def attack_transform_overlap(
attack_name,
transform_name,
transforms,
reduce_mode="none",
result_dir="result/lenet/9transform_attack_overlap"):
name = attack_name + '_' + transform_name
lenet_mnist_class_trace = class_trace(
trace_name,
model_config=LENET,
data_config=data_config,
)
threshold = 0.5
# DeepFool will shutdown when num_gpu<0.2
num_gpus = 0.2
overlap_fn = calc_trace_side_overlap
per_channel = False
path = os.path.join(result_dir, f"{name}_overlap.csv")
# print(f"Computing {name}")
# lenet_overlap_ratio = attack_reduced_edge_detection_count_violation(
lenet_overlap_ratio = attack_reduced_edge_detection(
attack_name=attack_name,
attack_fn=attacks[attack_name][0],
generate_adversarial_fn=cw_generate_adversarial_example if
attack_name.startswith("CW") else foolbox_generate_adversarial_example,
class_trace_fn=lambda class_id: lenet_mnist_class_trace(
class_id,
threshold,
label=model_label,
trace_dir=trace_dir,
),
select_fn=lambda input: arg_approx(input, threshold),
overlap_fn=overlap_fn,
path=path,
per_channel=per_channel,
preprocessing=(0.1307, 0.3081),
image_size=28,
class_num=10,
norm_fn=mnist.normalize,
data_format="channels_first",
**(attacks[attack_name][1] if len(attacks[attack_name]) == 2 else {}),
images_per_class=images_per_class,
model_dir=model_dir,
num_gpus=num_gpus,
transforms=transforms,
transform_name=transform_name,
reduce_mode=reduce_mode,
)
lenet_overlap_ratio.save()
return lenet_overlap_ratio.load()
def transform_overlap(name, transforms=None):
label = "augmentation"
result_dir = "result/lenet/transform_overlap"
if not os.path.exists(result_dir):
os.makedirs(result_dir, exist_ok=True)
threshold = 0.5
overlap_fn = calc_trace_side_overlap
per_channel = False
path = os.path.join(result_dir, f"{name}_overlap.csv")
print(f"Computing {name}")
lenet_overlap_ratio = translation_overlap_ratio(
class_trace_fn=lambda class_id: lenet_mnist_class_trace(
class_id, threshold, label=label),
# class_trace_fn=lambda class_id: lenet_mnist_class_trace(class_id, threshold),
select_fn=lambda input: arg_approx(input, threshold),
overlap_fn=overlap_fn,
path=path,
per_channel=per_channel,
preprocessing=(0.1307, 0.3081),
image_size=28,
class_num=10,
norm_fn=mnist.normalize,
data_format="channels_first",
transforms=transforms,
name=name,
num_gpus=0.1,
images_per_class=100)
lenet_overlap_ratio.save(header=1)
return lenet_overlap_ratio.load(header=0, index_col=0)
def clean_overlap():
label = "dropout"
result_dir = "result/lenet"
images_per_class = 1
num_gpus = 0.2
threshold = 0.5
overlap_fn = calc_trace_side_overlap
# overlap_fn = calc_trace_side_overlap_both_compact
per_channel = False
# per_channel = True
path_template = (os.path.join(result_dir,
"class_overlap_ratio_{0:.1f}_{1}.csv"))
lenet_overlap_ratio = clean_overlap_ratio(
class_trace_fn=lambda class_id: lenet_mnist_class_trace(
class_id, threshold, label=label),
# class_trace_fn=lambda class_id: lenet_mnist_class_trace(class_id, threshold),
select_fn=lambda input: arg_approx(input, threshold),
overlap_fn=overlap_fn,
path=path_template.format(threshold, label),
per_channel=per_channel,
preprocessing=(0.1307, 0.3081),
image_size=28,
class_num=10,
norm_fn=mnist.normalize,
data_format="channels_first",
num_gpus=num_gpus,
images_per_class=images_per_class,
)
lenet_overlap_ratio.save()
summary_mean_path_template = os.path.join(
result_dir, "mean_overlap_ratio_summary_{threshold:.1f}.csv")
key = TraceKey.EDGE
# key = TraceKey.WEIGHT
overlay_ratio_mean = lenet_overlap_ratio.load().mean()
summary_file = summary_mean_path_template.format(
threshold=threshold,
label=label,
)
overlay_ratio_mean.to_csv(summary_file)
def visualize_features(
transforms,
result_dir = "result/lenet/model_augmentation/visualize"
):
name = attack_name
lenet_mnist_class_trace = class_trace(
trace_name,
model_config=LENET,
data_config=data_config,
)
threshold = 0.5
# DeepFool will shutdown when num_gpu<0.2
num_gpus = 0.2
overlap_fn = calc_trace_side_overlap
per_channel = False
path = os.path.join(result_dir, f"{name}_overlap.csv")
print(f"Computing {name}")
lenet_overlap_ratio = test_data_forward(
class_trace_fn=lambda class_id: lenet_mnist_class_trace(
class_id,
threshold,
label=model_label,
trace_dir = trace_dir,
),
select_fn=lambda input: arg_approx(input, threshold),
overlap_fn=overlap_fn,
path=path,
per_channel=per_channel,
preprocessing=(0.1307, 0.3081),
image_size=28,
class_num=10,
norm_fn=mnist.normalize,
data_format="channels_first",
**(attacks[attack_name][1] if len(attacks[attack_name]) == 2 else {}),
images_per_class=images_per_class,
model_dir=model_dir,
num_gpus = num_gpus,
transforms = transforms,
)
num_gpus = 0.5
per_channel = False
use_class_trace = True
if use_class_trace:
sum_precision_error = 0.2
threshold_bar_ratio = 0.01
select_fn = lambda input, output: arg_abs_approx(
input,
output,
sum_precision_error=sum_precision_error,
threshold_bar_ratio=threshold_bar_ratio,
)
else:
select_fn = lambda input: arg_approx(input, threshold)
# Compute the mean overlap ratio of attacked image
def save_trace(
class_id,
image_id,
select_fn: Callable[[np.ndarray], np.ndarray],
class_dir,
graph_dir,
create_model,
graph,
per_node: bool = False,
images_per_class: int = 1,
num_gpus: float = 1,
model_dir=resnet18_dir,
