def __init__(self, total_privacy_budget: PrivacyBudget):
"""
:param total_privacy_budget: The total privacy budget that can be consumed by the private table.
When is there is no privacy budget left, stop answering queries.
"""
self.total_privacy_budget = total_privacy_budget
self.consumed_privacy_budget = PrivacyBudget(0., 0.)
def example_private_table():
adult_data = pd.read_csv(os.path.join("dataset", "adult_data.txt"),
names=[
"Age", "Workclass", "fnlwgt", "Education",
"Education-Num", "Martial Status",
"Occupation", "Relationship", "Race", "Sex",
"Capital Gain", "Capital Loss",
"Hours per week", "Country", "Target"
])
data = {
'Age': adult_data["Age"].tolist(),
'Education': adult_data["Education"].tolist()
}
df = pd.DataFrame(data)
domains = {
'Age':
RealDataDomain(17., 90.),
'Education':
CategoricalDataDomain([
' Bachelors', ' HS-grad', ' 11th', ' Masters', ' 9th',
' Some-college', ' Assoc-acdm', ' Assoc-voc', ' 7th-8th',
' Doctorate', ' Prof-school', ' 5th-6th', ' 10th', ' 1st-4th',
' Preschool', ' 12th'
])
}
return PrivateTable(df, domains, PrivacyBudget(100000.0, 1.))
def test_private_gaussian_mean(example_private_table: PrivateTable):
"""check private guassian mean implementation."""
noisy_mean = np.mean([
example_private_table.gaussian_mean('Age', PrivacyBudget(0.99, 0.5))
for i in range(100)
])
check_absolute_error(noisy_mean, 33.2, 10.)
def update_privacy_loss(self,
sampling_ratio: float,
sigma: float,
steps: int,
moment_order: int = 32,
target_eps: Union[float, None] = None,
target_delta: Union[float, None] = None):
"""Calculate and update privacy loss. Must specify exactly either one of `target_eps` or `target_delta`.
:param sampling_ratio: Ratio of data used to total data in one step
:param sigma: Noise scale
:param steps: Number of update performed
:param moment_order: Maximum order of moment to calculate privacy budget, defaults to 32
:param target_eps: Target value of :math:`\epsilon`, defaults to None
:param target_delta: Target value of :math:`\delta`, defaults to None
"""
assert (target_eps is None) or (
target_eps > 0), "Value of epsilon should be positive"
assert (target_delta is None) or (
target_delta > 0), "Value of delta should be positive"
log_moments = [(i, compute_log_moment(sampling_ratio, sigma, steps, i))
for i in range(1, moment_order + 1)]
privacy = get_privacy_spent(log_moments, target_eps, target_delta)
privacy_budget = PrivacyBudget(privacy[0], privacy[1])
e = self.consumed_privacy_budget + privacy_budget
assert e <= self.total_privacy_budget, "there is not enough privacy budget."
self.consumed_privacy_budget = e
def test_private_categorical_hist(example_private_table: PrivateTable):
"""check private hist implementation for categorical column."""
noisy_hist = example_private_table.cat_hist('Name', PrivacyBudget(10000.))
err = [1, 1, 1]
noisy_hist.sort()
assert all(np.abs(noisy_hist - [1, 1, 2]) < err)
del noisy_hist
def test_private_SGD(data):
train_data, test_data = data[:800], data[800:]
param = np.random.rand(2) # y = param[0]*x+param[1]
def gradient_function(batch_data):
x, y = batch_data
y_pred = param[0] * x + param[1]
d0 = -2.0 * x * (y - y_pred)
d1 = -2.0 * (y - y_pred)
return [d0, d1]
def get_weights_function():
return np.copy(param)
def learning_rate_function(step):
if step < 10:
return 0.1
elif step < 50:
return 0.01
else:
return 0.005
def update_weights_function(new_weight):
param[:] = new_weight
def test_function():
n = len(test_data)
x = np.array([i[0] for i in test_data])
y = np.array([i[1] for i in test_data])
y_pred = param[0] * x + param[1]
loss = 1.0 / n * np.sum((y_pred - y)**2)
check_absolute_error(loss, 0., 20.)
moment_accountant = MomentPrivacyBudgetTracker(PrivacyBudget(10, 0.001))
private_SGD(gradient_function=gradient_function,
get_weights_function=get_weights_function,
update_weights_function=update_weights_function,
learning_rate_function=learning_rate_function,
train_data=train_data,
group_size=100,
gradient_norm_bound=10,
number_of_steps=100,
sigma=1,
moment_privacy_budget_tracker=moment_accountant,
test_interval=100,
test_function=test_function)
check_absolute_error(moment_accountant.consumed_privacy_budget.epsilon,
8.805554, 1e-6)
check_absolute_error(moment_accountant.consumed_privacy_budget.delta,
0.000625, 1e-6)
def test_private_categorical_hist(example_private_table: PrivateTable):
"""check private hist implementation for categorical column of Classes in iris dataset.
bins: Iris-setosa, Iris-versicolor, Iris-virginica
"""
noisy_hist = example_private_table.cat_hist('Class', PrivacyBudget(10000.))
err = [1, 1, 1]
noisy_hist.sort()
assert all(np.abs(noisy_hist-[50, 50, 50]) < err)
del noisy_hist
def example_private_table():
data = {'Name': ['Tom', 'Jack', 'Steve', 'Jack'], 'Age': [28, 34, 29, 42]}
df = pd.DataFrame(data)
domains = {
'Name':
CategoricalDataDomain(
['Tom', 'Jack', 'Steve', 'Eve', 'Adam', 'Lucifer']),
'Age':
RealDataDomain(0., 130.)
}
return PrivateTable(df, domains, PrivacyBudget(100000.0, 1000.))
def test_column_names(example_table: DataFrame):
domains = {
'Name':
CategoricalDataDomain(
['Tom', 'Jack', 'Steve', 'Eve', 'Adam', 'Lucifer']),
'Age':
RealDataDomain(0., 130.)
}
t = PrivateTable(example_table, domains, PrivacyBudget(1.0, 0.))
assert 'Age' in t._columns
assert 'Name' in t._columns
def test_column_names(example_table: DataFrame):
"""check to ensure column names correspond to the domains"""
domains = {'Sepal Length': RealDataDomain(0., 10.),
'Sepal Width': RealDataDomain(0., 10.),
'Petal Length': RealDataDomain(0., 10.),
'Petal Width': RealDataDomain(0., 10.),
'Class': CategoricalDataDomain(['Iris-setosa', 'Iris-versicolor', 'Iris-virginica'])}
t = PrivateTable(example_table, domains, PrivacyBudget(1.0, 0.))
assert 'Sepal Length' in t._columns
assert 'Sepal Width' in t._columns
assert 'Petal Length' in t._columns
assert 'Petal Width' in t._columns
assert 'Class' in t._columns
def test_private_numerical_hist(example_private_table: PrivateTable):
"""check private hist implementation for numerical column.
bins: |.......|.......|.......|
boundaries: a0 a1 a2 a3
"""
bins: List[float] = [20, 30, 40, 50] # [a0, a1, a2, a3]
noisy_hist = example_private_table.num_hist('Age', bins,
PrivacyBudget(10000.))
err = [1, 1, 1]
noisy_hist.sort()
assert all(np.abs(noisy_hist - [1, 1, 2]) < err)
del noisy_hist, bins
def test_column_names(example_table: DataFrame):
"""check to ensure column names correspond to the domains"""
domains = {
'Age':
RealDataDomain(0., 130.),
'Education':
CategoricalDataDomain([
' Bachelors', ' HS-grad', ' 11th', ' Masters', ' 9th',
' Some-college', ' Assoc-acdm', ' Assoc-voc', ' 7th-8th',
' Doctorate', ' Prof-school', ' 5th-6th', ' 10th', ' 1st-4th',
' Preschool', ' 12th'
])
}
t = PrivateTable(example_table, domains, PrivacyBudget(1.0, 0.))
assert 'Age' in t._columns
assert 'Education' in t._columns
def example_private_table():
iris_data = pd.read_csv(os.path.join("dataset", "iris_data.txt"),
names=["Sepal Length", "Sepal Width", "Petal Length",
"Petal Width", "Class"])
data = {'Sepal Length': iris_data["Sepal Length"].tolist(),
'Sepal Width': iris_data["Sepal Width"].tolist(),
'Petal Length': iris_data["Petal Length"].tolist(),
'Petal Width': iris_data["Petal Width"].tolist(),
'Class': iris_data["Class"].tolist()}
df = pd.DataFrame(data)
domains = {'Sepal Length': RealDataDomain(0., 10.),
'Sepal Width': RealDataDomain(0., 10.),
'Petal Length': RealDataDomain(0., 10.),
'Petal Width': RealDataDomain(0., 10.),
'Class': CategoricalDataDomain(['Iris-setosa', 'Iris-versicolor', 'Iris-virginica'])}
return PrivateTable(df, domains, PrivacyBudget(100000.0, 1.))
def test_private_categorical_hist(example_private_table: PrivateTable):
"""check private hist implementation for categorical column of Education in adult dataset.
bins: HS-grad, Bachelors etc
"""
noisy_hist = example_private_table.cat_hist('Education',
PrivacyBudget(10000.))
err = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
noisy_hist.sort()
assert all(
np.abs(noisy_hist - [
51, 168, 333, 413, 433, 514, 576, 646, 933, 1067, 1175, 1382, 1723,
5355, 7291, 10501
]) < err)
del noisy_hist
def test_private_numerical_hist(example_private_table: PrivateTable):
"""check private hist implementation for numerical column of Age in adult dataset.
bins: 17, 18, 19 ... 90
"""
bins: List[float] = [int(i) for i in range(17, 91)]
noisy_hist = example_private_table.num_hist('Age', bins,
PrivacyBudget(10000.))
err = [int(1) for i in range(1, 74)]
noisy_hist.sort()
assert all(
np.abs(noisy_hist - [
1, 1, 3, 3, 6, 10, 12, 20, 22, 22, 23, 29, 43, 45, 46, 51, 64, 67,
72, 89, 108, 120, 150, 151, 178, 208, 230, 258, 300, 312, 355, 358,
366, 366, 395, 415, 419, 464, 478, 543, 550, 577, 595, 602, 708,
712, 720, 724, 734, 737, 753, 765, 770, 780, 785, 794, 798, 808,
813, 816, 827, 828, 835, 841, 858, 861, 867, 875, 876, 877, 886,
888, 898
]) < err)
del noisy_hist, bins
def update_privacy_loss(self,
privacy_budget: PrivacyBudget,
delta_prime: float,
k: int = 1):
"""Calculate and update privacy loss of multiple query with same privacy_budget.
:param privacy_budget: Privacy budget of query
:param delta_prime: Value of :math:`\epsilon'`
:param k: Number of query, defaults to 1
"""
assert delta_prime > 0, "Value of delta should be positive"
kfold_privacy_budget = PrivacyBudget(
np.sqrt(2 * k * np.log(1 / delta_prime)) * privacy_budget.epsilon +
k * privacy_budget.epsilon * (np.exp(privacy_budget.epsilon) - 1),
k * privacy_budget.delta + delta_prime)
e = self.consumed_privacy_budget + kfold_privacy_budget
assert e <= self.total_privacy_budget, "there is not enough privacy budget."
self.consumed_privacy_budget = e
def test_private_std(example_private_table: PrivateTable):
"""check private std implementation."""
noisy_std = example_private_table.std('Age', PrivacyBudget(10000.))
check_absolute_error(noisy_std, 5.54, 1.)
del noisy_std
def test_private_var(example_private_table: PrivateTable):
"""check private var implementation using Age in adult dataset."""
noisy_var = example_private_table.var('Age', PrivacyBudget(10000.))
check_absolute_error(noisy_var, 186.06140024879625, 2.)
del noisy_var
def test_private_std(example_private_table: PrivateTable):
"""check private std implementation using Age in adult dataset."""
noisy_std = example_private_table.std('Age', PrivacyBudget(10000.))
check_absolute_error(noisy_std, 13.640432553581146, 1.)
del noisy_std
def test_private_mean(example_private_table: PrivateTable):
"""check private mean implementation."""
noisy_mean = example_private_table.mean('Age', PrivacyBudget(10000.))
check_absolute_error(noisy_mean, 33.2, 1.)
def test_privacy_budget_class():
e1 = PrivacyBudget(1., 0.01)
e2 = PrivacyBudget(0.2, 0.004)
e3 = PrivacyBudget(1 + 0.2, 0.01 + 0.004)
assert e3 == e1 + e2
def test_private_mode(example_private_table: PrivateTable):
"""check private mode implementation using Education in adult dataset."""
noisy_mode = example_private_table.mode('Education', PrivacyBudget(10000.))
assert noisy_mode == " HS-grad"
del noisy_mode
def test_private_min(example_private_table: PrivateTable):
"""check private min implementation using Age in adult dataset."""
noisy_min = example_private_table.min('Age', PrivacyBudget(10000.))
check_absolute_error(noisy_min, 17., 1.)
del noisy_min
def test_private_mode(example_private_table: PrivateTable):
"""check private mode implementation."""
noisy_mode = example_private_table.mode('Name', PrivacyBudget(10000.))
assert noisy_mode == "Jack"
del noisy_mode
def test_private_min(example_private_table: PrivateTable):
"""check private min implementation."""
noisy_min = example_private_table.min('Age', PrivacyBudget(10000.))
check_absolute_error(noisy_min, 28., 1.)
del noisy_min
def test_combine_privacy_losses():
e1 = PrivacyBudget(1., 0.01)
e2 = PrivacyBudget(0.2, 0.004)
e3 = combine_privacy_losses([e1, e2])
expected_e3 = PrivacyBudget(1. + 0.2, 0.01 + 0.004)
assert e3 == expected_e3
def test_private_var(example_private_table: PrivateTable):
"""check private var implementation."""
noisy_var = example_private_table.var('Age', PrivacyBudget(10000.))
check_absolute_error(noisy_var, 30.69, 2.)
del noisy_var
def test_private_gaussian_mean(example_private_table: PrivateTable):
"""check private guassian mean implementation using Age in adult dataset."""
noisy_mean = example_private_table.gaussian_mean('Age',
PrivacyBudget(0.99, 0.5))
check_absolute_error(noisy_mean, 38.58164675532078, 1.)
def test_private_mean(example_private_table: PrivateTable):
"""check private mean implementation using Age in adult dataset."""
noisy_mean = example_private_table.mean('Age', PrivacyBudget(10000.))
check_absolute_error(noisy_mean, 38.58164675532078, 1.)
def test_private_max(example_private_table: PrivateTable):
"""check private max implementation."""
noisy_max = example_private_table.max('Age', PrivacyBudget(10000.))
check_absolute_error(noisy_max, 42., 1.)
del noisy_max