I've a small decision tree code and I believe I convert everything to int, and I have checked my train/test data with isnan, max etc.
I genuinely have no idea why its giving that error.
So I'm trying to pass Mnist dataset from Decision Tree and then I'll attack using a class.
Here is the code:
from AttackUtils import Attack
from AttackUtils import calc_output_weighted_weights, targeted_gradient, non_targeted_gradient, non_targeted_sign_gradient
(X_train_woae, y_train_woae), (X_test_woae, y_test_woae) = mnist.load_data()
X_train_woae = X_train_woae.reshape((len(X_train_woae), np.prod(X_train_woae.shape[1:])))
X_test_woae = X_test_woae.reshape((len(X_test_woae), np.prod(X_test_woae.shape[1:])))
from sklearn import tree
#model_woae = LogisticRegression(multi_class='multinomial', solver='lbfgs', fit_intercept=False)
model_woae = tree.DecisionTreeClassifier(class_weight='balanced')
model_woae.fit(X_train_woae, y_train_woae)
#model_woae.coef_ = model_woae.feature_importances_
coef_int = np.round(model_woae.tree_.compute_feature_importances(normalize=False) * X_train_woae.size).astype(int)
attack_woae = Attack(model_woae)
attack_woae.prepare(X_train_woae, y_train_woae, X_test_woae, y_test_woae)
weights_woae = attack_woae.weights
num_classes_woae = len(np.unique(y_train_woae))
attack_woae.create_one_hot_targets(y_test_woae)
attack_woae.attack_to_max_epsilon(non_targeted_gradient, 50)
non_targeted_scores_woae = attack_woae.scores
So the attack class does perturbation and non-targeted gradient attack. And here is the attack class:
import numpy as np
from sklearn.metrics import accuracy_score
def calc_output_weighted_weights(output, w):
for c in range(len(output)):
if c == 0:
weighted_weights = output[c] * w[c]
else:
weighted_weights += output[c] * w[c]
return weighted_weights
def targeted_gradient(foolingtarget, output, w):
ww = calc_output_weighted_weights(output, w)
for k in range(len(output)):
if k == 0:
gradient = foolingtarget[k] * (w[k]-ww)
else:
gradient += foolingtarget[k] * (w[k]-ww)
return gradient
def non_targeted_gradient(target, output, w):
ww = calc_output_weighted_weights(output, w)
for k in range(len(target)):
if k == 0:
gradient = (1-target[k]) * (w[k]-ww)
else:
gradient += (1-target[k]) * (w[k]-ww)
return gradient
def non_targeted_sign_gradient(target, output, w):
gradient = non_targeted_gradient(target, output, w)
return np.sign(gradient)
class Attack:
def __init__(self, model):
self.fooling_targets = None
self.model = model
def prepare(self, X_train, y_train, X_test, y_test):
self.images = X_test
self.true_targets = y_test
self.num_samples = X_test.shape[0]
self.train(X_train, y_train)
print("Model training finished.")
self.test(X_test, y_test)
print("Model testing finished. Initial accuracy score: " + str(self.initial_score))
def set_fooling_targets(self, fooling_targets):
self.fooling_targets = fooling_targets
def train(self, X_train, y_train):
self.model.fit(X_train, y_train)
self.weights = self.model.coef_
self.num_classes = self.weights.shape[0]
def test(self, X_test, y_test):
self.preds = self.model.predict(X_test)
self.preds_proba = self.model.predict_proba(X_test)
self.initial_score = accuracy_score(y_test, self.preds)
def create_one_hot_targets(self, targets):
self.one_hot_targets = np.zeros(self.preds_proba.shape)
for n in range(targets.shape[0]):
self.one_hot_targets[n, targets[n]] = 1
def attack(self, attackmethod, epsilon):
perturbed_images, highest_epsilon = self.perturb_images(epsilon, attackmethod)
perturbed_preds = self.model.predict(perturbed_images)
score = accuracy_score(self.true_targets, perturbed_preds)
return perturbed_images, perturbed_preds, score, highest_epsilon
def perturb_images(self, epsilon, gradient_method):
perturbed = np.zeros(self.images.shape)
max_perturbations = []
for n in range(self.images.shape[0]):
perturbation = self.get_perturbation(epsilon, gradient_method, self.one_hot_targets[n], self.preds_proba[n])
perturbed[n] = self.images[n] + perturbation
max_perturbations.append(np.max(perturbation))
highest_epsilon = np.max(np.array(max_perturbations))
return perturbed, highest_epsilon
def get_perturbation(self, epsilon, gradient_method, target, pred_proba):
gradient = gradient_method(target, pred_proba, self.weights)
inf_norm = np.max(gradient)
perturbation = epsilon / inf_norm * gradient
return perturbation
def attack_to_max_epsilon(self, attackmethod, max_epsilon):
self.max_epsilon = max_epsilon
self.scores = []
self.epsilons = []
self.perturbed_images_per_epsilon = []
self.perturbed_outputs_per_epsilon = []
for epsilon in range(0, self.max_epsilon):
perturbed_images, perturbed_preds, score, highest_epsilon = self.attack(attackmethod, epsilon)
self.epsilons.append(highest_epsilon)
self.scores.append(score)
self.perturbed_images_per_epsilon.append(perturbed_images)
self.perturbed_outputs_per_epsilon.append(perturbed_preds)
And this is the traceback it gives:
ValueError
Traceback (most recent call last) in 4 num_classes_woae = len(np.unique(y_train_woae)) 5 attack_woae.create_one_hot_targets(y_test_woae) ----> 6 attack_woae.attack_to_max_epsilon(non_targeted_gradient, 50) 7 non_targeted_scores_woae = attack_woae.scores
~\MULTIATTACK\AttackUtils.py in attack_to_max_epsilon(self, attackmethod, max_epsilon) 106 self.perturbed_outputs_per_epsilon = [] 107 for epsilon in range(0, self.max_epsilon): --> 108 perturbed_images, perturbed_preds, score, highest_epsilon = self.attack(attackmethod, epsilon) 109 self.epsilons.append(highest_epsilon) 110 self.scores.append(score)
~\MULTIATTACK\AttackUtils.py in attack(self, attackmethod, epsilon) 79 def attack(self, attackmethod, epsilon): 80 perturbed_images, highest_epsilon = self.perturb_images(epsilon, attackmethod) ---> 81 perturbed_preds = self.model.predict(perturbed_images) 82 score = accuracy_score(self.true_targets, perturbed_preds) 83 return perturbed_images, perturbed_preds, score, highest_epsilon
...\appdata\local\programs\python\python35\lib\site-packages\sklearn\tree\tree.py in predict(self, X, check_input) 413 """ 414 check_is_fitted(self, 'tree_') --> 415 X = self._validate_X_predict(X, check_input) 416 proba = self.tree_.predict(X) 417 n_samples = X.shape[0]
...\appdata\local\programs\python\python35\lib\site-packages\sklearn\tree\tree.py in _validate_X_predict(self, X, check_input) 374 """Validate X whenever one tries to predict, apply, predict_proba""" 375 if check_input: --> 376 X = check_array(X, dtype=DTYPE, accept_sparse="csr") 377 if issparse(X) and (X.indices.dtype != np.intc or 378 X.indptr.dtype != np.intc):
...\appdata\local\programs\python\python35\lib\site-packages\sklearn\utils\validation.py in check_array(array, accept_sparse, accept_large_sparse, dtype, order, copy, force_all_finite, ensure_2d, allow_nd, ensure_min_samples, ensure_min_features, warn_on_dtype, estimator) 566 if force_all_finite: 567 _assert_all_finite(array, --> 568 allow_nan=force_all_finite == 'allow-nan') 569 570 shape_repr = _shape_repr(array.shape)
...\appdata\local\programs\python\python35\lib\site-packages\sklearn\utils\validation.py in _assert_all_finite(X, allow_nan) 54 not allow_nan and not np.isfinite(X).all()): 55 type_err = 'infinity' if allow_nan else 'NaN, infinity' ---> 56 raise ValueError(msg_err.format(type_err, X.dtype)) 57 58
ValueError: Input contains NaN, infinity or a value too large for dtype('float32').
EDIT:
I've added coefficient numbers as 0 and it now gives the same error just below the line, at attack.attack_to_max_epsilon(non_targeted_gradient, epsilon_number)
