Steganografiya quyidagi sohalarda qo'llaniladi, lekin ular bilan cheklanmaydi

self.original_model = original_model
self.args = args

def __is_prediction_dawn(self, item):
""" Verify if prediction should be
correct as in the DAWN paper.

Args:
item (array): Single query

Returns:
(bool): Should the input be correctly classified ?

"""
bound = floor((2 ** self.args.precision_dawn)
* self.args.probability_dawn)
hashed = hmac.new(
self.args.key_dawn.encode("utf-8"),
item.tobytes(),
hashlib.sha256).hexdigest()
bits = bitstring.BitArray(hex=hashed).bin
if int(bits[:self.args.precision_dawn], 2) <= bound:
return False
else:
return True

def __prediction_dawn_batch(self, batch):
""" Prediction batches as in DAWN Paper

Args:
batch (array): Batch of queries

Returns:
(array): Predictions probabilities

"""
preds = self.original_model(batch)
results = torch.argmax(preds, 1)
classes = self.args.nbr_classes
probs = []
for x_item, pred, y_true in zip(batch, preds, results):
# Case when query should be treated like DAWN
if not self.__is_prediction_dawn(x_item.cpu().numpy()):
# Compute fake probabilities array
list_labels = [k for k in range(0, classes) if k != y_true]
choose_label = random.choice(list_labels)
fake_probs = [random.random() for k in range(classes)]
fake_probs[choose_label] = 1

norm_fake_probs = [k / sum(fake_probs) for k in fake_probs]
probs.append(torch.tensor([norm_fake_probs])[0])

else:
probs.append(pred)

return torch.stack(probs)