Removing Label Consistent Attack

modules/base_trainer/run_module.py

@@ -48,12 +48,6 @@ def run(experiment_name, module_name, **kwargs):
  Path(output_path[:output_path.rfind('/')]).mkdir(parents=True,
  exist_ok=True)
 
- reduce_amplitude = variant = None
- if "reduce_amplitude" in args:
- reduce_amplitude = None if args['reduce_amplitude'] < 0\
- else args['reduce_amplitude']
- variant = args['variant']
-
  # TODO: make this more extensible
  if dataset_flag == "cifar_100":
  model = load_model(model_flag, 20)
@@ -77,9 +71,8 @@ def run(experiment_name, module_name, **kwargs):
  print("Building datasets...")
 
  poisoner, _ = pick_poisoner(poisoner_flag,
- dataset_flag,
- target_label,
- reduce_amplitude)
+ dataset_flag,
+ target_label)
 
  # TODO: Hyperparam this 0.3!
  if slurm_id is None:

modules/base_utils/datasets.py

@@ -158,10 +158,6 @@
 }
 
 
-LABEL_CONSISTENT_PATH = Path("./data/label_consistent_poison")
-LABEL_CONSISTENT_TRANSFORM_XY = lambda xy: (transforms.functional.to_pil_image(xy[0].permute(2,0,1)), xy[1].item())
-
-
 class LabelSortedDataset(ConcatDataset):
  def __init__(self, dataset: Dataset):
  self.orig_dataset = dataset
@@ -348,11 +344,9 @@ class TurnerPoisoner(Poisoner):
  def __init__(
  self,
  *,
- method="bottom-right",
- reduce_amplitude=None
+ method="bottom-right"
  ):
  self.method = method
- self.reduce_amplitude = reduce_amplitude
  self.trigger_mask = [
  ((-1, -1), 1),
  ((-1, -2), -1),
@@ -370,7 +364,7 @@ def poison(self, x: Image.Image) -> Image.Image:
  px = ret_x.load()
 
  for (x, y), sign in self.trigger_mask:
- shift = int((self.reduce_amplitude or 1) * sign * 255)
+ shift = int(sign * 255)
  r, g, b = px[x, y]
  shifted = (r + shift, g + shift, b + shift)
  px[x, y] = shifted
@@ -436,12 +430,10 @@ def seed(self, i):
  self.poisoner.seed(i)
 
 
-def load_dataset(dataset_flag, train=True, variant=None):
+def load_dataset(dataset_flag, train=True):
  path = PATH[dataset_flag]
  if dataset_flag == 'cifar':
- if variant is None:
- return load_cifar_dataset(path, train)
- return load_label_consistent_dataset(path, variant)
+ return load_cifar_dataset(path, train)
  if dataset_flag == 'cifar_100':
  return load_cifar_100_dataset(path, train)
  elif dataset_flag == 'mnist':
@@ -509,21 +501,6 @@ def load_tiny_imagenet_dataset(path, train=True):
  print(np.unique(dataset.targets))
  return dataset
 
-
-def load_label_consistent_dataset(path, variant='gan_0_2'):
- cifar = load_cifar_dataset(path)
- labels = torch.tensor([xy[1] for xy in cifar])
- if not LABEL_CONSISTENT_PATH.is_dir():
- command = ["./modules/base_utils/label_consistent_setup.sh"]
- print("Downloading Label Consistent Dataset...")
- process = subprocess.Popen(command, shell=True, stdout=subprocess.DEVNULL)
- process.wait()
- images = torch.tensor(np.load(LABEL_CONSISTENT_PATH / (variant + '.npy')) / 255)
- dataset = TensorDataset(images, labels)
-
- return MappedDataset(dataset, LABEL_CONSISTENT_TRANSFORM_XY)
-
-
 def make_dataloader(
  dataset: Dataset,
  batch_size,
@@ -542,13 +519,13 @@ def make_dataloader(
  return dataloader
 
 
-def pick_poisoner(poisoner_flag, dataset_flag, target_label, reduce_amplitude=None):
+def pick_poisoner(poisoner_flag, dataset_flag, target_label):
  if dataset_flag == "cifar" or dataset_flag == "cifar_100":
- x_poisoner, all_x_poisoner = pick_cifar_poisoner(poisoner_flag, reduce_amplitude)
+ x_poisoner, all_x_poisoner = pick_cifar_poisoner(poisoner_flag)
  elif dataset_flag == "mnist" or dataset_flag == "fmnist":
  x_poisoner, all_x_poisoner = pick_mnist_poisoner(poisoner_flag)
  elif dataset_flag == "tiny_imagenet":
- x_poisoner, all_x_poisoner = pick_tiny_imagenet_poisoner(poisoner_flag, reduce_amplitude)
+ x_poisoner, all_x_poisoner = pick_tiny_imagenet_poisoner(poisoner_flag)
  else:
  raise NotImplementedError()
 
@@ -558,7 +535,7 @@ def pick_poisoner(poisoner_flag, dataset_flag, target_label, reduce_amplitude=No
  return x_label_poisoner, all_x_label_poisoner
 
 
-def pick_cifar_poisoner(poisoner_flag, reduce_amplitude):
+def pick_cifar_poisoner(poisoner_flag):
  if poisoner_flag == "1xp":
  x_poisoner = PixelPoisoner()
  all_x_poisoner = PixelPoisoner()
@@ -612,22 +589,20 @@ def pick_cifar_poisoner(poisoner_flag, reduce_amplitude):
  )
 
  elif poisoner_flag == "1xl":
- x_poisoner = TurnerPoisoner(reduce_amplitude=reduce_amplitude)
- all_x_poisoner = TurnerPoisoner(reduce_amplitude=reduce_amplitude)
+ x_poisoner = TurnerPoisoner()
+ all_x_poisoner = TurnerPoisoner()
 
  elif poisoner_flag == "4xl":
- x_poisoner = TurnerPoisoner(method="all-corners",
- reduce_amplitude=reduce_amplitude)
- all_x_poisoner = TurnerPoisoner(method="all-corners",
- reduce_amplitude=reduce_amplitude)
+ x_poisoner = TurnerPoisoner(method="all-corners")
+ all_x_poisoner = TurnerPoisoner(method="all-corners")
 
  else:
  raise NotImplementedError()
 
  return x_poisoner, all_x_poisoner
 
 
-def pick_tiny_imagenet_poisoner(poisoner_flag, reduce_amplitude):
+def pick_tiny_imagenet_poisoner(poisoner_flag):
  if poisoner_flag == "1xp":
  x_poisoner = PixelPoisoner(pos=(22, 32), col=(101, 0, 25))
  all_x_poisoner = PixelPoisoner(pos=(22, 32), col=(101, 0, 25))
@@ -681,14 +656,12 @@ def pick_tiny_imagenet_poisoner(poisoner_flag, reduce_amplitude):
  )
 
  elif poisoner_flag == "1xl":
- x_poisoner = TurnerPoisoner(reduce_amplitude=reduce_amplitude)
- all_x_poisoner = TurnerPoisoner(reduce_amplitude=reduce_amplitude)
+ x_poisoner = TurnerPoisoner()
+ all_x_poisoner = TurnerPoisoner()
 
  elif poisoner_flag == "4xl":
- x_poisoner = TurnerPoisoner(method="all-corners",
- reduce_amplitude=reduce_amplitude)
- all_x_poisoner = TurnerPoisoner(method="all-corners",
- reduce_amplitude=reduce_amplitude)
+ x_poisoner = TurnerPoisoner(method="all-corners")
+ all_x_poisoner = TurnerPoisoner(method="all-corners")
 
  else:
  raise NotImplementedError()
@@ -756,7 +729,6 @@ def generate_datasets(
  clean_label,
  target_label,
  target_mask_ind,
- variant=None,
  big=False
 ):
 
@@ -765,19 +737,13 @@ def generate_datasets(
 
  n_classes = len(test_dataset.classes)
 
- label_consistent_dataset = None
- if variant:
- label_consistent_dataset = load_dataset(dataset_flag,
- train=False,
- variant=variant)
-
  poison_train = PoisonedDataset(
  train_dataset,
  poisoner,
  eps=eps,
- label=target_label if label_consistent_dataset else clean_label,
+ label=clean_label,
  transform=TRANSFORM_TRAIN_XY[dataset_flag + ('_big' if big else '')],
- poison_dataset=label_consistent_dataset
+ poison_dataset=None
  )
 
  if target_mask_ind is not None:

schemas/base_trainer.toml

@@ -19,6 +19,4 @@ batch_size = "int: {0,1,...,infty}. Specifies batch size. Set to default for tra
 epochs = "int: {0,1,...,infty}. Specifies number of epochs. Set to default for trainer if omitted."
 train_pct = "TODO"
 optim_kwargs = "dict. Optional keywords for Pytorch SGD / Adam optimizer. See sever example."
-scheduler_kwargs = "dict. Optional keywords for Pytorch learning rate optimizer (with SGD). See sever example."
-variant = "string: Form: {model}_{hyperparam string}. Represents .npy file to consume from data/ when doing label consistent"
-reduce_amplitude = "float: {[0, 1], -1.0}. Specifies amplitude reduction for label consistent. -1.0 if none"
+scheduler_kwargs = "dict. Optional keywords for Pytorch learning rate optimizer (with SGD). See sever example."