def test_anonymize_count_with_budget_multiple(self):
expected_values = np.array([87.5864551, 437.701297])
epsilon = 1.0
self.set_seed()
anonymized = DiffPrivLaplaceMechanism.anonymize_count_with_budget(
[87.0, 435.0], epsilon)
np.testing.assert_almost_equal(anonymized, expected_values)
def test_anonymize_count_with_budget_single(self):
expected_value = 87.58645513850368
epsilon = 1.0
self.set_seed()
anonymized = DiffPrivLaplaceMechanism.anonymize_count_with_budget(
87.0, epsilon)
np.testing.assert_almost_equal(anonymized, expected_value)
def test_anonymize_count_with_budget_single_many(self):
expected_values = np.array([87.5864551, 89.701297, 86.4519884])
epsilon = 1.0
self.set_seed()
anonymized = DiffPrivLaplaceMechanism.anonymize_count_with_budget(
87.0, epsilon, size=3)
np.testing.assert_almost_equal(anonymized, expected_values)
def count(cls, data, epsilon, condition=None, axis=None, postprocess=True):
"""
Performs the count operation and anonymizes the value(s) using the provided
privacy budget.
The operation counts non-zero or non-False values. When a condition
function is provided, it is invoked so that the data can be transformed
before counting.
Parameters
----------
data : list|ndarray
The data to retrieve the count(s) from.
epsilon : float
The privacy budget.
[condition] : function
A condition function that receives the data as a parameter and returns
a list or ndarray of the same size in which for each element is either
a zero or False when it should not be counted and non-zero or True when
it should.
[axis] : int|tuple
Axis or tuple of axes along which to count non-zeros or non-False value(s).
[postprocess] : bool
Indicates whether or not to post-process the count values so that the
returned value is rounded and within the [0, n] range, where n is the total
number of observations.
Returns
-------
float|ndarray
The anonymized count(s).
"""
if condition:
data = condition(data)
value = np.count_nonzero(data, axis=axis)
anonymized = DiffPrivLaplaceMechanism.anonymize_count_with_budget(
value, epsilon)
if postprocess:
n = np.size(data, axis=axis)
anonymized = np.clip(anonymized, 0.0, n)
anonymized = np.round(anonymized)
return anonymized