Added TensorFlow v2.0 implementation of FGSM Attack

examples/fgsm_tensorflow_mnist.py

@@ -0,0 +1,76 @@
+"""
+This tutorial uses the stable release of TensorFlow v2.0 dated 4 October 2019. Here, a simple classifier 
+is built using the functional API as TensorFlow v2.0 is now tightly-knit with Keras and provides well-rounded 
+support. Adversarial examples are generated using the gradients of the model. 
+
+Note: The hyper-parameters are not fully optimised for accuracy. Eager executing is disabled so as to use the
+KerasClassifier instance
+
+Author: Rishabh Anand (GitHub: @rish-16)
+"""
+
+import numpy as np
+import tensorflow as tf
+
+from tensorflow.keras.models import Sequential, Model
+from tensorflow.keras.layers import Dense, Conv2D, MaxPooling2D, Flatten, Input
+from tensorflow.keras.utils import to_categorical
+
+from art.attacks.fast_gradient import FastGradientMethod
+from art.classifiers import KerasClassifier
+from art.utils import load_mnist
+
+# disabling eager execution
+print ('TensorFlow version: {}'.format(tf.__version__))
+tf.compat.v1.disable_v2_behavior()
+
+# preprocessing MNIST
+(x_train, y_train), (x_test, y_test), min_pixel, max_pixel = load_mnist()
+
+# building a ConvNet
+conv_input = Input(shape=[28, 28, 1])
+x = Conv2D(32, (3,3), activation="relu")(conv_input)
+x = MaxPooling2D(pool_size=(2,2))(conv_input)
+x = Conv2D(64, (3,3), activation="relu")(x)
+x = MaxPooling2D(pool_size=(2,2))(x)
+x = Flatten()(x)
+x = Dense(128, activation="relu")(x)
+conv_out = Dense(10, activation="softmax")(x)
+
+# compiling the model. Tracking the "accuracy" metric
+model = Model(inputs=conv_input, outputs=conv_out)
+model.compile(loss="categorical_crossentropy", optimizer="sgd", metrics=["acc"])
+
+# visualising model architecture
+model.summary()
+
+# building the classifier using ART
+mnist_net = KerasClassifier(model, clip_values=(min_pixel, max_pixel))
+
+# training the classifier
+training_history = mnist_net.fit(x_train, y_train, batch_size=128, nb_epochs=10)
+
+# getting accuracy of classifier
+def evaluate(predictions, labels):
+	accuracy = 0
+	for i in range(len(predictions)):
+		if np.argmax(predictions[i]) == np.argmax(labels[i]):
+			accuracy += 1
+
+	return float(accuracy / len(predictions)) * 100
+
+# testing the classifier
+predictions = mnist_net.predict(x_test)
+acc_before = evaluate(predictions, y_test)
+print ('Accuracy before attack: {}'.format(acc_before))
+
+# creating adversarial examples using the classifier's gradients
+epsilon = 0.2
+adversarial_crafter = FastGradientMethod(mnist_net, eps=epsilon)
+adv_x_test = adversarial_crafter.generate(x=x_test)
+
+# testing model on adversarial examples
+print ('Implementing attack on KerasClassifier model...')
+adv_preds = mnist_net.predict(adv_x_test)
+acc_after = evaluate(adv_preds, y_test)
+print ('Accuracy after attack: {}'.format(acc_after))