def fit_attack(X, y, outlier_distance_threshold=1):
attack = GeneralTikhonovAttack(step_size=5, max_steps=5000,
lasso_lambda=alpha_lambda,
boundary=boundary,
outlier_method='distancethreshold',
outlier_distance_threshold=outlier_distance_threshold)
attack.fit(X, y, starting_position=attack_starting_position)
return attack
def fit_attack(X, y):
attack = GeneralTikhonovAttack(step_size=5, max_steps=5000,
lasso_lambda=0.02,
boundary=boundary,
outlier_method='distancethreshold',
outlier_distance_threshold=1)
attack.fit(X, y)
return attack
def fit_attack(X, y):
attack = GeneralTikhonovAttack(step_size=150,
max_steps=5000,
lasso_lambda=0.02,
boundary=boundary,
outlier_method='kthdistance',
outlier_weight=0.005,
outlier_power=2,
outlier_k=3)
attack.fit(X, y)
return attack
x_train = dataset['train'][:, :-1]
y_train = (dataset['train'][:, -1] * 2) - 1
x_test_full = dataset['test'][:, :-1]
use = np.random.choice(x_test_full.shape[0], num_test)
x_test = x_test_full[use, :]
# Generate the Attack Points
boundary = np.concatenate((boundary_mins[i] * np.ones(
(1, x_test.shape[1])), boundary_maxs[i] * np.ones(
(1, x_test.shape[1]))))
if outlier_method == 'distancethreshold':
attack = GeneralTikhonovAttack(
boundary=boundary,
step_size=step_sizes[i],
max_steps=num_steps[i],
outlier_method='distancethreshold',
outlier_weight=outlier_weight,
outlier_distance_threshold=outlier_distance_threshold,
outlier_k=outlier_k)
else:
attack = GeneralTikhonovAttack(boundary=boundary,
step_size=step_sizes[i],
max_steps=num_steps[i])
print('fitting attack to training data')
attack.fit(x_train, y_train)
successful_load = True
if LOAD_ATTACK_POINTS:
print('loading attack points')
try:
attack_points = np.loadtxt(
for i, (dataset_name, dataset) in enumerate(datasets.items()):
x_train = dataset['train'][:, :-2]
y_train = (dataset['train'][:, -1] * 2) - 1
x_test_full = dataset['test'][:, :-2]
use = np.random.choice(x_test_full.shape[0], num_test)
x_test = x_test_full[use, :]
# Generate the Attack Points
boundary = np.concatenate((boundary_mins[i] * np.ones((1, x_test.shape[1])),
boundary_maxs[i] * np.ones((1, x_test.shape[1]))))
if outlier_method == 'distancethreshold':
attack = GeneralTikhonovAttack(boundary=boundary,
step_size=step_sizes[i],
max_steps=num_steps[i],
lasso_lambda=lasso_lambda,
outlier_method='distancethreshold',
outlier_weight=outlier_weight,
outlier_distance_threshold=outlier_distance_threshold,
outlier_k=outlier_k)
else:
attack = GeneralTikhonovAttack(boundary=boundary, step_size=step_sizes[i],
max_steps=num_steps[i], lasso_lambda=lasso_lambda)
attack.fit(x_train, y_train)
attack_points = np.zeros((num_attack, x_train.shape[1]))
for a in range(num_attack):
print('generating attack point {} of {} for {}'.format(a+1, num_attack, dataset_name))
x_attack, y_attack = attack.get_attack_point()
attack_points[a, :] = x_attack
print(attack_points)
def fit_attack(X, y):
attack = GeneralTikhonovAttack(lasso_lambda=0.02, boundary=boundary)
attack.fit(X, y)
return attack
boundary_max = 20
elif dataset_name is 'congressional-voting':
data_train_x = congressional_voting_train[:, :-2]
data_train_y = (congressional_voting_train[:, -1] * 2) - 1
data_test_x = congressional_voting_test[:, :-2]
data_test_y = (congressional_voting_test[:, -1] * 2) - 1
step_size = 500
max_steps = 1000
boundary_min = -20
boundary_max = 20
boundary = np.concatenate((boundary_min * np.ones(
(1, data_test_x.shape[1])), boundary_max * np.ones(
(1, data_test_x.shape[1]))))
attack = GeneralTikhonovAttack(boundary=boundary,
step_size=step_size,
max_steps=max_steps)
attack.fit(data_test_x, data_test_y)
attack_x, attack_y = attack.get_attack_point_trajectory()
print(attack_x[-1, :])
plt.figure(figsize=(12, 11))
for i in range(num_d):
for j in range(num_d):
print((i * num_d) + j + 1)
subplot = plt.subplot(num_d, num_d, (i * num_d) + j + 1)
if i == j:
plt.hist(data_test_x[:, i])
else:
plt_color = np.where(data_test_y == 1, 'b', 'r')
data_train_y = (datasets[dataset_name]['train'][:, -1] * 2) - 1
data_test_x = datasets[dataset_name]['test'][:, :-2]
data_test_y = (datasets[dataset_name]['test'][:, -1] * 2) - 1
# Setup Defender
defender = OnlineLasso(alpha=alpha_lambda)
# Setup Classifiers
lasso = OnlineLasso(alpha=alpha_lambda)
steps = [('KthNeighbor', KthNeighbor(outlier_distance_threshold=distance_threshold)),
('Lasso', OnlineLasso(alpha=alpha_lambda))]
lasso_outlier = Pipeline(steps)
# Setup Attacks
attack_lasso = GeneralTikhonovAttack(lasso_lambda=alpha_lambda,
step_size=5000,
max_steps=1000)
attack_lasso_outlier = GeneralTikhonovAttack(lasso_lambda=alpha_lambda,
step_size=5000,
max_steps=1000,
outlier_method='distancethreshold',
outlier_distance_threshold=distance_threshold)
# Setup Attack Pairs
attack_pairs = AttackPairs()
attack_pairs.add(lasso, attack_lasso, 1)
attack_pairs.add(lasso_outlier, attack_lasso_outlier, 1)
attack_pairs.fit_all(data_train_x, data_train_y)
# Setup Ensemble
ensemble = PoisonEnsemble(attack_pairs, data_test_x, defender=defender)