def run_attack(args, summary):
"""Run the attack."""
model = get_model(args.model, args.framework, summary)
distance = get_distance(args.distance)
criteria = get_criteria(args.criteria, args.target_class)
channel_axis = 'channels_first' if args.framework == 'pytorch' else 'channels_last'
image = get_image(args.image, args.framework, args.model, channel_axis)
if args.target_class is None and args.framework != 'cloud':
label = np.argmax(model.predictions(image))
elif args.framework == 'cloud':
label = model.predictions(image)
else:
label = args.target_class
metric = get_metric(args.metric, model, criteria, distance)
print(bcolors.BOLD + 'Process start' + bcolors.ENDC)
if args.model in ["yolo_v3", "keras_ssd300", "retina_resnet_50"]:
adversary = metric(image, unpack=False, binary_search_steps=1)
elif args.metric not in ["carlini_wagner_l2", "carlini_wagner_linf"]:
if args.metric in summary['verifiable_metrics']:
adversary = metric(image, label, unpack=False, verify=args.verify)
else:
adversary = metric(image, label, unpack=False, epsilons=1000)
else:
adversary = metric(image,
label,
unpack=False,
binary_search_steps=10,
max_iterations=5)
print(bcolors.BOLD + 'Process finished' + bcolors.ENDC)
if adversary.image is None:
print(bcolors.WARNING + 'Warning: Cannot find an adversary!' +
bcolors.ENDC)
return adversary
################### print summary info #####################################
keywords = [args.framework, args.model, args.criteria, args.metric]
if args.model not in ["yolo_v3", "keras_ssd300",
"retina_resnet_50"]: # classification
# interpret the label as human language
with open('perceptron/utils/labels.txt') as info:
imagenet_dict = eval(info.read())
if args.framework != 'cloud':
true_label = imagenet_dict[np.argmax(model.predictions(image))]
fake_label = imagenet_dict[np.argmax(
model.predictions(adversary.image))]
else:
print(args.framework)
true_label = str(model.predictions(image))
fake_label = str(model.predictions(image))
print(bcolors.HEADER + bcolors.UNDERLINE + 'Summary:' + bcolors.ENDC)
print('Configuration:' + bcolors.CYAN + ' --framework %s '
'--model %s --criterion %s '
'--metric %s' % tuple(keywords) + bcolors.ENDC)
print('The predicted label of original image is ' + bcolors.GREEN +
true_label + bcolors.ENDC)
print('The predicted label of adversary image is ' + bcolors.RED +
fake_label + bcolors.ENDC)
print('Minimum perturbation required: %s' % bcolors.BLUE +
str(adversary.distance) + bcolors.ENDC)
if args.metric in [
"brightness", "rotation", "horizontal_translation",
"vertical_translation"
]:
print('Verifiable bound: %s' % bcolors.BLUE +
str(adversary.verifiable_bounds) + bcolors.ENDC)
print('\n')
plot_image(adversary,
title=', '.join(keywords),
figname='examples/images/%s.png' % '_'.join(keywords))
else: # object detection
print(bcolors.HEADER + bcolors.UNDERLINE + 'Summary:' + bcolors.ENDC)
print('Configuration:' + bcolors.CYAN + ' --framework %s '
'--model %s --criterion %s '
'--metric %s' % tuple(keywords) + bcolors.ENDC)
print('Minimum perturbation required: %s' % bcolors.BLUE +
str(adversary.distance) + bcolors.ENDC)
print('\n')
plot_image_objectdetection(adversary,
model,
title=", ".join(keywords),
figname='examples/images/%s.png' %
'_'.join(keywords))
if args.framework == 'keras':
from perceptron.utils.func import clear_keras_session
clear_keras_session()
print(bcolors.BOLD + 'Process start' + bcolors.ENDC)
adversary = metric(image, unpack=False)
print(bcolors.BOLD + 'Process finished' + bcolors.ENDC)
if adversary.image is None:
print(bcolors.WARNING + 'Warning: Cannot find an adversary!' +
bcolors.ENDC)
exit(-1)
################### print summary info #####################################
keywords = ['Keras', 'SSD300', 'TargetClassMiss', 'BrightnessMetric']
print(bcolors.HEADER + bcolors.UNDERLINE + 'Summary:' + bcolors.ENDC)
print('Configuration:' + bcolors.CYAN + ' --framework %s '
'--model %s --criterion %s '
'--metric %s' % tuple(keywords) + bcolors.ENDC)
print('Minimum perturbation required: %s' % bcolors.BLUE +
str(adversary.distance) + bcolors.ENDC)
print('\n')
# print the original image and the adversary
plot_image_objectdetection(adversary,
kmodel,
bounds=(0, 255),
title=", ".join(keywords),
figname='examples/images/%s.png' %
'_'.join(keywords))