ["a", "b", "c"]],
weights=[[1, 2], [3, 7, 2]])
sn.sum(educ)
sn.variance(educ)
if run:
analysis.release()
return analysis