Bugfixes:

1. Child classes of 'TrainedAttacker' now have a 'backend' parameter, so require an __init__() method.

PiperOrigin-RevId: 451005298

tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/models.py

@@ -127,18 +127,18 @@ def _column_stack(logits, loss):
   return np.column_stack((logits, loss))
 
 
-class TrainedAttacker:
-    """Base class for training attack models.
+class TrainedAttacker(object):
+  """Base class for training attack models.
 
-    Attributes:
-      backend: Name of Scikit-Learn parallel backend to use for this attack
-        model. The default value of `None` performs single-threaded training.
-      model: The trained attack model.
-      ctx_mgr: The backend context manager within which to perform training.
-        Defaults to the null context manager for single-threaded training.
-      n_jobs: Number of jobs that can run in parallel when using a backend.
-        Set to `1` for single-threading, and to `-1` for all parallel
-        backends.
+  Attributes:
+    backend: Name of Scikit-Learn parallel backend to use for this attack
+      model. The default value of `None` performs single-threaded training.
+    model: The trained attack model.
+    ctx_mgr: The backend context manager within which to perform training.
+      Defaults to the null context manager for single-threaded training.
+    n_jobs: Number of jobs that can run in parallel when using a backend.
+      Set to `1` for single-threading, and to `-1` for all parallel
+      backends.
     """
 
   def __init__(self, backend: Optional[str] = None):
@@ -148,6 +148,7 @@ class TrainedAttacker:
       # Default value of `None` will perform single-threaded training.
       self.ctx_mgr = contextlib.nullcontext()
       self.n_jobs = 1
+      logging.info('Using single-threaded backend for training.')
     else:
       self.n_jobs = -1
       self.ctx_mgr = parallel_backend(
@@ -189,6 +190,9 @@ class TrainedAttacker:
 class LogisticRegressionAttacker(TrainedAttacker):
   """Logistic regression attacker."""
 
+  def __init__(self, backend: Optional[str] = None):
+    super().__init__(backend=backend)
+
   def train_model(self, input_features, is_training_labels):
     with self.ctx_mgr:
       lr = linear_model.LogisticRegression(solver='lbfgs', n_jobs=self.n_jobs)
@@ -204,6 +208,9 @@ class LogisticRegressionAttacker(TrainedAttacker):
 class MultilayerPerceptronAttacker(TrainedAttacker):
   """Multilayer perceptron attacker."""
 
+  def __init__(self, backend: Optional[str] = None):
+    super().__init__(backend=backend)
+
   def train_model(self, input_features, is_training_labels):
     with self.ctx_mgr:
       mlp_model = neural_network.MLPClassifier()
@@ -221,6 +228,9 @@ class MultilayerPerceptronAttacker(TrainedAttacker):
 class RandomForestAttacker(TrainedAttacker):
   """Random forest attacker."""
 
+  def __init__(self, backend: Optional[str] = None):
+    super().__init__(backend=backend)
+
   def train_model(self, input_features, is_training_labels):
     """Setup a random forest pipeline with cross-validation."""
     with self.ctx_mgr:
@@ -242,6 +252,9 @@ class RandomForestAttacker(TrainedAttacker):
 class KNearestNeighborsAttacker(TrainedAttacker):
   """K nearest neighbor attacker."""
 
+  def __init__(self, backend: Optional[str] = None):
+    super().__init__(backend=backend)
+
   def train_model(self, input_features, is_training_labels):
     with self.ctx_mgr:
       knn_model = neighbors.KNeighborsClassifier(n_jobs=self.n_jobs)

tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/models_test.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 from absl.testing import absltest
+from absl.testing import parameterized
 import numpy as np
 
 from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack import models
@@ -20,7 +21,7 @@ from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack.data_s
 from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack.data_structures import AttackType
 
 
-class TrainedAttackerTest(absltest.TestCase):
+class TrainedAttackerTest(parameterized.TestCase):
 
   def test_base_attacker_train_and_predict(self):
     base_attacker = models.TrainedAttacker()
@@ -90,15 +91,19 @@ class TrainedAttackerTest(absltest.TestCase):
     self.assertLen(attacker_data.fold_indices, 6)
     self.assertEmpty(attacker_data.left_out_indices)
 
-  def test_training_with_threading_backend(self):
+  # Parameters for testing: backend.
+  @parameterized.named_parameters(
+      ('threading_backend', 'threading'),
+      ('None_backend', None),
+  )
+  def test_training_with_backends(self, backend):
     with self.assertLogs(level='INFO') as log:
-      attacker = models.create_attacker(AttackType.LOGISTIC_REGRESSION,
-                                        'threading')
-    self.assertIsInstance(attacker, models.LogisticRegressionAttacker)
+      attacker = models.create_attacker(
+          AttackType.MULTI_LAYERED_PERCEPTRON, backend=backend)
+    self.assertIsInstance(attacker, models.MultilayerPerceptronAttacker)
     self.assertLen(log.output, 1)
     self.assertLen(log.records, 1)
     self.assertRegex(log.output[0], r'.+?Using .+? backend for training.')
 
-
 if __name__ == '__main__':
   absltest.main()