Merge branch 'dev_1.15.0' into trades_adv

Author: beat-buesser

art/attacks/evasion/projected_gradient_descent/projected_gradient_descent_pytorch.py

@@ -269,7 +269,7 @@ def _generate_batch(
         inputs = x.to(self.estimator.device)
         targets = targets.to(self.estimator.device)
         adv_x = torch.clone(inputs)
-        momentum = torch.zeros(inputs.shape)
+        momentum = torch.zeros(inputs.shape).to(self.estimator.device)
 
         if mask is not None:
             mask = mask.to(self.estimator.device)

art/attacks/evasion/sign_opt.py

@@ -306,14 +306,14 @@ def _fine_grained_binary_search_local(
         lbd = initial_lbd
         # For targeted: we want to expand(x1.01) boundary away from targeted dataset
         # For untargeted, we want to slim(x0.99) the boundary toward the original dataset
-        if (not self._is_label(x_0 + lbd * theta, target) and self.targeted) or (
-            self._is_label(x_0 + lbd * theta, y_0) and not self.targeted
+        if (self.targeted and not self._is_label(x_0 + lbd * theta, target)) or (
+            not self.targeted and self._is_label(x_0 + lbd * theta, y_0)
         ):
             lbd_lo = lbd
             lbd_hi = lbd * 1.01
             nquery += 1
-            while (not self._is_label(x_0 + lbd_hi * theta, target) and self.targeted) or (
-                self._is_label(x_0 + lbd_hi * theta, y_0) and not self.targeted
+            while (self.targeted and not self._is_label(x_0 + lbd_hi * theta, target)) or (
+                not self.targeted and self._is_label(x_0 + lbd_hi * theta, y_0)
             ):
                 lbd_hi = lbd_hi * 1.01
                 nquery += 1
@@ -323,17 +323,17 @@ def _fine_grained_binary_search_local(
             lbd_hi = lbd
             lbd_lo = lbd * 0.99
             nquery += 1
-            while (self._is_label(x_0 + lbd_lo * theta, target) and self.targeted) or (
-                not self._is_label(x_0 + lbd_lo * theta, y_0) and not self.targeted
+            while (self.targeted and self._is_label(x_0 + lbd_lo * theta, target)) or (
+                not self.targeted and not self._is_label(x_0 + lbd_lo * theta, y_0)
             ):
                 lbd_lo = lbd_lo * 0.99
                 nquery += 1
 
         while (lbd_hi - lbd_lo) > tol:
             lbd_mid = (lbd_lo + lbd_hi) / 2.0
             nquery += 1
-            if (self._is_label(x_0 + lbd_mid * theta, target) and self.targeted) or (
-                not self._is_label(x_0 + lbd_mid * theta, y_0) and not self.targeted
+            if (self.targeted and self._is_label(x_0 + lbd_mid * theta, target)) or (
+                not self.targeted and not self._is_label(x_0 + lbd_mid * theta, y_0)
             ):
                 lbd_hi = lbd_mid
             else:

art/attacks/poisoning/perturbations/image_perturbations.py

@@ -127,7 +127,7 @@ def insert_image(
     original_dtype = x.dtype
     data = np.copy(x)
     if channels_first:
-        data = data.transpose([1, 2, 0])
+        data = np.transpose(data, (1, 2, 0))
 
     height, width, num_channels = data.shape
 
@@ -136,37 +136,36 @@ def insert_image(
     backdoored_img = Image.new("RGBA", (width, height), 0)  # height and width are swapped for PIL
 
     if no_color:
-        backdoored_input = Image.fromarray((data * 255).astype("uint8").squeeze(axis=2), mode=mode)
+        backdoored_input = Image.fromarray((data * 255).astype(np.uint8).squeeze(axis=2), mode=mode)
     else:
-        backdoored_input = Image.fromarray((data * 255).astype("uint8"), mode=mode)
+        backdoored_input = Image.fromarray((data * 255).astype(np.uint8), mode=mode)
 
     orig_img.paste(backdoored_input)
 
     trigger = Image.open(backdoor_path).convert("RGBA")
-    if size:
-        trigger = trigger.resize(size)
+    if size is not None:
+        trigger = trigger.resize((size[1], size[0]))  # height and width are swapped for PIL
 
     backdoor_width, backdoor_height = trigger.size  # height and width are swapped for PIL
 
     if backdoor_width > width or backdoor_height > height:
         raise ValueError("Backdoor does not fit inside original image")
 
     if random:
-        x_shift = np.random.randint(width - backdoor_width)
-        y_shift = np.random.randint(height - backdoor_height)
+        x_shift = np.random.randint(width - backdoor_width + 1)
+        y_shift = np.random.randint(height - backdoor_height + 1)
 
     backdoored_img.paste(trigger, (x_shift, y_shift), mask=trigger)
     composite = Image.alpha_composite(orig_img, backdoored_img)
     backdoored_img = Image.blend(orig_img, composite, blend)
-
     backdoored_img = backdoored_img.convert(mode)
 
-    res = np.array(backdoored_img) / 255.0
+    res = np.asarray(backdoored_img) / 255.0
 
     if no_color:
         res = np.expand_dims(res, 2)
 
     if channels_first:
-        res = res.transpose([2, 0, 1])
+        res = np.transpose(res, (2, 0, 1))
 
     return res.astype(original_dtype)

requirements_test.txt

@@ -12,7 +12,7 @@ pydub==0.25.1
 resampy==0.4.2
 ffmpeg-python==0.2.0
 cma==3.3.0
-pandas==1.5.3
+pandas==2.0.1
 librosa==0.10.0.post2
 numba~=0.56.4
 opencv-python
@@ -51,7 +51,7 @@ jax[cpu]==0.4.8
 # lingvo==0.6.4
 
 # other
-pytest~=7.2.2
+pytest~=7.3.1
 pytest-flake8~=1.1.1
 flake8~=4.0.1
 pytest-mock~=3.10.0