Paul-Lez / adversarial_example_gen

A short beginner level tutorial on the generation adversarial examples to fool a neural network.

Generating adversarial examples for leNet5 type classifiers

This repository contains code to train a variant of the leNet5 classifier on the MNIST dataset and generate some adversarial examples using PGD. It also contains code to train a different CNN on the same dataset (performing the same classification task), in order to demonstrate a phenemenon called the "transferability of adversarial examples", which takes place when adversarial examples crafted to fool one model also succeed in fooling another model with a potentially different architecture.

This contains

• An implementation of a variant of LeNet and code to train it on the MNIST dataset.
• An implementation of a different CNN and code to train it on the same dataset.
• Some code to generate adversarial examples (based on https://adversarial-ml-tutorial.org/introduction/) for the leNet5 model and test them on the other CNN.
• Sample weights for both models
• Some test photos of digits

How to use this:

• Download the notebook adversarial_examples_notebook and open it on Google Colab.
• Read through and enjoy.