example / app.py
Kieran Fraser
Updated poisoning attack with clean prediction gallery
4ae31ed
'''
ART-JATIC Gradio Example App
To run:
- clone the repository
- execute: gradio examples/gradio_app.py or python examples/gradio_app.py
- navigate to local URL e.g. http://127.0.0.1:7860
'''
import gradio as gr
import numpy as np
from carbon_theme import Carbon
import numpy as np
import torch
import transformers
from art.estimators.classification.hugging_face import HuggingFaceClassifierPyTorch
from art.attacks.evasion import ProjectedGradientDescentPyTorch, AdversarialPatchPyTorch
from art.utils import load_dataset
from art.attacks.poisoning import PoisoningAttackBackdoor
from art.attacks.poisoning.perturbations import insert_image
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
css = """
.input-image { margin: auto !important }
.plot-padding { padding: 20px; }
"""
def clf_evasion_evaluate(*args):
'''
Run a classification task evaluation
'''
attack = args[0]
model_type = args[1]
model_url = args[2]
model_channels = args[3]
model_height = args[4]
model_width = args[5]
model_classes = args[6]
model_clip = args[7]
model_upsample = args[8]
attack_max_iter = args[9]
attack_eps = args[10]
attack_eps_steps = args[11]
x_location = args[12]
y_location = args[13]
patch_height = args[14]
patch_width = args[15]
data_type = args[-1]
if model_type == "Example":
model = transformers.AutoModelForImageClassification.from_pretrained(
'facebook/deit-tiny-distilled-patch16-224',
ignore_mismatched_sizes=True,
num_labels=10
)
upsampler = torch.nn.Upsample(scale_factor=7, mode='nearest')
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
loss_fn = torch.nn.CrossEntropyLoss()
hf_model = HuggingFaceClassifierPyTorch(
model=model,
loss=loss_fn,
optimizer=optimizer,
input_shape=(3, 32, 32),
nb_classes=10,
clip_values=(0, 1),
processor=upsampler
)
model_checkpoint_path = './state_dicts/deit_cifar_base_model.pt'
hf_model.model.load_state_dict(torch.load(model_checkpoint_path, map_location=device))
if data_type == "Example":
(x_train, y_train), (_, _), _, _ = load_dataset('cifar10')
x_train = np.transpose(x_train, (0, 3, 1, 2)).astype(np.float32)
y_train = np.argmax(y_train, axis=1)
classes = np.unique(y_train)
samples_per_class = 1
x_subset = []
y_subset = []
for c in classes:
indices = y_train == c
x_subset.append(x_train[indices][:samples_per_class])
y_subset.append(y_train[indices][:samples_per_class])
x_subset = np.concatenate(x_subset)
y_subset = np.concatenate(y_subset)
label_names = [
'airplane',
'automobile',
'bird',
'cat',
'deer',
'dog',
'frog',
'horse',
'ship',
'truck',
]
outputs = hf_model.predict(x_subset)
clean_preds = np.argmax(outputs, axis=1)
clean_acc = np.mean(clean_preds == y_subset)
benign_gallery_out = []
for i, im in enumerate(x_subset):
benign_gallery_out.append(( im.transpose(1,2,0), label_names[np.argmax(outputs[i])] ))
if attack == "PGD":
attacker = ProjectedGradientDescentPyTorch(hf_model, max_iter=attack_max_iter,
eps=attack_eps, eps_step=attack_eps_steps)
x_adv = attacker.generate(x_subset)
outputs = hf_model.predict(x_adv)
adv_preds = np.argmax(outputs, axis=1)
adv_acc = np.mean(adv_preds == y_subset)
adv_gallery_out = []
for i, im in enumerate(x_adv):
adv_gallery_out.append(( im.transpose(1,2,0), label_names[np.argmax(outputs[i])] ))
delta = ((x_subset - x_adv) + 8/255) * 10
delta_gallery_out = delta.transpose(0, 2, 3, 1)
if attack == "Adversarial Patch":
scale_min = 0.3
scale_max = 1.0
rotation_max = 0
learning_rate = 5000.
attacker = AdversarialPatchPyTorch(hf_model, scale_max=scale_max,
scale_min=scale_min,
rotation_max=rotation_max,
learning_rate=learning_rate,
max_iter=attack_max_iter, patch_type='square',
patch_location=(x_location, y_location),
patch_shape=(3, patch_height, patch_width))
patch, _ = attacker.generate(x_subset)
x_adv = attacker.apply_patch(x_subset, scale=0.3)
outputs = hf_model.predict(x_adv)
adv_preds = np.argmax(outputs, axis=1)
adv_acc = np.mean(adv_preds == y_subset)
adv_gallery_out = []
for i, im in enumerate(x_adv):
adv_gallery_out.append(( im.transpose(1,2,0), label_names[np.argmax(outputs[i])] ))
delta_gallery_out = np.expand_dims(patch, 0).transpose(0,2,3,1)
return benign_gallery_out, adv_gallery_out, delta_gallery_out, clean_acc, adv_acc
def clf_poison_evaluate(*args):
attack = args[0]
model_type = args[1]
trigger_image = args[2]
target_class = args[3]
data_type = args[-1]
if model_type == "Example":
model = transformers.AutoModelForImageClassification.from_pretrained(
'facebook/deit-tiny-distilled-patch16-224',
ignore_mismatched_sizes=True,
num_labels=10
)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
loss_fn = torch.nn.CrossEntropyLoss()
hf_model = HuggingFaceClassifierPyTorch(
model=model,
loss=loss_fn,
optimizer=optimizer,
input_shape=(3, 224, 224),
nb_classes=10,
clip_values=(0, 1),
)
if data_type == "Example":
import torchvision
transform = torchvision.transforms.Compose([
torchvision.transforms.Resize((224, 224)),
torchvision.transforms.ToTensor(),
])
train_dataset = torchvision.datasets.ImageFolder(root="./data/imagenette2-320/train", transform=transform)
labels = np.asarray(train_dataset.targets)
classes = np.unique(labels)
samples_per_class = 100
x_subset = []
y_subset = []
for c in classes:
indices = np.where(labels == c)[0][:samples_per_class]
for i in indices:
x_subset.append(train_dataset[i][0])
y_subset.append(train_dataset[i][1])
x_subset = np.stack(x_subset)
y_subset = np.asarray(y_subset)
label_names = [
'fish',
'dog',
'cassette player',
'chainsaw',
'church',
'french horn',
'garbage truck',
'gas pump',
'golf ball',
'parachutte',
]
if attack == "Backdoor":
from PIL import Image
im = Image.fromarray(trigger_image)
im.save("./tmp.png")
def poison_func(x):
return insert_image(
x,
backdoor_path='./tmp.png',
channels_first=True,
random=False,
x_shift=0,
y_shift=0,
size=(32, 32),
mode='RGB',
blend=0.8
)
backdoor = PoisoningAttackBackdoor(poison_func)
source_class = 0
target_class = label_names.index(target_class)
poison_percent = 0.5
x_poison = np.copy(x_subset)
y_poison = np.copy(y_subset)
is_poison = np.zeros(len(x_subset)).astype(bool)
indices = np.where(y_subset == source_class)[0]
num_poison = int(poison_percent * len(indices))
for i in indices[:num_poison]:
x_poison[i], _ = backdoor.poison(x_poison[i], [])
y_poison[i] = target_class
is_poison[i] = True
poison_indices = np.where(is_poison)[0]
hf_model.fit(x_poison, y_poison, nb_epochs=2)
clean_x = x_poison[~is_poison]
clean_y = y_poison[~is_poison]
outputs = hf_model.predict(clean_x)
clean_preds = np.argmax(outputs, axis=1)
clean_acc = np.mean(clean_preds == clean_y)
clean_out = []
for i, im in enumerate(clean_x):
clean_out.append( (im.transpose(1,2,0), label_names[clean_preds[i]]) )
poison_x = x_poison[is_poison]
poison_y = y_poison[is_poison]
outputs = hf_model.predict(poison_x)
poison_preds = np.argmax(outputs, axis=1)
poison_acc = np.mean(poison_preds == poison_y)
poison_out = []
for i, im in enumerate(poison_x):
poison_out.append( (im.transpose(1,2,0), label_names[poison_preds[i]]) )
return clean_out, poison_out, clean_acc, poison_acc
def show_params(type):
'''
Show model parameters based on selected model type
'''
if type!="Example":
return gr.Column(visible=True)
return gr.Column(visible=False)
def run_inference(*args):
model_type = args[0]
model_url = args[1]
model_channels = args[2]
model_height = args[3]
model_width = args[4]
model_classes = args[5]
model_clip = args[6]
model_upsample = args[7]
data_type = args[8]
if model_type == "Example":
model = transformers.AutoModelForImageClassification.from_pretrained(
'facebook/deit-tiny-distilled-patch16-224',
ignore_mismatched_sizes=True,
num_labels=10
)
upsampler = torch.nn.Upsample(scale_factor=7, mode='nearest')
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
loss_fn = torch.nn.CrossEntropyLoss()
hf_model = HuggingFaceClassifierPyTorch(
model=model,
loss=loss_fn,
optimizer=optimizer,
input_shape=(3, 32, 32),
nb_classes=10,
clip_values=(0, 1),
processor=upsampler
)
model_checkpoint_path = './state_dicts/deit_cifar_base_model.pt'
hf_model.model.load_state_dict(torch.load(model_checkpoint_path, map_location=device))
if data_type == "Example":
(x_train, y_train), (_, _), _, _ = load_dataset('cifar10')
x_train = np.transpose(x_train, (0, 3, 1, 2)).astype(np.float32)
y_train = np.argmax(y_train, axis=1)
classes = np.unique(y_train)
samples_per_class = 5
x_subset = []
y_subset = []
for c in classes:
indices = y_train == c
x_subset.append(x_train[indices][:samples_per_class])
y_subset.append(y_train[indices][:samples_per_class])
x_subset = np.concatenate(x_subset)
y_subset = np.concatenate(y_subset)
label_names = [
'airplane',
'automobile',
'bird',
'cat',
'deer',
'dog',
'frog',
'horse',
'ship',
'truck',
]
outputs = hf_model.predict(x_subset)
clean_preds = np.argmax(outputs, axis=1)
clean_acc = np.mean(clean_preds == y_subset)
gallery_out = []
for i, im in enumerate(x_subset):
gallery_out.append(( im.transpose(1,2,0), label_names[np.argmax(outputs[i])] ))
return gallery_out, clean_acc
# e.g. To use a local alternative theme: carbon_theme = Carbon()
carbon_theme = Carbon()
with gr.Blocks(css=css, theme=gr.themes.Base()) as demo:
import art
text = art.__version__
with gr.Row():
with gr.Column(scale=1):
gr.Image(value="./art_lfai.png", show_label=False, show_download_button=False, width=100)
with gr.Column(scale=20):
gr.Markdown(f"<h1>Red-teaming HuggingFace with ART (v{text})</h1>", elem_classes="plot-padding")
gr.Markdown('''This app guides you through a common workflow for assessing the robustness
of HuggingFace models using standard datasets and state-of-the-art adversarial attacks
found within the Adversarial Robustness Toolbox (ART).<br/><br/>Follow the instructions in each
step below to carry out your own evaluation and determine the risks associated with using
some of your favorite models! <b>#redteaming</b> <b>#trustworthyAI</b>''')
# Model and Dataset Selection
with gr.Accordion("1. Model selection", open=False):
gr.Markdown("Select a Hugging Face model to launch an adversarial attack against.")
model_type = gr.Radio(label="Hugging Face Model", choices=["Example", "Other"], value="Example")
with gr.Column(visible=False) as other_model:
model_url = gr.Text(label="Model URL",
placeholder="e.g. facebook/deit-tiny-distilled-patch16-224",
value='facebook/deit-tiny-distilled-patch16-224')
model_input_channels = gr.Text(label="Input channels", value=3)
model_input_height = gr.Text(label="Input height", value=32)
model_input_width = gr.Text(label="Input width", value=32)
model_num_classes = gr.Text(label="Number of classes", value=10)
model_clip_values = gr.Radio(label="Clip values", choices=[1, 255], value=1)
model_upsample_scaling = gr.Slider(label="Upsample scale factor", minimum=1, maximum=10, value=7)
model_type.change(show_params, model_type, other_model)
with gr.Accordion("2. Data selection", open=False):
gr.Markdown("This section enables you to select a dataset for evaluation or upload your own image.")
data_type = gr.Radio(label="Hugging Face dataset", choices=["Example", "URL", "Local"], value="Example")
with gr.Column(visible=False) as other_dataset:
gr.Markdown("Coming soon.")
data_type.change(show_params, data_type, other_dataset)
with gr.Accordion("3. Model inference", open=False):
with gr.Row():
with gr.Column(scale=1):
preds_gallery = gr.Gallery(label="Predictions", preview=False, show_download_button=True)
with gr.Column(scale=2):
clean_accuracy = gr.Number(label="Clean accuracy",
info="The accuracy achieved by the model in normal (non-adversarial) conditions.")
bt_run_inference = gr.Button("Run inference")
bt_clear = gr.ClearButton(components=[preds_gallery, clean_accuracy])
bt_run_inference.click(run_inference, inputs=[model_type, model_url, model_input_channels, model_input_height, model_input_width,
model_num_classes, model_clip_values, model_upsample_scaling, data_type],
outputs=[preds_gallery, clean_accuracy])
# Attack Selection
with gr.Accordion("4. Run attack", open=False):
gr.Markdown("In this section you can select the type of adversarial attack you wish to deploy against your selected model.")
with gr.Accordion("Evasion", open=False):
gr.Markdown("Evasion attacks are deployed to cause a model to incorrectly classify or detect items/objects in an image.")
with gr.Accordion("Projected Gradient Descent", open=False):
gr.Markdown("This attack uses PGD to identify adversarial examples.")
with gr.Row():
with gr.Column(scale=1):
attack = gr.Textbox(visible=True, value="PGD", label="Attack", interactive=False)
max_iter = gr.Slider(minimum=1, maximum=1000, label="Max iterations", value=10)
eps = gr.Slider(minimum=0.0001, maximum=255, label="Epslion", value=8/255)
eps_steps = gr.Slider(minimum=0.0001, maximum=255, label="Epsilon steps", value=1/255)
bt_eval_pgd = gr.Button("Evaluate")
# Evaluation Output. Visualisations of success/failures of running evaluation attacks.
with gr.Column(scale=3):
with gr.Row():
with gr.Column():
original_gallery = gr.Gallery(label="Original", preview=False, show_download_button=True)
benign_output = gr.Label(num_top_classes=3, visible=False)
clean_accuracy = gr.Number(label="Clean Accuracy", precision=2)
quality_plot = gr.LinePlot(label="Gradient Quality", x='iteration', y='value', color='metric',
x_title='Iteration', y_title='Avg in Gradients (%)',
caption="""Illustrates the average percent of zero, infinity
or NaN gradients identified in images
across all batches.""", elem_classes="plot-padding", visible=False)
with gr.Column():
adversarial_gallery = gr.Gallery(label="Adversarial", preview=False, show_download_button=True)
adversarial_output = gr.Label(num_top_classes=3, visible=False)
robust_accuracy = gr.Number(label="Robust Accuracy", precision=2)
with gr.Column():
delta_gallery = gr.Gallery(label="Added perturbation", preview=False, show_download_button=True)
bt_eval_pgd.click(clf_evasion_evaluate, inputs=[attack, model_type, model_url, model_input_channels, model_input_height, model_input_width,
model_num_classes, model_clip_values, model_upsample_scaling,
max_iter, eps, eps_steps, attack, attack, attack, attack, data_type],
outputs=[original_gallery, adversarial_gallery, delta_gallery, clean_accuracy,
robust_accuracy])
with gr.Accordion("Adversarial Patch", open=False):
gr.Markdown("This attack crafts an adversarial patch that facilitates evasion.")
with gr.Row():
with gr.Column(scale=1):
attack = gr.Textbox(visible=True, value="Adversarial Patch", label="Attack", interactive=False)
max_iter = gr.Slider(minimum=1, maximum=1000, label="Max iterations", value=10)
x_location = gr.Slider(minimum=1, maximum=32, label="Location (x)", value=1)
y_location = gr.Slider(minimum=1, maximum=32, label="Location (y)", value=1)
patch_height = gr.Slider(minimum=1, maximum=32, label="Patch height", value=12)
patch_width = gr.Slider(minimum=1, maximum=32, label="Patch width", value=12)
eval_btn_patch = gr.Button("Evaluate")
# Evaluation Output. Visualisations of success/failures of running evaluation attacks.
with gr.Column(scale=3):
with gr.Row():
with gr.Column():
original_gallery = gr.Gallery(label="Original", preview=False, show_download_button=True)
clean_accuracy = gr.Number(label="Clean Accuracy", precision=2)
with gr.Column():
adversarial_gallery = gr.Gallery(label="Adversarial", preview=False, show_download_button=True)
robust_accuracy = gr.Number(label="Robust Accuracy", precision=2)
with gr.Column():
delta_gallery = gr.Gallery(label="Patches", preview=False, show_download_button=True)
eval_btn_patch.click(clf_evasion_evaluate, inputs=[attack, model_type, model_url, model_input_channels, model_input_height, model_input_width,
model_num_classes, model_clip_values, model_upsample_scaling,
max_iter, eps, eps_steps, x_location, y_location, patch_height, patch_width, data_type],
outputs=[original_gallery, adversarial_gallery, delta_gallery, clean_accuracy,
robust_accuracy])
with gr.Accordion("Poisoning", open=False):
with gr.Accordion("Backdoor"):
with gr.Row():
with gr.Column(scale=1):
attack = gr.Textbox(visible=True, value="Backdoor", label="Attack", interactive=False)
target_class = gr.Radio(label="Target class", info="The class you wish to force the model to predict.",
choices=['dog',
'cassette player',
'chainsaw',
'church',
'french horn',
'garbage truck',
'gas pump',
'golf ball',
'parachutte',], value='dog')
trigger_image = gr.Image(label="Trigger Image", value="./baby-on-board.png")
eval_btn_patch = gr.Button("Evaluate")
with gr.Column(scale=2):
clean_gallery = gr.Gallery(label="Clean", preview=False, show_download_button=True)
clean_accuracy = gr.Number(label="Clean Accuracy", precision=2)
with gr.Column(scale=2):
poison_gallery = gr.Gallery(label="Poisoned", preview=False, show_download_button=True)
poison_success = gr.Number(label="Poison Success", precision=2)
eval_btn_patch.click(clf_poison_evaluate, inputs=[attack, model_type, trigger_image, target_class, data_type],
outputs=[clean_gallery, poison_gallery, clean_accuracy, poison_success])
if __name__ == "__main__":
# For development
'''demo.launch(show_api=False, debug=True, share=False,
server_name="0.0.0.0",
server_port=7777,
ssl_verify=False,
max_threads=20)'''
# For deployment
demo.launch(share=True, ssl_verify=False)