Add group-wise inference risks

src/anonymeter/evaluators/inference_evaluator.py

@@ -14,16 +14,16 @@
 
 
 def _run_attack(
-        target: pd.DataFrame,
-        syn: pd.DataFrame,
-        n_attacks: int,
-        aux_cols: list[str],
-        secret: str,
-        n_jobs: int,
-        naive: bool,
-        regression: Optional[bool],
-        inference_model: Optional[InferencePredictor],
-) -> int:
+    target: pd.DataFrame,
+    syn: pd.DataFrame,
+    n_attacks: int,
+    aux_cols: list[str],
+    secret: str,
+    n_jobs: int,
+    naive: bool,
+    regression: Optional[bool],
+    inference_model: Optional[InferencePredictor],
+) -> tuple[int, pd.Series]:
     if regression is None:
         regression = pd.api.types.is_numeric_dtype(target[secret])
 
@@ -35,9 +35,11 @@ def _run_attack(
         # Instantiate the default KNN model if no other model is passed through `inference_model`.
         if inference_model is None:
             inference_model = KNNInferencePredictor(data=syn, columns=aux_cols, target_col=secret, n_jobs=n_jobs)
+
         guesses = inference_model.predict(targets)
+    guesses = guesses.reindex_like(targets)
 
-    return evaluate_inference_guesses(guesses=guesses, secrets=targets[secret], regression=regression).sum()
+    return evaluate_inference_guesses(guesses=guesses, secrets=targets[secret], regression=regression).sum(), guesses
 
 
 def evaluate_inference_guesses(
@@ -72,6 +74,9 @@ def evaluate_inference_guesses(
         Array of boolean values indicating the correcteness of each guess.
 
     """
+    if not guesses.index.equals(secrets.index):
+        raise RuntimeError("The predictions indices do not match the target indices. Check your inference model.")
+
     guesses_np = guesses.to_numpy()
     secrets_np = secrets.to_numpy()
 
@@ -152,7 +157,7 @@ def __init__(
             syn: pd.DataFrame,
             aux_cols: list[str],
             secret: str,
-            regression: Optional[bool] = None,
+            regression: bool = False,
             n_attacks: int = 500,
             control: Optional[pd.DataFrame] = None,
             inference_model: Optional[InferencePredictor] = None
@@ -180,7 +185,7 @@ def __init__(
         self._aux_cols = aux_cols
         self._evaluated = False
 
-    def _attack(self, target: pd.DataFrame, naive: bool, n_jobs: int, n_attacks: int) -> int:
+    def _attack(self, target: pd.DataFrame, naive: bool, n_jobs: int, n_attacks: int) -> tuple[int, pd.Series]:
         return _run_attack(
             target=target,
             syn=self._syn,
@@ -207,14 +212,17 @@ def evaluate(self, n_jobs: int = -2) -> "InferenceEvaluator":
             The evaluated ``InferenceEvaluator`` object.
 
         """
-        self._n_baseline = self._attack(target=self._ori, naive=True, n_jobs=n_jobs,
-                                        n_attacks=self._n_attacks_baseline)
-        self._n_success = self._attack(target=self._ori, naive=False, n_jobs=n_jobs,
-                                       n_attacks=self._n_attacks_ori)
-        self._n_control = (
-            None if self._control is None else self._attack(target=self._control, naive=False, n_jobs=n_jobs,
-                                                            n_attacks=self._n_attacks_control)
-            )
+        self._n_baseline, self._guesses_baseline = self._attack(
+            target=self._ori, naive=True, n_jobs=n_jobs, n_attacks=self._n_attacks_baseline
+        )
+        self._n_success, self._guesses_success = self._attack(
+            target=self._ori, naive=False, n_jobs=n_jobs, n_attacks=self._n_attacks_ori
+        )
+        self._n_control, self._guesses_control = (
+            (None, None)
+            if self._control is None
+            else self._attack(target=self._control, naive=False, n_jobs=n_jobs, n_attacks=self._n_attacks_control)
+        )
 
         self._evaluated = True
         return self
@@ -269,3 +277,67 @@ def risk(self, confidence_level: float = 0.95, baseline: bool = False) -> Privac
         """
         results = self.results(confidence_level=confidence_level)
         return results.risk(baseline=baseline)
+
+    def risk_for_groups(self, confidence_level: float = 0.95) -> dict[str, EvaluationResults]:
+        """Compute the inference risk for each group of targets with the same value of the secret attribute.
+
+        Parameters
+        ----------
+        confidence_level : float, default is 0.95
+            Confidence level for the error bound calculation.
+
+        Returns
+        -------
+        dict[str, tuple[EvaluationResults | PrivacyRisk]
+            The group as a key, and then for every group the results (EvaluationResults),
+            and the risks (PrivacyRisk) as a tuple.
+
+        """
+        if not self._evaluated:
+            raise RuntimeError("The inference evaluator wasn't evaluated yet. Please, run `evaluate()` first.")
+
+        all_results = {}
+
+        # For every unique group in `self._secret`
+        for group, data_ori in self._ori.groupby(self._secret):
+            # Get the targets for the current group
+            common_indices = data_ori.index.intersection(self._guesses_success.index)
+            # Get the guesses for the current group
+            target_group = data_ori.loc[common_indices]
+            n_attacks_ori = len(target_group)
+
+            # Count the number of success attacks
+            n_success = evaluate_inference_guesses(
+                guesses=self._guesses_success.loc[common_indices],
+                secrets=target_group[self._secret],
+                regression=self._regression,
+            ).sum()
+
+            if self._control is not None:
+                # Get the targets for the current control group
+                data_control = self._control[self._control[self._secret] == group]
+                n_attacks_control = len(data_control)
+
+                # Get the guesses for the current control group
+                common_indices = data_control.index.intersection(self._guesses_control.index)
+
+                # Count the number of success control attacks
+                n_control = evaluate_inference_guesses(
+                    guesses=self._guesses_control.loc[common_indices],
+                    secrets=data_control[self._secret],
+                    regression=self._regression,
+                ).sum()
+            else:
+                n_control = None
+                n_attacks_control = -1
+
+            # Recreate the EvaluationResults for the current group
+            all_results[group] = EvaluationResults(
+                n_attacks=(n_attacks_ori, self._n_attacks_baseline, n_attacks_control),
+                n_success=n_success,
+                n_baseline=self._n_baseline,  # The baseline risk should be the same independent of the group
+                n_control=n_control,
+                confidence_level=confidence_level,
+            )
+
+        return all_results

tests/fixtures.py

@@ -11,7 +11,7 @@
 TEST_DIR_PATH = os.path.dirname(os.path.realpath(__file__))
 
 
-def get_adult(which: str, n_samples: Optional[int] = None) -> pd.DataFrame:
+def get_adult(which: str, deduplicate_on: Optional[list[str]] = None, n_samples: Optional[int] = None) -> pd.DataFrame:
     """Fixture for the adult dataset.
 
     For details see:
@@ -21,6 +21,8 @@ def get_adult(which: str, n_samples: Optional[int] = None) -> pd.DataFrame:
     ----------
     which : str, in ['ori', 'syn']
         Whether to return the "original" or "synthetic" samples.
+    deduplicate_on: list of str
+        A list of columns based on which we'd drop duplicates in the samples.
     n_samples : int
         Number of sample records to return.
         If `None` - return all samples.
@@ -37,4 +39,5 @@ def get_adult(which: str, n_samples: Optional[int] = None) -> pd.DataFrame:
     else:
         raise ValueError(f"Invalid value {which} for parameter `which`. Available are: 'ori' or 'syn'.")
 
-    return pd.read_csv(os.path.join(TEST_DIR_PATH, "datasets", fname), nrows=n_samples)
+    samples = pd.read_csv(os.path.join(TEST_DIR_PATH, "datasets", fname), nrows=n_samples)
+    return samples.drop_duplicates(subset=deduplicate_on) if deduplicate_on else samples

tests/test_inference_evaluator.py

@@ -29,6 +29,21 @@ def test_evaluate_inference_guesses_classification(guesses, secrets, expected):
     np.testing.assert_equal(out, expected)
 
 
+@pytest.mark.parametrize(
+    "guesses, secrets, expected",
+    [
+        (("a", "b"), ("a", "b"), (True, True)),
+        ((np.nan, "b"), (np.nan, "b"), (True, True))
+    ],
+)
+def test_evaluate_inference_guesses_secrets_indices(guesses, secrets, expected):
+    secrets = pd.Series(secrets).sort_index(ascending=False)
+    with pytest.raises(Exception) as runtime_error:
+        evaluate_inference_guesses(guesses=pd.Series(guesses), secrets=secrets, regression=False)
+    assert runtime_error.type is RuntimeError
+    assert "The predictions indices do not match the target indices" in str(runtime_error.value)
+
+
 @pytest.mark.parametrize(
     "guesses, secrets, expected",
     [
@@ -103,8 +118,10 @@ def test_inference_evaluator_rates(
 )
 @pytest.mark.parametrize("secret", ["education", "marital", "capital_gain"])
 def test_inference_evaluator_leaks(aux_cols, secret):
-    ori = get_adult("ori", n_samples=10)
-    evaluator = InferenceEvaluator(ori=ori, syn=ori, control=ori, aux_cols=aux_cols, secret=secret, n_attacks=10)
+    ori = get_adult("ori", deduplicate_on=aux_cols, n_samples=10)
+    evaluator = InferenceEvaluator(
+        ori=ori, syn=ori, control=ori, aux_cols=aux_cols, secret=secret, n_attacks=ori.shape[0]
+    )
     evaluator.evaluate(n_jobs=1)
     results = evaluator.results(confidence_level=0)
 
@@ -113,7 +130,7 @@ def test_inference_evaluator_leaks(aux_cols, secret):
 
 
 def test_evaluator_not_evaluated():
-    df = get_adult("ori", n_samples=10)
+    df = get_adult("ori", deduplicate_on=None, n_samples=10)
     evaluator = InferenceEvaluator(
         ori=df,
         syn=df,
@@ -123,3 +140,25 @@ def test_evaluator_not_evaluated():
     )
     with pytest.raises(RuntimeError):
         evaluator.risk()
+
+
+@pytest.mark.parametrize(
+    "aux_cols",
+    [
+        ["type_employer", "capital_loss", "hr_per_week", "age"],
+        ["education_num", "marital", "capital_loss"],
+    ],
+)
+@pytest.mark.parametrize("secret", ["education", "marital"])
+def test_inference_evaluator_group_wise(aux_cols, secret):
+    ori = get_adult("ori", deduplicate_on=aux_cols, n_samples=10)
+    evaluator = InferenceEvaluator(
+        ori=ori, syn=ori, control=ori, aux_cols=aux_cols, secret=secret, n_attacks=ori.shape[0]
+    )
+    evaluator.evaluate(n_jobs=1)
+
+    group_wise = evaluator.risk_for_groups(confidence_level=0)
+
+    for _, results in group_wise.items():
+        np.testing.assert_equal(results.attack_rate, (1, 0))
+        np.testing.assert_equal(results.control_rate, (1, 0))

tests/test_mixed_types_kneigbors.py

@@ -13,7 +13,7 @@
 
 
 def test_mixed_type_kNN():
-    df = get_adult("ori", n_samples=10)
+    df = get_adult("ori", deduplicate_on=None, n_samples=10)
     nn = MixedTypeKNeighbors().fit(df)
     shuffled_idx = rng.integers(10, size=10)
     dist, ids = nn.kneighbors(df.iloc[shuffled_idx], n_neighbors=1, return_distance=True)
@@ -43,7 +43,7 @@ def test_mixed_type_kNN_numerical_scaling():
 
 @pytest.mark.parametrize("n_neighbors, n_queries", [(1, 10), (3, 5)])
 def test_mixed_type_kNN_shape(n_neighbors, n_queries):
-    df = get_adult("ori", n_samples=10)
+    df = get_adult("ori", deduplicate_on=None, n_samples=10)
     nn = MixedTypeKNeighbors(n_neighbors=n_neighbors).fit(df)
     ids = nn.kneighbors(df.head(n_queries))
     assert isinstance(ids, np.ndarray)

tests/test_singling_out_evaluator.py

@@ -23,8 +23,8 @@
 
 @pytest.mark.parametrize("mode", ["univariate", "multivariate"])
 def test_so_general(mode: str) -> None:
-    ori = get_adult("ori", n_samples=10)
-    syn = get_adult("syn", n_samples=10)
+    ori = get_adult("ori", deduplicate_on=None, n_samples=10)
+    syn = get_adult("syn", deduplicate_on=None, n_samples=10)
     soe = SinglingOutEvaluator(ori=ori, syn=syn, n_attacks=5).evaluate(mode=mode)
 
     for q in soe.queries():
@@ -150,7 +150,7 @@ def test_singling_out_query_generator() -> None:
 @pytest.mark.parametrize("confidence_level", [0.5, 0.68, 0.95, 0.99])
 @pytest.mark.parametrize("mode", ["univariate", "multivariate"])
 def test_singling_out_risk_estimate(confidence_level: float, mode: str) -> None:
-    ori = get_adult("ori", 10)
+    ori = get_adult("ori", deduplicate_on=None, n_samples=10)
     soe = SinglingOutEvaluator(ori=ori, syn=ori, n_attacks=5)
     soe.evaluate(mode=mode)
     _, ci = soe.risk(confidence_level=confidence_level)
@@ -176,7 +176,7 @@ def _so_probability(n: int, w: float):
 
 @pytest.mark.parametrize("max_attempts", [1, 2, 3])
 def test_so_evaluator_max_attempts(max_attempts: int) -> None:
-    ori = get_adult("ori", 10)
+    ori = get_adult("ori", deduplicate_on=None, n_samples=10)
     soe = SinglingOutEvaluator(ori=ori, syn=ori, n_attacks=10, max_attempts=max_attempts)
     soe.evaluate(mode="multivariate")
 

tests/test_sklearn_inference_model.py

@@ -25,7 +25,7 @@
 )
 @pytest.mark.parametrize("secret", ["capital_gain", "capital_loss"])
 def test_inference_evaluator_custom_model_regressor(aux_cols, secret):
-    ori = get_adult("ori", n_samples=10)
+    ori = get_adult("ori", deduplicate_on=aux_cols, n_samples=10)
 
     # Inference model prep
     categorical_cols = ori[aux_cols].select_dtypes(include=["object"]).columns