Update application of eps_step

Signed-off-by: Beat Buesser <[email protected]>

Author: Beat Buesser

art/attacks/evasion/fast_gradient.py

@@ -377,15 +377,12 @@ def _apply_perturbation(
         self, batch: np.ndarray, perturbation: np.ndarray, eps_step: Union[int, float, np.ndarray]
     ) -> np.ndarray:
 
-        if isinstance(eps_step, (int, float)) and eps_step == np.inf:
+        perturbation_step = eps_step * perturbation
+        perturbation_step[np.isnan(perturbation_step)] = 0
+        batch = batch + perturbation_step
+        if self.estimator.clip_values is not None:
             clip_min, clip_max = self.estimator.clip_values
-            batch[perturbation < 0.0] = clip_min
-            batch[perturbation > 0.0] = clip_max
-        else:
-            batch = batch + eps_step * perturbation
-            if self.estimator.clip_values is not None:
-                clip_min, clip_max = self.estimator.clip_values
-                batch = np.clip(batch, clip_min, clip_max)
+            batch = np.clip(batch, clip_min, clip_max)
 
         return batch
 

art/attacks/evasion/projected_gradient_descent/projected_gradient_descent_pytorch.py

@@ -311,20 +311,16 @@ def _apply_perturbation(
         """
         import torch  # lgtm [py/repeated-import]
 
-        if isinstance(eps_step, (int, float)) and eps_step == np.inf:
+        eps_step = np.array(eps_step, dtype=ART_NUMPY_DTYPE)
+        perturbation_step = torch.tensor(eps_step).to(self.estimator.device) * perturbation
+        perturbation_step[torch.isnan(perturbation_step)] = 0
+        x = x + perturbation_step
+        if self.estimator.clip_values is not None:
             clip_min, clip_max = self.estimator.clip_values
-            x[perturbation < 0.0] = float(clip_min)
-            x[perturbation > 0.0] = float(clip_max)
-        else:
-            eps_step = np.array(eps_step, dtype=ART_NUMPY_DTYPE)
-            x = x + torch.tensor(eps_step).to(self.estimator.device) * perturbation
-
-            if self.estimator.clip_values is not None:
-                clip_min, clip_max = self.estimator.clip_values
-                x = torch.max(
-                    torch.min(x, torch.tensor(clip_max).to(self.estimator.device)),
-                    torch.tensor(clip_min).to(self.estimator.device),
-                )
+            x = torch.max(
+                torch.min(x, torch.tensor(clip_max).to(self.estimator.device)),
+                torch.tensor(clip_min).to(self.estimator.device),
+            )
 
         return x
 

art/attacks/evasion/projected_gradient_descent/projected_gradient_descent_tensorflow_v2.py

@@ -304,16 +304,12 @@ def _apply_perturbation(
         """
         import tensorflow as tf  # lgtm [py/repeated-import]
 
-        if isinstance(eps_step, (int, float)) and eps_step == np.inf:
+        perturbation_step = eps_step * perturbation
+        perturbation_step = tf.where(tf.math.is_nan(perturbation_step), 0, perturbation_step)
+        x = x + perturbation_step
+        if self.estimator.clip_values is not None:
             clip_min, clip_max = self.estimator.clip_values
-            x = tf.where(perturbation < 0.0, clip_min, x)
-            x = tf.where(perturbation > 0.0, clip_max, x)
-        else:
-            x = x + tf.constant(eps_step, dtype=ART_NUMPY_DTYPE) * perturbation
-
-            if self.estimator.clip_values is not None:
-                clip_min, clip_max = self.estimator.clip_values
-                x = tf.clip_by_value(x, clip_min, clip_max)
+            x = tf.clip_by_value(x, clip_value_min=clip_min, clip_value_max=clip_max)
 
         return x