Fix missing perturbation update stage 2

This bug was introduced while wrapping up PR 719. The gradient descrent
steps for loss theta and loss net were previously separated and merged
into one line. During that simplifiation of the gradient descent, the
perturbation update was lost.

Signed-off-by: Hessel Tuinhof <[email protected]>

Author: hesseltuinhof

art/attacks/evasion/imperceptible_asr/imperceptible_asr.py

@@ -254,10 +254,12 @@ def _create_imperceptible(self, x: np.ndarray, x_adversarial: np.ndarray, y: np.
         if x.ndim != 1:
             alpha = np.expand_dims(alpha, axis=-1)
 
-        perturbation = x_adversarial - x
         x_perturbed = x_adversarial.copy()
 
         for i in range(self.max_iter_2):
+            # get perturbation
+            perturbation = x_perturbed - x
+
             # get loss gradients of both losses
             gradients_net = self.estimator.loss_gradient(x_perturbed, y, batch_mode=True)
             gradients_theta, loss_theta = self._loss_gradient_masking_threshold(perturbation, x)