Fix bug in random sampling of patch locations in masks for adversarial patch attacks

.github/workflows/ci-pytorch-object-detectors.yml

@@ -41,11 +41,6 @@ jobs:
           python -m pip install --upgrade pip setuptools wheel
           pip3 install -q -r requirements_test.txt
           pip list
-      - name: Pre-install torch
-        run: |
-          pip install torch==1.12.1+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
-          pip install torchvision==0.13.1+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
-          pip install torchaudio==0.12.1+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
       - name: Run Test Action - test_pytorch_object_detector
         run: pytest --cov-report=xml --cov=art --cov-append -q -vv tests/estimators/object_detection/test_pytorch_object_detector.py --framework=pytorch --durations=0
       - name: Run Test Action - test_pytorch_faster_rcnn

art/attacks/evasion/adversarial_patch/adversarial_patch_pytorch.py

@@ -381,23 +381,23 @@ def _random_overlay(
             else:
                 mask_2d = mask[i_sample, :, :]
 
-                edge_x_0 = int(im_scale * padded_patch.shape[self.i_w + 1]) // 2
-                edge_x_1 = int(im_scale * padded_patch.shape[self.i_w + 1]) - edge_x_0
-                edge_y_0 = int(im_scale * padded_patch.shape[self.i_h + 1]) // 2
-                edge_y_1 = int(im_scale * padded_patch.shape[self.i_h + 1]) - edge_y_0
-
-                mask_2d[0:edge_x_0, :] = False
-                if edge_x_1 > 0:
-                    mask_2d[-edge_x_1:, :] = False
-                mask_2d[:, 0:edge_y_0] = False
-                if edge_y_1 > 0:
-                    mask_2d[:, -edge_y_1:] = False
-
-                num_pos = np.argwhere(mask_2d).shape[0]
-                pos_id = np.random.choice(num_pos, size=1)
-                pos = np.argwhere(mask_2d)[pos_id[0]]
-                x_shift = pos[1] - self.image_shape[self.i_w] // 2
+                edge_h_0 = int(im_scale * padded_patch.shape[self.i_h + 1]) // 2
+                edge_h_1 = int(im_scale * padded_patch.shape[self.i_h + 1]) - edge_h_0
+                edge_w_0 = int(im_scale * padded_patch.shape[self.i_w + 1]) // 2
+                edge_w_1 = int(im_scale * padded_patch.shape[self.i_w + 1]) - edge_w_0
+
+                mask_2d[0:edge_h_0, :] = False
+                if edge_h_1 > 0:
+                    mask_2d[-edge_h_1:, :] = False
+                mask_2d[:, 0:edge_w_0] = False
+                if edge_w_1 > 0:
+                    mask_2d[:, -edge_w_1:] = False
+
+                num_pos = np.nonzero(mask_2d.int())
+                pos_id = np.random.choice(num_pos.shape[0], size=1, replace=False)  # type: ignore
+                pos = num_pos[pos_id[0]]
                 y_shift = pos[0] - self.image_shape[self.i_h] // 2
+                x_shift = pos[1] - self.image_shape[self.i_w] // 2
 
             phi_rotate = float(np.random.uniform(-self.rotation_max, self.rotation_max))
 

art/attacks/evasion/imperceptible_asr/imperceptible_asr_pytorch.py

@@ -567,7 +567,7 @@
                 if decoded_output[local_batch_size_idx] == y[local_batch_size_idx]:
                     if loss_2nd_stage[local_batch_size_idx] < best_loss_2nd_stage[local_batch_size_idx]:
                         # Update the best loss at 2nd stage
-                        best_loss_2nd_stage[local_batch_size_idx] = (
+                        best_loss_2nd_stage[local_batch_size_idx] = (  # type: ignore
                             loss_2nd_stage[local_batch_size_idx].detach().cpu().numpy()
                         )
 
@@ -734,7 +734,7 @@
 
         theta_array = np.array(theta)
 
-        return theta_array, original_max_psd
+        return theta_array, original_max_psd  # type: ignore
 
     def _psd_transform(self, delta: "torch.Tensor", original_max_psd: np.ndarray) -> "torch.Tensor":
         """

art/attacks/evasion/saliency_map.py

@@ -88,7 +88,7 @@ def generate(self, x: np.ndarray, y: np.ndarray | None = None, **kwargs) -> np.n
 
         # Initialize variables
         dims = list(x.shape[1:])
-        self._nb_features = np.product(dims)
+        self._nb_features = np.prod(dims)
         x_adv = np.reshape(x.astype(ART_NUMPY_DTYPE), (-1, self._nb_features))
         preds = np.argmax(self.estimator.predict(x, batch_size=self.batch_size), axis=1)
 

art/estimators/classification/pytorch.py

@@ -855,6 +855,8 @@ def loss_gradient(
         else:
             loss.backward()
 
+        grads: torch.Tensor | np.ndarray
+
         if x_grad.grad is not None:
             if isinstance(x, torch.Tensor):
                 grads = x_grad.grad

art/estimators/object_detection/pytorch_object_detector.py

@@ -333,7 +333,7 @@ def _get_losses(
 
     def loss_gradient(
         self, x: np.ndarray | "torch.Tensor", y: list[dict[str, np.ndarray | "torch.Tensor"]], **kwargs
-    ) -> np.ndarray:
+    ) -> np.ndarray | "torch.Tensor":
         """
         Compute the gradient of the loss function w.r.t. `x`.
 
@@ -365,6 +365,8 @@ def loss_gradient(
         # Compute gradients
         loss.backward(retain_graph=True)  # type: ignore
 
+        grads: torch.Tensor | np.ndarray
+
         if x_grad.grad is not None:
             if isinstance(x, np.ndarray):
                 grads = x_grad.grad.cpu().numpy()
@@ -382,7 +384,8 @@ def loss_gradient(
         if not self.channels_first:
             if isinstance(x, np.ndarray):
                 grads = np.transpose(grads, (0, 2, 3, 1))
-            else:
+            elif isinstance(grads, torch.Tensor):
+                # grads_tensor: torch.Tensor = grads
                 grads = torch.permute(grads, (0, 2, 3, 1))
 
         assert grads.shape == x.shape

art/estimators/regression/pytorch.py

@@ -682,6 +682,8 @@ def loss_gradient(
         else:
             loss.backward()
 
+        grads: torch.Tensor | np.ndarray
+
         if x_grad.grad is not None:
             if isinstance(x, torch.Tensor):
                 grads = x_grad.grad

requirements_test.txt

@@ -31,9 +31,9 @@ mxnet-native==1.8.0.post0
 
 # PyTorch
 --find-links https://download.pytorch.org/whl/cpu/torch_stable.html
-torch==2.2.1
-torchaudio==2.2.1
-torchvision==0.17.1+cpu
+torch==2.5.0
+torchaudio==2.5.0
+torchvision==0.20.0
 
 # PyTorch image transformers
 timm==0.9.2