Merge pull request #883 from Trusted-AI/development_issue_824

Check for NaN in loss gradient of PGD attacks and replace with 0.0

Author: beat-buesser

art/attacks/evasion/fast_gradient.py

@@ -340,12 +340,20 @@ def _compute_perturbation(
         # Get gradient wrt loss; invert it if attack is targeted
         grad = self.estimator.loss_gradient(batch, batch_labels) * (1 - 2 * int(self.targeted))
 
+        # Check for NaN before normalisation an replace with 0
+        if np.isnan(grad).any():
+            logger.warning("Elements of the loss gradient are NaN and have been replaced with 0.0.")
+            grad = np.where(np.isnan(grad), 0.0, grad)
+
         # Apply mask
         if mask is not None:
             grad = np.where(mask == 0.0, 0.0, grad)
 
         # Apply norm bound
         def _apply_norm(grad, object_type=False):
+            if np.isinf(grad).any():
+                logger.info("The loss gradient array contains at least one positive or negative infinity.")
+
             if self.norm in [np.inf, "inf"]:
                 grad = np.sign(grad)
             elif self.norm == 1:

art/attacks/evasion/projected_gradient_descent/projected_gradient_descent_pytorch.py

@@ -278,6 +278,11 @@ def _compute_perturbation(
         # Get gradient wrt loss; invert it if attack is targeted
         grad = self.estimator.loss_gradient(x=x, y=y) * (1 - 2 * int(self.targeted))
 
+        # Check for nan before normalisation an replace with 0
+        if torch.any(grad.isnan()):
+            logger.warning("Elements of the loss gradient are NaN and have been replaced with 0.0.")
+            grad[grad.isnan()] = 0.0
+
         # Apply mask
         if mask is not None:
             grad = torch.where(mask == 0.0, torch.tensor(0.0), grad)

art/attacks/evasion/projected_gradient_descent/projected_gradient_descent_tensorflow_v2.py

@@ -269,6 +269,11 @@ def _compute_perturbation(self, x: "tf.Tensor", y: "tf.Tensor", mask: Optional["
             1 - 2 * int(self.targeted), dtype=ART_NUMPY_DTYPE
         )
 
+        # Check for NaN before normalisation an replace with 0
+        if tf.reduce_any(tf.math.is_nan(grad)):
+            logger.warning("Elements of the loss gradient are NaN and have been replaced with 0.0.")
+            grad = tf.where(tf.math.is_nan(grad), tf.zeros_like(grad), grad)
+
         # Apply mask
         if mask is not None:
             grad = tf.where(mask == 0.0, 0.0, grad)