Merge pull request #2214 from Trusted-AI/development_issue_2165

Add support for arbitrary-shaped input in APGD and ACG attacks

Author: beat-buesser

art/attacks/evasion/auto_conjugate_gradient.py

@@ -463,7 +463,9 @@ def generate(self, x: np.ndarray, y: Optional[np.ndarray] = None, **kwargs) -> n
 
                 # self.eta = np.full((self.batch_size, 1, 1, 1), 2 * self.eps_step).astype(ART_NUMPY_DTYPE)
                 _batch_size = x_k.shape[0]
-                eta = np.full((_batch_size, 1, 1, 1), self.eps_step).astype(ART_NUMPY_DTYPE)
+                eta = np.full((_batch_size,) + (1,) * len(self.estimator.input_shape), self.eps_step).astype(
+                    ART_NUMPY_DTYPE
+                )
                 self.count_condition_1 = np.zeros(shape=(_batch_size,))
                 gradk_1 = np.zeros_like(x_k)
                 cgradk_1 = np.zeros_like(x_k)
@@ -650,4 +652,4 @@ def get_beta(gradk, gradk_1, cgradk_1):
     betak = -(_gradk * delta_gradk).sum(axis=1) / (
         (_cgradk_1 * delta_gradk).sum(axis=1) + np.finfo(ART_NUMPY_DTYPE).eps
     )
-    return betak.reshape((_batch_size, 1, 1, 1))
+    return betak.reshape((_batch_size,) + (1,) * (len(gradk.shape) - 1))

art/attacks/evasion/auto_projected_gradient_descent.py

@@ -458,7 +458,9 @@ def generate(self, x: np.ndarray, y: Optional[np.ndarray] = None, **kwargs) -> n
 
                 # modification for image-wise stepsize update
                 _batch_size = x_k.shape[0]
-                eta = np.full((_batch_size, 1, 1, 1), self.eps_step).astype(ART_NUMPY_DTYPE)
+                eta = np.full((_batch_size,) + (1,) * len(self.estimator.input_shape), self.eps_step).astype(
+                    ART_NUMPY_DTYPE
+                )
                 self.count_condition_1 = np.zeros(shape=(_batch_size,))
 
                 for k_iter in trange(self.max_iter, desc="AutoPGD - iteration", leave=False, disable=not self.verbose):