This repository contains code for White Box adversarial attack on Neural Classifier trained on CIFAR-10 dataset and Intel Classification Competition Dataset.
Here, Fast Gradient Sign Method is used. The idea is to take the image we want to misclassify and run it through the model as usual, which gives us an output tensor. Typically for predictions, we’d look to see which of the tensor’s values was the highest and use that as the index into our classes, using argmax(). But this time we’re going to pretend that we’re training the network again and backpropagate that result back through the model, giving us the gradient changes of the model with respect to the original input. (https://learning.oreilly.com/library/view/programming-pytorch-for/9781492045342/ch09.html)
-
Install required modules from requirements.txt
-
Run main.py
- Intel Classification Competition: https://www.kaggle.com/puneet6060/intel-image-classification/version/2
- CIFAR 10 Dataset: https://www.cs.toronto.edu/~kriz/cifar.html
