def load(config, model, nb_epochs=100, cnndim=None, **kwargs):
from armory.utils.config_loading import load_dataset
model_config = config["model"]
if model in ["cnn", "cnn1d", "cnn2d"]:
classifier = cnn_model.make_audio_model(cnndim=cnndim, **kwargs)
preprocessing_fn = cnn_model.preprocessing_fn
elif model == "sincnet":
classifier, preprocessing_fn = make_sincnet_smooth_wrapper(**kwargs)
train_data = load_dataset(
config["dataset"],
epochs=nb_epochs,
split_type="train",
preprocessing_fn=preprocessing_fn,
)
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
)
val_data = load_dataset(
config["dataset"],
epochs=nb_epochs,
split_type="validation",
preprocessing_fn=preprocessing_fn,
)
classifier.to(DEVICE)
return classifier, (train_data, val_data, test_data)
def load_data(self, config_dataset, nb_epochs, preprocessing_fn):
train_data = load_dataset(
config_dataset,
epochs=nb_epochs,
split_type="train",
preprocessing_fn=preprocessing_fn,
)
test_data = load_dataset(
config_dataset,
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
)
val_data = load_dataset(
config_dataset,
epochs=nb_epochs,
split_type="validation",
preprocessing_fn=preprocessing_fn,
)
return train_data, val_data, test_data
def test_pytorch_xview_pretrained():
detector_module = import_module(
"armory.baseline_models.pytorch.xview_frcnn")
detector_fn = getattr(detector_module, "get_art_model")
weights_path = maybe_download_weights_from_s3(
"xview_model_state_dict_epoch_99_loss_0p67")
detector = detector_fn(
model_kwargs={},
wrapper_kwargs={},
weights_path=weights_path,
)
NUM_TEST_SAMPLES = 250
dataset_config = {
"batch_size": 1,
"framework": "numpy",
"module": "armory.data.datasets",
"name": "xview",
}
test_dataset = load_dataset(
dataset_config,
epochs=1,
split="test",
num_batches=NUM_TEST_SAMPLES,
shuffle_files=False,
)
list_of_ys = []
list_of_ypreds = []
for x, y in test_dataset:
y_pred = detector.predict(x)
list_of_ys.extend(y)
list_of_ypreds.extend(y_pred)
average_precision_by_class = object_detection_AP_per_class(
list_of_ys, list_of_ypreds)
mAP = np.fromiter(average_precision_by_class.values(), dtype=float).mean()
for class_id in [4, 23, 33, 39]:
assert average_precision_by_class[class_id] > 0.9
assert mAP > 0.25
def _evaluate(self, config: dict) -> dict:
"""
Evaluate a config file for classification robustness against attack.
"""
model_config = config["model"]
# Scenario assumes preprocessing_fn makes images all same size
classifier, preprocessing_fn = load_model(model_config)
config_adhoc = config.get("adhoc") or {}
train_epochs = config_adhoc["train_epochs"]
src_class = config_adhoc["source_class"]
tgt_class = config_adhoc["target_class"]
fit_batch_size = config_adhoc.get("fit_batch_size",
config["dataset"]["batch_size"])
# Set random seed due to large variance in attack and defense success
np.random.seed(config_adhoc["np_seed"])
set_random_seed(config_adhoc["tf_seed"])
use_poison_filtering_defense = config_adhoc.get(
"use_poison_filtering_defense", True)
if self.check_run:
# filtering defense requires more than a single batch to run properly
use_poison_filtering_defense = False
logger.info(f"Loading dataset {config['dataset']['name']}...")
clean_data = load_dataset(
config["dataset"],
epochs=1,
split_type="train",
preprocessing_fn=preprocessing_fn,
shuffle_files=False,
)
attack_config = config["attack"]
attack_type = attack_config.get("type")
if attack_type == "preloaded":
num_images_tgt_class = config_adhoc["num_images_target_class"]
logger.info(
f"Loading poison dataset {config_adhoc['poison_samples']['name']}..."
)
num_poisoned = int(config_adhoc["fraction_poisoned"] *
num_images_tgt_class)
if num_poisoned == 0:
raise ValueError(
"For the preloaded attack, fraction_poisoned must be set so that at least on data point is poisoned."
)
config_adhoc["poison_samples"]["batch_size"] = num_poisoned
poison_data = load_dataset(
config["adhoc"]["poison_samples"],
epochs=1,
split_type="poison",
preprocessing_fn=None,
)
else:
attack = load(attack_config)
logger.info(
"Building in-memory dataset for poisoning detection and training")
fraction_poisoned = config["adhoc"]["fraction_poisoned"]
poison_dataset_flag = config["adhoc"]["poison_dataset"]
# detect_poison does not currently support data generators
# therefore, make in memory dataset
x_train_all, y_train_all = [], []
if attack_type == "preloaded":
for x_clean, y_clean in clean_data:
x_poison, y_poison = poison_data.get_batch()
x_poison = np.array([xp for xp in x_poison], dtype=np.float)
x_train_all.append(x_clean)
y_train_all.append(y_clean)
x_train_all.append(x_poison)
y_train_all.append(y_poison)
x_train_all = np.concatenate(x_train_all, axis=0)
y_train_all = np.concatenate(y_train_all, axis=0)
else:
for x_train, y_train in clean_data:
x_train_all.append(x_train)
y_train_all.append(y_train)
x_train_all = np.concatenate(x_train_all, axis=0)
y_train_all = np.concatenate(y_train_all, axis=0)
if poison_dataset_flag:
total_count = np.bincount(y_train_all)[src_class]
poison_count = int(fraction_poisoned * total_count)
if poison_count == 0:
logger.warning(
f"No poisons generated with fraction_poisoned {fraction_poisoned} for class {src_class}."
)
src_indices = np.where(y_train_all == src_class)[0]
poisoned_indices = np.random.choice(src_indices,
size=poison_count,
replace=False)
x_train_all, y_train_all = poison_dataset(
x_train_all,
y_train_all,
src_class,
tgt_class,
y_train_all.shape[0],
attack,
poisoned_indices,
)
y_train_all_categorical = to_categorical(y_train_all)
if use_poison_filtering_defense:
defense_config = config["defense"]
defense_model_config = config_adhoc.get("defense_model",
model_config)
defense_train_epochs = config_adhoc.get("defense_train_epochs",
train_epochs)
classifier_for_defense, _ = load_model(defense_model_config)
logger.info(
f"Fitting model {defense_model_config['module']}.{defense_model_config['name']} "
f"for defense {defense_config['name']}...")
classifier_for_defense.fit(
x_train_all,
y_train_all_categorical,
batch_size=fit_batch_size,
nb_epochs=defense_train_epochs,
verbose=False,
)
defense_fn = load_fn(defense_config)
defense = defense_fn(classifier_for_defense, x_train_all,
y_train_all_categorical)
_, is_clean = defense.detect_poison(nb_clusters=2,
nb_dims=43,
reduce="PCA")
is_clean = np.array(is_clean)
logger.info(f"Total clean data points: {np.sum(is_clean)}")
logger.info("Filtering out detected poisoned samples")
indices_to_keep = is_clean == 1
x_train_final = x_train_all[indices_to_keep]
y_train_final = y_train_all_categorical[indices_to_keep]
else:
logger.info(
"Defense does not require filtering. Model fitting will use all data."
)
x_train_final = x_train_all
y_train_final = y_train_all_categorical
if len(x_train_final):
logger.info(
f"Fitting model of {model_config['module']}.{model_config['name']}..."
)
classifier.fit(
x_train_final,
y_train_final,
batch_size=fit_batch_size,
nb_epochs=train_epochs,
verbose=False,
)
else:
logger.warning(
"All data points filtered by defense. Skipping training")
logger.info("Validating on clean test data")
config["dataset"]["batch_size"] = fit_batch_size
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
shuffle_files=False,
)
validation_metric = metrics.MetricList("categorical_accuracy")
target_class_benign_metric = metrics.MetricList("categorical_accuracy")
for x, y in tqdm(test_data, desc="Testing"):
y_pred = classifier.predict(x)
validation_metric.append(y, y_pred)
y_pred_tgt_class = y_pred[y == src_class]
if len(y_pred_tgt_class):
target_class_benign_metric.append(
[src_class] * len(y_pred_tgt_class), y_pred_tgt_class)
logger.info(
f"Unpoisoned validation accuracy: {validation_metric.mean():.2%}")
logger.info(
f"Unpoisoned validation accuracy on targeted class: {target_class_benign_metric.mean():.2%}"
)
results = {
"validation_accuracy":
validation_metric.mean(),
"validation_accuracy_targeted_class":
target_class_benign_metric.mean(),
}
test_metric = metrics.MetricList("categorical_accuracy")
targeted_test_metric = metrics.MetricList("categorical_accuracy")
logger.info("Testing on poisoned test data")
if attack_type == "preloaded":
test_data_poison = load_dataset(
config_adhoc["poison_samples"],
epochs=1,
split_type="poison_test",
preprocessing_fn=None,
)
for x_poison_test, y_poison_test in tqdm(test_data_poison,
desc="Testing poison"):
x_poison_test = np.array([xp for xp in x_poison_test],
dtype=np.float)
y_pred = classifier.predict(x_poison_test)
y_true = [src_class] * len(y_pred)
targeted_test_metric.append(y_poison_test, y_pred)
test_metric.append(y_true, y_pred)
test_data_clean = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
shuffle_files=False,
)
for x_clean_test, y_clean_test in tqdm(test_data_clean,
desc="Testing clean"):
x_clean_test = np.array([xp for xp in x_clean_test],
dtype=np.float)
y_pred = classifier.predict(x_clean_test)
test_metric.append(y_clean_test, y_pred)
elif poison_dataset_flag:
logger.info("Testing on poisoned test data")
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
shuffle_files=False,
)
for x_test, y_test in tqdm(test_data, desc="Testing"):
src_indices = np.where(y_test == src_class)[0]
poisoned_indices = src_indices # Poison entire class
x_test, _ = poison_dataset(
x_test,
y_test,
src_class,
tgt_class,
len(y_test),
attack,
poisoned_indices,
)
y_pred = classifier.predict(x_test)
test_metric.append(y_test, y_pred)
y_pred_targeted = y_pred[y_test == src_class]
if not len(y_pred_targeted):
continue
targeted_test_metric.append([tgt_class] * len(y_pred_targeted),
y_pred_targeted)
if poison_dataset_flag or attack_type == "preloaded":
results["test_accuracy"] = test_metric.mean()
results[
"targeted_misclassification_accuracy"] = targeted_test_metric.mean(
)
logger.info(f"Test accuracy: {test_metric.mean():.2%}")
logger.info(
f"Test targeted misclassification accuracy: {targeted_test_metric.mean():.2%}"
)
return results
def _evaluate(self, config: dict, num_eval_batches: Optional[int],
skip_benign: Optional[bool]) -> dict:
"""
Evaluate the config and return a results dict
"""
model_config = config["model"]
classifier, preprocessing_fn = load_model(model_config)
defense_config = config.get("defense") or {}
defense_type = defense_config.get("type")
if defense_type in ["Preprocessor", "Postprocessor"]:
logger.info(
f"Applying internal {defense_type} defense to classifier")
classifier = load_defense_internal(config["defense"], classifier)
if model_config["fit"]:
classifier.set_learning_phase(True)
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
train_epochs = config["model"]["fit_kwargs"]["nb_epochs"]
batch_size = config["dataset"]["batch_size"]
logger.info(
f"Loading train dataset {config['dataset']['name']}...")
train_data = load_dataset(
config["dataset"],
epochs=train_epochs,
split_type="train",
preprocessing_fn=preprocessing_fn,
shuffle_files=True,
)
if defense_type == "Trainer":
logger.info(f"Training with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
else:
logger.info("Fitting classifier on clean train dataset...")
for epoch in range(train_epochs):
classifier.set_learning_phase(True)
for _ in tqdm(
range(train_data.batches_per_epoch),
desc=f"Epoch: {epoch}/{train_epochs}",
):
x, y = train_data.get_batch()
# x_trains consists of one or more videos, each represented as an
# ndarray of shape (n_stacks, 3, 16, 112, 112).
# To train, randomly sample a batch of stacks
x = np.stack(
[x_i[np.random.randint(x_i.shape[0])] for x_i in x])
if defense_type == "Trainer":
defense.fit(x, y, batch_size=batch_size, nb_epochs=1)
else:
classifier.fit(x,
y,
batch_size=batch_size,
nb_epochs=1)
if defense_type == "Transform":
# NOTE: Transform currently not supported
logger.info(
f"Transforming classifier with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
classifier = defense()
classifier.set_learning_phase(False)
metrics_logger = metrics.MetricsLogger.from_config(
config["metric"], skip_benign=skip_benign)
if skip_benign:
logger.info("Skipping benign classification...")
else:
# Evaluate the ART classifier on benign test examples
logger.info(f"Loading test dataset {config['dataset']['name']}...")
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
num_batches=num_eval_batches,
shuffle_files=False,
)
logger.info("Running inference on benign examples...")
for x_batch, y_batch in tqdm(test_data, desc="Benign"):
for x, y in zip(x_batch, y_batch):
# combine predictions across all stacks
with metrics.resource_context(
name="Inference",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.
computational_resource_dict,
):
y_pred = np.mean(classifier.predict(x, batch_size=1),
axis=0)
metrics_logger.update_task(y, y_pred)
metrics_logger.log_task()
# Evaluate the ART classifier on adversarial test examples
logger.info("Generating or loading / testing adversarial examples...")
attack_config = config["attack"]
attack_type = attack_config.get("type")
targeted = bool(attack_config.get("kwargs", {}).get("targeted"))
if targeted and attack_config.get("use_label"):
raise ValueError("Targeted attacks cannot have 'use_label'")
if attack_type == "preloaded":
test_data = load_adversarial_dataset(
attack_config,
epochs=1,
split_type="adversarial",
preprocessing_fn=preprocessing_fn,
num_batches=num_eval_batches,
shuffle_files=False,
)
else:
attack = load_attack(attack_config, classifier)
if targeted != getattr(attack, "targeted", False):
logger.warning(
f"targeted config {targeted} != attack field {getattr(attack, 'targeted', False)}"
)
attack.set_params(batch_size=1)
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
num_batches=num_eval_batches,
shuffle_files=False,
)
if targeted:
label_targeter = load_label_targeter(
attack_config["targeted_labels"])
for x_batch, y_batch in tqdm(test_data, desc="Attack"):
if attack_type == "preloaded":
x_batch = list(zip(*x_batch))
if targeted:
y_batch = list(zip(*y_batch))
for x, y in zip(x_batch, y_batch):
with metrics.resource_context(
name="Attack",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.
computational_resource_dict,
):
if attack_type == "preloaded":
x, x_adv = x
if targeted:
y, y_target = y
else:
# each x is of shape (n_stack, 3, 16, 112, 112)
# n_stack varies
if attack_config.get("use_label"):
# expansion required due to preprocessing
y_input = np.repeat(y, x.shape[0])
x_adv = attack.generate(x=x, y=y_input)
elif targeted:
y_target = label_targeter.generate(y)
y_input = np.repeat(y_target, x.shape[0])
x_adv = attack.generate(x=x, y=y_input)
else:
x_adv = attack.generate(x=x)
# combine predictions across all stacks
y_pred_adv = np.mean(classifier.predict(x_adv, batch_size=1),
axis=0)
if targeted:
metrics_logger.update_task(y_target,
y_pred_adv,
adversarial=True)
else:
metrics_logger.update_task(y, y_pred_adv, adversarial=True)
metrics_logger.update_perturbation([x], [x_adv])
metrics_logger.log_task(adversarial=True, targeted=targeted)
return metrics_logger.results()
def _evaluate(self, config: dict) -> dict:
"""
Evaluate the config and return a results dict
"""
model_config = config["model"]
classifier, preprocessing_fn = load_model(model_config)
defense_config = config.get("defense") or {}
defense_type = defense_config.get("type")
if defense_type in ["Preprocessor", "Postprocessor"]:
logger.info(f"Applying internal {defense_type} defense to classifier")
classifier = load_defense_internal(config["defense"], classifier)
if model_config["fit"]:
classifier.set_learning_phase(True)
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
fit_kwargs = model_config["fit_kwargs"]
logger.info(f"Loading train dataset {config['dataset']['name']}...")
train_data = load_dataset(
config["dataset"],
epochs=fit_kwargs["nb_epochs"],
split_type="train",
preprocessing_fn=preprocessing_fn,
)
if defense_type == "Trainer":
logger.info(f"Training with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
defense.fit_generator(train_data, **fit_kwargs)
else:
logger.info(f"Fitting classifier on clean train dataset...")
classifier.fit_generator(train_data, **fit_kwargs)
if defense_type == "Transform":
# NOTE: Transform currently not supported
logger.info(f"Transforming classifier with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
classifier = defense()
classifier.set_learning_phase(False)
# Evaluate the ART classifier on benign test examples
logger.info(f"Loading test dataset {config['dataset']['name']}...")
test_data_generator = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
)
logger.info("Running inference on benign examples...")
metrics_logger = metrics.MetricsLogger.from_config(config["metric"])
for x, y in tqdm(test_data_generator, desc="Benign"):
y_pred = classifier.predict(x)
metrics_logger.update_task(y, y_pred)
metrics_logger.log_task()
# Evaluate the ART classifier on adversarial test examples
logger.info("Generating / testing adversarial examples...")
attack = load_attack(config["attack"], classifier)
test_data_generator = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
)
for x, y in tqdm(test_data_generator, desc="Attack"):
x_adv = attack.generate(x=x)
y_pred_adv = classifier.predict(x_adv)
metrics_logger.update_task(y, y_pred_adv, adversarial=True)
metrics_logger.update_perturbation(x, x_adv)
metrics_logger.log_task(adversarial=True)
return metrics_logger.results()
def _evaluate(
self,
config: dict,
num_eval_batches: Optional[int],
skip_benign: Optional[bool],
skip_attack: Optional[bool],
skip_misclassified: Optional[bool],
) -> dict:
"""
Evaluate a config file for classification robustness against attack.
Note: num_eval_batches shouldn't be set for poisoning scenario and will raise an
error if it is
"""
if config["sysconfig"].get("use_gpu"):
os.environ["TF_CUDNN_DETERMINISM"] = "1"
if num_eval_batches:
raise ValueError(
"num_eval_batches shouldn't be set for poisoning scenario")
if skip_benign:
raise ValueError(
"skip_benign shouldn't be set for poisoning scenario")
if skip_attack:
raise ValueError(
"skip_attack shouldn't be set for poisoning scenario")
if skip_misclassified:
raise ValueError(
"skip_misclassified shouldn't be set for poisoning scenario")
model_config = config["model"]
# Scenario assumes canonical preprocessing_fn is used makes images all same size
classifier, _ = load_model(model_config)
config_adhoc = config.get("adhoc") or {}
train_epochs = config_adhoc["train_epochs"]
src_class = config_adhoc["source_class"]
tgt_class = config_adhoc["target_class"]
fit_batch_size = config_adhoc.get("fit_batch_size",
config["dataset"]["batch_size"])
if not config["sysconfig"].get("use_gpu"):
conf = ConfigProto(intra_op_parallelism_threads=1)
set_session(Session(config=conf))
# Set random seed due to large variance in attack and defense success
np.random.seed(config_adhoc["split_id"])
set_random_seed(config_adhoc["split_id"])
random.seed(config_adhoc["split_id"])
use_poison_filtering_defense = config_adhoc.get(
"use_poison_filtering_defense", True)
if self.check_run:
# filtering defense requires more than a single batch to run properly
use_poison_filtering_defense = False
logger.info(f"Loading dataset {config['dataset']['name']}...")
clean_data = load_dataset(
config["dataset"],
epochs=1,
split=config["dataset"].get("train_split", "train"),
preprocessing_fn=poison_scenario_preprocessing,
shuffle_files=False,
)
attack_config = config["attack"]
attack_type = attack_config.get("type")
fraction_poisoned = config["adhoc"]["fraction_poisoned"]
# Flag for whether to poison dataset -- used to evaluate
# performance of defense on clean data
poison_dataset_flag = config["adhoc"]["poison_dataset"]
# detect_poison does not currently support data generators
# therefore, make in memory dataset
x_train_all, y_train_all = [], []
if attack_type == "preloaded":
# Number of datapoints in train split of target clasc
num_images_tgt_class = config_adhoc["num_images_target_class"]
logger.info(
f"Loading poison dataset {config_adhoc['poison_samples']['name']}..."
)
num_poisoned = int(config_adhoc["fraction_poisoned"] *
num_images_tgt_class)
if num_poisoned == 0:
raise ValueError(
"For the preloaded attack, fraction_poisoned must be set so that at least on data point is poisoned."
)
# Set batch size to number of poisons -- read only one batch of preloaded poisons
config_adhoc["poison_samples"]["batch_size"] = num_poisoned
poison_data = load_dataset(
config["adhoc"]["poison_samples"],
epochs=1,
split="poison",
preprocessing_fn=None,
)
logger.info(
"Building in-memory dataset for poisoning detection and training"
)
for x_clean, y_clean in clean_data:
x_train_all.append(x_clean)
y_train_all.append(y_clean)
x_poison, y_poison = poison_data.get_batch()
x_poison = np.array([xp for xp in x_poison], dtype=np.float32)
x_train_all.append(x_poison)
y_train_all.append(y_poison)
x_train_all = np.concatenate(x_train_all, axis=0)
y_train_all = np.concatenate(y_train_all, axis=0)
else:
attack = load(attack_config)
logger.info(
"Building in-memory dataset for poisoning detection and training"
)
for x_train, y_train in clean_data:
x_train_all.append(x_train)
y_train_all.append(y_train)
x_train_all = np.concatenate(x_train_all, axis=0)
y_train_all = np.concatenate(y_train_all, axis=0)
if poison_dataset_flag:
total_count = np.bincount(y_train_all)[src_class]
poison_count = int(fraction_poisoned * total_count)
if poison_count == 0:
logger.warning(
f"No poisons generated with fraction_poisoned {fraction_poisoned} for class {src_class}."
)
src_indices = np.where(y_train_all == src_class)[0]
poisoned_indices = np.sort(
np.random.choice(src_indices,
size=poison_count,
replace=False))
x_train_all, y_train_all = poison_dataset(
x_train_all,
y_train_all,
src_class,
tgt_class,
y_train_all.shape[0],
attack,
poisoned_indices,
)
y_train_all_categorical = to_categorical(y_train_all)
# Flag to determine whether defense_classifier is trained directly
# (default API) or is trained as part of detect_poisons method
fit_defense_classifier_outside_defense = config_adhoc.get(
"fit_defense_classifier_outside_defense", True)
# Flag to determine whether defense_classifier uses sparse
# or categorical labels
defense_categorical_labels = config_adhoc.get(
"defense_categorical_labels", True)
if use_poison_filtering_defense:
if defense_categorical_labels:
y_train_defense = y_train_all_categorical
else:
y_train_defense = y_train_all
defense_config = config["defense"]
detection_kwargs = config_adhoc.get("detection_kwargs", dict())
defense_model_config = config_adhoc.get("defense_model",
model_config)
defense_train_epochs = config_adhoc.get("defense_train_epochs",
train_epochs)
# Assumes classifier_for_defense and classifier use same preprocessing function
classifier_for_defense, _ = load_model(defense_model_config)
logger.info(
f"Fitting model {defense_model_config['module']}.{defense_model_config['name']} "
f"for defense {defense_config['name']}...")
if fit_defense_classifier_outside_defense:
classifier_for_defense.fit(
x_train_all,
y_train_defense,
batch_size=fit_batch_size,
nb_epochs=defense_train_epochs,
verbose=False,
shuffle=True,
)
defense_fn = load_fn(defense_config)
defense = defense_fn(classifier_for_defense, x_train_all,
y_train_defense)
_, is_clean = defense.detect_poison(**detection_kwargs)
is_clean = np.array(is_clean)
logger.info(f"Total clean data points: {np.sum(is_clean)}")
logger.info("Filtering out detected poisoned samples")
indices_to_keep = is_clean == 1
x_train_final = x_train_all[indices_to_keep]
y_train_final = y_train_all_categorical[indices_to_keep]
else:
logger.info(
"Defense does not require filtering. Model fitting will use all data."
)
x_train_final = x_train_all
y_train_final = y_train_all_categorical
if len(x_train_final):
logger.info(
f"Fitting model of {model_config['module']}.{model_config['name']}..."
)
classifier.fit(
x_train_final,
y_train_final,
batch_size=fit_batch_size,
nb_epochs=train_epochs,
verbose=False,
shuffle=True,
)
else:
logger.warning(
"All data points filtered by defense. Skipping training")
logger.info("Validating on clean test data")
test_data = load_dataset(
config["dataset"],
epochs=1,
split=config["dataset"].get("eval_split", "test"),
preprocessing_fn=poison_scenario_preprocessing,
shuffle_files=False,
)
benign_validation_metric = metrics.MetricList("categorical_accuracy")
target_class_benign_metric = metrics.MetricList("categorical_accuracy")
for x, y in tqdm(test_data, desc="Testing"):
# Ensure that input sample isn't overwritten by classifier
x.flags.writeable = False
y_pred = classifier.predict(x)
benign_validation_metric.add_results(y, y_pred)
y_pred_tgt_class = y_pred[y == src_class]
if len(y_pred_tgt_class):
target_class_benign_metric.add_results(
[src_class] * len(y_pred_tgt_class), y_pred_tgt_class)
logger.info(
f"Unpoisoned validation accuracy: {benign_validation_metric.mean():.2%}"
)
logger.info(
f"Unpoisoned validation accuracy on targeted class: {target_class_benign_metric.mean():.2%}"
)
results = {
"benign_validation_accuracy":
benign_validation_metric.mean(),
"benign_validation_accuracy_targeted_class":
target_class_benign_metric.mean(),
}
poisoned_test_metric = metrics.MetricList("categorical_accuracy")
poisoned_targeted_test_metric = metrics.MetricList(
"categorical_accuracy")
logger.info("Testing on poisoned test data")
if attack_type == "preloaded":
test_data_poison = load_dataset(
config_adhoc["poison_samples"],
epochs=1,
split="poison_test",
preprocessing_fn=None,
)
for x_poison_test, y_poison_test in tqdm(test_data_poison,
desc="Testing poison"):
x_poison_test = np.array([xp for xp in x_poison_test],
dtype=np.float32)
y_pred = classifier.predict(x_poison_test)
y_true = [src_class] * len(y_pred)
poisoned_targeted_test_metric.add_results(
y_poison_test, y_pred)
poisoned_test_metric.add_results(y_true, y_pred)
test_data_clean = load_dataset(
config["dataset"],
epochs=1,
split=config["dataset"].get("eval_split", "test"),
preprocessing_fn=poison_scenario_preprocessing,
shuffle_files=False,
)
for x_clean_test, y_clean_test in tqdm(test_data_clean,
desc="Testing clean"):
x_clean_test = np.array([xp for xp in x_clean_test],
dtype=np.float32)
y_pred = classifier.predict(x_clean_test)
poisoned_test_metric.add_results(y_clean_test, y_pred)
elif poison_dataset_flag:
logger.info("Testing on poisoned test data")
test_data = load_dataset(
config["dataset"],
epochs=1,
split=config["dataset"].get("eval_split", "test"),
preprocessing_fn=poison_scenario_preprocessing,
shuffle_files=False,
)
for x_test, y_test in tqdm(test_data, desc="Testing"):
src_indices = np.where(y_test == src_class)[0]
poisoned_indices = src_indices # Poison entire class
x_test, _ = poison_dataset(
x_test,
y_test,
src_class,
tgt_class,
len(y_test),
attack,
poisoned_indices,
)
y_pred = classifier.predict(x_test)
poisoned_test_metric.add_results(y_test, y_pred)
y_pred_targeted = y_pred[y_test == src_class]
if not len(y_pred_targeted):
continue
poisoned_targeted_test_metric.add_results(
[tgt_class] * len(y_pred_targeted), y_pred_targeted)
if poison_dataset_flag or attack_type == "preloaded":
results["poisoned_test_accuracy"] = poisoned_test_metric.mean()
results[
"poisoned_targeted_misclassification_accuracy"] = poisoned_targeted_test_metric.mean(
)
logger.info(f"Test accuracy: {poisoned_test_metric.mean():.2%}")
logger.info(
f"Test targeted misclassification accuracy: {poisoned_targeted_test_metric.mean():.2%}"
)
return results
def _evaluate(
self,
config: dict,
num_eval_batches: Optional[int],
skip_benign: Optional[bool],
skip_attack: Optional[bool],
) -> dict:
"""
Evaluate the config and return a results dict
"""
model_config = config["model"]
estimator, _ = load_model(model_config)
defense_config = config.get("defense") or {}
defense_type = defense_config.get("type")
if defense_type in ["Preprocessor", "Postprocessor"]:
logger.info(f"Applying internal {defense_type} defense to estimator")
estimator = load_defense_internal(config["defense"], estimator)
attack_config = config["attack"]
attack_channels = attack_config.get("generate_kwargs", {}).get("channels")
if attack_channels is None:
if self.attack_modality == "sar":
logger.info("No mask configured. Attacking all SAR channels")
attack_channels = range(4)
elif self.attack_modality == "eo":
logger.info("No mask configured. Attacking all EO channels")
attack_channels = range(4, 14)
elif self.attack_modality == "both":
logger.info("No mask configured. Attacking all SAR and EO channels")
attack_channels = range(14)
else:
assert isinstance(
attack_channels, list
), "Mask is specified, but incorrect format. Expected list"
attack_channels = np.array(attack_channels)
if self.attack_modality == "sar":
assert np.all(
np.logical_and(attack_channels >= 0, attack_channels < 4)
), "Selected SAR-only attack modality, but specify non-SAR channels"
elif self.attack_modality == "eo":
assert np.all(
np.logical_and(attack_channels >= 4, attack_channels < 14)
), "Selected EO-only attack modality, but specify non-EO channels"
elif self.attack_modality == "both":
assert np.all(
np.logical_and(attack_channels >= 0, attack_channels < 14)
), "Selected channels are out-of-bounds"
if model_config["fit"]:
try:
estimator.set_learning_phase(True)
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
fit_kwargs = model_config["fit_kwargs"]
logger.info(f"Loading train dataset {config['dataset']['name']}...")
train_data = load_dataset(
config["dataset"],
epochs=fit_kwargs["nb_epochs"],
split=config["dataset"].get("train_split", "train"),
shuffle_files=True,
)
if defense_type == "Trainer":
logger.info(f"Training with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], estimator)
defense.fit_generator(train_data, **fit_kwargs)
else:
logger.info("Fitting estimator on clean train dataset...")
estimator.fit_generator(train_data, **fit_kwargs)
except NotImplementedError:
raise NotImplementedError(
"Training has not yet been implemented for object detectors"
)
if defense_type == "Transform":
# NOTE: Transform currently not supported
logger.info(f"Transforming estimator with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], estimator)
estimator = defense()
try:
estimator.set_learning_phase(False)
except NotImplementedError:
logger.warning(
"Unable to set estimator's learning phase. As of ART 1.4.1, "
"this is not yet supported for object detectors."
)
attack_type = attack_config.get("type")
targeted = bool(attack_config.get("kwargs", {}).get("targeted"))
performance_metrics = deepcopy(config["metric"])
performance_metrics.pop("perturbation")
performance_logger = metrics.MetricsLogger.from_config(
performance_metrics,
skip_benign=skip_benign,
skip_attack=skip_attack,
targeted=targeted,
)
eval_split = config["dataset"].get("eval_split", "test")
if skip_benign:
logger.info("Skipping benign classification...")
else:
# Evaluate the ART estimator on benign test examples
logger.info(f"Loading test dataset {config['dataset']['name']}...")
test_data = load_dataset(
config["dataset"],
epochs=1,
split=eval_split,
num_batches=num_eval_batches,
shuffle_files=False,
)
logger.info("Running inference on benign examples...")
for x, y in tqdm(test_data, desc="Benign"):
# Ensure that input sample isn't overwritten by estimator
x.flags.writeable = False
with metrics.resource_context(
name="Inference",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=performance_logger.computational_resource_dict,
):
y_pred = estimator.predict(x)
performance_logger.update_task(y, y_pred)
performance_logger.log_task()
if skip_attack:
logger.info("Skipping attack generation...")
return performance_logger.results()
# Evaluate the ART estimator on adversarial test examples
logger.info("Generating or loading / testing adversarial examples...")
perturbation_metrics = deepcopy(config["metric"])
perturbation_metrics.pop("task")
if self.attack_modality in ("sar", "both"):
sar_perturbation_logger = metrics.MetricsLogger.from_config(
perturbation_metrics,
skip_benign=True,
skip_attack=False,
targeted=targeted,
)
else:
sar_perturbation_logger = None
if self.attack_modality in ("eo", "both"):
eo_perturbation_logger = metrics.MetricsLogger.from_config(
perturbation_metrics,
skip_benign=True,
skip_attack=False,
targeted=targeted,
)
else:
eo_perturbation_logger = None
if targeted and attack_config.get("use_label"):
raise ValueError("Targeted attacks cannot have 'use_label'")
if attack_type == "preloaded":
test_data = load_adversarial_dataset(
attack_config,
epochs=1,
split="adversarial",
num_batches=num_eval_batches,
shuffle_files=False,
)
else:
attack = load_attack(attack_config, estimator)
if targeted != getattr(attack, "targeted", False):
logger.warning(
f"targeted config {targeted} != attack field {getattr(attack, 'targeted', False)}"
)
test_data = load_dataset(
config["dataset"],
epochs=1,
split=eval_split,
num_batches=num_eval_batches,
shuffle_files=False,
)
if targeted:
label_targeter = load_label_targeter(attack_config["targeted_labels"])
export_samples = config["scenario"].get("export_samples")
if export_samples is not None and export_samples > 0:
sample_exporter = SampleExporter(
self.scenario_output_dir, test_data.context, export_samples
)
else:
sample_exporter = None
for x, y in tqdm(test_data, desc="Attack"):
with metrics.resource_context(
name="Attack",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=performance_logger.computational_resource_dict,
):
if attack_type == "preloaded":
logger.warning(
"Specified preloaded attack. Ignoring `attack_modality` parameter"
)
if len(x) == 2:
x, x_adv = x
else:
x_adv = x
if targeted:
y, y_target = y
else:
generate_kwargs = deepcopy(attack_config.get("generate_kwargs", {}))
generate_kwargs["mask"] = attack_channels
if attack_config.get("use_label"):
generate_kwargs["y"] = y
elif targeted:
y_target = label_targeter.generate(y)
generate_kwargs["y"] = y_target
x_adv = attack.generate(x=x, **generate_kwargs)
# Ensure that input sample isn't overwritten by estimator
x_adv.flags.writeable = False
y_pred_adv = estimator.predict(x_adv)
performance_logger.update_task(y, y_pred_adv, adversarial=True)
if targeted:
performance_logger.update_task(
y_target, y_pred_adv, adversarial=True, targeted=True
)
# Update perturbation metrics for SAR/EO separately
x_sar = np.stack(
(x[..., 0] + 1j * x[..., 1], x[..., 2] + 1j * x[..., 3]), axis=3
)
x_adv_sar = np.stack(
(
x_adv[..., 0] + 1j * x_adv[..., 1],
x_adv[..., 2] + 1j * x_adv[..., 3],
),
axis=3,
)
x_eo = x[..., 4:]
x_adv_eo = x_adv[..., 4:]
if sar_perturbation_logger is not None:
sar_perturbation_logger.update_perturbation(x_sar, x_adv_sar)
if eo_perturbation_logger is not None:
eo_perturbation_logger.update_perturbation(x_eo, x_adv_eo)
if sample_exporter is not None:
sample_exporter.export(x, x_adv, y, y_pred_adv)
performance_logger.log_task(adversarial=True)
if targeted:
performance_logger.log_task(adversarial=True, targeted=True)
# Merge performance, SAR, EO results
combined_results = performance_logger.results()
if sar_perturbation_logger is not None:
combined_results.update(
{f"sar_{k}": v for k, v in sar_perturbation_logger.results().items()}
)
if eo_perturbation_logger is not None:
combined_results.update(
{f"eo_{k}": v for k, v in eo_perturbation_logger.results().items()}
)
return combined_results
def _evaluate(
self,
config: dict,
num_eval_batches: Optional[int],
skip_benign: Optional[bool],
skip_attack: Optional[bool],
) -> dict:
"""
Evaluate the config and return a results dict
"""
if config["dataset"]["batch_size"] != 1:
raise ValueError(
"batch_size must be 1 for evaluation, due to variable length inputs.\n"
" If training, set config['model']['fit_kwargs']['fit_batch_size']"
)
model_config = config["model"]
classifier, fit_preprocessing_fn = load_model(model_config)
defense_config = config.get("defense") or {}
defense_type = defense_config.get("type")
if defense_type in ["Preprocessor", "Postprocessor"]:
logger.info(f"Applying internal {defense_type} defense to classifier")
classifier = load_defense_internal(config["defense"], classifier)
if model_config["fit"]:
classifier.set_learning_phase(True)
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
fit_kwargs = model_config["fit_kwargs"]
logger.info(f"Loading train dataset {config['dataset']['name']}...")
batch_size = config["dataset"].pop("batch_size")
config["dataset"]["batch_size"] = fit_kwargs.get(
"fit_batch_size", batch_size
)
train_data = load_dataset(
config["dataset"],
epochs=fit_kwargs["nb_epochs"],
split=config["dataset"].get("train_split", "train"),
preprocessing_fn=fit_preprocessing_fn,
shuffle_files=True,
)
config["dataset"]["batch_size"] = batch_size
if defense_type == "Trainer":
logger.info(f"Training with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
defense.fit_generator(train_data, **fit_kwargs)
else:
logger.info("Fitting classifier on clean train dataset...")
classifier.fit_generator(train_data, **fit_kwargs)
if defense_type == "Transform":
# NOTE: Transform currently not supported
logger.info(f"Transforming classifier with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
classifier = defense()
classifier.set_learning_phase(False)
attack_config = config["attack"]
attack_type = attack_config.get("type")
targeted = bool(attack_config.get("kwargs", {}).get("targeted"))
metrics_logger = metrics.MetricsLogger.from_config(
config["metric"],
skip_benign=skip_benign,
skip_attack=skip_attack,
targeted=targeted,
)
if config["dataset"]["batch_size"] != 1:
logger.warning("Evaluation batch_size != 1 may not be supported.")
eval_split = config["dataset"].get("eval_split", "test")
if skip_benign:
logger.info("Skipping benign classification...")
else:
# Evaluate the ART classifier on benign test examples
logger.info(f"Loading test dataset {config['dataset']['name']}...")
test_data = load_dataset(
config["dataset"],
epochs=1,
split=eval_split,
num_batches=num_eval_batches,
shuffle_files=False,
)
logger.info("Running inference on benign examples...")
for x, y in tqdm(test_data, desc="Benign"):
# Ensure that input sample isn't overwritten by classifier
x.flags.writeable = False
with metrics.resource_context(
name="Inference",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.computational_resource_dict,
):
y_pred = classifier.predict(x)
metrics_logger.update_task(y, y_pred)
metrics_logger.log_task()
if skip_attack:
logger.info("Skipping attack generation...")
return metrics_logger.results()
# Evaluate the ART classifier on adversarial test examples
logger.info("Generating or loading / testing adversarial examples...")
if targeted and attack_config.get("use_label"):
raise ValueError("Targeted attacks cannot have 'use_label'")
if attack_type == "preloaded":
test_data = load_adversarial_dataset(
attack_config,
epochs=1,
split="adversarial",
num_batches=num_eval_batches,
shuffle_files=False,
)
else:
attack = load_attack(attack_config, classifier)
if targeted != getattr(attack, "targeted", False):
logger.warning(
f"targeted config {targeted} != attack field {getattr(attack, 'targeted', False)}"
)
test_data = load_dataset(
config["dataset"],
epochs=1,
split=eval_split,
num_batches=num_eval_batches,
shuffle_files=False,
)
if targeted:
label_targeter = load_label_targeter(attack_config["targeted_labels"])
export_samples = config["scenario"].get("export_samples")
if export_samples is not None and export_samples > 0:
sample_exporter = SampleExporter(
self.scenario_output_dir, test_data.context, export_samples
)
else:
sample_exporter = None
for x, y in tqdm(test_data, desc="Attack"):
with metrics.resource_context(
name="Attack",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.computational_resource_dict,
):
if attack_type == "preloaded":
x, x_adv = x
if targeted:
y, y_target = y
else:
generate_kwargs = deepcopy(attack_config.get("generate_kwargs", {}))
if attack_config.get("use_label"):
generate_kwargs["y"] = y
elif targeted:
y_target = label_targeter.generate(y)
generate_kwargs["y"] = y_target
x_adv = attack.generate(x=x, **generate_kwargs)
# Ensure that input sample isn't overwritten by classifier
x_adv.flags.writeable = False
y_pred_adv = classifier.predict(x_adv)
metrics_logger.update_task(y, y_pred_adv, adversarial=True)
if targeted:
metrics_logger.update_task(
y_target, y_pred_adv, adversarial=True, targeted=True
)
metrics_logger.update_perturbation(x, x_adv)
if sample_exporter is not None:
sample_exporter.export(x, x_adv, y, y_pred_adv)
metrics_logger.log_task(adversarial=True)
if targeted:
metrics_logger.log_task(adversarial=True, targeted=True)
return metrics_logger.results()
def _evaluate(self, config: dict) -> dict:
"""
Evaluate the config and return a results dict
"""
model_config = config["model"]
classifier, preprocessing_fn = load_model(model_config)
if isinstance(preprocessing_fn, tuple):
fit_preprocessing_fn, predict_preprocessing_fn = preprocessing_fn
else:
fit_preprocessing_fn = predict_preprocessing_fn = preprocessing_fn
defense_config = config.get("defense") or {}
defense_type = defense_config.get("type")
if defense_type in ["Preprocessor", "Postprocessor"]:
logger.info(
f"Applying internal {defense_type} defense to classifier")
classifier = load_defense_internal(config["defense"], classifier)
if model_config["fit"]:
classifier.set_learning_phase(True)
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
fit_kwargs = model_config["fit_kwargs"]
logger.info(
f"Loading train dataset {config['dataset']['name']}...")
batch_size = config["dataset"].pop("batch_size")
config["dataset"]["batch_size"] = config.get("adhoc", {}).get(
"fit_batch_size", batch_size)
train_data = load_dataset(
config["dataset"],
epochs=fit_kwargs["nb_epochs"],
split_type="train",
preprocessing_fn=fit_preprocessing_fn,
)
config["dataset"]["batch_size"] = batch_size
if defense_type == "Trainer":
logger.info(f"Training with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
defense.fit_generator(train_data, **fit_kwargs)
else:
logger.info("Fitting classifier on clean train dataset...")
classifier.fit_generator(train_data, **fit_kwargs)
if defense_type == "Transform":
# NOTE: Transform currently not supported
logger.info(
f"Transforming classifier with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
classifier = defense()
#HACK: to save model -- currently commenting it out
#SAIL-JATI ----------------------------------
#ts = time.time()
#st = datetime.datetime.fromtimestamp(ts).strftime('%Y-%m-%d-%H-%M-%S')
#model_save_dir_ = os.path.join("/nas/home/ajati/work/codes/SAIL_ALR_models/", st+"/")
#os.system("mkdir -p "+model_save_dir_)
#torch.save(classifier._model._model.state_dict(), model_save_dir_+"/sail_alr_model_state_dict.pt")
#torch.save(classifier._model._model, model_save_dir_+"/sail_alr_model.pt")
#torch.save(classifier._optimizer.state_dict(), model_save_dir_+"/sail_alr_optim_state_dict.pt")
#torch.save(classifier._optimizer, model_save_dir_+"/sail_alr_optim.pt")
##-------------------------------------------
classifier.set_learning_phase(False)
# Evaluate the ART classifier on benign test examples
logger.info(f"Loading test dataset {config['dataset']['name']}...")
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=predict_preprocessing_fn,
)
logger.info("Running inference on benign examples...")
metrics_logger = metrics.MetricsLogger.from_config(config["metric"])
for x, y in tqdm(test_data, desc="Benign"):
y_pred = classifier.predict(x)
metrics_logger.update_task(y, y_pred)
metrics_logger.log_task()
# Evaluate the ART classifier on adversarial test examples
logger.info("Generating or loading / testing adversarial examples...")
attack_config = config["attack"]
attack_type = attack_config.get("type")
targeted = bool(attack_config.get("kwargs", {}).get("targeted"))
if targeted and attack_config.get("use_label"):
raise ValueError("Targeted attacks cannot have 'use_label'")
if attack_type == "preloaded":
test_data = load_adversarial_dataset(
attack_config,
epochs=1,
split_type="adversarial",
preprocessing_fn=predict_preprocessing_fn,
)
else:
attack = load_attack(attack_config, classifier)
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=predict_preprocessing_fn,
)
#JATI -- snr
snrs = []
for x, y in tqdm(test_data, desc="Attack"):
if attack_type == "preloaded":
x, x_adv = x
if targeted:
y, y_target = y
elif attack_config.get("use_label"):
x_adv = attack.generate(x=x, y=y)
elif targeted:
raise NotImplementedError(
"Requires generation of target labels")
# x_adv = attack.generate(x=x, y=y_target)
else:
x_adv = attack.generate(x=x)
#JATI - snr
noise = x_adv - x
snr = 10 * np.log10(np.mean(x**2) / np.mean(noise**2))
snrs.append(snr)
y_pred_adv = classifier.predict(x_adv)
if targeted:
# NOTE: does not remove data points where y == y_target
metrics_logger.update_task(y_target,
y_pred_adv,
adversarial=True)
else:
metrics_logger.update_task(y, y_pred_adv, adversarial=True)
metrics_logger.update_perturbation(x, x_adv)
metrics_logger.log_task(adversarial=True, targeted=targeted)
mean_snr = np.mean(snrs)
logging.info(f"MEAN SNR of adversarial samples = {mean_snr}")
return metrics_logger.results()
def _evaluate(self, config: dict, num_eval_batches: Optional[int],
skip_benign: Optional[bool]) -> dict:
"""
Evaluate the config and return a results dict
"""
model_config = config["model"]
classifier, preprocessing_fn = load_model(model_config)
defense_config = config.get("defense") or {}
defense_type = defense_config.get("type")
if defense_type in ["Preprocessor", "Postprocessor"]:
logger.info(
f"Applying internal {defense_type} defense to classifier")
classifier = load_defense_internal(config["defense"], classifier)
if model_config["fit"]:
classifier.set_learning_phase(True)
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
fit_kwargs = model_config["fit_kwargs"]
logger.info(
f"Loading train dataset {config['dataset']['name']}...")
train_data = load_dataset(
config["dataset"],
epochs=fit_kwargs["nb_epochs"],
split_type="train",
preprocessing_fn=preprocessing_fn,
shuffle_files=True,
)
if defense_type == "Trainer":
logger.info(f"Training with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
defense.fit_generator(train_data, **fit_kwargs)
else:
logger.info("Fitting classifier on clean train dataset...")
classifier.fit_generator(train_data, **fit_kwargs)
if defense_type == "Transform":
# NOTE: Transform currently not supported
logger.info(
f"Transforming classifier with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
classifier = defense()
classifier.set_learning_phase(False)
metrics_logger = metrics.MetricsLogger.from_config(
config["metric"], skip_benign=skip_benign)
if skip_benign:
logger.info("Skipping benign classification...")
else:
# Evaluate the ART classifier on benign test examples
logger.info(f"Loading test dataset {config['dataset']['name']}...")
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
num_batches=num_eval_batches,
shuffle_files=False,
)
logger.info("Running inference on benign examples...")
for x, y in tqdm(test_data, desc="Benign"):
with metrics.resource_context(
name="Inference",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.
computational_resource_dict,
):
y_pred = classifier.predict(x)
metrics_logger.update_task(y, y_pred)
metrics_logger.log_task()
# Evaluate the ART classifier on adversarial test examples
logger.info("Generating or loading / testing adversarial examples...")
attack_config = config["attack"]
attack_type = attack_config.get("type")
targeted = bool(attack_config.get("kwargs", {}).get("targeted"))
if targeted and attack_config.get("use_label"):
raise ValueError("Targeted attacks cannot have 'use_label'")
if attack_type == "preloaded":
test_data = load_adversarial_dataset(
attack_config,
epochs=1,
split_type="adversarial",
preprocessing_fn=preprocessing_fn,
num_batches=num_eval_batches,
shuffle_files=False,
)
else:
attack = load_attack(attack_config, classifier)
if targeted != getattr(attack, "targeted", False):
logger.warning(
f"targeted config {targeted} != attack field {getattr(attack, 'targeted', False)}"
)
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
num_batches=num_eval_batches,
shuffle_files=False,
)
if targeted:
label_targeter = load_label_targeter(
attack_config["targeted_labels"])
for x, y in tqdm(test_data, desc="Attack"):
with metrics.resource_context(
name="Attack",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.
computational_resource_dict,
):
if attack_type == "preloaded":
x, x_adv = x
if targeted:
y, y_target = y
elif attack_config.get("use_label"):
x_adv = attack.generate(x=x, y=y)
elif targeted:
y_target = label_targeter.generate(y)
x_adv = attack.generate(x=x, y=y_target)
else:
x_adv = attack.generate(x=x)
y_pred_adv = classifier.predict(x_adv)
if targeted:
metrics_logger.update_task(y_target,
y_pred_adv,
adversarial=True)
else:
metrics_logger.update_task(y, y_pred_adv, adversarial=True)
metrics_logger.update_perturbation(x, x_adv)
metrics_logger.log_task(adversarial=True, targeted=targeted)
return metrics_logger.results()
def _evaluate(
self,
config: dict,
num_eval_batches: Optional[int],
skip_benign: Optional[bool],
skip_attack: Optional[bool],
skip_misclassified: Optional[bool],
) -> dict:
"""
Evaluate the config and return a results dict
"""
if skip_misclassified:
raise ValueError(
"skip_misclassified shouldn't be set for D-APRICOT scenario")
if skip_attack:
raise ValueError(
"--skip-attack should not be set for D-APRICOT scenario.")
if skip_benign:
logger.warning("--skip-benign is being ignored since the D-APRICOT"
" scenario doesn't include benign evaluation.")
attack_config = config["attack"]
attack_type = attack_config.get("type")
if attack_type == "preloaded":
raise ValueError(
"D-APRICOT scenario should not have preloaded set to True in attack config"
)
elif "targeted_labels" not in attack_config:
raise ValueError(
"Attack config must have 'targeted_labels' key, as the "
"D-APRICOT threat model is targeted.")
elif attack_config.get("use_label"):
raise ValueError(
"The D-APRICOT scenario threat model is targeted, and"
" thus 'use_label' should be set to false.")
if config["dataset"].get("batch_size") != 1:
raise ValueError(
"batch_size of 1 is required for D-APRICOT scenario")
model_config = config["model"]
estimator, _ = load_model(model_config)
defense_config = config.get("defense") or {}
defense_type = defense_config.get("type")
label_targeter = load_label_targeter(attack_config["targeted_labels"])
if defense_type in ["Preprocessor", "Postprocessor"]:
logger.info(
f"Applying internal {defense_type} defense to estimator")
estimator = load_defense_internal(config["defense"], estimator)
if model_config["fit"]:
try:
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
fit_kwargs = model_config["fit_kwargs"]
logger.info(
f"Loading train dataset {config['dataset']['name']}...")
train_data = load_dataset(
config["dataset"],
epochs=fit_kwargs["nb_epochs"],
split=config["dataset"].get("train_split", "train"),
shuffle_files=True,
)
if defense_type == "Trainer":
logger.info(f"Training with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"],
estimator)
defense.fit_generator(train_data, **fit_kwargs)
else:
logger.info("Fitting estimator on clean train dataset...")
estimator.fit_generator(train_data, **fit_kwargs)
except NotImplementedError:
raise NotImplementedError(
"Training has not yet been implemented for object detectors"
)
if defense_type == "Transform":
# NOTE: Transform currently not supported
logger.info(
f"Transforming estimator with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], estimator)
estimator = defense()
metrics_logger = metrics.MetricsLogger.from_config(
config["metric"],
skip_benign=True,
skip_attack=False,
targeted=True,
)
eval_split = config["dataset"].get("eval_split", "test")
# Evaluate the ART estimator on adversarial test examples
logger.info("Generating or loading / testing adversarial examples...")
attack = load_attack(attack_config, estimator)
test_data = load_dataset(
config["dataset"],
epochs=1,
split=eval_split,
num_batches=num_eval_batches,
shuffle_files=False,
)
export_samples = config["scenario"].get("export_samples")
if export_samples is not None and export_samples > 0:
sample_exporter = SampleExporter(self.scenario_output_dir,
test_data.context, export_samples)
else:
sample_exporter = None
for x, y in tqdm(test_data, desc="Attack"):
with metrics.resource_context(
name="Attack",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.
computational_resource_dict,
):
if x.shape[0] != 1:
raise ValueError("D-APRICOT batch size must be set to 1")
# (nb=1, num_cameras, h, w, c) --> (num_cameras, h, w, c)
x = x[0]
y_object, y_patch_metadata = y
generate_kwargs = deepcopy(
attack_config.get("generate_kwargs", {}))
generate_kwargs["y_patch_metadata"] = y_patch_metadata
y_target = label_targeter.generate(y_object)
generate_kwargs["y_object"] = y_target
x_adv = attack.generate(x=x, **generate_kwargs)
# Ensure that input sample isn't overwritten by estimator
x_adv.flags.writeable = False
y_pred_adv = estimator.predict(x_adv)
for img_idx in range(len(y_object)):
y_i_target = y_target[img_idx]
y_i_pred = y_pred_adv[img_idx]
metrics_logger.update_task([y_i_target], [y_i_pred],
adversarial=True,
targeted=True)
metrics_logger.update_perturbation(x, x_adv)
if sample_exporter is not None:
sample_exporter.export(x, x_adv, y_object, y_pred_adv)
metrics_logger.log_task(adversarial=True, targeted=True)
return metrics_logger.results()
def _evaluate(
self,
config: dict,
num_eval_batches: Optional[int],
skip_benign: Optional[bool],
skip_attack: Optional[bool],
skip_misclassified: Optional[bool],
) -> dict:
"""
Evaluate the config and return a results dict
"""
if skip_misclassified:
raise ValueError(
"skip_misclassified shouldn't be set for ASR scenario")
model_config = config["model"]
estimator, fit_preprocessing_fn = load_model(model_config)
audio_channel_config = config.get("adhoc", {}).get("audio_channel")
if audio_channel_config is not None:
logger.info("loading audio channel")
for k in "delay", "attenuation":
if k not in audio_channel_config:
raise ValueError(f"audio_channel must have key {k}")
audio_channel = load_audio_channel(**audio_channel_config)
if estimator.preprocessing_defences:
estimator.preprocessing_defences.insert(0, audio_channel)
else:
estimator.preprocessing_defences = [audio_channel]
estimator._update_preprocessing_operations()
defense_config = config.get("defense") or {}
defense_type = defense_config.get("type")
if defense_type in ["Preprocessor", "Postprocessor"]:
logger.info(
f"Applying internal {defense_type} defense to estimator")
estimator = load_defense_internal(config["defense"], estimator)
if model_config["fit"]:
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
fit_kwargs = model_config["fit_kwargs"]
logger.info(
f"Loading train dataset {config['dataset']['name']}...")
batch_size = config["dataset"].pop("batch_size")
config["dataset"]["batch_size"] = fit_kwargs.get(
"fit_batch_size", batch_size)
train_data = load_dataset(
config["dataset"],
epochs=fit_kwargs["nb_epochs"],
split=config["dataset"].get("train_split", "train_clean100"),
preprocessing_fn=fit_preprocessing_fn,
shuffle_files=True,
)
config["dataset"]["batch_size"] = batch_size
if defense_type == "Trainer":
logger.info(f"Training with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], estimator)
defense.fit_generator(train_data, **fit_kwargs)
else:
logger.info("Fitting estimator on clean train dataset...")
estimator.fit_generator(train_data, **fit_kwargs)
if defense_type == "Transform":
# NOTE: Transform currently not supported
logger.info(
f"Transforming estimator with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], estimator)
estimator = defense()
attack_config = config["attack"]
attack_type = attack_config.get("type")
targeted = bool(attack_config.get("targeted"))
metrics_logger = metrics.MetricsLogger.from_config(
config["metric"],
skip_benign=skip_benign,
skip_attack=skip_attack,
targeted=targeted,
)
if config["dataset"]["batch_size"] != 1:
logger.warning("Evaluation batch_size != 1 may not be supported.")
predict_kwargs = config["model"].get("predict_kwargs", {})
eval_split = config["dataset"].get("eval_split", "test_clean")
if skip_benign:
logger.info("Skipping benign classification...")
else:
# Evaluate the ART estimator on benign test examples
logger.info(f"Loading test dataset {config['dataset']['name']}...")
test_data = load_dataset(
config["dataset"],
epochs=1,
split=eval_split,
num_batches=num_eval_batches,
shuffle_files=False,
)
logger.info("Running inference on benign examples...")
for x, y in tqdm(test_data, desc="Benign"):
# Ensure that input sample isn't overwritten by estimator
x.flags.writeable = False
with metrics.resource_context(
name="Inference",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.
computational_resource_dict,
):
y_pred = estimator.predict(x, **predict_kwargs)
metrics_logger.update_task(y, y_pred)
metrics_logger.log_task()
if skip_attack:
logger.info("Skipping attack generation...")
return metrics_logger.results()
# Imperceptible attack still WIP
if (config.get("adhoc") or {}).get("skip_adversarial"):
logger.info("Skipping adversarial classification...")
return metrics_logger.results()
# Evaluate the ART estimator on adversarial test examples
logger.info("Generating or loading / testing adversarial examples...")
if attack_type == "preloaded":
test_data = load_adversarial_dataset(
attack_config,
epochs=1,
split="adversarial",
num_batches=num_eval_batches,
shuffle_files=False,
)
else:
attack = load_attack(attack_config, estimator)
if targeted != attack.targeted:
logger.warning(
f"targeted config {targeted} != attack field {attack.targeted}"
)
test_data = load_dataset(
config["dataset"],
epochs=1,
split=eval_split,
num_batches=num_eval_batches,
shuffle_files=False,
)
if targeted:
label_targeter = load_label_targeter(
attack_config["targeted_labels"])
export_samples = config["scenario"].get("export_samples")
if export_samples is not None and export_samples > 0:
sample_exporter = SampleExporter(self.scenario_output_dir,
test_data.context, export_samples)
else:
sample_exporter = None
for x, y in tqdm(test_data, desc="Attack"):
with metrics.resource_context(
name="Attack",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.
computational_resource_dict,
):
if attack_type == "preloaded":
x, x_adv = x
if targeted:
y, y_target = y
elif attack_config.get("use_label"):
x_adv = attack.generate(x=x, y=y)
elif targeted:
y_target = label_targeter.generate(y)
x_adv = attack.generate(x=x, y=y_target)
else:
x_adv = attack.generate(x=x)
# Ensure that input sample isn't overwritten by estimator
x_adv.flags.writeable = False
y_pred_adv = estimator.predict(x_adv, **predict_kwargs)
metrics_logger.update_task(y, y_pred_adv, adversarial=True)
if targeted:
metrics_logger.update_task(
y_target,
y_pred_adv,
adversarial=True,
targeted=True,
)
metrics_logger.update_perturbation(x, x_adv)
if sample_exporter is not None:
sample_exporter.export(x, x_adv, y, y_pred_adv)
metrics_logger.log_task(adversarial=True)
if targeted:
metrics_logger.log_task(adversarial=True, targeted=True)
return metrics_logger.results()
def _evaluate(
self,
config: dict,
num_eval_batches: Optional[int],
skip_benign: Optional[bool],
skip_attack: Optional[bool],
) -> dict:
"""
Evaluate a config file for classification robustness against attack.
Note: num_eval_batches shouldn't be set for poisoning scenario and will raise an
error if it is
"""
if config["sysconfig"].get("use_gpu"):
os.environ["TF_CUDNN_DETERMINISM"] = "1"
if num_eval_batches:
raise ValueError(
"num_eval_batches shouldn't be set for poisoning scenario")
if skip_benign:
raise ValueError(
"skip_benign shouldn't be set for poisoning scenario")
if skip_attack:
raise ValueError(
"skip_attack shouldn't be set for poisoning scenario")
model_config = config["model"]
# Scenario assumes canonical preprocessing_fn is used makes images all same size
classifier, _ = load_model(model_config)
proxy_classifier, _ = load_model(model_config)
config_adhoc = config.get("adhoc") or {}
train_epochs = config_adhoc["train_epochs"]
src_class = config_adhoc["source_class"]
tgt_class = config_adhoc["target_class"]
fit_batch_size = config_adhoc.get("fit_batch_size",
config["dataset"]["batch_size"])
if not config["sysconfig"].get("use_gpu"):
conf = ConfigProto(intra_op_parallelism_threads=1)
set_session(Session(config=conf))
# Set random seed due to large variance in attack and defense success
np.random.seed(config_adhoc["split_id"])
set_random_seed(config_adhoc["split_id"])
random.seed(config_adhoc["split_id"])
use_poison_filtering_defense = config_adhoc.get(
"use_poison_filtering_defense", True)
if self.check_run:
# filtering defense requires more than a single batch to run properly
use_poison_filtering_defense = False
logger.info(f"Loading dataset {config['dataset']['name']}...")
clean_data = load_dataset(
config["dataset"],
epochs=1,
split=config["dataset"].get("train_split", "train"),
preprocessing_fn=poison_scenario_preprocessing,
shuffle_files=False,
)
# Flag for whether to poison dataset -- used to evaluate
# performance of defense on clean data
poison_dataset_flag = config["adhoc"]["poison_dataset"]
# detect_poison does not currently support data generators
# therefore, make in memory dataset
x_train_all, y_train_all = [], []
logger.info(
"Building in-memory dataset for poisoning detection and training")
for x_train, y_train in clean_data:
x_train_all.append(x_train)
y_train_all.append(y_train)
x_train_all = np.concatenate(x_train_all, axis=0)
y_train_all = np.concatenate(y_train_all, axis=0)
if poison_dataset_flag:
y_train_all_categorical = to_categorical(y_train_all)
attack_train_epochs = train_epochs
attack_config = deepcopy(config["attack"])
use_adversarial_trainer_flag = attack_config.get(
"use_adversarial_trainer", False)
proxy_classifier_fit_kwargs = {
"batch_size": fit_batch_size,
"nb_epochs": attack_train_epochs,
}
logger.info("Fitting proxy classifier...")
if use_adversarial_trainer_flag:
logger.info("Using adversarial trainer...")
adversarial_trainer_kwargs = attack_config.pop(
"adversarial_trainer_kwargs", {})
for k, v in proxy_classifier_fit_kwargs.items():
adversarial_trainer_kwargs[k] = v
proxy_classifier = AdversarialTrainerMadryPGD(
proxy_classifier, **adversarial_trainer_kwargs)
proxy_classifier.fit(x_train_all, y_train_all)
attack_config["kwargs"][
"proxy_classifier"] = proxy_classifier.get_classifier()
else:
proxy_classifier_fit_kwargs["verbose"] = False
proxy_classifier_fit_kwargs["shuffle"] = True
proxy_classifier.fit(x_train_all, y_train_all,
**proxy_classifier_fit_kwargs)
attack_config["kwargs"]["proxy_classifier"] = proxy_classifier
attack, backdoor = load(attack_config)
x_train_all, y_train_all_categorical = attack.poison(
x_train_all, y_train_all_categorical)
y_train_all = np.argmax(y_train_all_categorical, axis=1)
if use_poison_filtering_defense:
y_train_defense = to_categorical(y_train_all)
defense_config = config["defense"]
detection_kwargs = config_adhoc.get("detection_kwargs", dict())
defense_model_config = config_adhoc.get("defense_model",
model_config)
# Assumes classifier_for_defense and classifier use same preprocessing function
classifier_for_defense, _ = load_model(defense_model_config)
# ART/Armory API requires that classifier_for_defense trains inside defense_fn
defense_fn = load_fn(defense_config)
defense = defense_fn(classifier_for_defense, x_train_all,
y_train_defense)
_, is_clean = defense.detect_poison(**detection_kwargs)
is_clean = np.array(is_clean)
logger.info(f"Total clean data points: {np.sum(is_clean)}")
logger.info("Filtering out detected poisoned samples")
indices_to_keep = is_clean == 1
x_train_final = x_train_all[indices_to_keep]
y_train_final = y_train_all[indices_to_keep]
else:
logger.info(
"Defense does not require filtering. Model fitting will use all data."
)
x_train_final = x_train_all
y_train_final = y_train_all
if len(x_train_final):
logger.info(
f"Fitting model of {model_config['module']}.{model_config['name']}..."
)
classifier.fit(
x_train_final,
y_train_final,
batch_size=fit_batch_size,
nb_epochs=train_epochs,
verbose=False,
shuffle=True,
)
else:
logger.warning(
"All data points filtered by defense. Skipping training")
logger.info("Validating on clean test data")
test_data = load_dataset(
config["dataset"],
epochs=1,
split=config["dataset"].get("eval_split", "test"),
preprocessing_fn=poison_scenario_preprocessing,
shuffle_files=False,
)
benign_validation_metric = metrics.MetricList("categorical_accuracy")
target_class_benign_metric = metrics.MetricList("categorical_accuracy")
for x, y in tqdm(test_data, desc="Testing"):
# Ensure that input sample isn't overwritten by classifier
x.flags.writeable = False
y_pred = classifier.predict(x)
benign_validation_metric.append(y, y_pred)
y_pred_tgt_class = y_pred[y == src_class]
if len(y_pred_tgt_class):
target_class_benign_metric.append(
[src_class] * len(y_pred_tgt_class), y_pred_tgt_class)
logger.info(
f"Unpoisoned validation accuracy: {benign_validation_metric.mean():.2%}"
)
logger.info(
f"Unpoisoned validation accuracy on targeted class: {target_class_benign_metric.mean():.2%}"
)
results = {
"benign_validation_accuracy":
benign_validation_metric.mean(),
"benign_validation_accuracy_targeted_class":
target_class_benign_metric.mean(),
}
poisoned_test_metric = metrics.MetricList("categorical_accuracy")
poisoned_targeted_test_metric = metrics.MetricList(
"categorical_accuracy")
if poison_dataset_flag:
logger.info("Testing on poisoned test data")
test_data = load_dataset(
config["dataset"],
epochs=1,
split=config["dataset"].get("eval_split", "test"),
preprocessing_fn=poison_scenario_preprocessing,
shuffle_files=False,
)
for x_test, y_test in tqdm(test_data, desc="Testing"):
src_indices = np.where(y_test == src_class)[0]
poisoned_indices = src_indices # Poison entire class
x_test, _ = poison_dataset(
x_test,
y_test,
src_class,
tgt_class,
len(y_test),
backdoor,
poisoned_indices,
)
y_pred = classifier.predict(x_test)
poisoned_test_metric.append(y_test, y_pred)
y_pred_targeted = y_pred[y_test == src_class]
if len(y_pred_targeted):
poisoned_targeted_test_metric.append(
[tgt_class] * len(y_pred_targeted), y_pred_targeted)
results["poisoned_test_accuracy"] = poisoned_test_metric.mean()
results[
"poisoned_targeted_misclassification_accuracy"] = poisoned_targeted_test_metric.mean(
)
logger.info(f"Test accuracy: {poisoned_test_metric.mean():.2%}")
logger.info(
f"Test targeted misclassification accuracy: {poisoned_targeted_test_metric.mean():.2%}"
)
return results
def _evaluate(self, config: dict) -> dict:
"""
Evaluate the config and return a results dict
"""
model_config = config["model"]
classifier, preprocessing_fn = load_model(model_config)
defense_config = config.get("defense") or {}
defense_type = defense_config.get("type")
if defense_type in ["Preprocessor", "Postprocessor"]:
logger.info(
f"Applying internal {defense_type} defense to classifier")
classifier = load_defense_internal(config["defense"], classifier)
if model_config["fit"]:
classifier.set_learning_phase(True)
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
fit_kwargs = model_config["fit_kwargs"]
logger.info(
f"Loading train dataset {config['dataset']['name']}...")
train_data = load_dataset(
config["dataset"],
epochs=fit_kwargs["nb_epochs"],
split_type="train",
preprocessing_fn=preprocessing_fn,
)
if defense_type == "Trainer":
logger.info(f"Training with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
defense.fit_generator(train_data, **fit_kwargs)
else:
logger.info("Fitting classifier on clean train dataset...")
classifier.fit_generator(train_data, **fit_kwargs)
################################################################
#### Save weights at the end of training
################################################################
ckpt_name = model_config['module'].replace('.', '_')
ckpt_name += '_pretrained' if model_config['model_kwargs'][
'pretrained'] else ''
ckpt_name += '_epochs%d.pth' % model_config['fit_kwargs'][
'nb_epochs']
classifier.save(
osp.join(paths.runtime_paths().saved_model_dir, ckpt_name))
logger.info(f"Saved classifier {ckpt_name} ...")
if defense_type == "Transform":
# NOTE: Transform currently not supported
logger.info(
f"Transforming classifier with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
classifier = defense()
classifier.set_learning_phase(False)
# Evaluate the ART classifier on benign test examples
logger.info(f"Loading test dataset {config['dataset']['name']}...")
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
)
logger.info("Running inference on benign examples...")
metrics_logger = metrics.MetricsLogger.from_config(config["metric"])
for x, y in tqdm(test_data, desc="Benign"):
y_pred = classifier.predict(x)
metrics_logger.update_task(y, y_pred)
metrics_logger.log_task()
# Evaluate the ART classifier on adversarial test examples
logger.info("Generating or loading / testing adversarial examples...")
attack_config = config["attack"]
attack_type = attack_config.get("type")
targeted = bool(attack_config.get("kwargs", {}).get("targeted"))
if targeted and attack_config.get("use_label"):
raise ValueError("Targeted attacks cannot have 'use_label'")
if attack_type == "preloaded":
test_data = load_adversarial_dataset(
attack_config,
epochs=1,
split_type="adversarial",
preprocessing_fn=preprocessing_fn,
)
else:
attack = load_attack(attack_config, classifier)
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
)
for x, y in tqdm(test_data, desc="Attack"):
if attack_type == "preloaded":
x, x_adv = x
if targeted:
y, y_target = y
elif attack_config.get("use_label"):
x_adv = attack.generate(x=x, y=y)
elif targeted:
raise NotImplementedError(
"Requires generation of target labels")
# x_adv = attack.generate(x=x, y=y_target)
else:
x_adv = attack.generate(x=x)
y_pred_adv = classifier.predict(x_adv)
if targeted:
# NOTE: does not remove data points where y == y_target
metrics_logger.update_task(y_target,
y_pred_adv,
adversarial=True)
else:
metrics_logger.update_task(y, y_pred_adv, adversarial=True)
metrics_logger.update_perturbation(x, x_adv)
metrics_logger.log_task(adversarial=True, targeted=targeted)
return metrics_logger.results()
def _evaluate(self, config: dict) -> dict:
"""
Evaluate a config file for classification robustness against attack.
"""
model_config = config["model"]
classifier, preprocessing_fn = load_model(model_config)
classifier_for_defense, _ = load_model(model_config)
train_epochs = config["adhoc"]["train_epochs"]
src_class = config["adhoc"]["source_class"]
tgt_class = config["adhoc"]["target_class"]
# Set random seed due to large variance in attack and defense success
np.random.seed(config["adhoc"]["np_seed"])
set_random_seed(config["adhoc"]["tf_seed"])
logger.info(f"Loading dataset {config['dataset']['name']}...")
batch_size = config["dataset"]["batch_size"]
train_data = load_dataset(
config["dataset"],
epochs=1,
split_type="train",
preprocessing_fn=preprocessing_fn,
)
logger.info(
"Building in-memory dataset for poisoning detection and training")
attack_config = config["attack"]
attack = load(attack_config)
fraction_poisoned = config["adhoc"]["fraction_poisoned"]
poison_dataset_flag = config["adhoc"]["poison_dataset"]
# detect_poison does not currently support data generators
# therefore, make in memory dataset
x_train_all, y_train_all = [], []
for x_train, y_train in train_data:
if poison_dataset_flag and np.random.rand() < fraction_poisoned:
x_train, y_train = poison_batch(x_train, y_train,
src_class, tgt_class,
len(y_train), attack)
x_train_all.append(x_train)
y_train_all.append(y_train)
x_train_all = np.concatenate(x_train_all, axis=0)
y_train_all = np.concatenate(y_train_all, axis=0)
y_train_all_categorical = to_categorical(y_train_all)
defense_config = config["defense"]
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']} "
f"for defense {defense_config['name']}...")
classifier_for_defense.fit(
x_train_all,
y_train_all_categorical,
batch_size=batch_size,
nb_epochs=train_epochs,
verbose=False,
)
defense_fn = load_fn(defense_config)
defense = defense_fn(classifier_for_defense, x_train_all,
y_train_all_categorical)
_, is_clean = defense.detect_poison(nb_clusters=2,
nb_dims=43,
reduce="PCA")
is_clean = np.array(is_clean)
logger.info(f"Total clean data points: {np.sum(is_clean)}")
logger.info("Filtering out detected poisoned samples")
indices_to_keep = is_clean == 1
x_train_filter = x_train_all[indices_to_keep]
y_train_filter = y_train_all_categorical[indices_to_keep]
if len(x_train_filter):
logger.info(
f"Fitting model of {model_config['module']}.{model_config['name']}..."
)
classifier.fit(
x_train_filter,
y_train_filter,
batch_size=batch_size,
nb_epochs=train_epochs,
verbose=False,
)
else:
logger.warning(
"All data points filtered by defense. Skipping training")
logger.info(f"Validating on clean test data")
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
)
validation_metric = metrics.MetricList("categorical_accuracy")
for x, y in tqdm(test_data, desc="Testing"):
y_pred = classifier.predict(x)
validation_metric.append(y, y_pred)
logger.info(
f"Unpoisoned validation accuracy: {validation_metric.mean():.2%}")
results = {"validation_accuracy": validation_metric.mean()}
if poison_dataset_flag:
logger.info(f"Testing on poisoned test data")
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
)
test_metric = metrics.MetricList("categorical_accuracy")
targeted_test_metric = metrics.MetricList("categorical_accuracy")
for x_test, y_test in tqdm(test_data, desc="Testing"):
x_test, _ = poison_batch(x_test, y_test, src_class, tgt_class,
len(y_test), attack)
y_pred = classifier.predict(x_test)
test_metric.append(y_test, y_pred)
y_pred_targeted = y_pred[y_test == src_class]
if not len(y_pred_targeted):
continue
targeted_test_metric.append([tgt_class] * len(y_pred_targeted),
y_pred_targeted)
results["test_accuracy"] = test_metric.mean()
results[
"targeted_misclassification_accuracy"] = targeted_test_metric.mean(
)
logger.info(f"Test accuracy: {test_metric.mean():.2%}")
logger.info(
f"Test targeted misclassification accuracy: {targeted_test_metric.mean():.2%}"
)
return results
def _evaluate(
self,
config: dict,
num_eval_batches: Optional[int],
skip_benign: Optional[bool],
skip_attack: Optional[bool],
skip_misclassified: Optional[bool],
) -> dict:
"""
Evaluate the config and return a results dict
"""
model_config = config["model"]
estimator, _ = load_model(model_config)
defense_config = config.get("defense") or {}
defense_type = defense_config.get("type")
if defense_type in ["Preprocessor", "Postprocessor"]:
logger.info(
f"Applying internal {defense_type} defense to estimator")
estimator = load_defense_internal(config["defense"], estimator)
if model_config["fit"]:
try:
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
fit_kwargs = model_config["fit_kwargs"]
logger.info(
f"Loading train dataset {config['dataset']['name']}...")
train_data = load_dataset(
config["dataset"],
epochs=fit_kwargs["nb_epochs"],
split=config["dataset"].get("train_split", "train"),
shuffle_files=True,
)
if defense_type == "Trainer":
logger.info(f"Training with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"],
estimator)
defense.fit_generator(train_data, **fit_kwargs)
else:
logger.info("Fitting estimator on clean train dataset...")
estimator.fit_generator(train_data, **fit_kwargs)
except NotImplementedError:
raise NotImplementedError(
"Training has not yet been implemented for object detectors"
)
if defense_type == "Transform":
# NOTE: Transform currently not supported
logger.info(
f"Transforming estimator with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], estimator)
estimator = defense()
attack_config = config["attack"]
attack_type = attack_config.get("type")
targeted = bool(attack_config.get("kwargs", {}).get("targeted"))
metrics_logger = metrics.MetricsLogger.from_config(
config["metric"],
skip_benign=skip_benign,
skip_attack=skip_attack,
targeted=targeted,
)
eval_split = config["dataset"].get("eval_split", "test")
if skip_benign:
logger.info("Skipping benign classification...")
else:
# Evaluate the ART estimator on benign test examples
logger.info(f"Loading test dataset {config['dataset']['name']}...")
test_data = load_dataset(
config["dataset"],
epochs=1,
split=eval_split,
num_batches=num_eval_batches,
shuffle_files=False,
)
logger.info("Running inference on benign examples...")
for x, y in tqdm(test_data, desc="Benign"):
# Ensure that input sample isn't overwritten by estimator
x.flags.writeable = False
with metrics.resource_context(
name="Inference",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.
computational_resource_dict,
):
y_pred = estimator.predict(x)
metrics_logger.update_task(y, y_pred)
metrics_logger.log_task()
if skip_attack:
logger.info("Skipping attack generation...")
return metrics_logger.results()
# Evaluate the ART estimator on adversarial test examples
logger.info("Generating or loading / testing adversarial examples...")
if skip_misclassified:
acc_task_idx = [i.name for i in metrics_logger.tasks
].index("categorical_accuracy")
benign_acc = metrics_logger.tasks[acc_task_idx].values()
if targeted and attack_config.get("use_label"):
raise ValueError("Targeted attacks cannot have 'use_label'")
if attack_type == "preloaded":
preloaded_split = attack_config.get("kwargs", {}).get(
"split", "adversarial")
test_data = load_adversarial_dataset(
attack_config,
epochs=1,
split=preloaded_split,
num_batches=num_eval_batches,
shuffle_files=False,
)
else:
attack = load_attack(attack_config, estimator)
if targeted != getattr(attack, "targeted", False):
logger.warning(
f"targeted config {targeted} != attack field {getattr(attack, 'targeted', False)}"
)
test_data = load_dataset(
config["dataset"],
epochs=1,
split=eval_split,
num_batches=num_eval_batches,
shuffle_files=False,
)
if targeted:
label_targeter = load_label_targeter(
attack_config["targeted_labels"])
export_samples = config["scenario"].get("export_samples")
if export_samples is not None and export_samples > 0:
sample_exporter = SampleExporter(self.scenario_output_dir,
test_data.context, export_samples)
else:
sample_exporter = None
for batch_idx, (x, y) in enumerate(tqdm(test_data, desc="Attack")):
with metrics.resource_context(
name="Attack",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.
computational_resource_dict,
):
if attack_type == "preloaded":
if len(x) == 2:
x, x_adv = x
else:
x_adv = x
if targeted:
y, y_target = y
else:
generate_kwargs = deepcopy(
attack_config.get("generate_kwargs", {}))
# Temporary workaround for ART code requirement of ndarray mask
if "mask" in generate_kwargs:
generate_kwargs["mask"] = np.array(
generate_kwargs["mask"])
if attack_config.get("use_label"):
generate_kwargs["y"] = y
elif targeted:
y_target = label_targeter.generate(y)
generate_kwargs["y"] = y_target
if skip_misclassified and benign_acc[batch_idx] == 0:
x_adv = x
else:
x_adv = attack.generate(x=x, **generate_kwargs)
# Ensure that input sample isn't overwritten by estimator
x_adv.flags.writeable = False
y_pred_adv = estimator.predict(x_adv)
metrics_logger.update_task(y, y_pred_adv, adversarial=True)
if targeted:
metrics_logger.update_task(y_target,
y_pred_adv,
adversarial=True,
targeted=True)
metrics_logger.update_perturbation(x, x_adv)
if sample_exporter is not None:
sample_exporter.export(x, x_adv, y, y_pred_adv)
metrics_logger.log_task(adversarial=True)
if targeted:
metrics_logger.log_task(adversarial=True, targeted=True)
return metrics_logger.results()
def _evaluate(self, config: dict, num_eval_batches: Optional[int],
skip_benign: Optional[bool]) -> dict:
"""
Evaluate the config and return a results dict
"""
model_config = config["model"]
estimator, fit_preprocessing_fn = load_model(model_config)
defense_config = config.get("defense") or {}
defense_type = defense_config.get("type")
if defense_type in ["Preprocessor", "Postprocessor"]:
logger.info(
f"Applying internal {defense_type} defense to estimator")
estimator = load_defense_internal(config["defense"], estimator)
if model_config["fit"]:
try:
estimator.set_learning_phase(True)
except NotImplementedError:
logger.exception(
"set_learning_phase error; training may not work.")
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
fit_kwargs = model_config["fit_kwargs"]
logger.info(
f"Loading train dataset {config['dataset']['name']}...")
batch_size = config["dataset"].pop("batch_size")
config["dataset"]["batch_size"] = fit_kwargs.get(
"fit_batch_size", batch_size)
train_data = load_dataset(
config["dataset"],
epochs=fit_kwargs["nb_epochs"],
split_type=config["dataset"].get("train_split",
"train_clean100"),
preprocessing_fn=fit_preprocessing_fn,
shuffle_files=True,
)
config["dataset"]["batch_size"] = batch_size
if defense_type == "Trainer":
logger.info(f"Training with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], estimator)
defense.fit_generator(train_data, **fit_kwargs)
else:
logger.info("Fitting estimator on clean train dataset...")
estimator.fit_generator(train_data, **fit_kwargs)
if defense_type == "Transform":
# NOTE: Transform currently not supported
logger.info(
f"Transforming estimator with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], estimator)
estimator = defense()
try:
estimator.set_learning_phase(False)
except NotImplementedError:
logger.warning(
"Unable to set estimator's learning phase. As of ART 1.4.1, "
"this is not yet supported for speech recognition models.")
metrics_logger = metrics.MetricsLogger.from_config(
config["metric"], skip_benign=skip_benign)
if config["dataset"]["batch_size"] != 1:
logger.warning("Evaluation batch_size != 1 may not be supported.")
predict_kwargs = config["model"].get("predict_kwargs", {})
eval_split = config["dataset"].get("eval_split", "test_clean")
if skip_benign:
logger.info("Skipping benign classification...")
else:
# Evaluate the ART estimator on benign test examples
logger.info(f"Loading test dataset {config['dataset']['name']}...")
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type=eval_split,
num_batches=num_eval_batches,
shuffle_files=False,
)
logger.info("Running inference on benign examples...")
for x, y in tqdm(test_data, desc="Benign"):
# Ensure that input sample isn't overwritten by estimator
x.flags.writeable = False
with metrics.resource_context(
name="Inference",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.
computational_resource_dict,
):
y_pred = estimator.predict(x, **predict_kwargs)
metrics_logger.update_task(y, y_pred)
metrics_logger.log_task()
# Imperceptible attack still WIP
if (config.get("adhoc") or {}).get("skip_adversarial"):
logger.info("Skipping adversarial classification...")
return metrics_logger.results()
# Evaluate the ART estimator on adversarial test examples
logger.info("Generating or loading / testing adversarial examples...")
attack_config = config["attack"]
attack_type = attack_config.get("type")
targeted = bool(attack_config.get("targeted"))
if attack_type == "preloaded":
test_data = load_adversarial_dataset(
attack_config,
epochs=1,
split_type="adversarial",
num_batches=num_eval_batches,
shuffle_files=False,
)
else:
attack = load_attack(attack_config, estimator)
if targeted != attack.targeted:
logger.warning(
f"targeted config {targeted} != attack field {attack.targeted}"
)
test_data = load_dataset(
config["dataset"],
epochs=1,
split_type=eval_split,
num_batches=num_eval_batches,
shuffle_files=False,
)
if targeted:
label_targeter = load_label_targeter(
attack_config["targeted_labels"])
for x, y in tqdm(test_data, desc="Attack"):
with metrics.resource_context(
name="Attack",
profiler=config["metric"].get("profiler_type"),
computational_resource_dict=metrics_logger.
computational_resource_dict,
):
if attack_type == "preloaded":
x, x_adv = x
if targeted:
y, y_target = y
elif attack_config.get("use_label"):
x_adv = attack.generate(x=x, y=y)
elif targeted:
y_target = label_targeter.generate(y)
x_adv = attack.generate(x=x, y=y_target)
else:
x_adv = attack.generate(x=x)
# Ensure that input sample isn't overwritten by estimator
x_adv.flags.writeable = False
y_pred_adv = estimator.predict(x_adv, **predict_kwargs)
metrics_logger.update_task(y, y_pred_adv, adversarial=True)
metrics_logger.update_perturbation(x, x_adv)
metrics_logger.log_task(adversarial=True, targeted=True)
return metrics_logger.results()
def _evaluate(self, config: dict) -> dict:
"""
Evaluate the config and return a results dict
"""
model_config = config["model"]
classifier, preprocessing_fn = load_model(model_config)
defense_config = config.get("defense") or {}
defense_type = defense_config.get("type")
if defense_type in ["Preprocessor", "Postprocessor"]:
logger.info(f"Applying internal {defense_type} defense to classifier")
classifier = load_defense_internal(config["defense"], classifier)
if model_config["fit"]:
classifier.set_learning_phase(True)
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
train_epochs = config["model"]["fit_kwargs"]["nb_epochs"]
batch_size = config["dataset"]["batch_size"]
logger.info(f"Loading train dataset {config['dataset']['name']}...")
train_data = load_dataset(
config["dataset"],
epochs=train_epochs,
split_type="train",
preprocessing_fn=preprocessing_fn,
)
if defense_type == "Trainer":
logger.info(f"Training with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
else:
logger.info(f"Fitting classifier on clean train dataset...")
for epoch in range(train_epochs):
classifier.set_learning_phase(True)
for _ in tqdm(
range(train_data.batches_per_epoch),
desc=f"Epoch: {epoch}/{train_epochs}",
):
x, y = train_data.get_batch()
# x_trains consists of one or more videos, each represented as an
# ndarray of shape (n_stacks, 3, 16, 112, 112).
# To train, randomly sample a batch of stacks
x = np.stack([x_i[np.random.randint(x_i.shape[0])] for x_i in x])
if defense_type == "Trainer":
defense.fit(x, y, batch_size=batch_size, nb_epochs=1)
else:
classifier.fit(x, y, batch_size=batch_size, nb_epochs=1)
if defense_type == "Transform":
# NOTE: Transform currently not supported
logger.info(f"Transforming classifier with {defense_type} defense...")
defense = load_defense_wrapper(config["defense"], classifier)
classifier = defense()
classifier.set_learning_phase(False)
# Evaluate the ART classifier on benign test examples
logger.info(f"Loading test dataset {config['dataset']['name']}...")
test_data_generator = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
)
logger.info("Running inference on benign examples...")
metrics_logger = metrics.MetricsLogger.from_config(config["metric"])
for x_batch, y_batch in tqdm(test_data_generator, desc="Benign"):
for x, y in zip(x_batch, y_batch):
# combine predictions across all stacks
y_pred = np.mean(classifier.predict(x), axis=0)
metrics_logger.update_task(y, y_pred)
metrics_logger.log_task()
# Evaluate the ART classifier on adversarial test examples
logger.info("Generating / testing adversarial examples...")
attack = load_attack(config["attack"], classifier)
test_data_generator = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
)
for x_batch, y_batch in tqdm(test_data_generator, desc="Attack"):
for x, y in zip(x_batch, y_batch):
# each x is of shape (n_stack, 3, 16, 112, 112)
# n_stack varies
attack.set_params(batch_size=x.shape[0])
x_adv = attack.generate(x=x)
# combine predictions across all stacks
y_pred = np.mean(classifier.predict(x), axis=0)
metrics_logger.update_task(y, y_pred, adversarial=True)
metrics_logger.update_perturbation([x], [x_adv])
metrics_logger.log_task(adversarial=True)
return metrics_logger.results()
def _evaluate(self, config: dict) -> dict:
"""
Evaluate a config file for classification robustness against attack.
"""
model_config = config["model"]
classifier, preprocessing_fn = load_model(model_config)
n_tbins = 100 # number of time bins in spectrogram input to model
task_metric = metrics.categorical_accuracy
# Train ART classifier
if not model_config["weights_file"]:
classifier.set_learning_phase(True)
logger.info(
f"Fitting model {model_config['module']}.{model_config['name']}..."
)
fit_kwargs = model_config["fit_kwargs"]
train_data_generator = load_dataset(
config["dataset"],
epochs=fit_kwargs["nb_epochs"],
split_type="train",
preprocessing_fn=preprocessing_fn,
)
for cnt, (x, y) in tqdm(enumerate(train_data_generator)):
x_seg, y_seg = segment(x, y, n_tbins)
classifier.fit(
x_seg,
y_seg,
batch_size=config["dataset"]["batch_size"],
nb_epochs=1,
verbose=True,
)
if (cnt + 1) % train_data_generator.batches_per_epoch == 0:
# evaluate on validation examples
val_data_generator = load_dataset(
config["dataset"],
epochs=1,
split_type="validation",
preprocessing_fn=preprocessing_fn,
)
cnt = 0
validation_accuracies = []
for x_val, y_val in tqdm(val_data_generator):
x_val_seg, y_val_seg = segment(x_val, y_val, n_tbins)
y_pred = classifier.predict(x_val_seg)
validation_accuracies.extend(
task_metric(y_val_seg, y_pred))
cnt += len(y_val_seg)
validation_accuracy = sum(validation_accuracies) / cnt
logger.info(
"Validation accuracy: {}".format(validation_accuracy))
classifier.set_learning_phase(False)
# Evaluate ART classifier on test examples
logger.info(f"Loading testing dataset {config['dataset']['name']}...")
test_data_generator = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
)
logger.info("Running inference on benign test examples...")
cnt = 0
benign_accuracies = []
for x, y in tqdm(test_data_generator, desc="Benign"):
x_seg, y_seg = segment(x, y, n_tbins)
y_pred = classifier.predict(x_seg)
benign_accuracies.extend(task_metric(y_seg, y_pred))
cnt += len(y_seg)
benign_accuracy = sum(benign_accuracies) / cnt
logger.info(f"Accuracy on benign test examples: {benign_accuracy:.2%}")
# Evaluate the ART classifier on adversarial test examples
logger.info("Generating / testing adversarial examples...")
attack = load_attack(config["attack"], classifier)
test_data_generator = load_dataset(
config["dataset"],
epochs=1,
split_type="test",
preprocessing_fn=preprocessing_fn,
)
cnt = 0
adversarial_accuracies = []
for x, y in tqdm(test_data_generator, desc="Attack"):
x_seg, y_seg = segment(x, y, n_tbins)
x_adv = attack.generate(x=x_seg)
y_pred = classifier.predict(x_adv)
adversarial_accuracies.extend(task_metric(y_seg, y_pred))
cnt += len(y_seg)
adversarial_accuracy = sum(adversarial_accuracies) / cnt
logger.info(
f"Accuracy on adversarial test examples: {adversarial_accuracy:.2%}"
)
results = {
"mean_benign_accuracy": benign_accuracy,
"mean_adversarial_accuracy": adversarial_accuracy,
}
return results