def attribute_inference(true_x, true_y, batch_size, classifier, train_loss,
features):
print('-' * 10 + 'ATTRIBUTE INFERENCE' + '-' * 10 + '\n')
attr_adv, attr_mem, attr_pred = [], [], []
for feature in features:
low_op, high_op = [], []
low_data, high_data, membership = getAttributeVariations(
true_x, feature)
pred_input_fn = tf.estimator.inputs.numpy_input_fn(x={'x': low_data},
num_epochs=1,
shuffle=False)
predictions = classifier.predict(input_fn=pred_input_fn)
_, low_op = get_predictions(predictions)
pred_input_fn = tf.estimator.inputs.numpy_input_fn(x={'x': high_data},
num_epochs=1,
shuffle=False)
predictions = classifier.predict(input_fn=pred_input_fn)
_, high_op = get_predictions(predictions)
low_op = low_op.astype('float32')
high_op = high_op.astype('float32')
low_op = log_loss(true_y, low_op)
high_op = log_loss(true_y, high_op)
pred_membership = np.where(
stats.norm(0, train_loss).pdf(low_op) >= stats.norm(
0, train_loss).pdf(high_op), 0, 1)
fpr, tpr, thresholds = roc_curve(membership,
pred_membership,
pos_label=1)
print(fpr, tpr, tpr - fpr)
attr_adv.append(tpr[1] - fpr[1])
#plt.plot(fpr, tpr)
# membership
fpr, tpr, thresholds = roc_curve(
membership,
stats.norm(0, train_loss).pdf(high_op) -
stats.norm(0, train_loss).pdf(low_op),
pos_label=1)
#plt.plot(fpr, tpr)
# non-membership
fpr, tpr, thresholds = roc_curve(
membership,
stats.norm(0, train_loss).pdf(low_op) -
stats.norm(0, train_loss).pdf(high_op),
pos_label=0)
#plt.show()
attr_mem.append(membership)
attr_pred.append(np.vstack((low_op, high_op)))
return attr_adv, attr_mem, attr_pred
def train_target_model(dataset,
epochs=100,
batch_size=100,
learning_rate=0.01,
l2_ratio=1e-7,
n_hidden=50,
model='nn',
save=True,
privacy='no_privacy',
dp='dp',
epsilon=0.5,
delta=1e-5):
train_x, train_y, test_x, test_y = dataset
classifier, _, _, train_loss, train_acc, test_acc = train_model(
dataset,
n_hidden=n_hidden,
epochs=epochs,
learning_rate=learning_rate,
batch_size=batch_size,
model=model,
l2_ratio=l2_ratio,
silent=False,
privacy=privacy,
dp=dp,
epsilon=epsilon,
delta=delta)
# test data for attack model
attack_x, attack_y = [], []
# data used in training, label is 1
pred_input_fn = tf.estimator.inputs.numpy_input_fn(x={'x': train_x},
num_epochs=1,
shuffle=False)
predictions = classifier.predict(input_fn=pred_input_fn)
_, pred_scores = get_predictions(predictions)
attack_x.append(pred_scores)
attack_y.append(np.ones(train_x.shape[0]))
# data not used in training, label is 0
pred_input_fn = tf.estimator.inputs.numpy_input_fn(x={'x': test_x},
num_epochs=1,
shuffle=False)
predictions = classifier.predict(input_fn=pred_input_fn)
_, pred_scores = get_predictions(predictions)
attack_x.append(pred_scores)
attack_y.append(np.zeros(test_x.shape[0]))
attack_x = np.vstack(attack_x)
attack_y = np.concatenate(attack_y)
attack_x = attack_x.astype('float32')
attack_y = attack_y.astype('int32')
if save:
np.savez(MODEL_PATH + 'attack_test_data.npz', attack_x, attack_y)
np.savez(MODEL_PATH + 'target_model.npz',
*lasagne.layers.get_all_param_values(output_layer))
classes = np.concatenate([train_y, test_y])
return attack_x, attack_y, classes, train_loss, classifier, train_acc, test_acc
def train_shadow_models(n_hidden=50,
epochs=100,
n_shadow=20,
learning_rate=0.05,
batch_size=100,
l2_ratio=1e-7,
model='nn',
save=True):
# for attack model
attack_x, attack_y = [], []
classes = []
for i in range(n_shadow):
#print('Training shadow model {}'.format(i))
data = load_data('shadow{}_data.npz'.format(i))
train_x, train_y, test_x, test_y = data
# train model
classifier, _, _, _, _, _ = train_model(data,
n_hidden=n_hidden,
epochs=epochs,
learning_rate=learning_rate,
batch_size=batch_size,
model=model,
l2_ratio=l2_ratio)
#print('Gather training data for attack model')
attack_i_x, attack_i_y = [], []
# data used in training, label is 1
pred_input_fn = tf.estimator.inputs.numpy_input_fn(x={'x': train_x},
num_epochs=1,
shuffle=False)
predictions = classifier.predict(input_fn=pred_input_fn)
_, pred_scores = get_predictions(predictions)
attack_i_x.append(pred_scores)
attack_i_y.append(np.ones(train_x.shape[0]))
# data not used in training, label is 0
pred_input_fn = tf.estimator.inputs.numpy_input_fn(x={'x': test_x},
num_epochs=1,
shuffle=False)
predictions = classifier.predict(input_fn=pred_input_fn)
_, pred_scores = get_predictions(predictions)
attack_i_x.append(pred_scores)
attack_i_y.append(np.zeros(test_x.shape[0]))
attack_x += attack_i_x
attack_y += attack_i_y
classes.append(np.concatenate([train_y, test_y]))
# train data for attack model
attack_x = np.vstack(attack_x)
attack_y = np.concatenate(attack_y)
attack_x = attack_x.astype('float32')
attack_y = attack_y.astype('int32')
classes = np.concatenate(classes)
if save:
np.savez(MODEL_PATH + 'attack_train_data.npz', attack_x, attack_y)
return attack_x, attack_y, classes