def test_fix_relabel_poison(self): (x_train, y_train), (_, _), (_, _) = self.mnist x_poison = x_train[:100] y_fix = y_train[:100] test_set_split = 0.7 n_train = int(len(x_poison) * test_set_split) x_test = x_poison[n_train:] y_test = y_fix[n_train:] predictions = np.argmax(self.classifier.predict(x_test), axis=1) ini_miss = 1 - np.sum(predictions == np.argmax(y_test, axis=1)) / y_test.shape[0] improvement, new_classifier = ActivationDefence.relabel_poison_ground_truth(self.classifier, x_poison, y_fix, test_set_split=test_set_split, tolerable_backdoor=0.01, max_epochs=5, batch_epochs=10) predictions = np.argmax(new_classifier.predict(x_test), axis=1) final_miss = 1 - np.sum(predictions == np.argmax(y_test, axis=1)) / y_test.shape[0] self.assertEqual(improvement, ini_miss - final_miss) # Other method (since it's cross validation we can't assert to a concrete number). improvement, _ = ActivationDefence.relabel_poison_cross_validation(self.classifier, x_poison, y_fix, n_splits=2, tolerable_backdoor=0.01, max_epochs=5, batch_epochs=10) self.assertGreaterEqual(improvement, 0)
def main(): try: print('See if poison model has been previously trained ') import pickle classifier = pickle.load(open('my_poison_classifier.p', 'rb')) print('Loaded model from pickle.... ') data_train = np.load('data_training.npz') x_train = data_train['x_train'] y_train = data_train['y_train'] is_poison_train = data_train['is_poison_train'] data_test = np.load('data_testing.npz') x_test = data_test['x_test'] y_test = data_test['y_test'] is_poison_test = data_test['is_poison_test'] except: # Read MNIST dataset (x_raw contains the original images): (x_raw, y_raw), (x_raw_test, y_raw_test), min_, max_ = load_mnist(raw=True) n_train = np.shape(x_raw)[0] num_selection = n_train random_selection_indices = np.random.choice(n_train, num_selection) x_raw = x_raw[random_selection_indices] y_raw = y_raw[random_selection_indices] # Poison training data perc_poison = .33 (is_poison_train, x_poisoned_raw, y_poisoned_raw) = generate_backdoor(x_raw, y_raw, perc_poison) x_train, y_train = preprocess(x_poisoned_raw, y_poisoned_raw) # Add channel axis: x_train = np.expand_dims(x_train, axis=3) # Poison test data (is_poison_test, x_poisoned_raw_test, y_poisoned_raw_test) = generate_backdoor(x_raw_test, y_raw_test, perc_poison) x_test, y_test = preprocess(x_poisoned_raw_test, y_poisoned_raw_test) # Add channel axis: x_test = np.expand_dims(x_test, axis=3) # Shuffle training data so poison is not together n_train = np.shape(y_train)[0] shuffled_indices = np.arange(n_train) np.random.shuffle(shuffled_indices) x_train = x_train[shuffled_indices] y_train = y_train[shuffled_indices] is_poison_train = is_poison_train[shuffled_indices] # Save data used for training and testing split: np.savez('data_training.npz', x_train=x_train, y_train=y_train, is_poison_train=is_poison_train, x_raw=x_poisoned_raw) np.savez('data_testing.npz', x_test=x_test, y_test=y_test, is_poison_test=is_poison_test, x_raw_test=x_poisoned_raw_test) # Create Keras convolutional neural network - basic architecture from Keras examples # Source here: https://github.com/keras-team/keras/blob/master/examples/mnist_cnn.py k.set_learning_phase(1) model = Sequential() model.add(Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=x_train.shape[1:])) model.add(Conv2D(64, (3, 3), activation='relu')) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.25)) model.add(Flatten()) model.add(Dense(128, activation='relu')) model.add(Dropout(0.5)) model.add(Dense(10, activation='softmax')) model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']) classifier = KerasClassifier((min_, max_), model=model) classifier.fit(x_train, y_train, nb_epochs=50, batch_size=128) print('Saving poisoned model: ') pickle.dump(classifier, open('my_poison_classifier.p', 'wb')) # Also saving for Anu: file_name = 'anu_poison_mnist' model.save(file_name + '.hdf5') model_json = model.to_json() with open(file_name + '.json', "w") as json_file: json_file.write(model_json) # Evaluate the classifier on the test set preds = np.argmax(classifier.predict(x_test), axis=1) acc = np.sum(preds == np.argmax(y_test, axis=1)) / y_test.shape[0] print("\nTest accuracy: %.2f%%" % (acc * 100)) # Evaluate the classifier on poisonous data preds = np.argmax(classifier.predict(x_test[is_poison_test]), axis=1) acc = np.sum(preds == np.argmax(y_test[is_poison_test], axis=1)) / y_test[is_poison_test].shape[0] print("\nPoisonous test set accuracy (i.e. effectiveness of poison): %.2f%%" % (acc * 100)) # Evaluate the classifier on clean data preds = np.argmax(classifier.predict(x_test[is_poison_test == 0]), axis=1) acc = np.sum(preds == np.argmax(y_test[is_poison_test == 0], axis=1)) / y_test[is_poison_test == 0].shape[0] print("\nClean test set accuracy: %.2f%%" % (acc * 100)) # Calling poisoning defence: defence = ActivationDefence(classifier, x_train, y_train) # End-to-end method: print("------------------- Results using size metric -------------------") print(defence.get_params()) defence.detect_poison(n_clusters=2, ndims=10, reduce="PCA") # Now fix the model x_new, y_fix = correct_poisoned_labels(x_train, y_train, is_poison_train) improvement = defence.relabel_poison_ground_truth(x_new, y_fix, test_set_split=0.7, tolerable_backdoor=0.001, max_epochs=5, batch_epochs=10) # Evaluate the classifier on poisonous data after backdoor fix: preds = np.argmax(classifier.predict(x_test[is_poison_test]), axis=1) acc_after = np.sum(preds == np.argmax(y_test[is_poison_test], axis=1)) / y_test[is_poison_test].shape[0] print("\nPoisonous test set accuracy (i.e. effectiveness of poison) after backdoor fix: %.2f%%" % (acc_after * 100)) print("\n Improvement after training: ", improvement) print('before: ', acc, ' after: ', acc_after) print("done :) ")