Simple implementation of the C&W attack on a pre-trained Keras InceptionV3 on Imagenet
To generate the adversarial image simply run:
Python adversarial_generation.py
To test the classification, Run:
Python Inception_v3.py
Adversarial examples are inputs that has been slightly modified to be imperceptible by the human and cause a misclassification Formalization often used: for a clean input x, an input x’ is an adversarial example if it is misclassified and d(x, x’) < eps.
For instance: For our example here is what we get using the C&W attack:
For more details about the C&W attack: