Steganografiya quyidagi sohalarda qo'llaniladi, lekin ular bilan cheklanmaydi

item_x = item_x.reshape(size_x, size_y)
for c in msg:
item_x = self.__write_letter(c, item_x, offset=(0, k))
k += 4
# Update trigger set and training data
reshape_item = item_x.reshape(batch, size_x, size_y)
triggerset.append((reshape_item, target))
watermarked_dataset.append((reshape_item, target))

# Update loaders
trainloader, valloader, testloader = self.loaders(watermarked_dataset)
triggerloader = torch.utils.data.DataLoader(
triggerset, batch_size=self.args.batch_size, shuffle=True)

return trainloader, valloader, testloader, triggerloader

def __fgsm_attack(self, data, epsilon, data_grad):
""" Compute adversarial example with the
"fast gradient sign" method.

Args:
data (object): original data to be perturbated
epsilon (float): epsilon parameter for perturbation
data_grad (object): gradient for perturbation

Returns:
perturbed_data (Object): data with inserted msg
"""
sign_data_grad = data_grad.sign()
perturbed_data = data + epsilon * sign_data_grad
perturbed_data = torch.clamp(perturbed_data, 0, 1)
return perturbed_data

def generate_trigger_merrer(self):
""" Generation trigger set, based on

Adversarial Frontier Stitching for
Remote Neural Network Watermarking

by Le Merrer et al. (2017)

Returns:
trainloader (Object): training loader
valloader (Object): validation loader
testloader (Object): test loader
triggerloader (Object): trigger loader
"""

epsilon = self.args.epsilon
triggerset = []
self.model.to(self.device)