Non-robust features transfer for a new adversarial attack against neural image classifiers

By Alexandre Variengien

Context

This repository contains the code used in the project of the CS-503 course "Visual Intelligence: machines and mind" at EPFL. The full report of the project can be found as a pdf here.

Instruction

This project implement a new adversarial attack again classifiers. The goal is to make them classify images humans cannot see.

The attack can be directly reproduced on the CIFAR10 dataset against 3 models in a Colab notebook. Try it now here!

I also included the notebook used to perform the ImageNet dataset. To be run, it need to be next to a folder called ImageNet containing images in folders corresponding to classes.