def test_digit(self):
train_x, train_y, test_x, test_y = datasets.load("digit")
self.assertEqual(train_x.shape[0], 1350)
self.assertEqual(train_y.shape[0], 1350)
self.assertEqual(test_x.shape[0], 150)
self.assertEqual(test_y.shape[0], 150)
for X in train_x, test_x:
for x in X:
self.assertTrue(1148 <= len(x) <= 18262)
def evaluate_classifier(config_path: str) -> None:
"""
Evaluate a config file for classiifcation robustness against attack.
"""
with open(config_path, "r") as fp:
config = json.load(fp)
model_config = config["model"]
classifier_module = import_module(model_config["module"])
classifier_fn = getattr(classifier_module, model_config["name"])
classifier = classifier_fn(model_config["model_kwargs"],
model_config["wrapper_kwargs"])
preprocessing_fn = getattr(classifier_module, "preprocessing_fn")
logger.info(f"Loading dataset {config['dataset']['name']}...")
clean_x, adv_x, labels = datasets.load(config["dataset"]["name"],
preprocessing_fn=preprocessing_fn)
# Evaluate the ART classifier on benign test examples
logger.info("Predicting on clean dataset...")
predictions = classifier.predict(clean_x)
benign_accuracy = np.sum(
np.argmax(predictions, axis=1) == labels) / len(labels)
logger.info("Accuracy on benign test examples: {}%".format(
benign_accuracy * 100))
# Evaluate the ART classifier on adversarial examples from transfer attack
logger.info("Predicting on adversarial dataset...")
predictions = classifier.predict(adv_x)
adversarial_accuracy = np.sum(
np.argmax(predictions, axis=1) == labels) / len(labels)
logger.info("Accuracy on adversarial test examples: {}%".format(
adversarial_accuracy * 100))
logger.info("Saving json output...")
filepath = os.path.join(paths.OUTPUTS, "evaluation-results.json")
with open(filepath, "w") as f:
output_dict = {
"config": config,
"results": {
"baseline_accuracy": str(benign_accuracy),
"adversarial_accuracy": str(adversarial_accuracy),
},
}
json.dump(output_dict, f, sort_keys=True, indent=4)
logger.info(f"Evaluation Results written to {filepath}")
def test_keras_imagenet(self):
classifier_module = import_module(
"armory.baseline_models.keras.keras_resnet50")
classifier_fn = getattr(classifier_module, "get_art_model")
classifier = classifier_fn(model_kwargs={}, wrapper_kwargs={})
preprocessing_fn = getattr(classifier_module, "preprocessing_fn")
clean_x, adv_x, labels = datasets.load("imagenet_adversarial",
preprocessing_fn)
predictions = classifier.predict(clean_x)
accuracy = np.sum(
np.argmax(predictions, axis=1) == labels) / len(labels)
self.assertGreater(accuracy, 0.65)
predictions = classifier.predict(adv_x)
accuracy = np.sum(
np.argmax(predictions, axis=1) == labels) / len(labels)
print(accuracy)
self.assertLess(accuracy, 0.02)
def test_keras_cifar10(self):
batch_size = 64
epochs = 2
classifier_module = import_module(
"armory.baseline_models.keras.keras_cifar")
classifier_fn = getattr(classifier_module, "get_art_model")
classifier = classifier_fn(model_kwargs={}, wrapper_kwargs={})
preprocessing_fn = getattr(classifier_module, "preprocessing_fn")
train_x, train_y, test_x, test_y = datasets.load(
"cifar10", preprocessing_fn=preprocessing_fn)
classifier.fit(train_x,
train_y,
batch_size=batch_size,
nb_epochs=epochs)
predictions = classifier.predict(test_x)
accuracy = np.sum(
np.argmax(predictions, axis=1) == test_y) / len(test_y)
self.assertGreater(accuracy, 0.4)
def evaluate_classifier(config_path: str) -> None:
"""
Evaluate a config file for classifcation robustness against attack.
"""
# Generate adversarial test examples
with open(config_path, "r") as fp:
config = json.load(fp)
model_config = config["model"]
classifier_module = import_module(model_config["module"])
classifier_fn = getattr(classifier_module, model_config["name"])
preprocessing_fn = getattr(classifier_module, "preprocessing_fn")
logger.info(f"Loading dataset {config['dataset']['name']}...")
x_clean_train, y_clean_train, x_clean_test, y_clean_test = datasets.load(
config["dataset"]["name"], preprocessing_fn=preprocessing_fn
)
batch_size = config["adhoc"]["batch_size"]
epochs = config["adhoc"]["epochs"]
n_trials = config["adhoc"]["n_trials"]
poison_frac_min = config["adhoc"]["poison_frac_min"]
poison_frac_max = config["adhoc"]["poison_frac_max"]
poison_frac_steps = config["adhoc"]["poison_frac_steps"]
source_class = config["adhoc"]["source_class"]
target_class = config["adhoc"]["target_class"]
fraction_poisons = np.linspace(poison_frac_min, poison_frac_max, poison_frac_steps)
# Test clean model accuracy to provide a benchmark to poisoned model accuracy
raw_metrics = {}
raw_metrics["undefended_backdoor_success_rate"] = init_metrics(
fraction_poisons, n_trials
)
raw_metrics["non_backdoored_accuracy"] = init_metrics(fraction_poisons, n_trials)
raw_metrics["clean_model_accuracy"] = [None for _ in range(n_trials)]
raw_metrics["defended_backdoor_success_rate"] = init_metrics(
fraction_poisons, n_trials
)
raw_metrics["delta_accuracy"] = init_metrics(fraction_poisons, n_trials)
for trial in range(n_trials):
classifier = classifier_fn(
model_config["model_kwargs"], model_config["wrapper_kwargs"]
)
logger.info(
f"Fitting clean unpoisoned model of {model_config['module']}.{model_config['name']}..."
)
classifier.fit(
x_clean_train, y_clean_train, batch_size=batch_size, nb_epochs=epochs
)
raw_metrics["clean_model_accuracy"][trial] = eval_targeted_fit(
classifier, x_clean_test, y_clean_test
)
for frac_poison in fraction_poisons:
# Need to retrain from scratch for each frac_poison value
classifier = classifier_fn(
model_config["model_kwargs"], model_config["wrapper_kwargs"]
)
classifier_defended = classifier_fn(
model_config["model_kwargs"], model_config["wrapper_kwargs"]
)
attack_config = config["attack"]
attack_module = import_module(attack_config["module"])
attack_fn = getattr(attack_module, attack_config["name"])
attack = attack_fn(
classifier=classifier,
x_train=x_clean_train,
y_train=y_clean_train,
pct_poison=frac_poison,
source_class=source_class,
target_class=target_class,
)
is_poison, x_poison, y_poison = attack.generate(
x_clean_train, y_clean_train
)
logger.info(f"Fitting poisoned model with poison fraction {frac_poison}...")
classifier.fit(x_poison, y_poison, batch_size=batch_size, nb_epochs=epochs)
x_test_targeted = x_clean_test[y_clean_test == source_class]
x_poison_test = attack.generate_target_test(x_test_targeted)
# Show targeted accuracy for poisoned classes is as expected
raw_metrics["undefended_backdoor_success_rate"][frac_poison][
trial
] = eval_targeted_fit(classifier, x_poison_test, target_class)
raw_metrics["non_backdoored_accuracy"][frac_poison][
trial
] = eval_targeted_fit(classifier, x_clean_test, y_clean_test)
defense_config = config["defense"]
defense_module = import_module(defense_config["module"])
defense_fn = getattr(defense_module, defense_config["name"])
defense = defense_fn(
classifier,
x_poison,
y_poison,
batch_size=batch_size,
ub_pct_poison=frac_poison,
**defense_config["kwargs"],
)
conf_matrix_json = defense.evaluate_defence(np.logical_not(is_poison))
logger.info(
f"Poison detection confusion matrix from defense {config['defense']['name']} "
f"with poison fraction {frac_poison}:"
)
logger.info(conf_matrix_json)
_, indices_to_keep = defense.detect_poison()
logger.info(
f"Fitting poisoned model with poisons filtered by defense {config['defense']['name']} "
f"with poison fraction {frac_poison}..."
)
classifier_defended.fit(
x_poison[indices_to_keep == 1],
y_poison[indices_to_keep == 1],
batch_size=batch_size,
nb_epochs=epochs,
)
defended_backdoor_success_rate = eval_targeted_fit(
classifier_defended, x_poison_test, target_class
)
raw_metrics["defended_backdoor_success_rate"][frac_poison][
trial
] = defended_backdoor_success_rate
logger.info(
f"Trial {trial+1} defended backdoor success rate {defended_backdoor_success_rate} "
f"with poisoning proportion of {frac_poison}"
)
defended_clean_accuracy = eval_targeted_fit(
classifier_defended, x_clean_test, y_clean_test
)
delta_accuracy = (
raw_metrics["non_backdoored_accuracy"][frac_poison][trial]
- defended_clean_accuracy
)
raw_metrics["delta_accuracy"][frac_poison][trial] = delta_accuracy
logger.info(
f"Trial {trial+1} delta accuracy of {delta_accuracy} "
f"with poisoning proportion of {frac_poison}"
)
logger.info(f"Trial {trial+1}/{n_trials} completed.")
summarized_metrics = summarize_metrics(raw_metrics)
logger.info("Saving json output...")
filepath = os.path.join(
paths.OUTPUTS, f"backdoor_performance_{int(time.time())}.json"
)
with open(filepath, "w") as f:
output_dict = {"config": config, "results": summarized_metrics}
json.dump(output_dict, f, sort_keys=True, indent=4)
shutil.copyfile(filepath, os.path.join(paths.OUTPUTS, "latest.json"))
classification_poisoning(filepath)
classification_poisoning(os.path.join(paths.OUTPUTS, "latest.json"))
def evaluate_classifier(config_path: str) -> None:
"""
Evaluate a config file for classification robustness against attack.
"""
with open(config_path) as fp:
config = json.load(fp)
model_config = config["model"]
classifier_module = import_module(model_config["module"])
classifier_fn = getattr(classifier_module, model_config["name"])
classifier = classifier_fn(
model_config["model_kwargs"], model_config["wrapper_kwargs"]
)
batch_size = config["adhoc"]["batch_size"]
preprocessing_fn = getattr(classifier_module, "preprocessing_fn")
# Defense
defense_config = config["defense"]
defense_module = import_module(defense_config["module"])
defense_fn = getattr(defense_module, defense_config["name"])
transformer = defense_fn(**defense_config.get("kwargs", {}))
if not isinstance(transformer, defences_ext.Transformer):
raise ValueError(
f'{defense_config["module"]}.{defense_config["name"]} is not an instance of '
f"{defences_ext.Transformer}"
)
defended_classifier = transformer.transform(classifier)
# retrofitted to work with existing code
logger.info(f"Loading dataset {config['dataset']['name']}...")
clean_x, adv_x, labels = datasets.load(
config["dataset"]["name"], preprocessing_fn=preprocessing_fn
)
logger.debug(f"Original model:\n{classifier}")
logger.info("Predicting on clean dataset...")
clean_y_pred = classifier.predict(clean_x, batch_size=batch_size)
clean_accuracy = np.sum(np.argmax(clean_y_pred, axis=1) == labels) / len(labels)
logger.info(f"Accuracy on benign test examples: {clean_accuracy * 100}%")
# Evaluate the ART classifier on adversarial examples from transfer attack
logger.info("Predicting on adversarial dataset...")
adv_y_pred = classifier.predict(adv_x, batch_size=batch_size)
adv_accuracy = np.sum(np.argmax(adv_y_pred, axis=1) == labels) / len(labels)
logger.info(f"Accuracy on adversarial test examples: {adv_accuracy * 100}%")
# Ee-evaluate on defended classifier
logger.debug(f"Defended classifier:\n{defended_classifier}")
logger.info(
f'Classifier defended by {defense_config["module"]}.{defense_config["name"]} transform'
)
logger.info("Predicting on clean dataset...")
def_clean_y_pred = defended_classifier.predict(clean_x, batch_size=batch_size)
def_clean_accuracy = np.sum(np.argmax(def_clean_y_pred, axis=1) == labels) / len(
labels
)
logger.info(f"Accuracy on benign test examples: {def_clean_accuracy * 100}%")
# Evaluate the ART classifier on adversarial examples from transfer attack
logger.info("Predicting on adversarial dataset...")
def_adv_y_pred = defended_classifier.predict(adv_x, batch_size=batch_size)
def_adv_accuracy = np.sum(np.argmax(def_adv_y_pred, axis=1) == labels) / len(labels)
logger.info(f"Accuracy on adversarial test examples: {def_adv_accuracy * 100}%")
def evaluate_classifier(config_path: str) -> None:
"""
Evaluate a config file for classiifcation robustness against attack.
"""
with open(config_path, "r") as fp:
config = json.load(fp)
model_config = config["model"]
classifier_module = import_module(model_config["module"])
classifier_fn = getattr(classifier_module, model_config["name"])
classifier = classifier_fn(model_config["model_kwargs"],
model_config["wrapper_kwargs"])
preprocessing_fn = getattr(classifier_module, "preprocessing_fn")
logger.info(f"Loading dataset {config['dataset']['name']}...")
x_train, y_train, x_test, y_test = datasets.load(
config["dataset"]["name"], preprocessing_fn=preprocessing_fn)
logger.info(
f"Fitting clean unpoisoned model of {model_config['module']}.{model_config['name']}..."
)
classifier.fit(
x_train,
y_train,
batch_size=config["adhoc"]["batch_size"],
nb_epochs=config["adhoc"]["epochs"],
)
# Generate adversarial test examples
attack_config = config["attack"]
attack_module = import_module(attack_config["module"])
attack_fn = getattr(attack_module, attack_config["name"])
attack = attack_fn(classifier=classifier, **attack_config["kwargs"])
norm = attack_config["budget"]["norm"][0]
if norm == "L2":
lp_norm = 2
elif norm == "Linf":
lp_norm = np.inf
else:
raise ValueError(
f"Adversarial budget must have a norm of L2 or Linf. Found {norm} in config"
)
y_target = (y_test + 1) % config["adhoc"]["num_classes"]
np.random.seed(config["adhoc"]["seed"])
indices = np.random.choice(x_test.shape[0],
config["adhoc"]["num_attacked_pts"])
x_test_sample = x_test[indices]
y_test_sample = y_test[indices]
y_target_sample = y_target[indices]
logger.info("Generating adversarial examples...")
x_test_adv = attack.generate(x=x_test_sample, y=y_target_sample)
diff = (x_test_adv - x_test_sample).reshape(x_test_adv.shape[0], -1)
epsilons = np.linalg.norm(diff, ord=lp_norm, axis=1)
y_clean_pred = np.argmax(classifier.predict(x_test_sample), axis=1)
y_adv_pred = np.argmax(classifier.predict(x_test_adv), axis=1)
# Evaluate the ART classifier on adversarial test examples and clean test examples
successful_attack_indices = (y_clean_pred != y_target_sample) & (
y_adv_pred == y_target_sample)
benign_misclassification_rate = np.sum(
y_clean_pred == y_target_sample) / float(y_clean_pred.shape[0])
logger.info(
f"Benign misclassification as targeted examples: {benign_misclassification_rate * 100}%"
)
targeted_attack_success_rate = np.sum(successful_attack_indices) / float(
y_clean_pred.shape[0])
clean_accuracy = np.sum(y_clean_pred == y_test_sample) / float(
y_clean_pred.shape[0])
logger.info(f"Accuracy on benign test examples: {clean_accuracy * 100}%")
epsilons = epsilons.astype(object)
epsilons[np.logical_not(successful_attack_indices)] = None
unique_epsilons, targeted_attack_success = roc_targeted_epsilon(epsilons)
results = {}
results[norm] = {
"epsilons": list(unique_epsilons),
"metric": "Targeted attack success rate",
"values": list(targeted_attack_success),
}
logger.info(
f"Finished attacking on norm {norm}. Attack success: {targeted_attack_success_rate * 100}%"
)
logger.info("Saving json output...")
filepath = os.path.join(
paths.OUTPUTS, f"carlini_wagner_attack_{norm}_targeted_output.json")
with open(filepath, "w") as f:
output_dict = {
"config": config,
"results": results,
}
json.dump(output_dict, f, sort_keys=True, indent=4)
logger.info("Plotting results...")
plot.classification(filepath)
def test_imagenet_adv(self):
clean_x, adv_x, labels = datasets.load("imagenet_adversarial")
self.assertEqual(clean_x.shape[0], 1000)
self.assertEqual(adv_x.shape[0], 1000)
self.assertEqual(labels.shape[0], 1000)
def test_cifar10(self):
train_x, train_y, test_x, test_y = datasets.load("cifar10")
self.assertEqual(train_x.shape[0], 50000)
self.assertEqual(train_y.shape[0], 50000)
self.assertEqual(test_x.shape[0], 10000)
self.assertEqual(test_y.shape[0], 10000)
def test_mnist(self):
train_x, train_y, test_x, test_y = datasets.load("mnist")
self.assertEqual(train_x.shape[0], 60000)
self.assertEqual(train_y.shape[0], 60000)
self.assertEqual(test_x.shape[0], 10000)
self.assertEqual(test_y.shape[0], 10000)
def evaluate_classifier(config_path: str) -> None:
"""
Evaluate a config file for classification robustness against attack.
"""
with open(config_path) as fp:
config = json.load(fp)
model_config = config["model"]
classifier_module = import_module(model_config["module"])
classifier_fn = getattr(classifier_module, model_config["name"])
classifier = classifier_fn(model_config["model_kwargs"],
model_config["wrapper_kwargs"])
preprocessing_fn = getattr(classifier_module, "preprocessing_fn")
logger.info(f"Loading dataset {config['dataset']['name']}...")
x_train, y_train, x_test, y_test = datasets.load(
config["dataset"]["name"], preprocessing_fn=preprocessing_fn)
logger.info(
f"Fitting clean unpoisoned model of {model_config['module']}.{model_config['name']}..."
)
classifier.fit(
x_train,
y_train,
batch_size=config["adhoc"]["batch_size"],
nb_epochs=config["adhoc"]["epochs"],
)
# Speeds up testing...
subsample = 100
x_test = x_test[::subsample]
y_test = y_test[::subsample]
# Evaluate the ART classifier on benign test examples
y_pred = classifier.predict(x_test)
benign_accuracy = np.sum(np.argmax(y_pred, axis=1) == y_test) / len(y_test)
logger.info("Accuracy on benign test examples: {}%".format(
benign_accuracy * 100))
attack_config = config["attack"]
attack_module = import_module(attack_config["module"])
attack_fn = getattr(attack_module, attack_config["name"])
budget = attack_config["budget"]
norms = budget["norm"]
results = {}
# Assume min_value = 0
max_value = 1.0
input_dim = np.product(x_test.shape[1:])
norm_map = { # from norm name to (fgm_input, max_epsilon)
"L0": (0, input_dim),
"L1": (1, input_dim * max_value),
"L2": (2, np.sqrt(input_dim) * max_value),
"Linf": (np.inf, max_value),
}
for norm in norms:
lp_norm, max_epsilon = norm_map[norm]
# Currently looking at untargeted attacks,
# where adversary accuracy ~ 1 - benign accuracy (except incorrect benign)
attack = attack_fn(
classifier=classifier,
norm=lp_norm,
eps=max_epsilon,
**attack_config["kwargs"],
)
logger.info(f"Generating adversarial examples for norm {norm}...")
x_test_adv = attack.generate(x=x_test)
# Map into the original input space (bound and quantize) and back to float
# NOTE: this step makes many of the attacks fail
x_test_adv = project_to_mnist_input(x_test_adv, preprocessing_fn)
diff = (x_test_adv - x_test).reshape(x_test.shape[0], -1)
epsilons = np.linalg.norm(diff, ord=lp_norm, axis=1)
if np.isnan(epsilons).any():
raise ValueError(f"Epsilons have nan values in norm {norm}")
min_epsilon = 0
if (epsilons < min_epsilon).any() or (epsilons > max_epsilon).any():
raise ValueError(
f"Epsilons have values outside bounds in norm {norm}")
y_pred_adv = classifier.predict(x_test_adv)
# Ignore benign misclassifications - no perturbation needed
epsilons[np.argmax(y_pred, axis=1) != y_test] = min_epsilon
# When all attacks fail, set perturbation to None
epsilons = epsilons.astype(object)
epsilons[(np.argmax(y_pred_adv, axis=1) == y_test)
& (np.argmax(y_pred, axis=1) == y_test)] = None
adv_acc = np.sum(np.argmax(y_pred_adv, axis=1) != y_test) / len(y_test)
# generate curve
unique_epsilons, accuracy = roc_epsilon(epsilons,
min_epsilon=min_epsilon,
max_epsilon=max_epsilon)
results[norm] = {
"epsilons": list(unique_epsilons),
"metric": "Categorical Accuracy",
"values": list(accuracy),
}
# Evaluate the ART classifier on adversarial test examples
logger.info(
f"Finished attacking on norm {norm}. Attack success: {adv_acc * 100}%"
)
logger.info("Saving json output...")
filepath = os.path.join(paths.OUTPUTS,
f"classifier_extended_{int(time.time())}.json")
with open(filepath, "w") as f:
output_dict = {
"config": config,
"results": results,
}
json.dump(output_dict, f, sort_keys=True, indent=4)
shutil.copyfile(filepath, os.path.join(paths.OUTPUTS, "latest.json"))
logger.info(f"Now plotting results...")
plot.classification(filepath)
plot.classification(os.path.join(paths.OUTPUTS, "latest.json"))
def evaluate_classifier(config_path: str) -> None:
"""
Evaluate a config file for classiifcation robustness against attack.
"""
with open(config_path, "r") as fp:
config = json.load(fp)
model_config = config["model"]
classifier_module = import_module(model_config["module"])
classifier_fn = getattr(classifier_module, model_config["name"])
classifier = classifier_fn(model_config["model_kwargs"],
model_config["wrapper_kwargs"])
preprocessing_fn = getattr(classifier_module, "preprocessing_fn")
logger.info(f"Loading dataset {config['dataset']['name']}...")
train_x, train_y, test_x, test_y = datasets.load(
config["dataset"]["name"], preprocessing_fn=preprocessing_fn)
logger.info(
f"Fitting clean unpoisoned model of {model_config['module']}.{model_config['name']}..."
)
classifier.fit(
train_x,
train_y,
batch_size=config["adhoc"]["batch_size"],
nb_epochs=config["adhoc"]["epochs"],
)
# Evaluate the ART classifier on benign test examples
logger.info("Running inference on benign examples...")
predictions = classifier.predict(test_x)
benign_accuracy = np.sum(
np.argmax(predictions, axis=1) == test_y) / len(test_y)
logger.info("Accuracy on benign test examples: {}%".format(
benign_accuracy * 100))
# Generate adversarial test examples
attack_config = config["attack"]
attack_module = import_module(attack_config["module"])
attack_fn = getattr(attack_module, attack_config["name"])
logger.info("Generating adversarial examples...")
attack = attack_fn(classifier=classifier, **attack_config["kwargs"])
test_x_adv = attack.generate(x=test_x)
# Evaluate the ART classifier on adversarial test examples
logger.info("Running inference on adversarial examples...")
predictions = classifier.predict(test_x_adv)
adversarial_accuracy = np.sum(
np.argmax(predictions, axis=1) == test_y) / len(test_y)
logger.info("Accuracy on adversarial test examples: {}%".format(
adversarial_accuracy * 100))
logger.info("Saving json output...")
filepath = os.path.join(paths.OUTPUTS, "evaluation-results.json")
with open(filepath, "w") as f:
output_dict = {
"config": config,
"results": {
"baseline_accuracy": str(benign_accuracy),
"adversarial_accuracy": str(adversarial_accuracy),
},
}
json.dump(output_dict, f, sort_keys=True, indent=4)
logger.info(f"Evaluation Results written to {filepath}")