Add utility function for least likely class targets

Author: Irina Nicolae

art/utils.py

@@ -124,6 +124,21 @@ def random_targets(labels, nb_classes):
     return to_categorical(result, nb_classes)
 
 
+def least_likely_class(x, classifier):
+    """
+    Compute the least likely class predictions for sample `x`. This strategy for choosing attack targets was used in
+    (Kurakin et al., 2016). See https://arxiv.org/abs/1607.02533.
+
+    :param x: A data sample of shape accepted by `classifier`.
+    :type x: `np.ndarray`
+    :param classifier: The classifier used for computing predictions.
+    :type classifier: `Classifier`
+    :return: Least-likely class predicted by `classifier` for sample `x` in one-hot encoding.
+    :rtype: `np.ndarray`
+    """
+    return to_categorical(np.argmin(classifier.predict(x), axis=1), nb_classes=classifier.nb_classes)
+
+
 def get_label_conf(y_vec):
     """
     Returns the confidence and the label of the most probable class given a vector of class confidences

art/utils_unittest.py

@@ -5,7 +5,7 @@
 
 import numpy as np
 
-from art.utils import load_mnist, projection, random_sphere, to_categorical
+from art.utils import load_mnist, projection, random_sphere, to_categorical, least_likely_class
 from art.utils import random_targets, get_label_conf, get_labels_np_array, preprocess
 
 logger = logging.getLogger('testLogger')
@@ -74,6 +74,25 @@ def test_random_targets(self):
         random_y = random_targets(y_, 10)
         self.assertTrue(np.all(y != random_y.argmax(axis=1)))
 
+    def test_least_likely_class(self):
+        class DummyClassifier():
+            @property
+            def nb_classes(self):
+                return 4
+
+            def predict(self, x):
+                fake_preds = [0.1, 0.2, 0.05, 0.65]
+                return np.array([fake_preds] * x.shape[0])
+
+        batch_size = 5
+        x = np.random.rand(batch_size, 10, 10, 1)
+        classifier = DummyClassifier()
+        preds = least_likely_class(x, classifier)
+        self.assertTrue(preds.shape == (batch_size, classifier.nb_classes))
+
+        expected_preds = np.array([[0, 0, 1, 0]] * batch_size)
+        self.assertTrue((preds == expected_preds).all())
+
     def test_get_label_conf(self):
         y = np.array([3, 1, 4, 1, 5, 9])
         y_ = to_categorical(y)