def setup_method(self, method):
if method.__name__.endswith("prob"):
global_seed(314159)
self.mech = Bingham
a = np.random.random((5, 3))
self.random_array = a.T.dot(a)
def test_different_results(self):
from sklearn.naive_bayes import GaussianNB as sk_nb
from sklearn import datasets
global_seed(12345)
dataset = datasets.load_iris()
x_train, x_test, y_train, y_test = train_test_split(dataset.data,
dataset.target,
test_size=.2)
bounds = ([4.3, 2.0, 1.0, 0.1], [7.9, 4.4, 6.9, 2.5])
clf_dp = GaussianNB(epsilon=1.0, bounds=bounds)
clf_non_private = sk_nb()
for clf in [clf_dp, clf_non_private]:
clf.fit(x_train, y_train)
# Todo: remove try...except when sklearn v1.0 is required
try:
nonprivate_var = clf_non_private.var_
except AttributeError:
nonprivate_var = clf_non_private.sigma_
theta_diff = (clf_dp.theta_ - clf_non_private.theta_)**2
self.assertGreater(theta_diff.sum(), 0)
var_diff = (clf_dp.var_ - nonprivate_var)**2
self.assertGreater(var_diff.sum(), 0)
def test_large_epsilon(self):
global_seed(12345)
a = np.random.random(1000)
res = float(np.std(a))
res_dp = std(a, epsilon=1, range=1)
self.assertAlmostEqual(res, res_dp, delta=0.01)
def test_large_epsilon(self):
global_seed(12345)
a = np.random.random(1000)
res = float(np.std(a))
res_dp = nanstd(a, epsilon=5, bounds=(0, 1))
self.assertAlmostEqual(res, res_dp, delta=0.01)
def test_density_2d(self):
global_seed(3141592653)
a = np.array([[1, 2, 3, 4, 5], [6, 7, 8, 9, 10]]).T
dp_hist, _ = histogramdd(a, epsilon=1, bins=3, range=[(0, 10), (0, 10)], density=True)
# print(dp_hist.sum())
self.assertAlmostEqual(dp_hist.sum(), 1.0 * (3 / 10) ** 2)
def test_large_epsilon_axis(self):
global_seed(12345)
a = np.random.random((1000, 5))
res = np.std(a, axis=0)
res_dp = nanstd(a, epsilon=15, bounds=(0, 1), axis=0)
for i in range(res.shape[0]):
self.assertAlmostEqual(res[i], res_dp[i], delta=0.01)
def test_different_result(self):
global_seed(3141592653)
a = np.array([1, 2, 3, 4, 5])
hist, _ = np.histogram(a, bins=3, range=(0, 10))
dp_hist, _ = histogram(a, epsilon=0.1, bins=3, range=(0, 10))
# print("Non-private histogram: %s" % hist)
# print("Private histogram: %s" % dp_hist)
self.assertTrue((hist != dp_hist).any())
def test_density(self):
global_seed(3141592653)
a = np.array([1, 2, 3, 4, 5])
dp_hist, _ = histogram(a,
epsilon=10,
bins=3,
range=(0, 10),
density=True)
self.assertAlmostEqual(dp_hist.sum(), 3 / 10)
def test_inf_epsilon(self):
global_seed(3141592653)
clf = KMeans(float("inf"), (0, 1), 3)
X = np.array([0.1, 0.1, 0.1, 0.1, 0.5, 0.5, 0.5, 0.5, 0.9, 0.9, 0.9]).reshape(-1, 1)
clf.fit(X)
centers = clf.cluster_centers_
self.assertIn(0.1, centers)
self.assertIn(0.5, centers)
self.assertIn(0.9, centers)
def setup_method(self, method):
if method.__name__.endswith("prob"):
global_seed(314159)
self.mech = Wishart
# with warnings.catch_warnings(record=True) as w:
# warnings.filterwarnings('always', '', DeprecationWarning)
# self.mech = Wishart()
# self.assertIs(w[0].category, DeprecationWarning)
self.random_array = np.random.randn(5, 5)
def test_density_1d(self):
global_seed(3141592653)
a = np.array([1, 2, 3, 4, 5])
dp_hist, _ = histogramdd(a,
epsilon=10,
bins=3,
range=[(0, 10)],
density=True)
# print(dp_hist.sum())
self.assertAlmostEqual(dp_hist.sum(), 1.0 * 3 / 10)
def test_different_result(self):
global_seed(3141592653)
x = np.array([1, 2, 3, 4, 5])
y = np.array([5, 7, 1, 5, 9])
hist, _, _ = np.histogram2d(x, y, bins=3, range=[(0, 10), (0, 10)])
dp_hist, _, _ = histogram2d(x,
y,
epsilon=0.1,
bins=3,
range=[(0, 10), (0, 10)])
# print("Non-private histogram: %s" % hist)
# print("Private histogram: %s" % dp_hist)
self.assertTrue((hist != dp_hist).any())
def test_density(self):
global_seed(3141592653)
x = np.array([1, 2, 3, 4, 5])
y = np.array([5, 7, 1, 5, 9])
dp_hist, _, _ = histogram2d(x,
y,
epsilon=1,
bins=3,
range=[(0, 10), (0, 10)],
density=True)
# print(dp_hist.sum())
self.assertAlmostEqual(dp_hist.sum(), 1.0 * (3 / 10)**2)
def test_simple(self):
global_seed(3141592653)
clf = KMeans(epsilon=10, bounds=(0, 1), n_clusters=3)
X = np.zeros(6000) + 0.1
X[:4000] = 0.5
X[:2000] = 0.9
X = X.reshape(-1, 1)
clf.fit(X)
centers = clf.cluster_centers_
self.assertAlmostEqual(np.min(centers), 0.1, delta=0.1)
self.assertAlmostEqual(np.median(centers), 0.5, delta=0.1)
self.assertAlmostEqual(np.max(centers), 0.9, delta=0.1)
def test_simple(self):
global_seed(3141592653)
clf = KMeans(5, (0, 1), 3)
X = np.zeros(1000) + 0.1
X[:666] = 0.5
X[:333] = 0.9
X = X.reshape(-1, 1)
clf.fit(X)
centers = clf.cluster_centers_
self.assertTrue(np.isclose(centers, 0.1, atol=0.05).any())
self.assertTrue(np.isclose(centers, 0.5, atol=0.05).any())
self.assertTrue(np.isclose(centers, 0.9, atol=0.05).any())
def test_with_iris(self):
global_seed(12345)
from sklearn import datasets
dataset = datasets.load_iris()
x_train, x_test, y_train, y_test = train_test_split(dataset.data, dataset.target, test_size=.2)
bounds = [(4.3, 7.9), (2.0, 4.4), (1.0, 6.9), (0.1, 2.5)]
clf = GaussianNB(epsilon=1.0, bounds=bounds)
clf.fit(x_train, y_train)
accuracy = sum(clf.predict(x_test) == y_test) / y_test.shape[0]
# print(accuracy)
self.assertGreater(accuracy, 0.5)
def test_similar_results(self):
global_seed(314159)
X = np.random.rand(100000, 5)
dp_ss = StandardScaler(bounds=(0, 1), epsilon=float("inf"))
dp_ss.fit(X)
sk_ss = sk_pp.StandardScaler()
sk_ss.fit(X)
self.assertTrue(
np.allclose(dp_ss.mean_, sk_ss.mean_, rtol=1, atol=1e-4),
"Arrays %s and %s should be close" % (dp_ss.mean_, sk_ss.mean_))
self.assertTrue(
np.allclose(dp_ss.var_, sk_ss.var_, rtol=1, atol=1e-4),
"Arrays %s and %s should be close" % (dp_ss.var_, sk_ss.var_))
self.assertTrue(np.all(dp_ss.n_samples_seen_ == sk_ss.n_samples_seen_))
def test_with_iris(self):
global_seed(12345)
from sklearn import datasets
dataset = datasets.load_iris()
x_train, x_test, y_train, y_test = train_test_split(dataset.data, dataset.target, test_size=.2)
bounds = ([4.3, 2.0, 1.0, 0.1], [7.9, 4.4, 6.9, 2.5])
clf = GaussianNB(epsilon=5.0, bounds=bounds)
clf.fit(x_train, y_train)
accuracy = clf.score(x_test, y_test)
counts = clf.class_count_.copy()
self.assertGreater(accuracy, 0.45)
clf.partial_fit(x_train, y_train)
new_counts = clf.class_count_
self.assertEqual(np.sum(new_counts), np.sum(counts) * 2)
def test_different_results(self):
from sklearn.naive_bayes import GaussianNB as sk_nb
from sklearn import datasets
global_seed(12345)
dataset = datasets.load_iris()
x_train, x_test, y_train, y_test = train_test_split(dataset.data, dataset.target, test_size=.2)
bounds = [(4.3, 7.9), (2.0, 4.4), (1.0, 6.9), (0.1, 2.5)]
clf_dp = GaussianNB(epsilon=1.0, bounds=bounds)
clf_non_private = sk_nb()
for clf in [clf_dp, clf_non_private]:
clf.fit(x_train, y_train)
same_prediction = clf_dp.predict(x_test) == clf_non_private.predict(x_test)
self.assertFalse(np.all(same_prediction))
def test_different_results(self):
from sklearn.naive_bayes import GaussianNB as sk_nb
from sklearn import datasets
global_seed(12345)
dataset = datasets.load_iris()
x_train, x_test, y_train, y_test = train_test_split(dataset.data, dataset.target, test_size=.2)
bounds = ([4.3, 2.0, 1.0, 0.1], [7.9, 4.4, 6.9, 2.5])
clf_dp = GaussianNB(epsilon=1.0, bounds=bounds)
clf_non_private = sk_nb()
for clf in [clf_dp, clf_non_private]:
clf.fit(x_train, y_train)
theta_diff = (clf_dp.theta_ - clf_non_private.theta_) ** 2
self.assertGreater(theta_diff.sum(), 0)
sigma_diff = (clf_dp.sigma_ - clf_non_private.sigma_) ** 2
self.assertGreater(sigma_diff.sum(), 0)
def setUp(self):
global_seed(2718281828)
def setup_method(self, method):
if method.__name__.endswith("prob"):
global_seed(314159)
self.mech = GaussianAnalytic
def setup_method(self, method):
if method.__name__.endswith("prob"):
global_seed(314159)
self.mech = Wishart()
self.random_array = np.random.randn(5, 5)
def setup_method(self, method):
global_seed(3141592653)
def setup_method(self, method):
if method.__name__ .endswith("prob"):
global_seed(314159)
self.mech = Exponential()
def setup_method(self, method):
if method.__name__.endswith("prob"):
global_seed(314159)
self.mech = Laplace()
def setup_method(self, method):
if method.__name__.endswith("prob"):
global_seed(314159)
self.mech = GeometricFolded()
import numpy as np
from unittest import TestCase
from diffprivlib.mechanisms import Geometric
from diffprivlib.mechanisms.transforms import StringToInt
from diffprivlib.utils import global_seed
global_seed(3141592653)
class TestStringToInt(TestCase):
def test_not_none(self):
mech = StringToInt(Geometric())
self.assertIsNotNone(mech)
_mech = mech.copy()
self.assertIsNotNone(_mech)
def test_class(self):
from diffprivlib.mechanisms import DPMachine
from diffprivlib.mechanisms.transforms import DPTransformer
self.assertTrue(issubclass(StringToInt, DPMachine))
self.assertTrue(issubclass(StringToInt, DPTransformer))
def test_no_parent(self):
with self.assertRaises(TypeError):
StringToInt()
def test_randomise(self):
mech = StringToInt(Geometric().set_sensitivity(1).set_epsilon(1))
self.assertIsInstance(mech.randomise("1"), str)
def setup_method(self, method):
if method.__name__.endswith("prob"):
global_seed(314159)
self.mech = GeneralRandomizedBinaryResponse()