def draw_kernels(start, end, ns, kernel_size):
global image
global tkImage
global labels
labels = [tk.Label(top) for _ in range(ns)]
global images
images = []
Label_visker = tk.Label(top, text="Used Gaussian Kernels:")
Label_visker.place(x=10, y=130)
for s in range(ns):
temp = template(kernel_size, start + s * (end - start) / ns)
temp = temp / temp.max() * 255
image = Image.fromarray(temp).resize((90, 90))
tkImage = ImageTk.PhotoImage(image=image)
images.append(tkImage)
# image.show()
for s in range(ns):
labels[s].config(image=images[s])
labels[s].place(x=10 + s * 90, y=160)
def run(data_source):
total_num = 100
run_once_button.config(state='disabled')
run_all_button.config(state='disabled')
start = float(s1.get())
end = float(s2.get())
ns = int(s3.get())
kernel_size = int(s4.get())
draw_kernels(start, end, ns, kernel_size)
ckpt = checkpoint.get()
if net_comboxlist.get() == 'LeNet':
net = Net()
params = torch.load(ckpt)
net.load_state_dict(params)
trans = transforms.Compose([transforms.ToTensor()])
if net_comboxlist.get() == 'ResNet32':
checkp = torch.load(ckpt, map_location='cpu')
net = checkp['net']
trans = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.4914, 0.4822, 0.4465],
std=[0.2023, 0.1994, 0.2010])
])
if use_cuda.get():
net.cuda()
net.eval()
Acc = 0
Count = 0
global line_label
line_label = tk.Label(top)
image_labels = [tk.Label(top) for _ in range(ns)]
pathdir = imagedir.get() + data_source
for name in os.listdir(pathdir):
filename = os.path.join(pathdir, name)
img = np.load(filename)
Count += 1
C = np.zeros((ns, 10))
global exam_images
exam_images = []
for i in range(ns):
# do gaussian blur
temp = template(2, i * (2 / ns))
input_image = img_conv2d(img, temp)
if i < 5:
image = Image.fromarray(np.uint8(input_image)).resize((90, 90))
exam_images.append(ImageTk.PhotoImage(image=image))
image_labels[i].config(image=exam_images[i])
image_labels[i].place(x=10 + i * 90, y=360)
if len(input_image.shape) == 2:
input_image = np.expand_dims(input_image, 2)
input_image = trans(input_image / 255)
input_image = input_image.view(-1, input_image.shape[0],
input_image.shape[1],
input_image.shape[2])
input_image = input_image.float()
if use_cuda.get():
input_image = input_image.cuda()
preds = F.softmax(net(input_image), dim=1)
C[i, :] = preds.cpu().detach().numpy()
draw_preds_lines(C)
draw_roc_curves(C, name, Image.fromarray(np.uint8(img)))
c = np.zeros(10)
for i in range(10):
c[i] = max(C[:, i]) - min(C[:, i])
fluc = max(c)
t = 0.5
if fluc > t:
Acc += 1
if Count >= total_num:
acc = Acc / total_num
log_text = "Done! The attack detection accuracy is {.2f}!".format(
acc)
send_information(Receive_window, log_text)
break
time.sleep(2)
run_once_button.config(state='normal')
run_all_button.config(state='normal')
def run():
run_button.config('disabled')
start = int(s1.get())
end = int(s2.get())
ns = int(s3.get())
kernel_size = int(s4.get())
draw_kernels(start, end, ns, kernel_size)
Progress = tk.Label(top, text="Running Progress", font=ft)
Progress.place(x=10, y=260)
global canvas
canvas = tk.Canvas(top, width=465, height=22)
canvas.place(x=10, y=280)
global out_rec
global fill_rec
out_rec = canvas.create_rectangle(2, 2, 440, 20, outline="blue", width=1)
fill_rec = canvas.create_rectangle(2,
2,
0,
20,
outline="",
width=0,
fill="green")
datasetname = comboxlist.get()
for idx, (data, target) in enumerate(train_loader):
if datasetname == 'MNIST':
data = data.numpy()[0][0] * 255
if datasetname == 'CIFAR10':
data = data.numpy()[0] * 255
data = data.transpose((1, 2, 0))
target = target.numpy()[0]
root = 'data/augemented_data/{}/{}'.format(datasetname, target)
if os.path.exists(root) is False:
os.makedirs(root)
for s in range(ns):
temp = template(kernel_size, start + s * (end - start) / ns)
img = img_conv2d(data, temp)
file_name = 'data/augemented_data/{}/{}/{}-{}.bmp'.format(
datasetname, target, idx, s)
if datasetname == 'MNIST':
im = Image.fromarray(np.uint8(img)).convert('L')
else:
im = Image.fromarray(np.uint8(img))
im.save(file_name)
if idx % int(len(train_loader) / 600) == 0:
change_schedule(idx, len(train_loader))
if idx > 10:
break
Finish_label = tk.Label(
top,
wraplength=520,
justify='left',
text=
"Done! The augmentated images are stored in \"data/augmente\nd_data/{}\" for next use in step2!"
.format(datasetname),
font=ft)
Finish_label.place(x=10, y=310)
draw_eample_results(ns)
run_button.config(state='normal')
load_button.config(state='normal')
def run(run_num):
run_once_button.config(state='disabled')
run_all_button.config(state='disabled')
start = int(s1.get())
end = int(s2.get())
ns = int(s3.get())
kernel_size = int(s4.get())
draw_kernels(start, end, ns, kernel_size)
# load model
ckpt = checkpoint.get()
if net_comboxlist.get() == 'LeNet':
net = Net()
params = torch.load(ckpt)
net.load_state_dict(params)
trans = transforms.Compose([transforms.ToTensor()])
if net_comboxlist.get() == 'ResNet32':
checkp = torch.load(ckpt, map_location='cpu')
net = checkp['net']
trans = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.4914, 0.4822, 0.4465],
std=[0.2023, 0.1994, 0.2010])
])
if use_cuda.get():
net.cuda()
net.eval()
C = 0
ns = 5
count_succ = np.zeros(ns + 1)
Label_visexa = tk.Label(top, text="Evolved images:")
Label_visexa.place(x=10, y=340)
image_labels = [tk.Label(top) for _ in range(ns)]
pathdir = imagedir.get()
Label_vispic = tk.Label(top, text="Histogram of Evolved Predictions:")
Label_vispic.place(x=480, y=10)
global bar_labels
bar_labels = [tk.Label(top) for _ in range(ns + 1)]
for name in os.listdir(pathdir):
filename = os.path.join(pathdir, name)
true_label = int(filename[-7])
wrong_label = int(filename[-5])
img = np.load(filename)
C += 1
scale_preds = np.zeros([1, 10])
global exam_images
exam_images = []
global bar_images
bar_images = []
for i in range(ns):
# do gaussian blur
temp = template(2, i * 0.4)
input_image = img_conv2d(img, temp)
# for visualization
image = Image.fromarray(np.uint8(input_image)).resize((90, 90))
exam_images.append(ImageTk.PhotoImage(image=image))
image_labels[i].config(image=exam_images[i])
image_labels[i].place(x=10 + i * 90, y=360)
if net_comboxlist.get() == 'LeNet':
input_image = np.expand_dims(input_image, 2)
input_image = trans(input_image / 255)
input_image = input_image.view(-1, input_image.shape[0],
input_image.shape[1],
input_image.shape[2])
input_image = input_image.float()
if use_cuda.get():
input_image = input_image.cuda()
preds = F.softmax(net(input_image), dim=1)
preds = preds.cpu().detach().numpy()
scale_preds += preds
draw_preds(preds, i)
pred_label = np.argmax(preds)
if pred_label == int(true_label):
count_succ[i] += 1
scale_preds = scale_preds / ns
Label_viscon = tk.Label(top, text="Multiscale confidences:")
Label_viscon.place(x=480, y=150)
draw_preds(scale_preds, ns)
scale_label = np.argmax(scale_preds)
logtext = "Classification result:\n"
send_information(Receive_window, logtext)
logtext = "True label: {}\n".format(true_label)
send_information(Receive_window, logtext)
if scale_label == int(true_label):
count_succ[ns] += 1
logtext = "Before multiscale defense {}, after multiscale defense {}, defense successes!\n".format(
wrong_label, scale_label)
send_information(Receive_window, logtext)
else:
logtext = "Before multiscale defense {}, after multiscale defense {}, defense fails!\n".format(
wrong_label, scale_label)
send_information(Receive_window, logtext)
if C % 50 == 0:
log_text = "processing: {} / {}\n".format(C, run_num)
send_information(Receive_window, log_text)
if C >= run_num:
break
time.sleep(1)
log_text = "Before defense: {} / {} \t Classification accuracy:{:.2f} \n".format(
0, run_num, 0)
send_information(Receive_window, log_text)
log_text = "After defense: {} / {} \t Classification accuracy:{:.2f} \n".format(
int(count_succ[ns]), run_num, count_succ[ns] / run_num)
send_information(Receive_window, log_text)
run_once_button.config(state='normal')
run_all_button.config(state='normal')