This code can be used to generate the adversarial image for any given image. Then, we also run that script to show the output of adversarial image on InceptionV3 architecture
Run this command to run on a new image called 1.jpg
python classify_image.py --image_file 1.jpg
If image file is not specified, it runs on imagenet/cropped_panda.jpg.
Adversarial image is stored in the folder adver_images.
By default it runs step_targeted_attack method for adversarial image geenration. There are two more methods step_ll_adversarial_images and step_fgsm which can be used for generating adversarial images.
To run other two methods, comment the line with call to step_targeted_attack(line 132) and Uncomment the appropriate methods below like step_fgsm(line 135).