Adversarial-Attacks-in-Machine-Learning

A brief study on Adversarial Attacks and python scripts to generate and study them.

Adversarial examples and attacks

Adversarial examples are morphed examples which lead a model into misclassifying them especially for relatively small changes in the input.
Various methods to generate them and defend them are studied here. Read the summary.

Generating adversarial examples

A CNN model is trained on cifar10 in model.py. Methods to generate adversarial examples are implemented in adversarial_generator.py.

• DIY: compare two identical networks, except one uses ReLU and the other Sigmoid. I predict the latter to be more robust to adversarial examples given how linear ReLU is compared to Sigmoid. This needs to be tested.