mirror of
https://github.com/IBM/ai-privacy-toolkit.git
synced 2026-04-26 05:16:22 +02:00
Support for one-hot encoded features in minimization (#87)
* Initial version with first working test * Make sure representative values in generalizations for 1-hot encoded features are consistent. * Updated notebooks for one-hot encoded data * Review comments Signed-off-by: abigailt <abigailt@il.ibm.com>
This commit is contained in:
abigailgold
GitHub
parent
5dce961092
commit
6d81cd8ed4
4 changed files with 26703 additions and 48 deletions
|
|
@ -181,8 +181,8 @@ def compare_generalizations(gener, expected_generalizations):
|
|||
== set(gener['range_representatives'][key]))
|
||||
if 'category_representatives' in expected_generalizations:
|
||||
for key in expected_generalizations['category_representatives']:
|
||||
assert (set([frozenset(sl) for sl in expected_generalizations['category_representatives'][key]])
|
||||
== set([frozenset(sl) for sl in gener['category_representatives'][key]]))
|
||||
assert (set(expected_generalizations['category_representatives'][key])
|
||||
== set(gener['category_representatives'][key]))
|
||||
|
||||
|
||||
def check_features(features, expected_generalizations, transformed, x, pandas=False):
|
||||
|
|
@ -200,9 +200,9 @@ def check_features(features, expected_generalizations, transformed, x, pandas=Fa
|
|||
if features[i] in modified_features:
|
||||
indexes.append(i)
|
||||
if len(indexes) != transformed.shape[1]:
|
||||
assert ((np.delete(transformed, indexes, axis=1) == np.delete(x, indexes, axis=1)).all())
|
||||
assert (np.array_equal(np.delete(transformed, indexes, axis=1), np.delete(x, indexes, axis=1)))
|
||||
if len(expected_generalizations['ranges'].keys()) > 0 or len(expected_generalizations['categories'].keys()) > 0:
|
||||
assert (((transformed[indexes]) != (x[indexes])).any())
|
||||
assert (not np.array_equal(transformed[:, indexes], x[:, indexes]))
|
||||
|
||||
|
||||
def check_ncp(ncp, expected_generalizations):
|
||||
|
|
@ -920,6 +920,233 @@ def test_BaseEstimator_regression(diabetes_dataset):
|
|||
assert ((rel_accuracy >= target_accuracy) or (target_accuracy - rel_accuracy) <= ACCURACY_DIFF)
|
||||
|
||||
|
||||
def test_minimizer_ndarray_one_hot():
|
||||
x_train = np.array([[23, 0, 1, 165],
|
||||
[45, 0, 1, 158],
|
||||
[56, 1, 0, 123],
|
||||
[67, 0, 1, 154],
|
||||
[45, 1, 0, 149],
|
||||
[42, 1, 0, 166],
|
||||
[73, 0, 1, 172],
|
||||
[94, 0, 1, 168],
|
||||
[69, 0, 1, 175],
|
||||
[24, 1, 0, 181],
|
||||
[18, 1, 0, 190]])
|
||||
y_train = np.array([1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 0])
|
||||
|
||||
model = DecisionTreeClassifier()
|
||||
model.fit(x_train, y_train)
|
||||
predictions = model.predict(x_train)
|
||||
|
||||
features = ['0', '1', '2', '3']
|
||||
QI = [0, 1, 2]
|
||||
QI_slices = [[1, 2]]
|
||||
target_accuracy = 0.7
|
||||
gen = GeneralizeToRepresentative(model, target_accuracy=target_accuracy, feature_slices=QI_slices,
|
||||
features_to_minimize=QI)
|
||||
gen.fit(dataset=ArrayDataset(x_train, predictions))
|
||||
transformed = gen.transform(dataset=ArrayDataset(x_train))
|
||||
gener = gen.generalizations
|
||||
expected_generalizations = {'categories': {}, 'category_representatives': {},
|
||||
'range_representatives': {'0': [34.5]},
|
||||
'ranges': {'0': [34.5]}, 'untouched': ['3', '1', '2']}
|
||||
|
||||
compare_generalizations(gener, expected_generalizations)
|
||||
|
||||
check_features(features, expected_generalizations, transformed, x_train)
|
||||
ncp = gen.ncp.transform_score
|
||||
check_ncp(ncp, expected_generalizations)
|
||||
|
||||
rel_accuracy = model.score(transformed, predictions)
|
||||
assert ((rel_accuracy >= target_accuracy) or (target_accuracy - rel_accuracy) <= ACCURACY_DIFF)
|
||||
transformed_slice = transformed[:, QI_slices[0]]
|
||||
assert ((np.sum(transformed_slice, axis=1) == 1).all())
|
||||
assert ((np.max(transformed_slice, axis=1) == 1).all())
|
||||
assert ((np.min(transformed_slice, axis=1) == 0).all())
|
||||
|
||||
|
||||
def test_minimizer_ndarray_one_hot_gen():
|
||||
x_train = np.array([[23, 0, 1, 165],
|
||||
[45, 0, 1, 158],
|
||||
[56, 1, 0, 123],
|
||||
[67, 0, 1, 154],
|
||||
[45, 1, 0, 149],
|
||||
[42, 1, 0, 166],
|
||||
[73, 0, 1, 172],
|
||||
[94, 0, 1, 168],
|
||||
[69, 0, 1, 175],
|
||||
[24, 1, 0, 181],
|
||||
[18, 1, 0, 190]])
|
||||
y_train = np.array([1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 0])
|
||||
|
||||
model = DecisionTreeClassifier()
|
||||
model.fit(x_train, y_train)
|
||||
predictions = model.predict(x_train)
|
||||
|
||||
features = ['0', '1', '2', '3']
|
||||
QI = [0, 1, 2]
|
||||
QI_slices = [[1, 2]]
|
||||
target_accuracy = 0.2
|
||||
gen = GeneralizeToRepresentative(model, target_accuracy=target_accuracy, feature_slices=QI_slices,
|
||||
features_to_minimize=QI)
|
||||
gen.fit(dataset=ArrayDataset(x_train, predictions))
|
||||
transformed = gen.transform(dataset=ArrayDataset(x_train))
|
||||
gener = gen.generalizations
|
||||
expected_generalizations = {'categories': {'1': [[0, 1]], '2': [[0, 1]]},
|
||||
'category_representatives': {'1': [0], '2': [1]},
|
||||
'range_representatives': {'0': []}, 'ranges': {'0': []}, 'untouched': ['3']}
|
||||
|
||||
compare_generalizations(gener, expected_generalizations)
|
||||
|
||||
check_features(features, expected_generalizations, transformed, x_train)
|
||||
ncp = gen.ncp.transform_score
|
||||
check_ncp(ncp, expected_generalizations)
|
||||
|
||||
rel_accuracy = model.score(transformed, predictions)
|
||||
assert ((rel_accuracy >= target_accuracy) or (target_accuracy - rel_accuracy) <= ACCURACY_DIFF)
|
||||
transformed_slice = transformed[:, QI_slices[0]]
|
||||
assert ((np.sum(transformed_slice, axis=1) == 1).all())
|
||||
assert ((np.max(transformed_slice, axis=1) == 1).all())
|
||||
assert ((np.min(transformed_slice, axis=1) == 0).all())
|
||||
|
||||
|
||||
def test_minimizer_ndarray_one_hot_multi():
|
||||
x_train = np.array([[23, 0, 1, 0, 0, 1, 165],
|
||||
[45, 0, 1, 0, 0, 1, 158],
|
||||
[56, 1, 0, 0, 0, 1, 123],
|
||||
[67, 0, 1, 1, 0, 0, 154],
|
||||
[45, 1, 0, 1, 0, 0, 149],
|
||||
[42, 1, 0, 1, 0, 0, 166],
|
||||
[73, 0, 1, 0, 0, 1, 172],
|
||||
[94, 0, 1, 0, 1, 0, 168],
|
||||
[69, 0, 1, 0, 1, 0, 175],
|
||||
[24, 1, 0, 0, 1, 0, 181],
|
||||
[18, 1, 0, 0, 0, 1, 190]])
|
||||
y_train = np.array([1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 0])
|
||||
|
||||
model = DecisionTreeClassifier()
|
||||
model.fit(x_train, y_train)
|
||||
predictions = model.predict(x_train)
|
||||
|
||||
features = ['0', '1', '2', '3', '4', '5', '6']
|
||||
QI = [0, 1, 2, 3, 4, 5]
|
||||
QI_slices = [[1, 2], [3, 4, 5]]
|
||||
target_accuracy = 0.2
|
||||
gen = GeneralizeToRepresentative(model, target_accuracy=target_accuracy, feature_slices=QI_slices,
|
||||
features_to_minimize=QI)
|
||||
gen.fit(dataset=ArrayDataset(x_train, predictions))
|
||||
transformed = gen.transform(dataset=ArrayDataset(x_train))
|
||||
gener = gen.generalizations
|
||||
expected_generalizations = {'categories':
|
||||
{'1': [[0, 1]], '2': [[0, 1]], '3': [[0, 1]], '4': [[0, 1]], '5': [[0, 1]]},
|
||||
'category_representatives': {'1': [0], '2': [1], '3': [0], '4': [1], '5': [0]},
|
||||
'range_representatives': {'0': []}, 'ranges': {'0': []}, 'untouched': ['6']}
|
||||
|
||||
compare_generalizations(gener, expected_generalizations)
|
||||
|
||||
check_features(features, expected_generalizations, transformed, x_train)
|
||||
ncp = gen.ncp.transform_score
|
||||
check_ncp(ncp, expected_generalizations)
|
||||
|
||||
rel_accuracy = model.score(transformed, predictions)
|
||||
assert ((rel_accuracy >= target_accuracy) or (target_accuracy - rel_accuracy) <= ACCURACY_DIFF)
|
||||
transformed_slice = transformed[:, QI_slices[0]]
|
||||
assert ((np.sum(transformed_slice, axis=1) == 1).all())
|
||||
assert ((np.max(transformed_slice, axis=1) == 1).all())
|
||||
assert ((np.min(transformed_slice, axis=1) == 0).all())
|
||||
transformed_slice = transformed[:, QI_slices[1]]
|
||||
assert ((np.sum(transformed_slice, axis=1) == 1).all())
|
||||
assert ((np.max(transformed_slice, axis=1) == 1).all())
|
||||
assert ((np.min(transformed_slice, axis=1) == 0).all())
|
||||
|
||||
|
||||
def test_minimizer_ndarray_one_hot_multi2():
|
||||
x_train = np.array([[0, 0, 1],
|
||||
[0, 0, 1],
|
||||
[0, 1, 0],
|
||||
[0, 1, 0],
|
||||
[1, 0, 0],
|
||||
[1, 0, 0]])
|
||||
y_train = np.array([1, 1, 2, 2, 0, 0])
|
||||
|
||||
model = DecisionTreeClassifier()
|
||||
model.fit(x_train, y_train)
|
||||
predictions = model.predict(x_train)
|
||||
|
||||
features = ['0', '1', '2']
|
||||
QI = [0, 1, 2]
|
||||
QI_slices = [[0, 1, 2]]
|
||||
target_accuracy = 0.2
|
||||
gen = GeneralizeToRepresentative(model, target_accuracy=target_accuracy, feature_slices=QI_slices,
|
||||
features_to_minimize=QI)
|
||||
gen.fit(dataset=ArrayDataset(x_train, predictions))
|
||||
transformed = gen.transform(dataset=ArrayDataset(x_train))
|
||||
gener = gen.generalizations
|
||||
expected_generalizations = {'categories': {'0': [[0, 1]], '1': [[0, 1]], '2': [[0, 1]]},
|
||||
'category_representatives': {'0': [0], '1': [0], '2': [1]}, 'range_representatives': {},
|
||||
'ranges': {}, 'untouched': []}
|
||||
|
||||
compare_generalizations(gener, expected_generalizations)
|
||||
|
||||
check_features(features, expected_generalizations, transformed, x_train)
|
||||
ncp = gen.ncp.transform_score
|
||||
check_ncp(ncp, expected_generalizations)
|
||||
|
||||
rel_accuracy = model.score(transformed, predictions)
|
||||
assert ((rel_accuracy >= target_accuracy) or (target_accuracy - rel_accuracy) <= ACCURACY_DIFF)
|
||||
transformed_slice = transformed[:, QI_slices[0]]
|
||||
assert ((np.sum(transformed_slice, axis=1) == 1).all())
|
||||
assert ((np.max(transformed_slice, axis=1) == 1).all())
|
||||
assert ((np.min(transformed_slice, axis=1) == 0).all())
|
||||
|
||||
|
||||
def test_anonymize_pandas_one_hot():
|
||||
features = ["age", "gender_M", "gender_F", "height"]
|
||||
x_train = np.array([[23, 0, 1, 165],
|
||||
[45, 0, 1, 158],
|
||||
[56, 1, 0, 123],
|
||||
[67, 0, 1, 154],
|
||||
[45, 1, 0, 149],
|
||||
[42, 1, 0, 166],
|
||||
[73, 0, 1, 172],
|
||||
[94, 0, 1, 168],
|
||||
[69, 0, 1, 175],
|
||||
[24, 1, 0, 181],
|
||||
[18, 1, 0, 190]])
|
||||
y_train = np.array([1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 0])
|
||||
x_train = pd.DataFrame(x_train, columns=features)
|
||||
y_train = pd.Series(y_train)
|
||||
|
||||
model = DecisionTreeClassifier()
|
||||
model.fit(x_train, y_train)
|
||||
predictions = model.predict(x_train)
|
||||
|
||||
QI = ["age", "gender_M", "gender_F"]
|
||||
QI_slices = [["gender_M", "gender_F"]]
|
||||
target_accuracy = 0.7
|
||||
gen = GeneralizeToRepresentative(model, target_accuracy=target_accuracy, feature_slices=QI_slices,
|
||||
features_to_minimize=QI)
|
||||
gen.fit(dataset=ArrayDataset(x_train, predictions))
|
||||
transformed = gen.transform(dataset=ArrayDataset(x_train))
|
||||
gener = gen.generalizations
|
||||
expected_generalizations = {'categories': {}, 'category_representatives': {},
|
||||
'range_representatives': {'age': [34.5]},
|
||||
'ranges': {'age': [34.5]}, 'untouched': ['height', 'gender_M', 'gender_F']}
|
||||
|
||||
compare_generalizations(gener, expected_generalizations)
|
||||
|
||||
check_features(features, expected_generalizations, transformed, x_train, True)
|
||||
ncp = gen.ncp.transform_score
|
||||
check_ncp(ncp, expected_generalizations)
|
||||
|
||||
rel_accuracy = model.score(transformed, predictions)
|
||||
assert ((rel_accuracy >= target_accuracy) or (target_accuracy - rel_accuracy) <= ACCURACY_DIFF)
|
||||
transformed_slice = transformed.loc[:, QI_slices[0]]
|
||||
assert ((np.sum(transformed_slice, axis=1) == 1).all())
|
||||
assert ((np.max(transformed_slice, axis=1) == 1).all())
|
||||
assert ((np.min(transformed_slice, axis=1) == 0).all())
|
||||
|
||||
|
||||
def test_keras_model():
|
||||
(x, y), (x_test, y_test) = get_iris_dataset_np()
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue