def train(epoch):
print('\nEpoch: %d' % epoch)
net.train()
train_loss = 0
correct = 0
total_epsilon = 0
total = 0
for batch_idx, (inputs, targets) in enumerate(trainloader):
inputs, targets = inputs.to(device), targets.to(device)
optimizer.zero_grad()
outputs = net(
utils.random_mask_batch_one_sample(inputs,
args.band_size,
reuse_noise=True))
_, predicted = outputs.max(1)
correct += predicted.eq(targets).sum().item()
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
total += targets.size(0)
train_loss += loss.item()
progress_bar(
batch_idx, len(trainloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)' %
(train_loss /
(batch_idx + 1), 100. * correct / total, correct, total))
def test():
global best_acc
correctclean = 0
correctattacked = 0
cert_correct = 0
total = 0
for batch_idx, (inputs, targets) in enumerate(testloader):
inputs, targets = inputs.to(device), targets.to(device)
total += targets.size(0)
save_image(make_grid(inputs, nrow=10),
"baseline_" + str(batch_idx) + ".jpg")
attacked = attacker.perturb(inputs,
targets,
float('inf'),
random_count=args.randomizations)
predictionsclean, certyn = utils.predict_and_certify(
inputs,
net,
args.band_size,
args.size_of_attack,
10,
threshold=args.threshhold)
predictionsattacked, certynx = utils.predict_and_certify(
attacked,
net,
args.band_size,
args.size_of_attack,
10,
threshold=args.threshhold)
correctclean += (predictionsclean.eq(targets)).sum().item()
correctattacked += (predictionsattacked.eq(targets)).sum().item()
cert_correct += (predictionsclean.eq(targets) & certyn).sum().item()
save_image(make_grid(attacked, nrow=10),
"attacks_" + str(batch_idx) + ".jpg")
progress_bar(
batch_idx, len(testloader),
'Clean Acc: %.3f%% (%d/%d) Cert: %.3f%% (%d/%d) Adv Acc: %.3f%% (%d/%d)'
% ((100. * correctclean) / total, correctclean, total,
(100. * cert_correct) / total, cert_correct, total,
(100. * correctattacked) / total, correctattacked, total))
print('Using band size ' + str(args.band_size) + ' with threshhold ' +
str(args.threshhold))
print('Size of Attack Patch ' + str(args.size_of_attack) + '*' +
str(args.size_of_attack))
print('Total images: ' + str(total))
print('Clean Correct: ' + str(correctclean) + ' (' +
str((100. * correctclean) / total) + '%)')
print('Certified: ' + str(cert_correct) + ' (' +
str((100. * cert_correct) / total) + '%)')
print('Attacked Correct: ' + str(correctattacked) + ' (' +
str((100. * correctattacked) / total) + '%)')
def test():
global best_acc
correct = 0
total = 0
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(testloader):
inputs, targets = inputs.to(device), targets.to(device)
total += targets.size(0)
outputs = net(
utils.random_mask_batch_one_sample(inputs,
args.block_size,
reuse_noise=False))
_, predicted = outputs.max(1)
correct += predicted.eq(targets).sum().item()
progress_bar(
batch_idx, len(testloader), ' Acc: %.3f%% (%d/%d)' %
(100. * correct / total, correct, total))
print('Correct: ' + str(correct) + ' out of ' + str(total))
def test():
global best_acc
correct = 0
cert_correct = 0
cert_incorrect = 0
total = 0
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(testloader):
inputs, targets = inputs.to(device), targets.to(device)
total += targets.size(0)
predictions, certyn = utils.predict_and_certify(
inputs,
net,
args.keep,
args.bandwidth,
args.size_to_certify,
10,
threshold=args.threshhold)
correct += (predictions.eq(targets)).sum().item()
cert_correct += (predictions.eq(targets) & certyn).sum().item()
cert_incorrect += (~predictions.eq(targets) & certyn).sum().item()
progress_bar(
batch_idx, len(testloader),
'Acc: %.3f%% (%d/%d) Cert: %.3f%% (%d/%d)' %
((100. * correct) / total, correct, total,
(100. * cert_correct) / total, cert_correct, total))
print('Band size: ' + str(args.bandwidth) + ' threshhold: ' +
str(args.threshhold) + ' num bands: ' + str(args.keep))
print('Total images: ' + str(total))
print('Correct: ' + str(correct) + ' (' + str((100. * correct) / total) +
'%)')
print('Certified Correct class: ' + str(cert_correct) + ' (' +
str((100. * cert_correct) / total) + '%)')
print('Certified Wrong class: ' + str(cert_incorrect) + ' (' +
str((100. * cert_incorrect) / total) + '%)')
def test():
global best_acc
correctclean = 0
correctattacked = 0
total = 0
for batch_idx, (inputs, targets) in enumerate(testloader):
inputs, targets = inputs.to(device), targets.to(device)
total += targets.size(0)
save_image(make_grid(inputs, nrow=10),
"baseline_baseline_" + str(batch_idx) + ".jpg")
attacked = attacker.perturb(inputs,
targets,
float('inf'),
random_count=args.randomizations)
_, predictionsclean = net(inputs).max(1)
_, predictionsattacked = net(attacked).max(1)
correctclean += (predictionsclean.eq(targets)).sum().item()
correctattacked += (predictionsattacked.eq(targets)).sum().item()
save_image(make_grid(attacked, nrow=10),
"attacks_baseline_" + str(batch_idx) + ".jpg")
progress_bar(
batch_idx, len(testloader),
'Clean Acc: %.3f%% (%d/%d) Adv Acc: %.3f%% (%d/%d)' %
((100. * correctclean) / total, correctclean, total,
(100. * correctattacked) / total, correctattacked, total))
print('Baseline')
print('Size of Attack Patch ' + str(args.size_of_attack) + '*' +
str(args.size_of_attack))
print('Total images: ' + str(total))
print('Clean Correct: ' + str(correctclean) + ' (' +
str((100. * correctclean) / total) + '%)')
print('Attacked Correct: ' + str(correctattacked) + ' (' +
str((100. * correctattacked) / total) + '%)')
if (args.model == 'resnet50'):
net = resnet.ResNet50()
elif (args.model == 'resnet18'):
net = resnet.ResNet18()
net = net.to(device)
if device == 'cuda':
net = torch.nn.DataParallel(net)
cudnn.benchmark = True
assert os.path.isdir(checkpoint_dir), 'Error: no checkpoint directory found!'
resume_file = '{}/{}'.format(checkpoint_dir, args.checkpoint)
assert os.path.isfile(resume_file)
checkpoint = torch.load(resume_file)
net.load_state_dict(checkpoint['net'])
net.eval()
all_batches = []
for batch_idx, (inputs, targets) in enumerate(testloader):
inputs, targets = inputs.to(device), targets.to(device)
with torch.no_grad():
#breakpoint()
batch_radii = utils.certify(inputs, targets, net, args.alpha,
args.band_size, args.attack_size,
args.predsamples, args.boundsamples)
all_batches.append(batch_radii)
progress_bar(batch_idx, len(testloader))
out = torch.cat(all_batches)
print('band size: ' + str(args.band_size))
print('certify correct: ' + str(float((out == 1).sum()) / out.shape[0]))
print('certify wrong: ' + str(float((out == -1).sum()) / out.shape[0]))