Uses PyTorch to generate adversarial images against Google's Inception v3 model. Code based off Roman Trusov's repo. Experiments were run in support of the paper Adversarial Attacks in Machine Learning.
An installation of PyTorch is assumed. Examples can then be generated by running adversarial.py. Input images as well as attack types (targeted vs. non-targeted) can be specified within this file.
$ python adversarial.pyFast gradient sign attack with epsilon=0.05.
Iterative non-targeted attack with epsilon=0.05.
Iterative targeted attack with epsilon=0.02. The attack was successful given the target class "cello."
