def standard_nn(X_train, y_train, X_test, y_test, X_extra, y_extra, ep):
n_test = len(y_test)
n_extra = len(y_extra)
standard_acc = []
neigh = KNeighborsClassifier(n_neighbors=1)
neigh.fit(X_train, y_train)
y_extra_pred = neigh.predict(X_extra)
clf = None
dt = None
mask = None
mapping = 'noMapping'
[clf, mask,
[acc_train, acc_extra,
acc_original]] = generate_sub(X_train, y_train, X_test, y_test, X_extra,
y_extra_pred, neigh, ep[-1])
for i in range(len(ep)):
eps = ep[i]
adv_test = generate_adversarial_examples(FLAG, eps, X_train, X_test,
y_train, y_test, X_extra,
y_extra_pred, mapping, clf,
mask)
print adv_test - X_test
y_adv = neigh.predict(adv_test)
adv = sum(abs(y_test - y_adv)) / 2
test_acc = 1 - (adv * 1.0 / n_test)
standard_acc.append(test_acc)
return [standard_acc, [acc_train, acc_extra, acc_original]]
def at_nn_all(X_train, y_train, X_test, y_test, X_extra, y_extra, ep):
n_test = len(y_test)
n_extra = len(y_extra)
atnn_acc = []
[new_train, y_new_train, neigh] = ATNN_all(X_train, y_train, eps_aug_all)
y_extra_pred = neigh.predict(X_extra)
clf = None
dt = None
mask = None
mapping = 'noMapping'
[clf, mask,
[acc_train, acc_extra,
acc_original]] = generate_sub(new_train, y_new_train, X_test, y_test,
X_extra, y_extra_pred, neigh, ep[-1])
for i in range(len(ep)):
eps = ep[i]
adv_test = generate_adversarial_examples(FLAG, eps, new_train, X_test,
y_new_train, y_test, X_extra,
y_extra_pred, mapping, clf,
mask)
y_adv = neigh.predict(adv_test)
adv = sum(abs(y_test - y_adv)) / 2
test_acc = 1 - (adv * 1.0 / n_test)
atnn_acc.append(test_acc)
if clf:
return [atnn_acc, [acc_train, acc_extra, acc_original]]
else:
return atnn_acc
def ATNN(X_train, y_train, eps, X_aug=None, y_aug=None):
neigh = KNeighborsClassifier(n_neighbors=1)
neigh.fit(X_train, y_train)
if X_aug and y_aug:
new_train = np.concatenate((X_train, X_aug), axis=0)
y_new_train = np.concatenate((y_train, y_aug), axis=0)
else:
clf = None
dt = None
mask = None
mapping = 'noMapping'
if FLAG == 'dt':
clf, dt = prepare_tree(X_train, y_train, MAX_DEPTH)
elif FLAG == 'lr':
clf = LogisticRegression()
clf.fit(X_train, y_train)
elif FLAG == 'kernel' or FLAG == 'wb_kernel':
clf = Kernel_Classifier(c=C)
clf.fit(X_train, y_train)
elif FLAG == 'nn':
clf = neural_net_classifier(task=TASK)
clf.fit(X_train,
y_train,
target_model=neigh,
model=model,
train_params=train_params,
shape=shape)
else:
clf = neigh
y_pred = clf.predict(X_train).reshape(y_train.shape)
mask = y_pred == y_train
mask = [True for i in range(len(y_train))]
if FLAG == 'dt':
clf = dt
aug = generate_adversarial_examples(FLAG, eps, X_train, X_train,
y_train, y_train, X_train, y_train,
mapping, clf, mask, 'attack')
new_train = np.concatenate((X_train, aug), axis=0)
y_new_train = np.concatenate((y_train, y_train), axis=0)
[new_train, y_new_train] = shuffle_data(new_train, y_new_train)
neigh.fit(new_train, y_new_train)
return [new_train, y_new_train, neigh]
def robust_nn(X_train, y_train, X_test, y_test, X_extra, y_extra, ep, eps_adv):
n_test = len(y_test)
n_extra = len(y_extra)
robustnn_acc = []
print ep
robust_1nn = Robust_1NN(X_train, y_train, 0.45, 0.1, eps_adv)
robust_1nn.find_confident_label()
robust_1nn.find_red_points()
robust_1nn.fit()
[new_train, y_new_train] = robust_1nn.get_data()
neigh = robust_1nn.get_clf()
y_extra_pred = neigh.predict(X_extra)
clf = None
dt = None
mask = None
mapping = 'noMapping'
[clf, mask,
[acc_train, acc_extra,
acc_original]] = generate_sub(new_train, y_new_train, X_test, y_test,
X_extra, y_extra_pred, neigh, ep[-1])
for i in range(len(ep)):
eps = ep[i]
adv_test = generate_adversarial_examples(FLAG, eps, new_train, X_test,
y_new_train, y_test, X_extra,
y_extra_pred, mapping, clf,
mask)
#print adv_test
y_adv = neigh.predict(adv_test)
adv = sum(abs(y_test - y_adv)) / 2
test_acc = 1 - (adv * 1.0 / n_test)
robustnn_acc.append(test_acc)
return [robustnn_acc, [acc_train, acc_extra, acc_original]]
def ATNN_all(X_train, y_train, eps, X_aug=None, y_aug=None):
neigh = KNeighborsClassifier(n_neighbors=1)
neigh.fit(X_train, y_train)
if X_aug and y_aug:
new_train = np.concatenate((X_train, X_aug), axis=0)
y_new_train = np.concatenate((y_train, y_aug), axis=0)
else:
clf = None
dt = None
mask = None
clfs = []
masks = []
#FLAGS = ['wb', 'dt', 'lr', 'kernel', 'nn']
FLAGS = ['wb', 'kernel', 'nn']
mapping = 'noMapping'
new_train = np.copy(X_train)
y_new_train = np.copy(y_train)
clf = neigh
y_pred = clf.predict(X_train).reshape(y_train.shape)
mask = y_pred == y_train
clfs.append(clf)
masks.append(mask)
'''
clf, dt = prepare_tree(X_train, y_train, MAX_DEPTH)
y_pred = clf.predict(X_train).reshape(y_train.shape)
mask = y_pred==y_train
clfs.append(dt)
masks.append(mask)
clf = LogisticRegression()
clf.fit(X_train, y_train)
y_pred = clf.predict(X_train).reshape(y_train.shape)
mask = y_pred==y_train
clfs.append(clf)
masks.append(mask)
'''
clf = Kernel_Classifier(c=C)
clf.fit(X_train, y_train)
y_pred = clf.predict(X_train).reshape(y_train.shape)
mask = y_pred == y_train
clfs.append(clf)
masks.append(mask)
clf = neural_net_classifier(task=TASK)
clf.fit(X_train,
y_train,
target_model=neigh,
model=model,
train_params=train_params,
shape=shape)
y_pred = clf.predict(X_train).reshape(y_train.shape)
mask = y_pred == y_train
clfs.append(clf)
masks.append(mask)
augs = []
for (clf, mask, flag) in zip(clfs, masks, FLAGS):
mask = [True for i in range(len(y_train))]
aug = generate_adversarial_examples(flag, eps, X_train, X_train,
y_train, y_train, X_train,
y_train, mapping, clf, mask,
'attack')
augs.append(aug)
new_train = np.concatenate((new_train, aug), axis=0)
y_new_train = np.concatenate((y_new_train, y_train), axis=0)
[new_train, y_new_train] = shuffle_data(new_train, y_new_train)
neigh.fit(new_train, y_new_train)
return [new_train, y_new_train, neigh]