img_index = args.index
# the target for deep leakage
gt_data = tp(dst[img_index][0]).to(device)
# Data range is from 0 to 1.
gt_data = gt_data.view(1, *gt_data.size())
print(gt_data.shape)
gt_label = torch.Tensor([dst[img_index][1]]).long().to(device)
print("Ground Truth Label is %d" % gt_label.item())
gt_label = gt_label.view(1, )
gt_onehot_label = label_to_onehot(gt_label, num_classes=10)
plt.imshow(tt(gt_data[0].cpu()))
from models.vision import LeNet, weights_init, ResNet18, ResNet34, AlexNet
if args.arch == 'LeNet5':
net = LeNet(nchannel, nclass, args.act).to(device)
net.apply(weights_init)
elif args.arch == 'ResNet18':
net = ResNet18(nclass, nchannel, args.act).to(device)
net.apply(weights_init)
elif args.arch == 'ResNet34':
net = ResNet34(nclass, nchannel, args.act).to(device)
net.apply(weights_init)
elif args.arch == 'AlexNet':
net = AlexNet(nclass=nclass, in_channels=nchannel, act=args.act).to(device)
else:
raise NotImplementedError(
"Only supports LeNet5, AlexNet, ResNet18 and ResNet34.")
torch.manual_seed(1234)
criterion = cross_entropy_for_onehot
gt_data = tp(dst[img_index][0]).to(device)
if len(args.image) > 1:
gt_data = Image.open(args.image)
gt_data = tp(gt_data).to(device)
gt_data = gt_data.view(1, *gt_data.size())
gt_label = torch.Tensor([dst[img_index][1]]).long().to(device)
gt_label = gt_label.view(1, )
gt_onehot_label = label_to_onehot(gt_label)
plt.imshow(tt(gt_data[0].cpu()))
from models.vision import LeNet, weights_init
net = LeNet().to(device)
net.apply(weights_init)
criterion = cross_entropy_for_onehot
# compute original gradient
pred = net(gt_data)
y = criterion(pred, gt_onehot_label)
dy_dx = torch.autograd.grad(y, net.parameters())
original_dy_dx = list((_.detach().clone() for _ in dy_dx))
# generate dummy data and label
dummy_data = torch.randn(gt_data.size()).to(device).requires_grad_(True)
dummy_label = torch.randn(gt_onehot_label.size()).to(device).requires_grad_(True)
# dst = datasets.CIFAR100("~/.torch", download=True)
dst = datasets.CIFAR10("~/.torch", download=True)
imageDataBatchLoad = torch.utils.data.DataLoader(dst, batch_size=batch, shuffle=True, num_workers=2)
tp = transforms.ToTensor()
tt = transforms.ToPILImage()
# img_index = args.index
# gt_data = tp(dst[img_index][0]).to(device)
if len(args.image) > 1:
gt_data1 = Image.open(args.image)
gt_data1 = tp(gt_data1).to(device)
# from models.vision import LeNet, weights_init
net = LeNet().to(device)
## Accuray computation!!
def AreDummyAndOriginalLabelNoMatch(gt_oneshot_label, dummy_onehot_label_copy, decimalRound=3):
originalLabelNo = int(torch.argmax(gt_oneshot_label[0]))
dummyLabelCopy = int(torch.argmax(dummy_onehot_label_copy[0]))
if(originalLabelNo == dummyLabelCopy):
print("Dummy has successfully been restored! The label no. is " + str(originalLabelNo))
return True
else:
print("Restored from Image from Dummy is incorrect! The original label no. is " + str(originalLabelNo) +". The dummy label no. is " + str(dummyLabelCopy))
return False
# Gausian Noise adder
def addGausianNoise(net, standardDeviation=[1.0]):