Adversarial Examples

What are Adversarial Examples?

Adversarial examples are input instances that have been intentionally perturbed to cause a machine learning model, particularly deep learning models, to misclassify them. These perturbations are often imperceptible to humans but can lead to significant changes in the model’s output. Adversarial examples pose security and reliability concerns, as they can be exploited to attack and manipulate the behavior of machine learning systems.

How are Adversarial Examples generated?

Adversarial examples can be generated using various methods, including:

  1. Fast Gradient Sign Method (FGSM): Computes the gradient of the loss function with respect to the input and adds a small perturbation in the direction of the gradient sign.

  2. Projected Gradient Descent (PGD): Performs an iterative version of FGSM, projecting the perturbed input back onto the feasible set at each step.

  3. Carlini & Wagner (C&W) attack: Solves an optimization problem to find the minimal perturbation that causes misclassification while staying close to the original input.

Example of generating an adversarial example using FGSM in Python with TensorFlow:

import tensorflow as tf
from tensorflow.keras.datasets import mnist
from tensorflow.keras.models import load_model

# Load the MNIST dataset and a pretrained model
(X_train, y_train), (X_test, y_test) = mnist.load_data()
X_test = X_test.astype('float32') / 255
model = load_model('mnist_model.h5')

# Select an example from the test set
x = X_test[0]
y = y_test[0]

# Calculate the gradient of the loss with respect to the input
x_tensor = tf.constant(x.reshape(1, 28, 28, 1))
with tf.GradientTape() as tape:
    y_pred = model(x_tensor)
    loss = tf.keras.losses.sparse_categorical_crossentropy(y, y_pred)

grad = tape.gradient(loss, x_tensor)

# Generate the adversarial example using FGSM
epsilon = 0.1
adv_example = x_tensor + epsilon * tf.sign(grad)
adv_example = tf.clip_by_value(adv_example, 0, 1)

More resources on Adversarial Examples: